Science and Engineering Careers in the United States: An Analysis of Markets and Employment 9780226261904

Citation preview

Science and Engineering Careers in the United States

A National Bureau of Economic Research Conference Report

Science and Engineering Careers in the United States An Analysis of Markets and Employment

Edited by

Richard B. Freeman and Daniel L. Goroff

The University of Chicago Press Chicago and London

RICHARD B. FREEMAN holds the Herbert Ascherman Chair in Economics at Harvard University and is currently serving as faculty director of the Labor and Worklife Program at Harvard Law School. He is director of the Labor Studies Program at the National Bureau of Economic Research. DANIEL L. GOROFF is professor of mathematics and economics at Harvey Mudd College, where he previously served as vice president for academic affairs and dean of the faculty. He codirects the Sloan Scientific and Engineering Workforce Project at the National Bureau of Economic Research.

The University of Chicago Press, Chicago 60637 The University of Chicago Press, Ltd., London © 2009 by the National Bureau of Economic Research All rights reserved. Published 2009 Printed in the United States of America 18 17 16 15 14 13 12 11 10 09

1 2 3 4 5

ISBN-13: 978-0-226-26189-8 (cloth) ISBN-10: 0-226-26189-1 (cloth) Library of Congress Cataloging-in-Publication Data Science and engineering careers in the United States : an analysis of markets and employment / edited by Richard B. Freeman and Daniel L. Goroff. p. cm.—(National Bureau of Economic Research conference report) Includes bibliographical references and index. ISBN-13: 978-0-226-26189-8 (cloth : alk. paper) ISBN-10: 0-226-26189-1 (cloth : alk. paper) 1. Scientists—United States. 2. Engineers—United States. 3. Science—Study and teaching (Higher)—United States. 4. Engineering—Study and teaching (Higher)—United States. I. Freeman, Richard B. (Richard Barry), 1943– II. Goroff, Daniel L. Q149.U5S3112 2009 331.761500973—dc22 2008032103 o The paper used in this publication meets the minimum requirements of the American National Standard for Information Sciences— Permanence of Paper for Printed Library Materials, ANSI Z39.48-1992.

National Bureau of Economic Research Officers John S. Clarkeson, chairman Kathleen B. Cooper, vice-chairman James M. Poterba, president and chief executive officer Robert Mednick, treasurer

Kelly Horak, controller and assistant corporate secretary Alterra Milone, corporate secretary Gerardine Johnson, assistant corporate secretary

Directors at Large Peter C. Aldrich Elizabeth E. Bailey Richard B. Berner John H. Biggs John S. Clarkeson Don R. Conlan Kathleen B. Cooper Charles H. Dallara George C. Eads

Jessica P. Einhorn Mohamed El-Erian Jacob A. Frenkel Judith M. Gueron Robert S. Hamada Karen N. Horn John Lipsky Laurence H. Meyer Michael H. Moskow

Alicia H. Munnell Rudolph A. Oswald Robert T. Parry James M. Poterba John S. Reed Marina v. N. Whitman Martin B. Zimmerman

Directors by University Appointment George Akerlof, California, Berkeley Jagdish Bhagwati, Columbia Glen G. Cain, Wisconsin Ray C. Fair, Yale Franklin Fisher, Massachusetts Institute of Technology Mark Grinblatt, California, Los Angeles Saul H. Hymans, Michigan Marjorie B. McElroy, Duke

Joel Mokyr, Northwestern Andrew Postlewaite, Pennsylvania Uwe E. Reinhardt, Princeton Nathan Rosenberg, Stanford Craig Swan, Minnesota David B. Yoffie, Harvard Arnold Zellner (Director Emeritus), Chicago

Directors by Appointment of Other Organizations Jean-Paul Chavas, American Agricultural Economics Association Gail D. Fosler, The Conference Board Martin Gruber, American Finance Association Timothy W. Guinnane, Economic History Association Arthur B. Kennickell, American Statistical Association Thea Lee, American Federation of Labor and Congress of Industrial Organizations

William W. Lewis, Committee for Economic Development Robert Mednick, American Institute of Certified Public Accountants Angelo Melino, Canadian Economics Association Harvey Rosenblum, National Association for Business Economics John J. Siegfried, American Economic Association

Directors Emeriti Andrew Brimmer Carl F. Christ George Hatsopoulos Lawrence R. Klein

Franklin A. Lindsay Paul W. McCracken Peter G. Peterson Richard N. Rosett

Eli Shapiro Arnold Zellner

Relation of the Directors to the Work and Publications of the National Bureau of Economic Research 1. The object of the NBER is to ascertain and present to the economics profession, and to the public more generally, important economic facts and their interpretation in a scientific manner without policy recommendations. The Board of Directors is charged with the responsibility of ensuring that the work of the NBER is carried on in strict conformity with this object. 2. The President shall establish an internal review process to ensure that book manuscripts proposed for publication DO NOT contain policy recommendations. This shall apply both to the proceedings of conferences and to manuscripts by a single author or by one or more coauthors but shall not apply to authors of comments at NBER conferences who are not NBER affiliates. 3. No book manuscript reporting research shall be published by the NBER until the President has sent to each member of the Board a notice that a manuscript is recommended for publication and that in the President’s opinion it is suitable for publication in accordance with the above principles of the NBER. Such notification will include a table of contents and an abstract or summary of the manuscript’s content, a list of contributors if applicable, and a response form for use by Directors who desire a copy of the manuscript for review. Each manuscript shall contain a summary drawing attention to the nature and treatment of the problem studied and the main conclusions reached. 4. No volume shall be published until forty-five days have elapsed from the above notification of intention to publish it. During this period a copy shall be sent to any Director requesting it, and if any Director objects to publication on the grounds that the manuscript contains policy recommendations, the objection will be presented to the author(s) or editor(s). In case of dispute, all members of the Board shall be notified, and the President shall appoint an ad hoc committee of the Board to decide the matter; thirty days additional shall be granted for this purpose. 5. The President shall present annually to the Board a report describing the internal manuscript review process, any objections made by Directors before publication or by anyone after publication, any disputes about such matters, and how they were handled. 6. Publications of the NBER issued for informational purposes concerning the work of the Bureau, or issued to inform the public of the activities at the Bureau, including but not limited to the NBER Digest and Reporter, shall be consistent with the object stated in paragraph 1. They shall contain a specific disclaimer noting that they have not passed through the review procedures required in this resolution. The Executive Committee of the Board is charged with the review of all such publications from time to time. 7. NBER working papers and manuscripts distributed on the Bureau’s web site are not deemed to be publications for the purpose of this resolution, but they shall be consistent with the object stated in paragraph 1. Working papers shall contain a specific disclaimer noting that they have not passed through the review procedures required in this resolution. The NBER’s web site shall contain a similar disclaimer. The President shall establish an internal review process to ensure that the working papers and the web site do not contain policy recommendations, and shall report annually to the Board on this process and any concerns raised in connection with it. 8. Unless otherwise determined by the Board or exempted by the terms of paragraphs 6 and 7, a copy of this resolution shall be printed in each NBER publication as described in paragraph 2 above.

Contents

Acknowledgments

ix

Introduction Richard B. Freeman and Daniel L. Goroff

1

I. Supply of Students and Postdoctoral Fellows to Science and Engineering 1. Supporting “The Best and Brightest” in Science and Engineering: NSF Graduate Research Fellowships Richard B. Freeman, Tanwin Chang, and Hanley Chiang 2. Internationalization of U.S. Doctorate Education John Bound, Sarah Turner, and Patrick Walsh 3. Improving the Postdoctoral Experience: An Empirical Approach Geoff Davis

19

59

99

II. Careers in Changing Markets 4. Immigration in High-Skill Labor Markets: The Impact of Foreign Students on the Earnings of Doctorates George J. Borjas

131

vii

viii

Contents

5. Does Science Promote Women? Evidence from Academia 1973–2001 Donna K. Ginther and Shulamit Kahn

163

6. Patterns of Male and Female Scientific Dissemination in Public and Private Science Kjersten Bunker Whittington

195

7. Educational Mismatch among Ph.D.s: Determinants and Consequences Keith A. Bender and John S. Heywood

229

8. Capturing Knowledge: The Location Decision of New Ph.D.s Working in Industry Albert J. Sumell, Paula E. Stephan, and James D. Adams

257

III. Creation and Use of KnowledgE 9. Instruments of Commerce and Knowledge: Probe Microscopy, 1980–2000 Cyrus C. M. Mody 10. International Knowledge Flows: Evidence from an Inventor-Firm Matched Data Set Jinyoung Kim, Sangjoon John Lee, and Gerald Marschke 11. The Growing Allocative Inefficiency of the U.S. Higher Education Sector James D. Adams and J. Roger Clemmons Contributors Author Index Subject Index

291

321

349 383 385 389

Acknowledgments

This study was supported by the Alfred P. Sloan Foundation as part of the Science and Engineering Workforce Project (http://www.nber.org/~sewp/). We benefitted greatly from many meetings and discussions with a wide variety of scientists, policymakers, and business and labor groups concerned with U.S. scientific and engineering work and the economic status of scientists and engineers. Michael Teitlebaum of the Sloan Foundation was particularly helpful in our work.

ix

Introduction Richard B. Freeman and Daniel L. Goroff

In the mid-2000s, when the research in this book was nearing completion, policy makers in the United States were expressing greater concern about the job market for scientists and engineers than they had since the 1950s, following the Soviet Union’s 1956 launch of Sputnik. National commissions and groups issued reports about the dangers that the weakening state of science and engineering posed to the country and called for new policies to increase the supply of scientific and engineering talent by improving education from grades K through 12 to undergraduate and graduate training, and by additional funding of research and development (see appendix). The most prominent report was the National Academy of Science’s Rising Above The Gathering Storm: Energizing and Employing America for a Brighter Economic Future. The panel that undertook this study worried that the United States was losing leadership in science and engineering and that this threatened the nation’s competitiveness in the global economy and future economic well-being and national security. Concurring with these assessments, in his 2006 State of the Union Address, President Bush announced the American Competitiveness Initiative. He stressed that “for the U.S. to maintain its global economic leadership, we must ensure a continuous supply of highly trained mathematicians, scientists, engineers, technicians, and scientific support staff.” The research in this book illuminates many of the issues underlying the studies and reports summarized in the appendix and that spurred the president’s initiative. It provides new information about the economics of sciRichard B. Freeman holds the Herbert Ascherman Chair in Economics at Harvard University, and is the director of the labor studies program at the NBER. Daniel L. Goroff is professor of mathematics and economics at Harvey Mudd College and codirects the Sloan Scientific and Engineering Workforce Project at the NBER.

1

2

Richard B. Freeman and Daniel L. Goroff

ence and engineering work in three broad areas: the determinants of the supply of scientists and engineers, the career patterns that they follow upon graduation, and the creation and transmission of scientific and engineering knowledge as reflected in patents, papers, and the mobility of doctorate workers between academe and industry. To do this, we used a wide variety of data sets, as indicated in table I.1. These include the National Science Foundation’s (NSF’s) Surveys of Earned Doctorates (which all Ph.D. recipients are encouraged to fill out prior to obtaining their degree), Sigma Xi’s special survey of postdoctoral graduates, a specially constructed inventor-firm matched data set, a university productivity data set, interviews with persons involved in the scanning tunneling innovation in scientific instrumentation, and so on. Each of the chapters gives a detailed report on the particulars of the data analyzed, the methodology used, and the findings. By way of introducing the reader to what they will learn in the chapters and of delineating the links among them, we summarize the findings organized around the three main areas in the book. Then we consider how the findings illuminate some of the concerns over the science and engineering job market expressed by the various commissions and studies. Supply of Students and Postdoctoral Fellows to Science and Engineering 1. The supply of U.S. science and engineering students responds to fellowship support. Between 1999 and 2005 the National Science Foundation doubled the value of its Graduate Research Fellowships (GRF), which created a pseudo-experiment to assess student responsiveness to the economic incentive of graduate fellowships. When the NSF developed the program, most of the applicants were in the physical sciences and mathematics, but as the labor market opportunities increased in other sciences and engineering, a growing proportion of applicants and GRF awardees came from life sciences, social sciences, and engineering. Similarly, in the 1950s men gained most of the awards but by 2004 women won over half of the awards, largely because the increased number of women in bachelor’s degree science and engineering led to more women seeking graduate study in these fields and consequently more applying for fellowship support. Because the NSF did not increase the number of awards over time, however, the ratio of awards per B.S. graduate fell. Because NSF changed the value of awards intermittently, the value of awards relative to earnings in the economy varied over time and fell markedly in the 1990s. To reverse this, the NSF decided to increase the value of fellowships from 1999 to 2005. This produced a commensurately large increase in the number of applicants. The estimated elasticity of the applicants to stipend value over the entire history of the GRF is on the order of 0.8 to 1.0. It is more difficult to link the supply of graduate students in total in the science and engineering fields to NSF stipend policies (since only 1,000 or so are supported by these awards), so that

National representative sample of science and engineering doctorate holders from U.S. institutions, with data on education and training, work experience, career development.

Census on applicants to the Graduate Research Fellowship (GRF) with information on field, demographic characteristics, GRE scores, GPA, undergraduate college, ratings of letters of . recommendation, and review panels.

Multi-campus survey conducted at 47 institutions. Largest detailed survey of postdocs ever, with over 8,500 respondents.

Survey of Doctorate Recipients, biannual, with longitudinal component 1993–2001

Cumulative Index of NSF Fellowship Applicants and Awardees and updates, 1952–2004

Sigma Xi Postdoctoral survey, 2003–2005

U.S. Patent and Trademark Office information on all patents.

Data on author, title, field.

Data on firms from 10-K reports.

History of changes in names of firms, acquisitions, and mergers.

Founding year of firm.

Patent Bibliographic Data, 1969–2002

Proquest Digital Abstracts, 1945–2003

Compact D/SEC Database, 1989–1997

Standard & Poor’s Guide to Stocks, Directory of Obsolete Securities

Thomas Register, Megent, Corptech data

Inventor-Firm Matched Data set:

Annual survey of all new U.S. research doctorate graduates with questions about their educational histories, funding sources, and post-doctoral plans.

Characteristics of survey

Survey of Earned Doctorates, annual

Name of survey

Table I.1 Data sets used in this volume

(continued)

Chapter 9, country residence of persons on patents by U.S. firms

Chapter 3, impact of compensation and organization of postdoctorate on productivity and success of the postdoctoral experience

Chapter 1, effect of academic skills and personal attributes on receipt of fellowship; and of the value and number of awards on the number of applicants and measured academic quality of fellowship recipients

Chapter 7, job satisfaction and mismatch Chapter 4, cohort earnings and quantity data Chapter 5, longitudinal career data; publications Chapter 6, patent and articles data

Chapter 2, degrees by national origin, field, university Chapter 4, degrees by national origin Chapter 8, firm of first job when employed in industry

Chapters that use survey

Papers and citations

Academic R&D, graduate students

University finances, degrees

Faculty by university

Graduate departments at U.S. higher education institutions.

NSF CASPAR database

Integrated postsecondary education data system completions survey

National Center for Education Statistics (NCES) faculty survey

NSF Survey of Graduate Students and Postdoctorates in Science and Engineering (GSS)

Information on the development of the scanning tunneling and atomic force microscopes.

Data on patents by 1,115 inventors in Boston area biotech firms and 1,003 inventors in 4 Boston area universities.

All citations made in patents.

Characteristics of survey

Institute for Scientific Information

University-Output Data set for 102 major universities:

Interviews with 150 Probe Microscopists

U.S. Patent and Trademark Office data

Bioscan report on firms in existence, 1988–1998, identifying 49 firms

Patent data for biotech firms and departments, Boston area

NBER Patent Citations Data, 1975–1999

Name of survey

Table I.1 (continued)

Chapter 1

Chapter 11, research and teaching productivity of public and private universities

Chapter 10, history of industry and university innovation in scientific instrumentation

Chapter 6, network connections among inventors in both settings and within the firms and universities separately

Chapters that use survey

Introduction

5

changes in their value can affect outcomes primarily by impacting other stipend providers as well. This appears to have occurred and there is a positive link between the NSF increase in the value of awards and first year enrollments by U.S. students in graduate studies, though with a lower estimated elasticity than that for applications for the awards (see Freeman, Chang, and Chiang, chapter 1, this volume). 2. The supply of foreign students to U.S. science and engineering programs responds to U.S. economic and educational opportunities. As noted earlier, an increasing proportion of U.S. Ph.D. graduates in science and engineering are foreign-born. In some fields, such as engineering and economics, on the order of two-thirds of Ph.D.s are granted to the foreignborn. The huge supply of the foreign-born to U.S. graduate programs can be attributed to three factors. The first is the extension of mass higher education throughout the world. Countries that had large increases in the number of bachelor’s graduates in science and engineering have also had large increases in the number of their nationals earning Ph.D.s in the United States. The second is the greater opportunity to pursue quality graduate training in the United States than in other countries. The foreignborn Ph.D. explosion in the United States has been fueled by students from China and India, where domestic opportunities for graduate education have lagged behind the growth of undergraduate degrees. The third is the potential for working as a Ph.D. scientist or engineer in the United States with a U.S. graduate degree. As long as working in the United States is more attractive than working in one’s home country, and a U.S. degree will open doors for jobs in the United States, students will flock to U.S. graduate studies. In fact, graduates from low income countries, where science and engineering pay and employment prospects are lower than in the United States, are far more likely to stay in the United States and work than students from more advanced countries, where pay and prospects are closer to those in the United States. Given that the supply of students from low income countries appears to be quite elastic with respect to opportunities to study in the United States, increased federal research funding has also contributed to the increase in the number of foreign graduate student and postdocs, since researchers fund students and postdocs to work in their labs (see Bound, Turner, and Walsh, chapter 2, this volume). 3. The U.S. university system expanded to meet changing demand for doctorate training largely through additional places at lower quality programs. In contrast to the trend upward in foreign-born doctorate degrees, the number of Ph.D.s granted to the U.S.-born has varied over time. It has varied with the changing number of bachelor’s degrees in science and engineering and with changes in the propensity of undergraduates to go on to graduate study in response to economic opportunities. It has risen greatly for women while falling for men. In 1964, the ratio of Ph.D.s to Bachelor’s degrees seven years earlier peaked at 5.6 percent, as many bachelor’s grad-

6

Richard B. Freeman and Daniel L. Goroff

uates responded to the booming job market for Ph.D. scientists and engineers and to the fellowship opportunities that followed Sputnik. The ratio then fell to 2.5 percent by 1974. While there is competition between international students and U.S.-born students within programs, the growing number of foreign-born from lower income countries has come largely in less highly ranked programs that would seem to have expanded to meet the demands of those students for U.S. graduate training. The highly elastic supply of places at lower quality programs—which increased threefold in the primary science fields in the 1960s and 1970s—argues against any important crowd-out of U.S. students in graduate programs in total (see Bound, Turner, and Walsh, chapter 2, this volume; also Freeman, Jin, and Shen 2004). 4. Structured plans and professional opportunities at the places of work of postdoctoral students have a larger impact on their productivity and satisfaction than higher pecuniary rewards. Postdoctoral work has become an increasingly important part of scientific careers, in part to give new Ph.D.s an apprenticeship research experience before becoming independent investigators. But the growing number of postdocs in the 1990s to mid-2000s resulted as much or more from a weak job market for new Ph.D.s as from the need for greater skill acquisition. Sluggish growth of academic positions limited full-time jobs in academe while research funding increased the demand for postdocs in labs. The growing supply of foreign-born graduates eager to get into the U.S. job market, moreover, produced a large supply of Ph.D.s for the available postdoc positions. Since many postdocs were sufficiently unhappy with their experiences, they formed the National Postdoctoral Association (http://www.nationalpostdoc.org) to lobby for better treatment in their labs. Using four measures of the success of a postdoctoral experience—the postdoc’s satisfaction, their relation with their supervisor, the presence/absence of problems at the laboratory, and research productivity reflected in papers and grants—Davis (chapter 3, this volume) finds that structured oversight and creation of professional opportunities are the main factors associated with positive outcomes. Postdocs who plan their fellowship experience with their advisors, for example, have a higher submission rate of papers to journals as well as higher satisfaction. Agreements that clarify obligations for both sides are important in organizing the postdoctoral experience in a mutually advantageous way and limiting the danger that one side will opportunistically take advantage of the other. Careers in Changing Markets 5. Increases in the supply of foreign-born students to a field increases new doctorates and lowers the earnings of graduates, in part because the increased supply leads to more low-paying postdoctoral appointments. The influx of

Introduction

7

immigrant scientists and engineers has helped the United States meet changing demands for these specialists, but at the cost of reducing the earnings and employment opportunities for U.S.-born as well as other graduates in those disciplines. Between 1968 and 2000, over 200,000 foreign-born persons obtained doctorates in the United States, largely in the sciences and engineering. Many chose to stay and work in the United States. They constitute about 90 percent of all foreign-born doctorates in the country. Using the fact that the number of immigrant doctorate students and immigrant scientists and engineers varies greatly among fields and among cohorts, Borjas (chapter 4, this volume) finds that increased numbers of foreign-born Ph.D. graduates in a cohort and field reduces earnings in the cohort and field and increases the probability of working as a postdoctoral fellow. The elasticity of annual earnings to the increase in supply is on the order of 0.3 to 0.4. About half of the impact of increased supply on earnings occurs through the increased likelihood that new graduates will end up doing postdoctorate work, whose annual earnings are markedly lower than those of otherwise comparable scientists and engineers with regular jobs. In the 1990s, native-born postdocs earned about 55 percent as much as those with regular employment. 6. In academe, women in the physical sciences, engineering, and life sciences have similar chances of receiving tenure track jobs and promotion to tenure as men, whereas women in the social sciences and humanities have lower chances. A rising female share of doctorates in science and engineering rises should reduce the gender gap in the number of faculty, as long as women get into tenure-track jobs and win promotions at the same rate as men. Analyzing career patterns in academe in the Survey of Doctorate Recipients files, Ginther and Kahn (chapter 5, this volume) find that in the physical sciences, engineering, and life sciences women have about the same probability of getting tenure-track jobs and being promoted as do men. By contrast, they find that women have lower promotion probabilities than men in the social sciences and humanities. Still, even in the physical sciences, engineering, and life sciences, family factors affect men and women differently. Marriage is associated with better career outcomes for men but not for women, whereas having young children is associated with poorer outcomes for women but not for men. And there remain significant differences in salaries by gender once scientists obtain tenure-track jobs, especially at the full professor rank. The problems of balancing work and family and attaining equality in pay notwithstanding, the female proportion of tenured-track faculty in the physical sciences, engineering, and life sciences has risen commensurate with the rising female proportion of doctorates in those fields. 7. In industry, the organizational structure of research affects the research productivity of women in ways that reduce gender disparities in producing

8

Richard B. Freeman and Daniel L. Goroff

measurable outcomes, such as patents. The percentage of science and engineering doctorates working in industry has been growing, which places more new Ph.D.s, both female and male, in an environment in which scientific activity is organized differently than in academe. In industry, patents are more important than papers in scientific journals, and work is more likely to involve collaborative teams than in academe. The Survey of Doctorate Recipients (http://www.nsf.gov/statistics/srvydoctoratework/) indicates that between 1990 and 1995 about 40 percent of doctorate scientists in industry patented, compared to 15 percent of doctorate scientists in university. By contrast, while 67 percent of the scientists in industry published articles, 95 percent of scientists in academe did. In industry women are about as likely to patent, publish, or publish and patent as men, whereas in academe they are less likely to do so. Looking at the way academic and industrial scientists in biotechnology collaborate in patenting, Whittington (chapter 6, this volume) finds that academic patenting is linked around tenured scientists while industrial patenting is linked to wider networks of researchers. The organization of research work in industry as collaborative networks as opposed to the organization of research in academe as competing labs may fit better with the work patterns of women and help explain the differences in patenting and publishing between female and male scientists in industry relative to academe. 8. Working in jobs unrelated or weakly related to their fields of study or doing work different from what they expected as graduate students reduces job satisfaction and earnings and raises the turnover of doctorate scientists. Ph.D.s in science and engineering spend six to seven years studying as graduate students and many spend another three or more years employed as postdoctoral fellows before obtaining a regular job. Most want to work in the area in which they were trained, doing what they trained to do. But Bender and Heywood (chapter 7, this volume) find that in the 1990s on the order of 15 to 30 percent of doctorate scientists and engineers report a mismatch between their careers and their training: 15 percent report that they would not choose a similar field if they could start over, 20 percent report their job was not what they expected, and 30 percent reported that their job was only somewhat related (23 percent) or not related (7 percent) to their education. These mismatches are associated with large differences in earnings and job satisfaction and with greater chances of job turnover across workers and for the same worker over time. 9. New Ph.D.s in science and engineering who choose to work in industry are less likely to stay in the state or local area in which they are trained than bachelor’s or master’s graduates. Geographic entities that support graduate education in science and engineering often do so in the hope that relatively many graduates will stay in the local area and transfer the knowledge that they obtain to industry, which they hope will spur local industrial growth. Examining the placement of newly-minted science and engineering Ph.D.s who obtained first jobs in industry in 1997 to 1999, Sumell, Stephan, and

Introduction

9

Adams (chapter 8, this volume) find that a bit over one-third stay in the same state as their doctoral institution. This rate falls far below the 60 percent or so stay rate for bachelor’s or master’s degree graduates in science and engineering or for graduates in law (one of the few nonscience fields for which such data is readily available). Individuals trained in top-rated departments are more likely to leave the area in which they obtained their degree than graduates in lower-rated departments. But there is huge crossstate variation in state retention and attraction of new Ph.D.s. California and New Jersey, in particular, are more likely to retain their Ph.D.s in science and engineering and to attract others than most other states. A major factor in the geographic location of Ph.D.s is the employment opportunity for graduates in a particular field in a given locality, so that industrial demand (rather than location of training) is the critical determinant of location, especially for graduates from top programs. Creation and Use of Knowledge 10. Industry and university innovations in instrumentation create communities of producers and users that connect corporate and academic worlds in ways beyond simple commercial transactions. Scientific instruments are the physical capital in the production of knowledge. The development of a new instrument for analysis—the telescope, microscope, computers, FMRI brain scan machines—can revolutionize the way a science operates. Tracing the development of the scanning tunneling microscope (STM) developed at IBM in the 1980s into a widely used instrument in the 1990s, Mody (chapter 9, this volume) shows how corporate and university innovators formed a research and development community that made the STM and the follow-up atomic force microscope such great successes. Given differing resources and goals, the firms and universities operated differently to design STMs for different audiences. The need for tacit knowledge in producing the microscopes meant that in universities and firms graduate students and postdoctoral fellows were important in spreading STM use in scientific laboratories. The links between the industrial and academic communities led to the atomic force microscope. The implication is that having both industry and universities working in the area led to hybrid forms and innovations that might never have happened in a highly managed single institution environment. 11. Innovative U.S. firms employ researchers with foreign experience living and working overseas to tap the spread of technological knowledge around the world. As scientific and engineering activity spreads around the globe, U.S. companies who seek to stay in the forefront of technology must find ways to use overseas talent and ideas. By identifying the addresses of inventors on U.S. patents, Kim, Lee, and Marschke (chapter 10, this volume) find that U.S. pharmaceutical and semiconductor firms have relied increasingly on inventors residing overseas and appear to use them to tap

10

Richard B. Freeman and Daniel L. Goroff

into patented knowledge from those countries. From 1989 to 1997 the percentage of inventors who were foreign residents on U.S. patents increased from 14 to 30 percent in pharmaceutical firms and from 9 to 15 percent in semiconductor firms. By contrast, the percent of inventors living in the United States who had foreign patenting experience is less than 2 percent and fell over the period. The implication is that U.S. pharmaceutical and semiconductor firms employ or collaborate with researchers overseas to tap foreign talent as opposed to hiring immigrant researchers to work for them in the United States. In addition, U.S. firms tap knowledge from overseas by making use of inventions from other countries in their own inventions. In pharmaceuticals, 55 percent of patents cite at least one patent for an innovation originating outside the United States, while in semiconductors, 48 percent do so. When one inventor is or has residence in the past in a foreign country, the U.S. patent is more likely to cite patents from that country, suggesting that employing or collaborating with researchers who have research experience abroad facilitates access to overseas knowledge. 12. The level and increase of research and teaching productivity of universities differed greatly between private and public universities and among universities in each sector in the 1980s and 1990s. Universities are multiproduct institutions that produce research output (reflected in papers and citations) and teaching (reflected in undergraduate and graduate degrees). In the 1980s and 1990s when the growth of full-time faculty was less than half the rate of growth of science and engineering workers in industry, research activity grew more rapidly in leading universities than did teaching activity. Research productivity was higher in private universities than in public universities, with the gap increasing over the period. Teaching productivity was similar between private and public universities, though it increased a bit more rapidly over time in the public universities. Universities with more rapid growth of research or teaching productivity expanded less than those with less rapid growth of productivity, suggesting a possible allocative inefficiency in the higher education market. Sector aside, the analysis highlighted wide variation among universities in papers or citations per faculty and in bachelor’s and graduate degrees per faculty, that implies that the United States has a highly variegated higher educational system even among top institutions (Adams and Clemmons, chapter 11, this volume).

Illuminating the Mid-2000s Concern Chicken Little was walking in the woods when—kerplunk—an acorn fell on her head. “Oh my goodness!” said Chicken Little. “The sky is falling! I must go and tell the king.” –Children’s fable (available from http://www.geocities.com/mjloundy)

Many of the findings summarized previously depict a science and engineering job market in the 2000s that changed greatly in the 1990s and 2000s

Introduction

11

(and in some cases earlier) compared to the market in the 1960s and 1970s. The chapters on supply show that a largely academic market dominated by native-born men changed into a market dependent greatly on international students and women, and where many doctorates came to work as postdocs in labs before obtaining regular jobs. The chapters on careers highlight science and engineering doctorates working in industry, the increased success of doctorate women in academe and industry, and the impact of the growing supply of foreign-born doctorates on the job market. They identify differences in the nature of work in industry and academe and the way this affects performance. The chapters on outcomes stress the interaction between universities and industry and the determinants of patents in industry, including the contribution of researchers overseas to U.S. patents. Overall, the volume shows the importance of both largely pecuniary/ economic factors and of nonpecuniary factors and the organization of work, and the connections between industry and academe on careers and productivity in supply and demand decisions and market outcomes. In contrast to the late 1950s, the upsurge of concern about the science and engineering job market in the mid-2000s was not sparked by a Sputnikstyle signature event. The country did not face a shortage of scientists or engineers. If rapidly rising pay is the primary signal of a market shortage, the United States lacked CEOs and financiers, professional athletes, and entertainers, not scientists and engineers. As chapter 2 indicates, the earnings of doctorate scientists and engineers increased modestly—less rapidly than the earnings of college graduates in general by most accounts. The postdoctoral experiences through which many young Ph.D.s went were of mixed quality (chapter 3). A substantial number of Ph.D.s found their skills mismatched with their training, producing job dissatisfaction (chapter 7). Employment in science and engineering grew at an annual rate of 3.2 percent from the 1990s to 2004, far above the rate of growth of the work force (Freeman 2006b, exhibit 3), with foreign-born graduates of U.S. institutions contributing greatly to this increase (chapter 4). A cynic might view the burst of concern about the need for more scientists and engineers in the appendix to be a replay of the late-1980s bogus claims that the United States had a shortage of scientists and engineers (Weinstein 1998). At that time the leadership of NSF proclaimed that the country faced an impending shortage of some 675,000 scientists and engineers. It based these claims on extrapolations that were not based on any remotely plausible assessment of the labor market. When the scientific community learned what had happened, there were angry articles and editorials in Science and Nature. The next director of the NSF apologized for the claims.1 The top officials seemingly proclaimed a shortage to induce 1. “[The NSF scarcity study] went on to project the Ph.D. replacement needs would double between the years 1988 and 2006. Based on a number of assumptions, these data were pretty widely interpreted as predictions of a shortage, while there was really no basis to predict a shortage.” (NSF Director Neal Lane, Congressional Testimony, July 13, 1995).

12

Richard B. Freeman and Daniel L. Goroff

more young Americans into science and engineering to lower the cost of scientists and engineers to large firms. There is undoubtedly some self-interest in the expressed concerns about the state of the science and engineering job market in the mid-2000s. The high tech firms who benefit from an ample supply of scientists and engineers were in the lead proclaiming a problem. The senior scientists who employ graduate students and postdocs in their laboratories were members of the various commissions. And the major research universities that would benefit from increased funding for R&D and graduate science education were also in the forefront of discussion. But the reports listed in the appendix are not based on misleading projections of the supply-demand balance for scientists and engineers or on claims of an impending shortage. Most of the reports recognize that scientific careers are less attractive to young Americans than they were in earlier decades. Still, they see problems that can be solved only with an infusion of more talent into science and engineering. Their analyses are based, as best we can tell, on a subtler picture of the role of science and engineering in the economy and in national security than the clear and compelling post-Sputnik shortages or the late 1980s erroneous forecasts—a picture that this volume illuminates. One reason the blue-ribbon commissions and panels worried about the market for scientists and engineers in the mid-2000s is that various metrics show the United States losing its dominance in science and engineering. The U.S. shares of world R&D, papers and citations in scientific journals, science and engineering graduates and workers, are all falling (Freeman 2006a). This is a near-inevitable trend. With 5 percent of the world’s population, the United States cannot maintain the 35 to 45 percent of science and engineering activity that it had at the end of the twentieth century, at least as long as other countries also invest in the modern knowledge economy, as they have done. The European Union has rebuilt and expanded its university system. In 2003 it graduated 56 percent more Ph.D.s in science and engineering than the United States and is on track to graduate twice as many Ph.D.s in the fields as the United States in 2010. The major Asian countries—China, India, Japan, Korea, Taiwan—also expanded their university system to graduate more S&E Ph.D.s in the early 2000s than the United States. By 2010 China will by itself graduate more Ph.D.s in science and engineering than the United States. Like the rest of economic and social life, science and engineering have become increasingly global. But as indicated in chapter 10’s evidence on the contribution of foreign residents to the patents of U.S. firms, the knowledge created by foreign resident scientists and engineers can be used by American firms and inventors to help make technological advances. And the increased number of science and engineering bachelor’s graduates overseas is also likely to mean a continued sizable flow of international students into the United States for graduate study. Whether Ph.D.s trained overseas will be as willing to come to the

Introduction

13

United States for work as U.S.-trained Ph.D.s are willing to move from the state in which they earned their degree to other parts of the United States (chapter 8) is, however, a matter which our data do not address. Some of the commissions and study groups fear that as other countries become more competitive in knowledge production and its application to the economy, the United States will lose comparative advantage in some high-tech sectors and thus gain less from exporting high-tech products. But greater competition should not translate into a fall in the average American’s living standard. There are advantages to increased science and engineering activity around the world that will in the long run benefit virtually everyone. A scientific advance in China, Germany, Brazil, wherever, adds to the stock of knowledge that allows firms to create new products or to reduce the price of existing products. It will benefit consumers regardless of where the ideas are generated or the product is made. Increased knowledge offers the best opportunity for solving the great problems of climate change, global warming, energy efficiency, and disease that affect all people around the world. And the United States has some distinct advantages in turning knowledge into innovation—the close links between universities and business (described in chapter 9 for the development of probe microscopy), and in the flow of Ph.D.s into industrial jobs (examined in chapter 8). The collaborative organization of industry research and development may be more suitable for the rising supply of women Ph.D. scientists and engineers than the academic tournament style model of research. The experience of the late 1990s/early 2000s shows, moreover, that simply increasing R&D spending does not improve the job market for scientists and engineers. Between 1998 and 2003 the U.S. government doubled spending on the National Institute of Health (NIH), but this did not create a boom in the job market for bioscientists. Most of the research awards went to senior scientists, who hired graduate students and newly minted Ph.D.s from the United States and overseas to work as postdocs in their labs. The chances that a young scientist would gain a grant on their own fell to negligible proportions. And with universities hiring few new tenured faculty, the chances for postdocs to move into independent research positions dropped as well. When NIH spending leveled off in the mid-1990s, it was a hard landing for senior scientists, who had greater difficulty than in the past funding their research projects, for postdoctorate fellows who worked in the labs, and for universities who had built up research facilities. The analysis in chapter 3 on the organization of the postdoctoral experience, in chapter 6 on network collaborations in industry and academic research, and in chapter 7 on mismatch bring to the fore the importance of the organization of work in creating a market that makes best use of talent. Going beyond economic issues, some of the commission reports were

14

Richard B. Freeman and Daniel L. Goroff

especially concerned about the number of U.S. citizens in the nation’s science and engineering workforce, particularly in areas where the government hires only citizens, such as the National Security Agency. The huge change in the demographic composition of the U.S. graduate student and doctorate graduate population in science and engineering analyzed in chapters 2 and 4 underlie these concerns. Foreign-born students accounted for slightly over half of U.S. Ph.D. graduates in science and engineering in 2003, which is more than double the proportion in 1966. If, for some reason, the United States became a less attractive place to work for foreignborn scientists and engineers, the United States could indeed face a supplyside problem. The analysis in chapter 1 shows that if the United States wants to increase the supply of citizens in science and engineering, it can do so by offering higher valued or more fellowships for graduate study. All told, the concerns expressed by the blue-chip commissions and study groups go far beyond special pleading by the scientific-education establishment. The concerns are based on interpretations of how science and the economy interconnect, of how globalization of science affects economic performance, and how the job market for scientists and engineers operates, all of which this book illuminates in various ways. The more knowledge we have on these issues the better will we be able to assess the concerns and proposed policies to deal with them. At the same time, as with any research, the volume raises new questions about the economics of the science and engineering workforce and the ways to organize their activities to stimulate innovation and economic growth, which the chapter authors lay out. From the data sets that we analyzed and others there is more to learn about this important job market and the work of scientists and engineers in increasing the stock of useful knowledge and innovation and economic growth.

Appendix Concern about the Science and Engineering Workforce, circa mid-2000s We must “enhance the science-technology enterprise so the U.S. can compete, prosper, and be secure.” National Academy of Sciences (NAS) (2007) The Department of Defense and the defense industry are “having difficulty attracting and retaining the best and brightest students to the science and engineering disciplines relevant to maintaining current and future strategic strike capabilities.” U.S. Department of Defense (2006) “To maintain our leadership amidst intensifying global economic competition, we must make the best use of talented and innovative individuals,

Introduction

15

including scientists, engineers, linguists, and cultural experts. . . . The nation must cultivate young talent and orient national economic, political, and educational systems to offer the greatest opportunities to the most gifted American and international students.” American Association of Universities (2006) “If trends in U.S. research and education continue, our nation will squander its economic leadership, and the result will be a lower standard of living for the American people.” National Summit on Competitiveness (2005) “Together, we must ensure that U.S. students and workers have the grounding in math and science that they need to succeed and that mathematicians, scientists and engineers do not become an endangered species in the United States.” Business Roundtable (2005) “It is essential that we act now; otherwise our global leadership will dwindle, and the talent pool required to support our high-tech economy will evaporate. . . . [n]ot only do our economy and quality of life depend critically on a vibrant R&D enterprise, but so too do our national and homeland security. . . . [a] robust educational system to support and train the best U.S. scientists and engineers and to attract outstanding students from other nations is essential for producing a world-class workforce and enabling the R&D enterprise it underpins.” Task Force on the future of American Innovation (2005) There is “a shortage of U.S. citizen scientists to work in sensitive national security programs.” Lewis (2005) “The message is clear. Today’s relentless search for global talent will reduce our national capacity to innovate unless we develop a science and engineering workforce that is second to none.” Building Engineering and Science Talent 2004 “The United States is facing a crisis in science and engineering talent and expertise . . . For the United States to remain competitive in a vibrant global innovative and research environment, it must . . . attract, educate, recruit, and retain the best S&E workers. Assuring that the nation has the number and quality of scientists and engineers is a national imperative upon which the nation’s security and prosperity rests entirely.” Jackson (2003). “The Federal Government and its agencies must step forward to ensure the adequacy of the U.S. science and engineering workforce. All stakeholders must mobilize and initiate efforts that increase the number of U.S. citizens pursuing science and engineering studies and careers.” National Science Board (2003) “The inadequacies of our systems of research and education pose a greater threat to U.S. national security over the next quarter century than any potential conventional war that we might imagine.” Hart-Rudman Commission on National Security (2001)

16

Richard B. Freeman and Daniel L. Goroff

References Association of American Universities. 2006. National defense education and innovation initiative: Meeting America’s economic and security challenges in the 21st century. Washington, D.C.: Association of American Universities. Building Engineering and Science Talent (BEST). 2004. The talent imperative: Meeting America’s challenge in science and engineering. San Diego, CA: BEST. Business Roundtable Task Force. 2005. Tapping america’s potential: The education for innovation initiative. Washington, D.C.: Business Roundtable. Freeman, R. 2006a. Does globalization of the scientific/engineering workforce threaten U.S. economic leadership? In Innovation Policy and the Economy, vol. 6, ed. A. B. Jaffe, J. Lerner, and S. Stern, 123–58. Cambridge, MA: MIT Press. ———. 2006b. What! Another science-engineering work force crisis? American Association for the Advancement of Science (AAAS) Meeting. Washington D.C.: AAAS. Freeman, R., E. Jin, and C.-Y. Shen. 2004. Where do new U.S.-trained scienceengineering PhDs come from? NBER Working Paper no. 10554. Cambridge, MA: National Bureau of Economic Research, June. Hart-Rudman Commission on National Security, Roadmap for National Security. Available at http://www.au.af.mil/au/awc/awcgate/nssg. Jackson, S. A. 2003. Envisioning a 21st century science and engineering workforce for the United States: Tasks for university, industry, and government. Report to the Government-University-Industry Research Roundtable (GUIRR). Washington, D.C.: The National Academies Press. Lewis, J. A. 2005. Waiting for Sputnik: Basic research and strategic competition. Prepublication copy (available at http://www.csis.org/media/csis/pubs/051028). Washington, D.C. Center for Strategic and International Studies. National Academy of Sciences. 2007. Rising Above the gathering storm: Energizing and employing America for a brighter economic future. Available at http:// www.nap.edu/catalog/11463.html. National Science Board. 2003. The science and engineering work force: Realizing America’s potential. Washington, D.C.: National Science Foundation. Available at http://www.nsf.gov/nsb/documents/2003/nsb0369/nsb0369.pdf. National Summit on Competitiveness. 2005. 6 December, Washington, D.C., p. 2. Available at http://www.nam.org/~/media/Files/s_nam/docs/235900/235820.pdf .ashx. Task Force on the Future of American Innovation. 2005. The knowledge economy: Is the United States losing its competitive edge? Washington, D.C.: Task Force on the Future of American Innovation, February. U.S. Department of Defense. 2006. Report of the defense science board task force on future strategic strike skills. Washington, D.C.: Office of the Under Secretary of Defense, March. Weinstein, E. 1998. How and why government, universities, and industry create domestic labor shortages of scientists and high-tech workers. Working Paper. Available at http://www.nber.org/~peat/PapersFolder/Papers/SG/NSF.html.

1 Supporting “The Best and Brightest” in Science and Engineering NSF Graduate Research Fellowships Richard B. Freeman, Tanwin Chang, and Hanley Chiang

Stipends—payments to students as part of their studies—are a major form of income for graduate students in science and engineering (S&E) and a potential determinant of decisions to undertake graduate study.1 Broadly defined to include fellowships, research assistantships, teaching assistantships, and postdoc awards, stipends account for upwards of one-fourth of the lifetime incomes of S&E Ph.D.s, depending on field, time to degree, and prevalence and length of postdocs.2 In addition, fellowships are potentially the most attractive stipend because they signal to students that the granting institution views them as top prospects for success in graduate study and is willing to support their studies rather than requiring them to work Richard B. Freeman holds the Herbert Ascherman Chair in Economics at Harvard University, and is the director of the labor studies program at the National Bureau of Economic Research. Tanwin Chang is an optical scientist at Adaptive Optics Associates. Hanley Chiang is a researcher at Mathematica Policy Research, Inc. Supported by a grant from the NSF in collaboration with the Council of Graduate Schools (CGS). While serving as an AAAS/NSF Science and Technology Fellow, Dan Stanzione was instrumental in facilitating this study. Jason Abaluck of Harvard assisted in preparation of the data. We thank Myles Boylan of the NSF for a thoughtful review of a short version of this work published in the AEA proceedings. 1. In the 1990s, about two-thirds of S&E grad students received stipend support of some form and thus did not have to rely on loans, personal finances, or money from family and friends as their primary means of support (National Science Board 2004). 2. Fellowships pay much less than faculty jobs, but the fact that they are earned early in the career greatly augments their value relative to earnings from full-time employment. Using a five percent discount rate, a $30,000 fellowship received at the first year of graduate study is comparable to $49,000 received ten years later, $79,500 received twenty years later, and $129,300 received thirty years later. The discounting is such that scientists who did ten years of study and postdoctoral work and then worked for thirty years would earn 29 percent of their lifetime income during their fellowship years if their annual salary was twice that of their fellowship and would earn 21 percent of their lifetime income during their fellowship years if their annual salary was three times that of their fellowship.

19

20

Richard B. Freeman, Tanwin Chang, and Hanley Chiang

to cover graduate school expenses. In 2001 13 percent of full-time graduate students reported that a fellowship or traineeship was their primary source of graduate support (National Science Foundation 2001a). Slightly more than one-third of those, or approximately 4 percent of full-time graduate students, obtained awards from federal government sources (National Science Board 2004). Since its establishment in 1952, the National Science Foundation’s (NSF) Graduate Research Fellowship (GRF) has been the United State’s premier award for science and engineering graduate students. The NSF’s stated purpose for the Graduate Research Fellowship is to “ensure the vitality of the scientific and technological workforce in the United States and to reinforce its diversity.” Because NSF fellowships are limited to citizens or permanent resident aliens, they are a potentially important policy tool for inducing citizens and residents, including those from underrepresented groups, to study S&E and thus to raise their representation in the S&E workforce. What determines the likelihood that an applicant wins a GRF award? Has the composition of applicants and awardees changed over time? What is the supply response of students to increases in the number or the amount of fellowships? How might alternative fellowship policies affect the numbers and measured attributes of awardees? Can we generalize from the response of students to GRFs to the possible impact of stipends on the supply of students to S&E more broadly? This chapter uses data from the NSF’s 1952 to 1993 Cumulative Index (CI) file, its ensuing updates, and other sources to examine these questions. The Cumulative Index plus updated data provides information on the over 200,000 individuals who applied for NSF graduate fellowships from 1952 through 2004, including the student’s graduate record exam scores (GREs), grade point average (GPA), ratings of reference letters, (and from 1976 on), the ratings of applicants by review panels. We supplement these data with information on the earnings of college graduates, unemployment rates, and the sizes of graduating bachelor’s cohorts. To determine the effects of changes in the number of GRFs and the amount of awards on the number of applicants and quality of awardees, we use time series fluctuations in the awards as the exogenous variation necessary to identify student supply responses.3 Section 1.1 describes the GRF program and how it has changed over 3. The 2002 report on the GRF by WestEd looks at what happens to fellows rather than what attracts students to the fellowships. It shows that in the time period leading up to 1993, around 70 percent of fellows received their baccalaureate from RU1 institutions (as defined by the Carnegie Classification), and more than 90 percent went on to perform graduate work at RU1 institutions. Some programs receive a disproportionately high number of fellows: 34 percent of Life Sciences Fellows enrolled in five institutions. The report shows that eleven years after beginning their Ph.D. program NSF Fellows had completion rates of around 70 percent, which exceeds the overall completion rate for entering students; that fellows pursued

Supporting “The Best and Brightest” in Science and Engineering

21

time. It finds that the ratio of GRFs to S&E bachelor’s degrees fell by nearly 75 percent from the 1950s to the 1990s/2000s; that the distribution of awards shifted from physics and chemistry to biological sciences, social sciences, and engineering, and that women and minorities have increased their share of awards. Section 1.2 uses a linear probability model to estimate the characteristics that help an applicant win a GRF. It finds that winning depends greatly on GREs, GPAs, and reference letters operating through panel ratings, and that, consistent with NSF’s diversity goal, women and minorities have higher chances of winning an award than white men with similar attributes. Section 1.3 shows that an increase in stipend value increases the number of applicants, including those with very high skills, and thus raises the average skill of awardees. By contrast, increases in the number of awards are associated with modest declines in the average quality of awardees. Section 1.4 explores new ways to determine the number and value of GRF awards, such as increasing the number and/or value of awards; indexing the number and value to measures of the supply of potential applicants and to the monetary attraction of alternative careers; and giving awards by the measured academic and personal characteristics of applicants, irrespective of field. 1.1 The GRF Program Over Time Congress established the National Science Foundation’s Graduate Research Fellowship program in 1952 to support the vitality of the scientific and technological workforce in the United States. The program was open to U.S. citizens only. It gave a $1,600 stipend and covered normal tuition and fees at the institution of choice. The NSF typically supported students for three years but this was not automatic; students had to reapply each year. The GRF has changed since its inception.4 In 1963 it capped the costof-education, and in 1972 students no longer had to reapply (Goldsmith, Presley, and Cooley 2002). Today awardees may choose three years of support over a five-year period. They must be pursuing a research-based graduate degree in an NSF-supported field. The fellowship may be used outside the United States and is portable across institutions. In 1997 the NSF introduced new selection criteria that put greater weight than in the past in its awarding fellowships to “broader impacts” such as integration of research and education, diversity, benefit to society, and enhancement of scientific and technical understanding. academic careers more than their peers, and that a slightly larger percentage reported having achieved traditional measures of accomplishment during graduate school (such as presentations, refereed articles, edited book parts, etc.). 4. National Science Foundation, Graduate Research Fellowship Program (2007) gives the details of the program as of 2007. The details change yearly with NSF policies.

22

Richard B. Freeman, Tanwin Chang, and Hanley Chiang

Fig. 1.1

The number and value of GRF awards vary over time.

Source: NSF, Division of Graduate Education, Cumulative Index of the GRF Program and related datasets.

The number of awards and the dollar value of stipends have changed over time with NSF budgets and decisions. Figure 1.1 shows the absolute number of awards and constant dollar value of the stipends from 1952 to 2004. In the 1950s the NSF gave around 500 to 600 awards. In the 1960s it gave around 1,000 awards, with a peak of 1,373 in 1966. In the 1970s and 1980s the number of awards fell to around 500, then rose in the late 1980s/ early 1990s toward 900 to 1,000. The big jumps in the value of awards reflect NSF decisions to raise nominal stipend amounts. The NSF raised the value of fellowships in the early 1980s to the late 1980s and again in 1999 to 2004.5 In 1999, the Committee of Visitor’s (COV) report evaluating the program noted that “the GRF awards are no longer as attractive as they once were” (Committee of Visitors 1999). At $15,000 per year for three years the awards were significantly lower than comparable fellowships. The COV speculated that this could (at least partially) explain the steady decline in applicants in the years leading up to 1999 and recommended that the stipend value be raised to $18,000. In the subsequent four years, the value of the stipend rose 83 percent to $27,500, which prompted the next COV (convened in 2003) to worry that the stipend was now too large. 5. The gradual declines in the stipend values in years when the NSF has not raised them reflect the impact of inflation. The NSF has never reduced the nominal value of awards.

Supporting “The Best and Brightest” in Science and Engineering

Fig. 1.2

23

The number of awards per S&E baccalaureate has shifted downwards.

Source: NSF-DGE, Cumulative Index and related Datasets. Bachelor’s degree data tabulated by National Science Foundation/Division of Science Resources Studies; data from Department of Education.

Universities complained about the negative effects of a support gap between NSF fellows and other students. The 2003 COV recommended that the “NSF thoroughly examine the external consequences of the $27,500 stipend and what intended and unintended consequences this might engender” (COV 2003, 13). For the 2005 fiscal year, the stipend has been set at $30,000, with an accompanying $10,500 cost of education allowance. The roller-coaster behavior of the award amount has prompted some in the graduate education community to seek a consistent basis for the determination of stipend levels. The absolute number of GRF awards of about 900 to 1,000 in recent years is comparable to the levels in the 1950s. Relative to the increased supply of S&E bachelor’s graduates, however, the number of awards has fallen sharply. In 1952 to 1956 NSF awarded 5.4 GRFs per thousand S&E bachelor’s degrees. In 2000 to 2004 NSF awarded 2.2 GRFs per thousand S&E bachelor’s degrees—a 60 percent drop. Figure 1.2, which graphs the number of awards per S&E bachelor’s, shows that the drop occurred discontinuously around 1970.6 It also shows sizable fluctuations in the ratio of 6. Since only about 7 percent of U.S. bachelor’s degrees in science and engineering go to the foreign-born, with little trend over time, it makes little difference if one uses all bachelor’s graduates or only nonforeign-born bachelor’s as the base for this calculation.

24

Richard B. Freeman, Tanwin Chang, and Hanley Chiang

awardees to BS graduates due to changing numbers of awards granted before 1970. 1.1.1 Field and Demographic Distribution of Awards In the early years of the program, GRF awards went largely to physical sciences and mathematics and disproportionately to white men. Today, larger shares go to biological sciences, social sciences, and engineering, and also to women and minorities. Figure 1.3 shows the huge change in the distribution of awards among fields. From 1952 to 1956 the physical sciences obtained 49 percent of the awards, math received 9 percent, and life sciences obtained 25 percent compared to 4 percent in psychology and social sciences and 13 percent in engineering. In 2004, by contrast, 15 percent of awards were given to the physical sciences, 11 percent to mathematics and computer sciences, and 27 percent to life sciences; 30 percent were given in engineering and 17 percent in social sciences and psychology. Given NSF policy to award approximately the same proportion of applicants among fields, the shift in awards among disciplines largely reflects a change in the distribution of applications among disciplines. Consistent with this, figure 1.4 shows that before 1990 the proportion of applicants who won awards in the major discipline categories were within a few percentage points of each other, although social science and psychology were

Fig. 1.3

The GRF disciplinary distribution over time of applicants

Source: NSF, Division of Graduate Education, Cumulative Index of the GRF Program and related datasets.

Supporting “The Best and Brightest” in Science and Engineering

Fig. 1.4

25

Percentage of applicants who gain GRF awards, selected years

Source: NSF, Division of Graduate Education, Cumulative Index of the GRF Program and related datasets.

slightly disfavored. This is still the case after 1990, except that applicants in engineering and computer science have a higher chance of getting an award due to the process by which the NSF seeks to attract women to these disciplines. Turning to the demographic characteristics of awardees, figure 1.5 shows a sharp rise in the proportion of women and underrepresented minorities7 winning the GRF since the program was initiated. The female proportion of awardees climbed from about 5 percent in the 1950s to 55 percent of awards in 2004, with considerable variation among fields. Women make up roughly three-quarters of the awardees in life sciences, social science, or psychology compared to 20 percent of awardees in physical science or computer science. Around 1960, the female proportion in life science was about 25 percent, and in physical science it was around 6 percent, while there were barely any female engineering fellows.8 To increase the female representation in these fields, NSF set up the Women in Engineering Program (WENG) in 1990, which became the Women in Engineering and 7. Underrepresented Minorities in S&E include African Americans, Hispanics, and Native Americans. Asian Americans are overrepresented in science and engineering. 8. In 1960 there were few computer science fellows, either male or female.

26

Richard B. Freeman, Tanwin Chang, and Hanley Chiang

Fig. 1.5

Percentages of fellows who are women or URMs

Source: NSF, Division of Graduate Education, Cumulative Index of the GRF Program and related datasets.

Computer Science (WECS) program in 1993. This program boosts the number of females by awarding fellowships to all or nearly all computer science or engineering women in NSF’s quality group 2, and by making women in quality group 3 also eligible for awards if enough slots are available.9 The data in figure 1.4 show that in 2001 applicants in computer science and engineering had a higher ratio of awards to applicants than other disciplines as a result of this program. The data in figure 1.5 on the proportion of underrepresented minorities (URM) among GRF fellows starts in 1976 due to lack of ethnicity data before then. The GRF data is further complicated by the development in 1978 of the Minority Graduate Fellowship Program (MGF), which gave awards solely to underrepresented minorities, and its dissolution in 1998. The MGF was officially a separate program, although executed in parallel with the GRF. In 1999, the first program year since 1975 without the MGF, the NSF emphasized procedural changes in its review panel process in the 9. As described following, NSF categorizes applicants by group depending on their measured scholastic skills. All quality group 1 applicants are offered GRFs, and a sizable proportion of group 2 applicants are also given GRFs.

Supporting “The Best and Brightest” in Science and Engineering

27

hope of maintaining a diverse awardee pool, partly by attracting more applicants through expanded outreach, and partly via other measures such as insuring a diverse group of panelists (COV 2003). Combining the fellows from the GRF and the MGF, the URM proportion is around 10 percent in the early 1980s, and climbed to 30 percent in 1998, the last year of the program. After that, the percentage dropped to the GRF-only proportion of around 10 percent, which is still much larger than the 1 to 2 percent two decades earlier. The disciplinary variation for URM fellows shows a disproportionately large proportion in social science and psychology. The increased diversity in GRFs reflects two factors. First is that as a result of these programs and efforts, the likelihood that a woman or minority applicant would get an award has risen. In the first decade of the GRF, the overall award rate10 was 17 percent while the award rate for women was 12 percent, as shown in figure 1.6. By the mid-1990s the female award rate exceeded the overall rate. The award rate for URMs in the GRF was low at 3 percent during the run of the MGF. The combined rate of URMs in both the GRF and MGF topped 18 percent from 1994 to 1998, which is actually greater than the MGF-only award rate, 15.3 percent, for the same time period.11 After the end of that program, the number of URM awardees per applicants approached the overall rate of awardees per applicant, though it is still noticeably lower by several percentage points. The second and more important reason for the increased diversity of GRFs is that an increasing share of bachelor’s degrees in science and engineering have gone to women and minorities. This has raised the proportion of applicants from these groups, which in turn produces an increased share of awards, coupled with the NSF’s efforts to improve diversity. 1.2 Determinants of GRF Awards The GRF is a merit-based award. Each year, the NSF assembles a panel of experts to evaluate and rate applications and arrange them into quality groups. Quality Group 1 (QG1) is the highest quality group, and all applicants deemed fit for this group are offered the fellowship. Quality Group 2 (QG2) applicants are regarded as sufficiently meritorious for the fellowship, but not all in this category will receive the award. Applicants in other quality groups rarely receive the fellowship, but may receive an honorable mention, which carries some nonmonetary prestige. As with other aspects of the program, the procedures for determining awardees have evolved. Presently, panelists meet in groups arranged by discipline for an intensive two to three 10. “Award rate” is synonymous with “awardee to applicant ratio.” 11. The explanation for this result lies in the high incidence of duplicate applications. Data was available from 1994 to 1998 to match applicants in the two programs, and we found that approximately 90 percent of those who applied for the MGF also submitted an application for the GRF.

28

Richard B. Freeman, Tanwin Chang, and Hanley Chiang

Fig. 1.6 Average percentage of applicants winning awards in selected time periods, by broad demographic group Source: NSF, Division of Graduate Education, Cumulative Index of the GRF Program and related datasets. Notes: Data on ethnic demography only available beginning in 1976. From 1994–1998, the available data allows the identification of duplicate applicants to both the MGF and GRF. To calculate “URMs in GRF+ MGF” the total number of unique fellows in either program was divided by the total number of unique applicants. The final year of the MGF program was 1998.

day session, which includes a briefing on the GRF program and a calibration exercise using applications from previous years. The categorization process consists of cycles during which the panelists score the applications on a scale from 1.00 to 4.99 (the lower number indicating higher quality), and then rank them. In the first cycle, at least two panelists score each application. Any applicants ranked below the sixty-fourth percentile are put in QG4 and no longer considered for the fellowship. A third panelist scores remaining applications and the applications are reranked. Approximately 55 percent of the awardees are from QG1, all of whom receive an award. Virtually all of the other awards are from QG2, with approximately 60 percent of QG2 applicants offered the fellowship (NSF 2002). From 1998 to the present, NSF has instructed panelists to evaluate applications based on the two broad criteria common to NSF solicitations: intellectual merit and broader impacts. The panelists are given rating sheets upon which intellectual merit is broken down into subcriteria: GPA, GRE scores, references, proposed research, and previous research experience. Broader impacts are based on integration of research and education, diversity, benefit to society, and enhancement of scientific and technical un-

Supporting “The Best and Brightest” in Science and Engineering

29

derstanding. The NSF emphasizes a comprehensive concept of merit, and advises panelists to avoid placing too much emphasis on the easily quantifiable quality measures such as GRE scores and GPA. In the latest program years, GRE scores are not required on the GRF application. Further, there is a new emphasis on the value of leadership skills in NSF fellows (COV 2003). It is possible that future applicants will be asked to provide evidence of leadership potential, and future panelists will include this in their assessment of merit. Because the NSF recognizes the value of a diverse workforce, the program implements measures to insure representation in geography, demographics, and discipline. Specifically, the applicants from Quality Group 2 who receive the fellowship are chosen based on an algorithm that includes their rank within QG2, but also other factors. The key policy with respect to discipline is that NSF allocates the number of QG1 and QG2 positions to review panels on the basis of the disciplinary distribution of all applicants and the number of awards NSF plans to offer. 1.2.1 Regression Analysis To see how the selection procedure uses the information on student skills to determine which applicants win awards, we estimated linear probability models in which the dependent variable is the 0/1 measure of obtaining an award and the explanatory variables consist of measures of student skills from the NSF’s Cumulative Index data set. The CI has recorded the scores that a panelist gives to an application starting in 1976, but does not record assessments of the subcriteria, which limits our choice of measures largely to the quantitative and verbal GRE scores, the GPA, and the scoring of reference letters. Table 1.1 lists the measures of scholastic skills in the data and gives their mean values and standard deviations, as well as the overall mean of the probability of winning the GRF award, and the scales used to measure each quantity. The GRE scores are coded on a 200 to 800 scale; the GPA uses a four point scale; reference scores were originally scaled from ten to seventy, with lower scores being better; finally the Panel Ratings12 (which reflect the GRE, GPA, and reference scores) were originally coded on a scale of 100 to 600, again with lower score representing higher quality applicants. The panel ratings and reference scores changed scales at various times in GRF history. To deal with this, we normalized the panel and reference data such that they have zero mean and standard deviation of one in the applicant pool of each year. We also reverse-coded the actual panel and reference scores, so that increases in all variables imply a more favorable situation. Regression coefficients for GRE scores are sometimes re12. In the original Cumulative Index data set, each applicant may have more than one panel rating if more than one panelist reviewed the application. An average panel rating was trivially generated for each applicant. In this chapter, “panel rating” means “average panel rating.” Similarly, “reference score” is synonymous with “average reference score.”

30

Richard B. Freeman, Tanwin Chang, and Hanley Chiang

Table 1.1

Means, standard deviations, and scaling of the quantitative measures of applicant scholastic skills Total time period

Records

Overall award rate

1952–2004

271,391

13.50%

Data available

Mean

STD

Full range

1976–1993 1994–1998 1999 2000–2004 1954–1993 1994–1998 1952–2004 1952–2004 1955–1998

328 315 3.11 2.90 24.2 1.91 691 606 3.49

115 118 1.02 0.85 8.7 0.65 100 104 0.40

100 to 600 100 to 600 1.00 to 6.00 1.00 to 4.99 10 to 70 1 to 7 200 to 800 200 to 800 1 to 4

Summary for key data in the GRF datafiles

Panel ratings

Reference score GRE quantitative GRE verbal GPA

Ethnicity

Gender

1976–2004 1981–1987 1988–2000 2000–2004 1952–2000 2001–2004

(1, 2, 3, 4, 5) (1, 2, 3, 4, 5, 6, 7, 8, 9) (A, B, C, D, E, F, H, I, J) (t, f) or (y, n, b) (m, f) (m, f, b)

Source: NSF, Division of Graduate Education, Cumulative Index of the GRF program and related data sets. Notes: For analysis, Panel ratings and Reference scores will be normalized such that they have mean zero and standard deviation one in the applicant pool of each year. These variables are also recoded such that higher values denote better ratings. For GRE quantitative and GRE verbal, values in original data set above 800 set to 800. For GPA, full range (0 to 3) before 1971, values shifted by  1. A separate variable codes each ethnic category. For Gender, value of “b” indicates “decline to answer.”

ported in tables with the scores scaled as GREQuant/100 or GREVerbal/ 100, which effectively puts them on a two to eight point scale.13 Note in the table that applicants have higher GRE quantitative scores than they have GRE verbal scores, so that while the standard deviations of scores are similar, the coefficient of variation is lower for the quantitative score than for the verbal score. Table 1.2 records linear regressions of the impact of scholastic attributes, demographic factors, and field dummies for nearly the full data set from 1955 to 1998, which lacks measures of minority status, panel scores, and reference letters; and for the 1976 to 1998 period when we have data on those measures as well as the others. The column (1) and (2) regressions do not include the panel rating, while the column (3) regression includes the panel rating. The column (1) estimates show that over the entire period, GRE Quantitative, GRE Verbal, and GPAs have sizable and significant 13. This allows us to report coefficients concisely as 0.XX rather than 0.00XX.

Supporting “The Best and Brightest” in Science and Engineering Table 1.2

31

Linear probability model for the impact of scholastic and demographic variables on GRF awards Offered Award

Panel rating Reference score GRE quantitative/100 GRE verbal/100 GPA Female Minority Field effects: Chemistry Computer science Engineering Earth/atmospheric Life science Math Physics/astronomy Psychology Social science Year effects Observations R2

1955–1998

1976–1998

1976–1998

—

—

0.082 (0.001) 0.034 (0.001) 0.060 (0.001) 0.089 (0.002) —

0.074 (0.001) 0.035 (0.001) 0.066 (0.001) 0.095 (0.003) 0.033 (0.002) 0.087 (0.004)

0.170 (0.001) 0.022 (0.001) 0.010 (0.001) 0.019 (0.001) 0.010 (0.003) 0.037 (0.002) 0.077 (0.003)

0.003 (0.003) 0.007 (0.004) 0.022 (0.003) 0.033 (0.004) 0.037 (0.002) 0.019 (0.003) 0.030 (0.003) 0.000 (0.003) —

0.011 (0.004) 0.009 (0.005) 0.048 (0.004) 0.034 (0.005) 0.034 (0.003) 0.012 (0.005) 0.006 (0.005) 0.005 (0.005) —

0.045 (0.004) 0.115 (0.005) 0.093 (0.003) 0.049 (0.005) 0.040 (0.003) 0.051 (0.005) 0.060 (0.004) 0.018 (0.004) —

Yes 207,498 0.198

Yes 107,658 0.1793

Yes 107,597 0.2731

—

Source: NSF, Division of Graduate Education, Cumulative Index of the GRF program and related data sets.

effects on the probability of obtaining the GRF. A 100 point increase in Verbal GRE score raises the probability of award by 6.4 percent. This might not seem large, but since just 13.5 percent of applicants get an award this raises the probability of getting an award by nearly 50 percent. The higher effect of the Verbal GRE than of the Quantitative GRE may seem

32

Richard B. Freeman, Tanwin Chang, and Hanley Chiang

surprising, given the presumably greater importance of quantitative thinking in science and engineering. It reflects the fact that the Verbal GRE is often the key distinguishing factor between applicants with similarly high Quantitative GREs. Finally, the calculations show that women and minorities have higher chances of getting an award than other applicants. The field dummy variables show some variation in the probability of getting an award among fields. Engineering, life science, and earth/atmospheric science have significantly higher probabilities of getting an award while mathematics has a lower one compared to the omitted social science group. The results in column (2), which include the measure of the quality of references and dummy variables for gender and minority status for the period 1976 to 1998 are similar: GREs, the GPA, and the reference scores are key determinants of the probability of getting an award, while women and minorities have a higher probability of gaining awards. The dummy variables for fields show moderate disciplinary variation. Column (3) adds the panel rating to the list of variables. Since the NSF panels weigh heavily the effect of GREs, GPAs, and reference scores in deciding which applicants should be given awards, inclusion of the panel rating alters the estimated coefficients markedly. The coefficients on factors on which panels place most weight in their evaluation fall relative to those that enter less highly in the panel rating. The GRE quantitative score and the GPA obtain negative coefficients conditional on the panel rating, implying that the panels weigh those factors especially heavily in their ratings. In addition, conditional on the panel ratings the field dummies become large and significant relative to the omitted social science group. This presumably reflects the initial allocation of the number of QG1 and QG2 positions on the basis of the disciplinary distribution of applicants, which leads panels in the quantitative and physical sciences to give lower ratings than those in the social sciences to applicants with comparable measured scholastic skills. From this perspective, the estimated positive coefficients on the field dummies in table 1.1 are a correction that takes account of the attributes of applicants in these fields. Consistent with this interpretation, table 1.3 shows that panel ratings are indeed lower for applicants with the same measured scholastic skills in physical sciences than in the social sciences. For instance, the regression shows that an applicant in chemistry obtains a –0.198 lower panel rating than an applicant in the omitted social science group with the same GREs, GPA, reference score, and demographics. The rank correlation of the ordering of fields between the table 1.3 estimates of the panel ratings and the ordering of fields in the column (3) of table 1.2 is –0.92 (inclusive of the omitted group). Thus, the modest estimated impacts of fields in columns (1) and (2) of table 1.2 reflect two offsetting factors: lower panel ratings in physical sciences but an offsetting higher probability of getting an award conditional on panel ratings in those disciplines.

Supporting “The Best and Brightest” in Science and Engineering Table 1.3

33

Regression coefficients and standard errors for scholastic and demographic determinants of average panel rating, 1976–1998 Average panel rating 1976–1998 Reference score GRE quantitative/100 GRE verbal/100 GPA Female Minority Field effects: Chemistry Computer science Engineering Earth/atmospheric Life science Math Physics/astronomy Psychology Social science Year effects Observations R2

0.304 (0.002) 0.268 (0.003) 0.279 (0.002) 0.614 (0.006) 0.024 (0.004) 0.057 (0.007) 0.198 (0.009) 0.629 (0.010) 0.263 (0.007) 0.092 (0.011) 0.038 (0.006) 0.367 (0.010) 0.386 (0.009) 0.079 (0.009) — Yes 107,597 0.6079

Source: NSF, Division of Graduate Education, Cumulative Index of the GRF program and related data sets.

1.2.2 The Effect of Demographics As noted, the table 1.2 regressions for obtaining a GRF show that women and minorities have a greater probability of getting an award than do majority men with similar measured scholastic skills. This is consistent with the goal of increasing the diversity of the S&E workforce. The table 1.3 regression of the panel ratings on scholastic skills and demographic characteristics show, however, that panels give lower scores to women than to majority men with the same measured attributes while they give higher

34

Richard B. Freeman, Tanwin Chang, and Hanley Chiang

scores to minorities than to majority men with the same measured attributes. Since the probability of getting an award is higher for both women and minorities than for majority men, the difference in coefficients in panel ratings suggests that the process of taking account of diversity differs between the two groups. To unpack this difference, we estimated a recursive model. First, we related the panel scores and the probability of being assigned to groups 1 and 2 to demographic characteristics. Then we examined the probability of getting an award conditional on being in groups 1 and 2 and on the panel rating. Table 1.4 gives the results of this analysis for 1990 to 1997 and 1998 to 2004. The data for 1990 to 1997 includes measures of GRE scores, GPA, Table 1.4

Regression coefficients for determinants of panel rating, quality group, and award offer, 1990–2004 1990–1997

Female Minority Panel rating Quantitative/100 Verbal/ 100 GPA Reference Field effects Year effects Observations R2

Panel rating (1)

Assigned quality group 1 or 2 (2)

0.012 (0.006) 0.003 (0.009) —

0.021 (0.003) 0.076 (0.004) —

0.263 (0.004) 0.235 (0.003) 0.634 (0.010) 0.289 (0.003) Yes Yes 47,821 0.5882

1998–2004 Award conditional on being group 1 or 2 (3)

Award conditional on being group 1 or 2 (6)

Panel rating (4)

Assigned quality group 1 or 2 (5)

0.135 (0.009) 0.161 (0.015) —

0.024 (0.004) 0.091 (0.006) —

0.035 (0.002) 0.074 (0.002) 0.118 (0.004) 0.093 (0.002)

0.38 (0.014) 0.072 (0.041) 0.275 (0.039) — — — — — — — —

0.334 (0.007) 0.285 (0.005) — — — —

0.081 (0.003) 0.093 (0.002) — — — —

0.25 (0.012) 0.073 (0.021) 0.546 (0.028) — — — — — — — —

Yes Yes 47,851 0.2109

Yes Yes 4,904 0.1445

Yes Yes 41,404 0.215

Yes Yes 41,404 0.0901

Yes Yes 5,695 0.1574

Source: NSF, Division of Graduate Education, Cumulative Index of the GRF program and related data sets. Notes: For comparison with the 1998–2004 analyses, we estimated the 1990–1997 equations excluding GPA and reference scores. The coefficients (standard error) on the dummy for female barely changed: it was 0.012 (.006) in the panel rating equation and 0.026 (.003) in the equation for being assigned to group 1 or 2. For minorities, the coefficient on the dummy for minority status in the panel rating became 0.241 (.011). The implication is that GPA and reference scores for minority applicants were much lower than for others. But the probability of being assigned to group 1 or 2 gave to minorities a positive, though smaller, coefficient than in the table, .018(.003).

Supporting “The Best and Brightest” in Science and Engineering

35

and a measure of the quality of reference letters. The 1998 to 2004 data do not include GPA or the quality of reference letters, and cover the period when NSF revised its merit criteria. Column (1) shows that in 1990 to 1997 women received a slightly lower score than men from review panels while minorities received comparable panel ratings as the majority group. Women also were less likely to be assigned to quality groups 1 and 2 than men, while by contrast minority applicants were more likely to be assigned to those groups. Among persons in groups 1 and 2, and conditional on panel ratings, women were much more likely to obtain awards than men while minority applicants had a modestly higher probability of obtaining awards. Thus, the higher probability of women getting awards occurs at the last stage, presumably in part because the WECS (WENG) program boosted the chances that women in quality group 2 would get GRFs, whereas the higher probability of minorities getting awards occurs because panels were more likely to assign them into groups 1 or 2. To make sure that this interpretation is not marred by the availability of GPA and reference scores in the 1998 to 2004 period but not earlier, we estimated the 1990 to 1997 equations excluding those variables. As noted in the table note, the coefficient on females barely changed with their exclusion. The 1998 to 2004 regressions show a very different pattern, indicating that the change in NSF’s policies in awarding GRFs had a substantial effect on the awards process. Column (4) shows that panels give both women and minorities higher ratings in 1998 to 2004 than in 1990 to 1997. Column (5) shows that panels are more likely to assign them to quality groups 1 or 2 than in 1990 to 1997. The coefficient on being female in column (6) shows that, conditional on being in groups 1 or 2, the boost given to women in the probability of getting an award was more moderate in 1998 to 2004 than in 1990 to 1997, presumably because they were getting higher panel ratings and having a higher chance of being assigned to quality groups 1 or 2. By contrast, the coefficient on being minority in column (6) is nearly identical to the coefficient in column (3). Finally, panel ratings have a stronger impact on awards in 1998 to 2004 than in 1990 to 1997. In sum, table 1.4 provides evidence for the change in NSF’s emphasis that was enacted, at least partly, to increase diversity in S&E fields. Finally, table 1.5 uses a regression model to examine the GRE scores of applicants with different demographic characteristics from 1976 to 2004— a period in which the proportion of applicants who were women or minority increased substantially. In 2004 there were 8,939 applicants, of whom 46 percent were women and 10 percent underrepresented minorities. In 1976, there were 5,366 applicants, of whom 30 percent were women and 4 percent were underrepresented minorities. To the extent that increased numbers of applicants draw persons with decreasing measured scholastic skills, one would expect that the increased proportion of women and minorities would reduce the average scores of those groups. The coefficients and standard errors on the demographic variables in the

Yes Yes 41,045 0.3688

36.3 (0.9) 166.5 (1.8) 98.0 (2.0) 73.9 (4.1)

41.7 (1.0) 175.9 (3.2) 77.5 (3.5) 49.2 (8.1)

Yes Yes 30,213 0.3256

1984–1991

1976–1983

Yes Yes 41,804 0.3778

35.8 (0.8) 146.3 (1.4) 87.1 (1.5) 69.1 (3.1)

1992–1998

GRE Quantitative test scores

Yes Yes 36,658 0.2536

28.9 (0.8) 100.1 (1.9) 54.7 (1.7) 44.7 (3.4)

1999–2004

Yes Yes 30,213 0.116

0.0 (1.2) 168.7 (3.8) 80.5 (4.2) 41.0 (9.5)

1976–1983

Yes Yes 41,045 0.1868

10.6 (1.0) 160.1 (2.1) 107.8 (2.3) 69.8 (4.9)

Yes Yes 41,804 0.2273

16.1 (1.0) 146.5 (1.7) 95.9 (1.9) 60.7 (3.8)

1992–1998

GRE Verbal test scores 1984–1991

Demographic correlates of test scores, regression coefficients and standard errors, 1976–2004

Source: NSF, Division of Graduate Education, Cumulative Index of the GRF program and related data sets.

Field effects Year effects Observations R2

Other minority

Hispanic

Black

Female

Table 1.5

Yes Yes 36,659 0.1279

7.2 (1.0) 100.4 (2.4) 70.6 (2.2) 38.7 (4.4)

1999–2004

Supporting “The Best and Brightest” in Science and Engineering

37

table show the opposite pattern, particularly for GRE quantitative scores. The scores of both women and minority applicants rise relative to those of majority men, though both groups still score lower. The coefficient on the dummy for women rises from –41.7 in 1976 to 1983 to –28.9 in 1999 to 2004. The coefficient on being black rises from –175.9 in 1976 to 1984 to –100.1 in 1999 to 2004. The negative coefficient on being Hispanic changes less evenly but is lower in 1999 to 2004 than in 1976 to 1983, while that for other minorities shows little change. The regression coefficients for GRE verbal scores tell a similar story, with the exception that women applicants go from having the same verbal scores as men in 1976 to 1983 to a modest 7.2 lower score in 1999 to 2004. The coefficient for being black, however, advances steadily from –168.7 to –100.4. Overall, the results show that the increased proportion of female and minority GRFs was associated with generally improved measured skills. The implication is that the GRF program is attracting increasingly qualified women and minorities. 1.3 Supply Responses of GRF Applicants The NSF has two policy levers that can affect the supply of applicants to the GRF program and, given the process of granting awards, that can determine the measured skills of awardees: the dollar amount of the awards and the number of awards. Increasing the dollar value of awards should increase the number of applicants. It could increase or decrease the average measured skill of applicants depending on whether they come largely from persons with the highest skills or from those with lower skills. But given the process of selecting awardees, increases in the value of awards should increase the average skill of awardees as long as the increased value of awards attract some persons with high skills. This is because review panels would be expected to select high skill candidates from the pool of newly attracted persons over less skilled candidates that they would have selected from the smaller initial pool of applicants. Increasing the number of awards, by contrast, is likely to reduce the measured quality of awardees since it means accepting applicants who would otherwise be rejected. The key issue here is the extent to which the increase in numbers reduces quality, which depends on the number of applicants on the margin of acceptance with similar skills. If there is a large number of persons on the margin with similar skills, then increasing the numbers of awards will reduce measured skills modestly. If there is a small number, an increased number of awards could reduce measured skills substantially. 1.3.1 Number of Applicants To examine the link between the value of stipends and the pool of GRF applicants, we calculated a relative value of stipends by taking the ratio of the

38

Richard B. Freeman, Tanwin Chang, and Hanley Chiang

dollar value of GRF awards to the average annual earnings of college graduates aged twenty-one to twenty-five14 and contrasted this with the relative number of applicants by dividing the number of applicants by the number of S&E bachelor graduates. As figure 1.7 shows, the two series track each other closely. For example, the sharp increase in the stipend value in the 1980s, when NSF increased stipend amounts from $3,900 (1979) to $14,000 (1991) is matched by a large rise in the number of applicants relative to S&E bachelor’s graduates. By contrast, from 1991 to 1999 the relative value of stipends fell, as the nominal stipend amount rose by $1,000 while college graduate earnings rose by about $7,600; and the relative number of applicants fell. Since 1999 the nominal amount of the stipend has doubled, and the number of applications has nearly doubled as well.15 Another factor that might affect the decision to seek a GRF and go on into graduate school is the ease of finding work with a bachelor’s degree, which varies with the business cycle. Figure 1.8 compares the relative number of applicants to the unemployment rate of college graduates.16 Since the late 1980s, the size of the GRF applicant pool has moved countercyclically: a weak labor market for college graduates generated a higher number of applicants as a share of bachelor’s degrees. But in the early to mid1980s, the unemployment rate fell while the number of applicants rose, presumably in part because of the increase in stipend amounts shown in figure 1.7. Thus, figure 1.8 suggests that unavailability of jobs for recent college graduates is a secondary factor influencing the decision to apply for a GRF stipend. Taking the analysis a step further, we amalgamated the data on GRF fellowships into nine major fields17 by year from 1952 to 2004, to obtain a cross-field time series panel with 477 observations. Using this sample, we did a multivariate regression analysis of the determinants of the number of GRF applicants in each field and year. Since the NSF changes the stipend value in particular years by policy decision rather than changing it in response to the previous years’ labor market, we take these changes as exogenous and use least squares to estimate the effect of relative stipend values on the relative number of applicants. To control for field differences and changes in market conditions by field over time, the regressions in14. We estimated the college graduate earnings from the Integrated Public Use Microdata Series (PUMS) of the March Current Population Survey. College graduates include those who obtain degrees higher than a bachelor’s degree, but the age range is restrictive enough to exclude most doctorate recipients. 15. Bachelor’s degree data for 2002 to 2004 are extrapolated because actual figures are not available. The key to the change in the ratio of applicants to bachelor’s graduates is that the number of applicants jumped from 4,852 in 1998 to 8,939 in 2004. 16. Estimated from the annual Current Population Survey (CPS) Outgoing Rotation Group. 17. The nine fields are: Chemistry, Computer Science, Earth/Atmospheric Science, Engineering, Life Science, Mathematics Physics, Psychology, and Social Science. The availability of earnings data for college graduates and degree data by field limits the sample to 1968 to 2001.

Supporting “The Best and Brightest” in Science and Engineering

39

Fig. 1.7 GRF applications relative to S&E bachelor’s degrees and relative value of GRF stipend, 1968–2004 Source: NSF, Division of Graduate Education, Cumulative Index of the GRF Program and related datasets. Bachelor’s degree data tabulated by National Science Foundation/Division of Science Resources Studies; data from Department of Education.

Fig. 1.8 GRF applications relative to S&E bachelor’s degrees and college graduate unemployment rate, 1979–2004 Source: NSF DGE, Cumulative Index of the GRF Program and related datasets. Bachelor’s degree data tabulated by National Science Foundation/Division of Science Resources Studies; data from Department of Education. Unemployment data are estimated from the annual Current Population Survey (CPS) Outgoing Rotation Group, Bureau of Labor Statistics.

clude field dummy variables and separate time trends for the nine fields. In addition, the regressions include the number of bachelor’s degrees earned in the nine fields in the previous year.18 18. We recognize that students may apply for a GRF in a different field than the field in which they earned their undergraduate degree, but still regard the number of undergraduate degrees as a reasonable indicator of interest in the field.

40

Richard B. Freeman, Tanwin Chang, and Hanley Chiang

Table 1.6

Determinants of the number of GRFP applicants, 1969–2004 Dependent variable: ln(applicants in academic field in current year)

Log(bachelor’s degrees) by field in current year Log(stipend/outside salary) in previous year Unemp. rate for college grads age 21–25 Unemp. rate for all college grads Log(awards/bachelor’s degrees) by field in previous year Field effects Field  time trend Observations R2

(1)

(2)

(3)

(4)

0.195 (0.057)

0.304 (0.063)

0.298 (0.062)

0.516 (0.066)

0.996 (0.084)

0.916 (0.060)

0.852 (0.059)

0.772 (0.056)

— — — —

0.049 (0.013) — —

— — 0.104 (0.024)

— — 0.094 (0.022)

— —

— —

— —

0.349 (0.054)

Yes Yes 324 0.8931

Yes Yes 234 0.955

Yes Yes 234 0.9561

Yes Yes 234 0.9634

Source: NSF, Division of Graduate Education, Cumulative Index of the GRF program and related data sets, as described in text. Notes: Outside salary are earnings of college graduates aged 21–25. Outside salaries and unemployment rates from Current Population Survey.

Table 1.6 displays the regression results. The basic specification in column (1) shows that a larger number of bachelor graduates in a field generates a larger number of applicants from that field; and that stipends in the previous year have a strongly positive effect on applications, with a near unitary elasticity. Column (2) adds to the set of explanatory variables the unemployment rate among college graduates twenty-one to twenty-five years of age as an indicator of the availability of alternative jobs for those contemplating going directly to the job market. The sample size shrinks because the unemployment variable is only available starting in 1979. Consistent with figure 1.8, the unemployment rate has a positive effect on the number of applications in column (2). In column (3) we replace the unemployment rate of twenty-one to twenty-five-year-olds with the unemployment rate for all college graduates regardless of age. The effect of unemployment nearly doubles: a one percentage point rise in the unemployment rate raises the GRFP applications by just over 10 percent, presumably because a one point increase in unemployment for all graduates is a bigger economic shock than a one point increase in unemployment for all graduates is a bigger economic shock than a one point increase for younger grad-

Supporting “The Best and Brightest” in Science and Engineering

41

uates (whose unemployment rate will be more volatile). Finally, in column (4) we include a measure of the relative availability of awards by field—the number of awards divided by the number of bachelor’s graduates in the field lagged one period. This shows that more students apply when the probability of obtaining an award is perceived to be higher. All of the specifications show that the coefficient on the log bachelor’s degrees is substantially below unity. One likely reason for this is that the expansion of degrees in the United States has occurred largely at institutions of lower academic quality, so that the marginal bachelor’s graduate is unlikely to be of GRF quality. In sum, the regression specifications in table 1.6 show that the number of GRF applications responds substantially to the relative value of the stipend and to the perceived availability of awards. For every 10 percent increase in the stipend value, the number of applications goes up by 8 to 10 percent. 1.3.2 Measured Skills of Awardees To see how the value of stipends affects the measured achievement of GRF awardees, we compare the GRE scores and GPAs of awardees over time and across fields. Figure 1.9 provides a first indication that stipend values affect the qualifications of the awardees. Figure 1.9, panel A shows that the average GRE quantitative score of awardees has been higher when the relative stipend value is higher. Figure 1.9, panel B shows a similar relation between stipend values and average GRE verbal scores of awardees. The time series of raw verbal scores trended downward beginning in 1987, so we examined the time series of GRE verbal scores adjusted for trend, as well as the raw time series. The trend-adjusted series shows a rise in scores with the large increases in relative stipend values in the first half of the decade, and a subsequent fall in verbal scores throughout the 1990s. However, the continuing decline in GRE verbal scores through the late 1990s/ early 2000s cannot be explained by the evolution of stipend values, which are rapidly rising during this time. To quantify the pictures in the figures, we regressed measures of the scholastic qualifications of awardees on the number and value of GRF awards, using a cross-field time series regression framework similar to that in table 1.6, with field fixed effects and field-specific time trends included in all regressions. Table 1.7 gives the results of this analysis.19 The first row of coefficients shows that the number of awardees has a significant but quantitatively modest negative effect on the measures of quality. An approximate 10 percent or 0.10 ln increase in the number of awards is associated with a decline in average quantitative GRE scores by 1.1 points and a decline in verbal GRE scores by 2.2 points on the 200 to 800 GRE scale, and a drop 19. Freeman (2005) finds similar results with a different specification that suggests that the basic findings in tables 1.6 and 1.7 are robust to some changes in specifications.

42

Richard B. Freeman, Tanwin Chang, and Hanley Chiang

A

B

Fig. 1.9 Mean GRE scores of awardees and relative stipend value, 1968–2004: A, quantitative scores; B, verbal scores Source: NSF DGE, Cumulative Index of the GRF Program and related datasets. Salary data estimated from the Integrated Public Use Microdata Series (IPUMS) of the March Current Population Survey.

in average GPA by 0.009 points (on a one to four scale). These findings are consistent with the notion that as the number of fellowships increases, the review panels move down the measured scholastic skill distribution to give stipends to marginal awardees that are less outstanding than the inframarginal winners. The larger drop in the verbal scores than in the quantitative scores presumably reflects the wider variation in those scores and the extent to which panels view high quantitative scores as more critical to success in the sciences. The second row of coefficients in table 1.7 shows that, with the number of awards fixed, increases in the relative value of fellowships increases two of our measures of the skill of awardees, but has no discernible

Supporting “The Best and Brightest” in Science and Engineering Table 1.7

43

Determinants of awardee achievement, 1969–2004 GRE quantitative

GRE verbal

GPA

Log(number of awards) by field in current year

10.6 (1.80)

21.8 (3.30)

0.087 (0.01)

Log(stipend/outside salary) in previous year

29 3.3

35.4 6.1

0.007 0.017

Yes Yes 324 0.8943

Yes Yes 324 0.684

Yes Yes 270 0.7354

Field effects Field  time trend Observations R2

Source: Tabulated from NSF, Division of Graduate Education, Cumulative Index of the GRF Program and related data sets, as described in text. Note: Outside salary are earnings of college graduates aged 21–25, tabulated from Current Population Survey.

effect on the third measure. An increase in the relative value of the stipend by 10 percent raises the average quantitative GRE score by 2.9 points and raises the average verbal GRE score by 3.5 points. However, the average GPA of awardees appears invariant to the stipend values. Overall, the table 1.7 results on the GRE scores suggest that at least some of the students on the margin of applying for a GRF fellowship are highachieving students who can be drawn away from other career or academic opportunities through the prospect of larger awards. In the context of the occupational choice model of Roy (1951), these results imply a positive correlation between ability in science and engineering and ability in alternative careers: the able S&E students who are drawn into the GRF program by higher stipends perform well on general measures of achievement (i.e., GRE scores) that are associated with success in other occupations. To test this interpretation, we regressed the number of applicants with relatively high measured skills (defined as those scoring 770 or over on the GRE quantitative exam, 680 on the verbal GRE, and those with GPAs above 3.88 on a four point scale) on the relative value of stipends and the number of awards given. By controlling for the number of awards, we get a measure of the relative number of applicants with high skills, which should largely determine the measured skills of awardees. Column (1) of table 1.8 shows that a 10 percent increase in the stipend value raises the number of applicants with 770 or over on the quantitative GRE by 12.2 percent. Column (2) shows a comparable increase in the number of applicants with 680 or over on the verbal GRE. The column (3) regression gives a 0.448 impact of stipend values on the number of students with very high GPAs.20 20. We also estimated models with the log of the number of applicants in the regression. This gave smaller positive coefficients to the relative stipend value in the regressions for the

44

Richard B. Freeman, Tanwin Chang, and Hanley Chiang

Table 1.8

Determinants of the number of applicants scoring above high thresholds Log(applicants with GRE quantitative  770) (1)

Log(applicants with GRE verbal  680) (2)

Log(applicants with GPA  3.88) (3)

Log(stipend/outside salary) in previous year

1.221 (0.104)

1.276 (0.086)

0.448 (0.100)

Log(number of awards) by field in current year

0.539 (0.090)

0.664 (0.074)

0.512 (0.077)

Field effects Field  time trend Observations R2

Yes Yes 261 0.9101

Yes Yes 261 0.9097

Yes Yes 207 0.9235

Source: Tabulated from NSF, Division of Graduate Education, Cumulative Index of the GRF program and related data sets, as described in text. Note: Outside salary are earnings of college graduates aged 21–25, tabulated from Current Population Survey.

In sum, higher stipend values incentivize enough high-achieving students to seek GRF offers to enable the NSF selection procedure to increase the measured skills of awardees when stipend values are higher. 1.3.3 The Economic Supply Link Overall, our results show that the resources NSF allocates to stipends, through the number of awards and their value, affects the supply of applicants, both in terms of numbers and measured scholastic skills. As a summary of the economic supply link, we have graphed in figure 1.10 our measure of the supply of students seeking GRFs—the number of applicants relative to the number of S&E baccalaureates—and a measure of NSF’s stipend budget (the number of awards offered in a particular year multiplied by the stipend amount) relative to GDP, which we take as the broadest possible measure of other economic activity. The stipend budget is an indicator of national resources spent to make graduate S&E study more financially attractive, whereas GDP represents national resources broadly. The tight link between the two series gives the bottom line message: if the United States wants more applicants for S&E fellowships and thus potentially more graduate students and Ph.D.s in these fields, the country has to spend more money on this objective. number of applicants with high math and verbal GREs, because the number of applicants is dependent on stipend values and associated with the number of applicants with high scores. It also raised the number with high GPAs, but conditional on the number of applicants, the relative value of stipends were modestly negatively associated with the number of applicants with high GPAs.

Supporting “The Best and Brightest” in Science and Engineering

45

Fig. 1.10 Fraction of bachelors choosing to apply to GRF vs. total GRF stipend budget/GDP Source: NSF DGE, Cumulative Index of the GRF Program and related datasets. Data on the Gross Domestic Product (GDP) from the Bureau of Economic Analysis, an agency of the U.S. Department of Commerce.

1.4 Alternative Policy Scenarios The analyses thus far have shown that the two policy levers—numbers of GRFs and the dollar value of the awards—affect the number and measured skills of persons seeking GRF awards and that coupled with NSF selection procedures, they determine the measured skills of awardees. In this section, we consider three scenarios that assess the way changes in the mode of determining stipend policy could affect the supply of students seeking and obtaining GRFs. 1. Agency determined changes in the number and value of awards. 2. Granting awards and setting values through fixed rules rather than policy discretion. 3. Allocating the number of quality group 1 and 2 positions among disciplines on the basis of quantifiable academic skills of applicants rather than on the number of applicants. 1.4.1 Agency Determined Changes in Stipend Numbers and Values For this scenario, we assume that the goal of fellowship policy is to increase the number of the “best and brightest” pursuing S&E graduate

46

Richard B. Freeman, Tanwin Chang, and Hanley Chiang

studies, and that both the numbers and qualifications of GRF awardees contribute to the goal, as described in the model in the appendix. The agency is assumed to face a fixed budget constraint to reflect alternative uses of resources. Our finding that increasing the number of awards reduces average quality while increasing the value of awards raises quality. This yields an optimal solution that depends on the supply response of students to number of awards and to the stipend amount. The critical factor that determines whether it is better to give more awards or raise stipends are the characteristics of applicants who are on the margin of winning awards. There are two ways to estimate the impact of changes in awards on the margin on the measured attributes of awardees. The first is to regress average GREs or other indicators of academic skills on the number of awardees, as we did in table 1.7, where the estimates suggested that average skills were not that sensitive to increased numbers of awards. The second way to assess skills on the margin is to go back to the micro records on individual attributes and rank applicants by the probability that they would get an award and then to examine the characteristics of those who would be given awards on the margin. To do this, we estimated the probability of getting awards in 2004 and then computed the quantitative and verbal GRE scores and GPA for awardees with the lowest probabilities in bins of fifty persons. We also estimated the average GRE and GPA for nonawardees with the highest probabilities of having gained an NSF, again in groups of fifty. The difference in attributes between these nonawardees and the awardees in the lowest group reflects the potential change in quality from increasing the number of awards. Figure 1.11 displays the results of our calculations in terms of the attributes of awardees and nonawardees. Persons in group 1 are awardees with the highest probability of getting an award. Persons in group 2 are awardees with the next highest probability of getting an award, and so on. Group 5 consists of awardees in the lowest probability group of getting an award. Groups 6 through 10 consist of persons who did not get an award, with those whom we estimated as most likely to get an award placed in group 6, followed by persons with less likelihood of getting an award in groups 7 through 10. The figure shows that there is very little difference in the GRE Quantitative measures around the cutoff point between groups 5 and 6. This indicates that the number of awards could have been increased along a substantial margin without greatly reducing quality, or could have been reduced along a substantial margin without greatly increasing quality. But the figure does show a noticeable drop in GRE Verbal scores between the marginal awardees and the marginal nonawardees. In part, this reflects the fact that while our model does a good job of predicting who gets an award, the errors in the model are for persons along the margin. Table 1.9 simulates what would happen if NSF has $10 million addi-

Supporting “The Best and Brightest” in Science and Engineering

Fig. 1.11

47

Quality of GRF applicants on the margin of getting an award

Notes: All persons to the left of the line were given awards. All persons to the right of the line did not get awards. The numbers related to groups of 50 persons, ordered by the estimated probability they would win an NSF award. The 5th group consists of 50 awardees with the lowest probability of getting an award, the 4th group consists of 50 awardees with the next lowest probability, and so on. The 6th group consists of the 50 non-awardees with the highest probability of getting an award. The probabilities are predicted values from an OLS regression of an award receipt dummy variable on panel rating, female dummy, underrepresented minority dummy, and eight field dummies.

tional funding to spend and divided the spending in three different ways: by increasing the number of awards, by increasing the value of awards, and by spending half of the additional money on increasing the number and half on increasing the value. Spending all of the money on increasing the number of awards at current stipend amounts would add 333 awardees and would give awards to persons with modestly lower GRE quantitative and verbal scores than current awardees—4.9 points lower in the quantitative GRE and 29.0 points lower in the verbal GRE. Because NSF gave 1,020 awards, the impact of the lower scores for the 333 awardees on the average for all awardees is considerably smaller. Spending all of the money on increasing stipend values would raise awards by 32.7 percent and would change the overall mean GRE scores of awardees modestly. If such a large increase in stipends were to attract more able persons, this improved quality does not readily show up in the GRE measures. Finally, dividing the 10 million dollars between additional awards and higher valued stipends would increase the number of awards by 146. The increased stipend attracts just enough higher scoring applicants to offset the mean quality decrease associated with more awardees. The result is virtually no change in the average measures of awardee skills.

48

Richard B. Freeman, Tanwin Chang, and Hanley Chiang

Table 1.9

All extra funds to increasing number of awards All extra funds to increasing value of stipends Half to additional awards and half to higher stipend

Impact of hypothetical $10 million increase in annual funding on selected outcomes in 2004, by alternative uses of funds Difference between mean quality of extra awardees and mean quality of original awardees

Change in overall mean quality of awardees

Change in annual number of awardees (1)

% Change in relative value of stipend (2)

GRE quantitative (3)

GRE verbal (4)

GRE quantitative (5)

GRE verbal (6)

333

0.00

4.9

29

1.2

7.1

0

32.70

n.a.

n.a.

8.2

146

14.30

n.a.

n.a.

3.3

10

0.7

Notes: Nonawardees in 2004 were ranked by their predicted probability of award receipt; see text for details. For a given increase of N in the number of awards, columns (3) and (4) are calculated by subtracting the mean quality of the actual 2004 awardees from the mean quality of the top N nonawardees. In column (5)/(6), the change in the overall mean quality induced by more awardees is calculated by multiplying column (3)/(4) by N/(1,020  N), given that 1,020 awards were actually granted in 2004. The change in the overall mean quality induced by higher stipend values is calculated from the coefficients in columns (1) and (3) of table 1.7. The change in overall mean quality induced by both more awards and higher stipend values is the sum of the two preceding changes. n.a.  not applicable.

It is easy to use our analysis to simulate the outcomes of other hypothesized changes in the NSF stipend budget and its allocation between numbers of awards and stipend values. 1.4.2 Using Fixed Rules to Determine the Number and/or Value of Stipends An alternative way to set the number and/or value of stipends is to use a fixed rule to set them. Monetary economists and officials have often discussed using fixed rules to set the growth of money supply. Some central banks target a range of rate of inflation as their goal while others focus on keeping interest rates within some band. Fixed rules increase social security payments with inflation and have been proposed to raise minimum wages and income tax brackets with inflation rather than through legislative means. If policymakers chose a given adjustment rule, the value or number of stipends would increase according to the rule. There are diverse rules that policymakers could use to set stipend values and the numbers of stipends. Taking the value of stipends, one possible

Supporting “The Best and Brightest” in Science and Engineering

49

rule would be to set stipends as a percentage of alternative earnings in the labor market, such as the earnings of young college graduates. Another would be to make stipends a given percentage of the earnings of doctorates. Yet another would be to adjust stipends to measures of employment opportunities, such as unemployment rates or, given low rates of joblessness among S&E doctorates, the length of time to obtaining a job upon completion of a Ph.D., or to the growth of R&D spending. Since any single indicator would be subject to measurement error, perhaps the most sensible target would be a weighted average of indicators: stipends would rise/ fall as the average of the indicators increased or decreased. To set the number of stipends by a fixed rule, policymakers could simply make them a given proportion of the number of BS graduates, thereby assuring students in different cohorts the same probability of gaining a stipend. When the supply of S&E baccalaureates increased, so too would stipend availability. Alternatively, one could use the same set of weighted indicators for setting the number of stipends as for setting the value of stipends. To get some sense of how a fixed rule would set the value and number of GRF stipends, we have simulated what would happen to the number of awards and measured scholastic achievement of GRF awardees under two conditions: (a) that NSF fixed the number of awards at 0.41 percent of S&E bachelor’s degrees, the level in 1968, before the ratio of awards to bachelor’s graduates began to fall; and (b) that NSF set stipends at the 2004 ratio of the value of stipends to the earnings of young college graduates from the current population survey (CPS), 115.8 percent. By setting numbers using a high awards per S&E baccalaureate and setting values at a high value to earnings of young college graduates, we get an upper bound on how a set of fixed rules would affect the supply of applicants to the field. Table 1.10 shows what would have happened from the 1970s to the 2000s if these fixed rules had been employed. The greater number of awards would have lowered measured scholastic skills, but the higher value of awards would have attracted enough applicants with greater skills to allow NSF to raise average quality. Bottom line, these rules would have given awards to a larger number of persons and, given our supply equations, would have raised the average quality of awardees, as well. Alternative sets of rules would, of course, give different outcomes. 1.4.3 Equilibrating Quality Across Disciplines Our final policy scenario considers what might happen if NSF changed its charge to panel committees from giving awards in approximately the same proportion to the number of applicants among fields to giving awards in a pure “measured scholastic” achievement, irrespective of the number of applicants from different fields. Thus, if physics received 100 applicants, all of whom had higher GREs, GPAs, and so on than, say social science,

921 511 504 460 520 807 794 820 915

Awards

0.35 0.16 0.17 0.15 0.16 0.25 0.22 0.21 0.23

Awards as % of S&E bachelor’s degrees

55.70 66.00 55.00 56.30 82.00 92.20 84.50 72.10 93.40

Stipend as % of alt. wage 745.2 751 753.1 746.9 759.6 766.3 762 751.1 753.2

Quantitative 692.8 693.8 703.6 693.1 696.5 692.9 681.2 660.3 633.9

Verbal

Average quality of awardees

Actual program characteristics (annual averages)

1,082 1,285 1,237 1,261 1,342 1,330 1,493 1,580 1,640

Awards

Verbal 3.5 20.1 19.6 22 20.7 10.9 13.8 14.3 12.7

Quantitative 1.7 9.8 9.5 10.7 10.1 5.3 6.7 7 6.2

Change in awardee quality from actual due to more awards

21.2 16.3 21.6 20.9 10 6.6 9.1 13.8 6.3

Quantitative

25.9 19.9 26.3 25.6 12.2 8.1 11.2 16.8 7.6

Verbal

Change in awardee quality from actual due to higher stipend

Hypothetical program characteristics

764.7 751 765.2 757.1 759.6 767.7 764.4 757.9 753.3

Quantitative

715.2 693.8 710.4 696.6 688 690.1 678.6 662.8 628.8

Verbal

of awardees

Predicted quality

Hypothetical GRF program characteristics from maintaining 1968 relative award availability and 2004 relative stipend value

Notes: “Hypothetical” characteristics denote those predicted to prevail if awards as percent of S&E bachelor’s degrees had been maintained at 0.41 percent (its 1968 value) and stipend as percent of alternative wage had been maintained at 115.8 percent (its 2004 value). Hypothetical changes in GRE quantitative and GRE verbal are based on coefficients in columns (1) and (3) of Table 1.7. In the calculations, actual numbers of bachelor’s degrees in 2002–2004 were imputed from 2000–2001 field-specific growth rates in earned bachelor’s degrees, and actual alternative wage in 2004 was imputed from 1983–2003 growth rate in salary of 21–25-year-olds with bachelor’s degrees.

1968–1971 1972–1975 1976–1979 1980–1983 1984–1987 1988–1991 1992–1995 1996–1999 2000–2004

Period

Table 1.10

Supporting “The Best and Brightest” in Science and Engineering

51

which had 500 applicants, all 100 physics applicants would receive awards before any social science applicant would receive an award. To see how awards would be given among disciplines under this hypothetical system, we estimated a linear probability regression of actual award receipt on GRE Quant, GRE Verbal, GPA, reference score (normalized), female indicator, and minority indicator in 1997, the last year for which we had a full set of scholastic qualification measures. In 1997, NSF awarded 850 GRFs, so under the hypothetical system, the applicants with the highest 850 predicted probabilities would be offered an award. Since physical science and mathematics and computer science students have the skills to shift to other fields and do well while students in the social sciences, psychology, and life sciences cannot easily shift into the physical and mathematical sciences, we anticipated that this change in policy would benefit the “harder” sciences. The calculations summarized in table 1.11 tell a different story. Column (1) gives the number and percentage of applicants in each field in 1997. Column (2) gives the number and percentage of awardees in each field. The percentages of awards by discipline differ moderately from the percentages of applicants by discipline, with engineering gaining more awards and life sciences gaining fewer awards, due in part to the WECS program. Column (3) gives our hypothetical distribution of awards by discipline on the basis of the measured scholastic attributes. The table note shows the coefficients of the equation used to generate the hypothetical distribution. It gives equal weight to the quantitative GRE and verbal GRE and a relatively heavy weight on the measure of references. The hypothetical distribution shows some differences from the 1997 actual distribution of awardees among natural science fields: mathematics, chemistry, and physics/astronomy gain a few awards while earth and atmospheric sciences and life sciences get fewer awards. The biggest change, however, is between engineering and the social sciences and psychology. Engineering loses a substantial number of awards, where social sciences and psychology gain. Why? The social sciences have the highest mean verbal scores whereas engineering has the lowest mean verbal scores. The social sciences and psychology score best on the reference letters. The GRE quantitative scores favor the physical and mathematical sciences, including engineering, but a larger proportion of students obtain top scores in the GRE quantitative test than in the GRE verbal test, which puts a bound on the extent to which our hypothetical allocation gives awards to those fields. It is possible that other simulations could give a somewhat different story: a high GPA in mathematics or physics may be more difficult to attain than one in the social sciences or biology, and reference letters do not compare students across fields. Still, it would take a very different set of calculations to support our initial expectation: that if NSF gave awards with no reference to fields, the physical sciences and mathematics would get more GRFs and the other sciences would get less awards.

52

Richard B. Freeman, Tanwin Chang, and Hanley Chiang

Table 1.11

Existing and hypothetical allocation of GRF awards across fields in 1997. Number of candidates (% in parentheses)

Field Chemistry Computer Science Engineering Earth/ Atmospheric Life Science Math Physics/ Astronomy Psychology Social Science Total

Mean applicant quality

Applicants

Awardees actual

Awardees hypothetical

Quantitative

Verbal

GPA

Reference

358 (7.0) 261 (5.1) 1,212 (23.6) 194 (3.8) 1,502 (29.3) 242 (4.7) 330 (6.4) 331 (6.4) 698 (13.6) 5,128 (100.0)

57 (6.7) 47 (5.5) 252 (29.6) 32 (3.8) 210 (24.7) 41 (4.8) 58 (6.8) 48 (5.6) 105 (12.4) 850 (100.0)

69 (8.1) 46 (5.4) 184 (21.7) 25 (2.9) 198 (23.3) 59 (6.9) 62 (7.3) 63 (7.4) 144 (16.9) 850 (100.0)

709

586

3.66

0.05

730

600

3.66

0.02

728

573

3.67

0.05

690

603

3.57

0.05

686

605

3.54

0.09

747

614

3.69

0.12

749

632

3.72

0.26

653

615

3.57

0.27

667

642

3.60

0.19

702

604

3.62

0.00

Notes: Awards under the “hypothetical” system are allocated as follows. A linear probability regression of actual award receipt on GRE quantitative, GRE verbal, GPA, reference score (normalized), female indicator, and minority indicator was used to predict the probability of being offered an award in 1997. Under the hypothetical system, applicants with the highest 850 predicted probabilities are designated as being offered an award. Regression with robust standard errors Number of obs  4566 R-squared  0.1772 gotaward Coef. Std. Err. t quant/100 .0541323 .0062852 8.61 verbal/100 .0582354 .0056672 10.28 gpa .1265555 .0153695 8.23 refscore .0879098 .0050868 17.28 female .059188 .011009 5.38 minority .0319515 .0126827 2.52 1.04754 .0670683 15.62 _cons

1.5 Conclusion Stipends to U.S. citizens/residents are a natural policy tool for increasing the incentive for Americans to enter S&E fields without directly impacting the flow of talent from overseas. Analysis of NSF Graduate Research Fellowships suggests that raising the value of awards increases the

Supporting “The Best and Brightest” in Science and Engineering

53

number of applicants and quality of awardees, while giving more awards increases the number of awardees, by definition, with only a modest reduction in measured academic skills. Because the analysis is limited to a single stipend program, it is uncertain whether the finding that the number of applicants responds to the relative value and relative number of stipends can be extrapolated to the market for graduate students as a whole. To the extent that changes in NSF fellowship policy induce changes in the policies of other stipend-granting groups, we suspect that the qualitative results, at least, can be extrapolated to the broader market. To see if such an extrapolation is at least consistent with the data, we examined the changing number of first-year first-time graduate students in science and engineering who were U.S. citizens and permanent residents relative to the number of S&E bachelor’s graduates in the United States. Appendix tables 1A.1 and 1A.2 give the results of our analysis. Columns (1) and (2) in appendix table 1A.1 show the number and percentage change in all first-year full-time students, including international students. Columns (3) and (4) give the number and percentage change in number of firstyear graduate students who were U.S. citizens/permanent residents and thus likely to be affected by higher valued awards for U.S. students. The final columns show an increased proportion of S&E bachelor’s graduates enrolling in graduate school over this period. These statistics show greater increase in U.S. citizen/permanent resident first year enrollments during the period when NSF increased the value of awards, and an increase in the number of graduate students relative to S&E bachelor’s graduates. The regression analyses in appendix table 1A.2 confirm this reading of the evidence. They relate the logarithm of the number of first-time full-time citizen/residents graduate enrollments to the value of stipends relative to college graduate salaries (column [1]) over the period and the log of the number of enrollments to the number of GRF applicants the previous spring. The estimates in column (1) give an elasticity of enrollments to the value of the stipends of nearly 0.40, which is less than half the elasticity of GRF applicants to stipends found in table 1.6. The estimates in column (2) show that indeed the increased enrollments are associated with increased applicants to the GRF program.21 Given the small sample sizes and limited time series variation, we view the appendix tables 1A.1 and 1A.2 analyses as suggestive. The results are consistent with the notion that more and better-paying stipends could raise the number of native-born/residents choosing S&E fields broadly. To go further than that would require a more extensive analysis of the decision of students to pursue S&E graduate studies as opposed to other options (such as simply applying for the GRF) and a careful study of the responses of

21. Analyses that eliminated social science and psychology bachelor’s graduates from the calculations gave similar results to those in the appendix tables.

54

Richard B. Freeman, Tanwin Chang, and Hanley Chiang

universities, other government agencies, and nonprofit groups to NSF stipend policies.

Appendix Model of Optimal Stipend Policy As a starting point for analyzing stipend policy, we consider the choices and constraints faced by an abstract granting agency. We assume that the granting agency cares about the number of awards (N) that it grants each year and the average quality (Q) of the students who receive the awards. The real-world analogue is that the NSF would like to support as many students as possible for graduate study in S&E, but at the same time the NSF would like its fellowship recipients to be as high-achieving as possible. The agency thus seeks to maximize

Table 1A.1

First-time, full-time graduate students in science and engineering

First-time, full-time graduate students in S&E

First-time, full-time graduate students in S&E (U.S. citizens & perm. res.)

Year

Number of students

Annual growth (%)

Number of students

Annual growth (%)

U.S. first-time, full-time graduate students as % of U.S. S&E bachelor degrees

1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004

71,183 70,894 74,400 76,597 81,484 83,159 79,366 77,968 74,384 73,479 73,544 74,431 75,447 78,332 82,411 86,827 89,331 86,565

— 0.4 4.9 3.0 6.4 2.1 4.6 1.8 4.6 1.2 0.1 1.2 1.4 3.8 5.2 5.4 2.9 3.1

48,121 46,478 48,309 49,145 52,186 54,603 54,027 54,318 52,378 51,260 50,375 49,828 48,362 46,316 48,207 54,625 59,649 58,853

— 3.4 3.9 1.7 6.2 4.6 1.1 0.5 3.6 2.1 1.7 1.1 2.9 4.2 4.1 13.3 9.2 1.3

15.6 15.2 15.9 15.7 16.1 15.9 15.3 15.0 14.4 13.9 13.5 13.3 12.7 12.1 12.5 13.7 14.1 13.5

Source: National Science Foundation, Survey of Graduate Students and Postdoctorates in Science and Engineering (GSS), various years; Integrated Postsecondary Education Data System Completions Survey, various years; and authors’ tabulations and imputations.

Supporting “The Best and Brightest” in Science and Engineering Table 1A.2

55

Determinants of first-time, full-time fall graduate enrollment in science and engineering Dependent variable: Logarithm of first-time, full-time graduate enrollment in S&E by U.S. citizens and permanent residents

Independent variables Ln(GRF stipend / outside salary) from previous calendar year Ln(GRF applicants) from most recent spring Ln(S&E bachelors degrees) from most recent spring Observations R-squared

(1)

(2)

0.397 (0.088)***

— —

— —

0.284 (0.038)***

0.462 (0.096)***

0.228 (0.072)***

18 0.714

18 0.854

Notes: Standard errors are in parentheses. Analysis period is 1987 through 2004. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

(A1)

U  aN  (1  a)Q.

a weighted average of N and Q where a is the weight placed on numbers. The agency has two policy levers to do this: the numbers of stipends (N) and the annual value of stipends (V ). The agency faces a budget constraint (A2)

NV  K,

where K is the (fixed) total annual budget allocated to stipends. The agency’s choice of N and V affects the quality of awardees. Average recipient quality is modeled as (A3)

Q  Q(V, N, other factors),

where Q/V  0 because higher-valued stipends are able to attract higherachieving candidates who would otherwise have pursued alternative fields and careers or seek other fellowships. We also expect that Q/N  0: average quality of the awardees declines with the number of awards because the highest-achieving students are generally the first to obtain awards. The optimal policy in this model, derived from maximizing the objective (A1) subject to the constraints (A2) and (A3), is characterized by the following condition:

56

(A4)

Richard B. Freeman, Tanwin Chang, and Hanley Chiang





Q Q V a  (1  a)   (1  a) . N V N

The optimal policy is to equate the marginal benefit and marginal cost of funding another student. The marginal benefit, given on the left-hand side of (A4), is simply the subjective value (a) that the agency places on being able to support an additional student. The marginal cost, on the righthand side, consists of two terms. The first term denotes the marginal decrease in quality resulting directly from giving more awards. The second term reflects the fact that, in the face of a fixed budget, a larger number of awards necessitates a lower annual stipend value, which in turn lowers the average quality of awardees. This model suggests that the relationship between stipend values and awardee quality is of central interest. How do higher stipend values bring higher-achieving students into the GRFP program? The main mechanism underlying this relationship is that higher stipend values attract a larger applicant pool, and at least some of the additional applicants are highly able candidates who are pulled away from alternative careers (or from immediate entry into the labor market). Essentially, larger applicant pools allow the agency to be more selective in granting awards. Thus, it is also of interest to examine the factors that determine the number of applicants to the GRFP program. We model the number of applicants (A) as (A5)

A  A(N1, V, Other factors),

where N–1 is the number of awards granted in the previous year. We predict that A/V  0 and A/N–1  0; that is, more students apply when stipend values increase and when the perceived probability of receiving an award rises. Other factors predicted to influence the applicant pool include the attractiveness of alternative careers. Thus, equations (A1) through (A5) capture the quality-quantity choice facing a stipend-granting agency: to spend its budget on funding many students versus funding a few “superstars.” Since identification of stars is difficult in most disciplines, the choice is in fact more complex. The key behavioral relations to be empirically estimated are: the change in the quality of students as the number of awardees is increased; the change in quality in response to the value of stipends; and the change in the number of applicants as stipend values rise. If the agencies responsible for stipend policy optimizes as in this model, then the preceding model reflects the basic tradeoffs factoring into the agencies’ decisions. In this case the number and monetary value of GRFP fellowships cannot be assumed to be exogenous for the purposes of estimating behavioral relationships. On the other hand, if the number and value of GRFP fellowships depend on vagaries of budgeting, then adjustments to the number and value of stipends constitute “natural experiments” for assessing the effect of stipend policy.

Supporting “The Best and Brightest” in Science and Engineering

57

References Committee of Visitors (COV). 1999. Report of the Committee of Visitors, National Science Foundation, The Graduate Fellowship Program. June 17–18. ———. 2003. Committee of Visitors (COV) report for the National Science Foundation (NSF) Graduate Research Fellowship Program (GRFP). June 17–18. Freeman, R. B. 2005. Fellowship stipend support and the supply of S&E students: NSF graduate research fellowships. Paper presented at Annual Meeting of the American Economic Association. 7–9 January, Philadelphia, PA. Goldsmith, S., J. Presley, and E. Cooley. 2002. National Science Foundation graduate research fellowship program: Final evaluation report. Arlington, VA: National Science Foundation. National Science Board. 2004. Science and engineering indicators, vol. 2., appendix table 2-15. VA: National Science Foundation. National Science Foundation, Division of graduate education. 2002. Guideline for panelists in the review and rating of National Science Foundation Graduate Research Fellowship applications. Arlington, VA: National Science Foundation. National Science Foundation, Graduate Research Fellowship Program (GRFP). 2007. Program solicitation, NSF 07-576. Arlington, VA: National Science Foundation. National Science Foundation, Division of Science Resource Statistics. 2001. Graduate students and postdoctorates in science and engineering: Fall 2001 tables 39, 40. Arlington, VA: National Science Foundation, Division of Science Resource Statistics. Roy, A. D. 1951. Some thoughts on the distribution of earnings. Oxford Economic Papers 3 (2): 135–46.

2 Internationalization of U.S. Doctorate Education John Bound, Sarah Turner, and Patrick Walsh

2.1 Introduction The representation of a large number of students born outside the United States among the ranks of Ph.D. recipients from U.S. universities is one of the most significant transformations in the international market for higher education in the last quarter century. Students from outside the United States accounted for 51 percent of Ph.D. recipients in science and engineering fields in 2003, up from 27 percent in 1973.1 The primary objective of this research is to understand the factors affecting this growth. We wish to understand the pattern of flows into U.S. Ph.D. programs both across countries and over time. John Bound is a professor of economics at the University of Michigan and a research associate of the National Bureau of Economic Research. Sarah Turner is a professor of education and economics at the University of Virginia and a research associate of the National Bureau of Economic Research. Patrick Walsh is an assistant professor of economics at Saint Michael’s College. Disclaimer: The use of NSF data does not imply NSF endorsement of the research methods or conclusions contained in this report. We would like to thank Richard Freeman, Daniel Goroff, Bill Kerr, and Michael Rothschild for helpful comments. We are grateful to our colleagues who helped us to understand particular country circumstances and locate international data including Michael Baker, Olivier Blanchard, Michael Elsby, Al Hermalin, Lutz Killian, Albert Park, Steve Pishke, and Yu Xie. Our research has been supported by grants from the Andrew W. Mellon Foundation and the Science and Engineering Workforce Project at NBER, funded by the Alfred P. Sloan Foundation. This work was finished while Bound was a fellow at the Center for Advanced Study in the Behavioral Sciences, Stanford University. 1. Tabulations presented in publications such as Science and Engineering Indicators (NSF 1996) show a somewhat lower representation of students from outside the United States among Ph.D. recipients in science and engineering for two reasons. First, we include only engineering, life sciences, physical sciences, and economics in our definition of science and engineering, excluding social science fields like sociology and political science, which have not drawn substantial number of foreign students. Secondly, we classify students as foreign if they

59

60

John Bound, Sarah Turner, and Patrick Walsh

Variation across countries and over time in the demand for graduate study in the United States affects the number and distribution of students by country of origin at universities in the United States. It is also the case that the representation of foreign students in U.S. Ph.D. programs is a function of the resources available to these programs or the “supply side” of the Ph.D. market. In the postwar years, substantial federal and state subsidies increased both the excellence and scale of U.S. graduate education. The growth of graduate education in the sciences at U.S. research universities has fundamentally changed international access to doctoratelevel training in the last half century. In motivating this analysis, we note that it is not uncommon to find rhetoric suggesting that the relative erosion in the quality of education afforded to young people in the United States is a primary cause of the decline in share of doctorate degrees in science and engineering awarded to U.S. students.2 Our interpretation of the available evidence is that such claims have little empirical basis. Natural economic forces of supply and demand, with these effects varying considerably in magnitude across countries, go a significant distance in explaining the observed changes in doctorate receipt among students from abroad and the United States. In the second section, we outline the basic trends in Ph.D. degree attainment and set forth the institutional context of doctorate education in the United States. The third section considers the differential cross-sectional representation of students by country at the graduate level in the United States. The analysis of the determinants of the growth over time in foreign participation in U.S. doctorate study in the sciences follows in the fourth section. The fifth section turns to the analysis of the determinants of participation of U.S. students in graduate education in the sciences. In understanding the substantial foreign share of doctorate recipients from U.S. institutions, we address two related questions. The first concerns the distribution of doctorate recipients by country of origin, as students from Asian countries tend to be overrepresented on a per capita basis and distributed somewhat differently by type of institution than students from Europe and other parts of the world. The second dimension of our analysis is to understand the determinants of changes over time in the number of did complete high school in the United States, which results in some overstatement of the aggregate counts of the foreign representation of doctorate recipients as respondents missing information on high school location are included in this count. The conclusions of the chapter and the statements about trends are invariant to the choice of classification of cases with unreported citizenship or high school location. 2. Bowen, Kurzweil, and Tobin (2005, 38) note that presidential and national commissions tend to urge policy changes “to counteract the alleged rising tide of mediocrity.” A recently released report (National Academy of Sciences 2005) notes, “Having reviewed trends in the United States and abroad, the committee is deeply concerned that the scientific and technical building blocks of our economic leadership are eroding at a time when many other nations are gathering strength” (3).

Internationalization of U.S. Doctorate Education

61

foreign students receiving doctorate degrees in the United States. Changes in demand—generated by increased undergraduate degree attainment abroad and political shocks—and changes in research support for universities affect the flow of foreign students to U.S. Ph.D. programs. Our interest in understanding the production of graduate education at universities is ultimately an input to the study of the link between the graduate education process and the employment of scientists and engineers in the United States and abroad. In turn, decisions to pursue graduate study reflect variation over time and countries of origin in labor market opportunities for the high-skill workers. A significant innovation of our work is the identification of trends in doctorate awards by country of origin. First, even countries that are relatively similar to the United States in socioeconomic circumstances and institutions (such as Canada and countries in Europe) send a considerable number of students to U.S. doctorate programs. Second, as baccalaureate degree receipt grows within countries so too does the attainment of Ph.D.s at U.S. universities, with these changes particularly marked among countries experiencing substantial changes in educational attainment. Finally, political transformations involving either the opening or closing of trade with the United States also lead to substantial changes in doctorate receipt in the sciences among students from foreign countries. 2.2 Basic Trends and Policy Context The U.S. education market has never been closed to foreign students, though the absolute number of students from other countries enrolling in U.S. colleges and universities was relatively modest until the 1970s. The post-World War II strengthening of U.S. universities—particularly in the sciences and engineering—made advanced study in the United States more attractive to foreign students. In the two decades between 1936 and 1956, foreign students accounted for 19 percent of Ph.D.s awarded by U.S. institutions in engineering, 10 percent of Ph.D.s awarded in the physical sciences, 12 percent of Ph.D.s in the life sciences, and 12 percent of Ph.D.s in economics (National Academy of Sciences 1958).3 Advances in air travel, global communication, and visa arrangements 3. There is a small representation of foreigners in U.S. undergraduate programs as well, with temporary residents representing about 3 percent of BA recipients from U.S. institutions. The distribution of foreign undergraduate and professional students studying in the United States is quite different than the distribution of students pursuing doctorate degrees, in large part because undergraduate students and students in professional programs are generally expected to pay their own way. As the size of the U.S. college-age population fell in the late 1970s many colleges and universities actively recruited students from foreign countries to increase revenues. A 1979 report from the American Council on Education identified foreign students as a potentially important market for undergraduate colleges facing declining enrollment demand with smaller high school cohorts (Maeroff 1979).

62

John Bound, Sarah Turner, and Patrick Walsh

no doubt also reduced the fixed barriers that might otherwise have limited the access of foreign students to U.S. universities. Immigration reform through the Immigration and Nationality Act of 1952 and subsequent amendments formalized the status of students attending U.S. institutions from abroad by creating categories of “nonimmigrant visas,” permitting temporary study in the United States. The most common designation is the F-1 visa, which is issued to students admitted to an approved institution of learning with the demonstration of sufficient financial support.4 Dramatic growth in doctorate education, as well as higher education more generally, characterized the immediate post-World War II decades in the United States. Doctorate degrees awarded increased from less than 10,000 in 1958 to nearly 35,000 in 1973 (fig. 2.1). Then, after a period of stagnation, the overall number of doctorate degrees expanded again in many fields during the 1980s and the number of doctorates awarded by U.S. institutions climbed to an historic peak of 42,652 in 1998. The rise in the share of degrees awarded to students born outside the United States is a distinguishing feature of the last quarter century (see fig. 2.1), particularly in scientific fields. Changes in federal funding for science, as well as direct public support for graduate education,5 are an important determinant of both opportunities for graduate education and the labor market demand for Ph.D.s. Figure 2.2 shows the overall trend in federal research funding to universities; the dramatic rise from the late 1950s to the late 1960s is followed by a period of stagnation in the 1970s, before increases in federal funding for the sciences resume in the 1980s. The Survey of Earned Doctorates provides a comprehensive picture of Ph.D.s produced by U.S. universities by country of origin from the late 1950s to the present. The Survey of Earned Doctorates is an individual-level census of recipients of doctorates at U.S. institutions. Because survey participation is often coupled with the formal process of degree receipt, response rates have been quite high. When we organize these data by country 4. To obtain a student visa, an individual submits a letter of admission from a university and a certificate of eligibility issued by the school (known as Form I-20) to the American embassy or consulate in the home country. The scope of education that F visas have historically included is not limited to degree-granting colleges and universities but also includes profitmaking technical training schools and proprietary language institutes. 5. Federal support for doctoral study came in the form of fellowships to individuals as well as project support to researchers and universities. In 1952, the National Science Foundation established the Graduate Research Fellowship program, which provided generous multi-year support for those pursuing doctorate study in the sciences and engineering. The annual number of awards grew from about 500 in the 1950s to a peak of 1,373 in 1966, with the number of awards offered then contracting back to about 500 in the 1970s and 1980s before rising to nearly 1,000 awards in the 1990s (Freeman, Chang, and Chiang, chapter 1, this volume). In addition, the National Defense Education Act (NDEA) Fellowships for graduate study were passed by Congress in 1958 as part of a broader package of legislation intended to improve funding of education in the sciences and other areas of national need (including foreign languages), partly in response to the launching of Sputnik.

Fig. 2.1 Ph.D. degrees awarded by U.S. universities and national origin, 1958–2003 Source: NSF, Survey of Earned Doctorates microdata and, before 1958, National Academy of Sciences (1958). Note: National origin is defined by the country in which an individual went to high school.

Fig. 2.2

Federal funds to universities for research

Source: National Science Foundation. Federal obligations for total research and development, by major agency and performer: fiscal years 1951–2001, http://www.nsf.gov/sbe/srs/ nsf01334/tables/histb.xls. University totals include Federally Funded Research and Development Centers.

64

John Bound, Sarah Turner, and Patrick Walsh

of origin, we focus on the country where an individual completed high school, as this measure does not include those immigrating to the United States at young ages. This method also serves to count individuals in the country where they resided at young ages.6 Access to the restricted microdata files is particularly important for the analysis that follows. Foreign students are more heavily concentrated in the sciences than in the humanities. Moreover, the broad area of the social sciences masks considerable diversity in representation of foreign students, as 51 percent of economics doctorates hold temporary visas though only 5 percent of psychology Ph.D.s are neither citizens nor permanent residents. The variation in choice of specialization at the undergraduate level importantly affects demand for U.S. Ph.D. programs by field. In Asian countries, the majority of undergraduate degrees are awarded in science and engineering fields with a reported share of 65 percent for Japan and 60 percent for China,7 while in the United States (32 percent) and European countries including the UK (35 percent) the share of BA degrees awarded in science and engineering fields is appreciably smaller (National Science Board 2004, table A2-34). Within science fields, the growth since the mid-1970s in doctorates awarded among those from outside the United States is particularly striking. Figure 2.3 shows the trend in doctorates awarded to U.S. residents and individuals from foreign countries in engineering, the life sciences, the physical sciences, and economics. In all but the life sciences, the foreign share now equals or exceeds the share of Ph.D. recipients from the United States. With some modest differences in timing across fields, the expansion in degrees awarded to foreign students commenced in the mid-1970s and slowed in the mid-1990s. Summarizing the broad developments from 1980 to 1996 (the peak year in recent Ph.D. awards to foreign students), the total number of Ph.D.s in science and engineering increased from 12,126 to 21,253. If we engage in the accounting exercise of holding constant the foreign share at the 1980 level, the total expansion in doctorates awarded 6. Country of origin was defined by the country in which the respondent attended high school (“hsplace” in the the Survey of Earned Degrees). Out of the 1.35 million observations, 88,709 (6.6 percent) listed no hsplace. Among respondents in fields classified as “science and engineering,” 6.0 percent listed no hsplace. Since “foreign” is defined as simply “not U.S.,” it is possible that people who went to high school in the United States but listed no high school country are classified as “foreign.” However, of all those listing no hsplace, 15 percent list “United States” as their country of birth (compared with 69 percent of the overall), while 75 percent list no birthplace. Since at most 6 percent (and probably much less) of the relevant sample can be misclassified this way, the foreign/U.S. treatment of these individuals should not materially affect this chapter’s results. 7. While there is no question that the scale of undergraduate education has grown very dramatically in China in the last decade, there is evidence that some of the widely reported data on the number and share of degrees awarded in science and engineering are overstated, counting sub-baccalaureate training in trades as engineering or science degrees. See Gereffi et al. (2008) for further discussion.

Internationalization of U.S. Doctorate Education

65

Fig. 2.3 Ph.D. degrees awarded by U.S. universities and national origin, 1958–2003 Source: NSF, Survey of Earned Doctorates microdata. Notes: National origin is defined by the country in which an individual went to high school. Fields defined using NSF classification, from SED annual reports.

would have been expected to be a more modest 2,619 degrees, relative to the observed change of 9,127 doctorate degrees awarded by U.S. institutions. 2.2.1 Institutional Context of U.S. Universities Universities in the United States award more Ph.D. degrees than those in any other country. In 2001, the United States awarded 40,744 Ph.D.s, relative to 24,769 awarded by Germany, 14,210 awarded by Great Britain, and 16,078 awarded by Japan. In the science and engineering fields, the United States continues to dominate but by a more modest margin, with the United States awarding 25,509 Ph.D.s relative to 11,803 awarded by Germany, 8,520 awarded by Great Britain, 7,401 awarded by Japan, and 8,153 awarded by China (National Science Board, 2004, table A2-36). Not only do U.S. institutions award more Ph.D. degrees than those in any other

66

John Bound, Sarah Turner, and Patrick Walsh

country, but U.S. universities also dominate at the highest levels of scientific accomplishment. At the same time, the rate of growth of doctorate education in the United States has lagged other counties (particularly those in Asia) over the last decade. The average annual rate of growth in doctorates awarded in science and engineering fields was 3.2 percent in the decade of the 1980s and 1.6 percent in the decade of the 1990s, relative to annual rates of growth in doctorate production exceeding 20 percent in South Korea and Taiwan in the 1980s and China during the 1990s.8 The U.S. market for doctorate education is also highly stratified. In 2002, 413 universities in the United States awarded doctorates, with the mean number of degrees per institution 97, and the median number 38 degrees. Overall, production is relatively concentrated, with twenty institutions awarding 27 percent of the 2002 total of 39,955 degrees.9 Substantial subsidies from state, federal, and institutional sources to research universities affect the quantity and quality of graduate education, while the concentration of federal support at a relatively small set of universities adds to the stratification in graduate education. The National Academy of Sciences’ rankings show the wide difference in faculty publications and research funding between the top and bottom quartile of graduate programs.10 Stratification is apparent in outcomes as well as funding, as high achieving scientists come from a relatively small number of graduate institutions. Top U.S. universities are often considered leaders at an international level, resulting in a comparative advantage in the production of doctorate education.11 That research and doctorate education are often complementary 8. See National Science Board (2004, appendix tables 2-38 and 2-39). 9. While this concentration is considerable, it is appreciably less than at the start of the century. Up until the mid-1920s, five institutions (Columbia, Harvard, Johns Hopkins, Yale, and Chicago, notably all private) awarded about one-half of the annual flow of doctorates; by the 1930s, there had been some dispersion as these five institutions awarded about one-third of new doctorates (Berelson 1960, 93). By 1950, there were at least thirty institutions, including many large public universities, awarding a significant number of Ph.D. degrees annually. Focusing on the interval between 1958 and 1972, Bowen and Rudenstine (1992) document the extraordinary growth in the number of institutions and departments operating Ph.D. programs. In economics, the number of Ph.D.-granting institutions increased nearly 90 percent from 57 to 108, while in mathematics the number of programs increased more than 130 percent from 60 to 139. 10. To give but one example, graduate faculties in the top quartile of doctorate-granting programs in economics averaged thirty-six faculty members and nearly thirteen citations per faculty member, relative to 17.3 faculty members and 1.36 citations in the bottom quartile. See National Research Council (1995). 11. At least one effort has been made to compare universities through the creation of an index including measures such as Nobel laureates, articles in major scientific publications, and citations. The result of this effort is that fifteen of the top twenty, as well as thirty-five of the top fifty, universities are in the United States (Shanghai Jiao Tong University 2003). While the strength of U.S. universities at the top of the international rankings is widely recognized, it should also be noted that there is considerable variance as well in the quality of U.S. doctorate programs. One British observer comparing the United States and the United Kingdom notes: “The U.S., with 4,000 institutions of higher education, probably has fifty of the best universities in the world and undoubtedly has 500 of the worst.” (Stevens [2004], as cited in Bowen, Kurzweil and Tobin [2005, 66]).

Internationalization of U.S. Doctorate Education

67

in production further strengthens the advantage of elite U.S. universities, contributing to potential agglomeration effects in the location of science. 2.3 Cross-Sectional Distribution by Country 2.3.1 Motivation There is considerable cross-country variation in doctorate attainment from U.S. institutions. Asian countries—particularly India (736), Taiwan (423), South Korea (842), and China (2,452)—accounted for more than one-half the doctorates in science awarded to those from outside the United States in 2003. Students from France (77), Germany (168), and Great Britain (76) were less than 3 percent of the foreign degree recipients. Why students from some countries are particularly likely to pursue doctorate education in the United States surely depends on opportunity costs. In general, demand for doctorate education will be lower for those students with more abundant home country opportunities and, in turn, students from countries with relatively substantial university systems will be unlikely to study in the United States unless they can attend top-tier doctorate programs. What matters for students potentially pursuing study in the United States is the expected return to a U.S. Ph.D. program relative to the best alternative in the home country. In the cross-section, individual students in each country face a choice based on the expected benefit to doctorate study in the United States and an expected return to persistence in the home country, which may include attending graduate school in the home country or pursuing some other vocation. It follows that the opportunity cost of pursuing a doctorate degree at a U.S. university varies among countries of origin. Alternative options for post-baccalaureate study as well as fixed costs of foreign study will vary by country. Two presumptions about graduate study in the United States and abroad have implications for who comes to the United States for graduate study. First, expected success in home country and anticipated benefits from graduate education in the United States are correlated, implying that people likely to have high returns from graduate study in the United States are also likely to have an absolute advantage in home country graduate education or alternative activity. Second, U.S. programs tend to be dominant in the top tail of the international distribution of program quality. For countries in which forgone opportunities are close to those in the United States (countries with large and well-established university sectors) only a select few individuals will pursue graduate studies in the United States. These individuals will be among those with relatively high ability and receive admission offers from some of the best programs in the United States. In contrast, individuals from countries with much more limited higher education systems will have fewer opportunities for graduate study

68

John Bound, Sarah Turner, and Patrick Walsh

in their home countries and will be much more likely to choose to pursue graduate study at a U.S. university. In turn, these individuals may choose to come to the United States to pursue studies at programs outside the most highly ranked departments. Moreover, part of the return to doctorate study in the United States may be future access to the U.S. labor market.12 Foreign doctorate recipients from U.S. universities may be particularly well-positioned to find employment in the United States and to receive H-1B work visas for employment in specialty occupations.13 By one estimate (Lowell 2000), nearly onequarter of H-1B visa recipients have changed from foreign student status. Completion of a Ph.D. may be particularly important to an individual’s prospects for receiving an H visa, as educational requirements are one way for firms to document that an individual has skills that are scarce and specialized in the application process. The previous considerations lead to two clear predictions. First, countries with relatively modest home country options for doctorate study will be represented in greater relative numbers in U.S. Ph.D. programs than countries with significant home country university options. Secondly, the average quality of students (and the graduate programs selected) receiving Ph.D.s in the United States is inversely related to the share of a country’s potential doctorate students completing advanced study in the United States. 2.3.2 Cross-Country Differences in Doctorate Degree Attainment In the cross-section, both the level of undergraduate degree attainment in foreign countries and the extent to which there are established doctorate-level programs in these countries has a substantial effect on the flow of Ph.D. students to U.S. institutions. The data in table 2.1 provide a crosssectional picture, combining undergraduate degree production in the early 1990s with doctorate production at the end of the decade across countries. The number of college-age individuals in each country receiving a science and engineering BA would seem to represent a reasonable measure of the potential demand for doctorate-level graduate study in science and 12. Finn (2003) estimates that about 71 percent of foreign citizens who received science/engineering doctorates from U.S. universities in 1999 were in the United States in 2001. For those receiving degrees in 1991, about 58 percent were still in the United States in 2001. The attractiveness of the transition from graduate study to employment with an H-status visa increased with the Immigration Act of 1990, allowing H-1B visa holders to also apply for permanent resident status, where formerly H-1B visa holders were required to declare an intention to return to their countries of residence. 13. The government defines a specialty occupation as: “A specialty occupation requires theoretical and practical application of a body of specialized knowledge along with at least a bachelor’s degree or its equivalent. For example, architecture, engineering, mathematics, physical sciences, social sciences, medicine and health, education, business specialties, accounting, law, theology, and the arts are specialty occupations.” Accessed at http://uscis.gov/ graphics/howdoi/h1b.htm (U.S. Citizenship and Immigration Services)

Table 2.1

Cross-sectional analysis of BA degrees and Ph.D.s by country U.SAwarded S&E Ph.D.s 1996–1998 avg. (5)

Ph.D. U.S./BA (6)

Ph.D. U.S./Ph.D. country (7)

S&E BA 1990 (1)

BA/ Pop 24 1992 (2)

S&E BA/ Pop 24 1992 (3)

21,159 169,726

0.296 0.306

North America 0.053 898 0.046 11,034

222 11,034

0.010 0.065

0.247 1.000

Argentina Brazil Mexico

10,032 28,379 35,443

0.067 0.082 0.084

Latin America 0.015 382 0.017 1,775 0.028 396

66 169 144

0.007 0.006 0.004

0.173 0.095 0.364

Belgium France Germany Greece Ireland Italy Netherlands Spain Sweden Switzerland UK

6,253 30,400 66,299 5,203 3,364 19,204 5,536 21,492 3,978 2,154 28,608

0.133 0.130 0.128 0.119 0.151 0.104 0.086 0.195 0.135 0.083 0.208

Western Europe 0.044 388 0.042 5,530 0.050 7,199 0.032 301 0.045 297 0.023 1,558 0.023 1,306 0.035 2,301 0.034 785 0.020 1,569 0.056 4,394

18 70 155 113 20 75 33 48 15 21 87

0.003 0.002 0.002 0.022 0.006 0004 0.006 0.002 0.004 0.010 0.003

0.046 0.013 0.022 0.375 0.067 0.048 0.025 0.021 0.019 0.013 0.020

Czechoslovakia Hungary Poland

14,589 2,369 14,415

0.124 0.095 0.106

Eastern Europe 0.057 471 0.017 600 0.028 —

30 31 47

0.002 0.013 0.003

0.064 0.052 —

Australiaa New Zealanda

14,049 1,500

0.359 0.337

Australian Cont. 0.080 1,584 0.061 —

39 22

0.003 0.015

0.025 —

149,607 91,221 36,585 2,498 11,431 168,000

0.012 0.234 0.205 0.115 0.150 0.048

0.006 0.062 0.067 0.048 0.059 0.011

5,036 4,311 2,410 — 765 4,890

2,537 100 761 37 1047 3,669

0.017 0.001 0.021 0.015 0.092 0.022

0.504 0.023 0.316 — 1.369 0.750

17,011 2,664 3,701 4,426

0.088 0.088 0.129 0.048

Middle East/Africa 0.012 — 0.017 — 0.033 499 0.006 —

82 59 38 29

0.005 0.022 0.010 0.007

— — 0.076 —

Country

Canada United States

Domestic S&E Ph.D.s (4)

Asia China Japan S. Korea Singapore Taiwan India Egypta Saudi Arabiaa Israela S. Africa

Sources: Column (1): NSF (1993, NSF 93-303, table A-9); NSF (1996, NSF 96-316, table A-16); NSF (2004); UNESCO (annual series, 1963–1999). Columns (2) and (3): National Science Foundation (1996). Column (4): National Science Foundation (2000). Column (5): NSF Survey of Earned Doctorates microdata (authors’ tabulations). Notes: Numbers in column (4) represent the total number of Ph.D.s earned from institutions in the country in question, including foreigners, except in the case of the United States, Germany, France, the UK, Japan, and Canada. In these six cases the numbers are net of foreign nationals obtaining Ph.D.s in the country in question. a Indicates rows for 1998 (and 1999 in the case of Australia) from NSF (2004).

70

John Bound, Sarah Turner, and Patrick Walsh

engineering. In column (1) of table 2.1, we report data on this number for 1990 or the most recent available year.14 In the third column we report the undergraduate degrees in science an engineering relative to the twenty-fouryear-old population in each country (column [2] reports all BA degrees relative to population). While there are some cross country differences, roughly 3 to 5 percent of the populations of North American and Western European countries received an undergraduate degree in science or engineering. In the early 1990s a comparable or even somewhat larger fraction of the population in Asian countries such as Japan, Taiwan, and South Korea did so. In contrast, just over 1 percent of those from India and 0.6 percent of those from China received a science and engineering BA, reflecting relatively low overall levels of undergraduate degree attainment in these countries. Much of this observed difference across countries in the representation of science and engineering BA degree recipients is representative of the scale of higher education; in countries where only a small fraction of the population receives a BA degree (column [2]), it follows that the overall number of science and engineering BA recipients will be limited.15 A second measure of the development of the higher education sector within a country is the size of the doctorate-granting sector of higher education—both in an absolute sense and relative to the BA sector. Column (4) of table 2.1 shows the number of Ph.D.s awarded in each of the listed countries.16 Column (5) presents the number of individuals from the country in question receiving a science or engineering Ph.D. from a U.S. institution. In the case of most of the countries listed in the table, somewhere between 5 to 10 percent of college graduates in the sciences and engineering go on to get a Ph.D., though the fraction of those awarded a Ph.D. from a U.S. institution varies dramatically.17 A clear hypothesis is that countries with low domestic Ph.D. production 14. Note that these numbers would not be qualitatively different if we were to include BA degrees received at U.S. institutions by foreign students. Particularly for countries sending large numbers of students to the United States for graduate study, the proportion of Ph.D. recipients who also received BA degrees from a U.S. institution was 4 percent for those from India and less than 2 percent for those from China measured over the last fifteen years. 15. Nevertheless, there are large differences across countries in the relative share of degrees awarded in the sciences. For example, less than one-fifth of U.S. undergraduate degrees are awarded in science and engineering fields while about one-half of degrees in China are awarded in science and engineering fields. The United States—as well as other countries with substantial service sectors—educates many people at the baccalaureate level in professional fields such as accounting and business, which are unlikely to provide the preparation for the pursuit of a Ph.D. degree in science. 16. In the cases where this is possible—the United States, the United Kingdom, Germany, France and Japan—we have netted out foreigners obtaining a Ph.D. in the country in question. In all other cases, the data refer to the total number of Ph.D.s granted, irrespective of whether the individual is or is not a foreigner. 17. The largely Asian countries in table 2.1 that send significant numbers of students to the United States to receive their Ph.D.s also send students to Canada, Australia, and several European countries. For these countries, the numbers in table 2.1, to some extent, underestimates the total number of individuals from these countries receiving Ph.D.s in the sciences.

Internationalization of U.S. Doctorate Education

71

relative to undergraduate degrees awarded and relatively less developed higher education systems will be among those most likely to send Ph.D. students to the United States. Columns (6) and (7) in table 2.1 underscore this point clearly, as European countries with long traditions in higher education send relatively few students to the United States, while Asian countries are much more likely to send students to the United States to pursue Ph.D. studies.18 In short, the international representation of students in U.S. doctorate education depends appreciably on home country undergraduate and graduate options. When we focus on top-ranked programs in the United States, the distribution of Ph.D. recipients by country of origin is much different than when the focus is on Ph.D. recipients in aggregate.19 Students from European countries are represented in far greater proportions among top institutions than in the overall pool of doctorate recipients. Moreover, within each country of origin, countries that send a relatively high fraction of potential doctorates to the United States for training have relatively lower concentrations of Ph.D. recipients among the top-ranked U.S. programs. Table 2.2 presents these data on the proportion of a country’s Ph.D. recipients receiving degrees from top-five programs and shows the distribution of degrees by country awarded by the most highly ranked programs. What is clear is that for a number of Asian countries—notably Taiwan, South Korea, and China—Ph.D. recipients in science are underrepresented in the top-five departments and are much less likely to receive their degrees from these programs than Ph.D. recipients from the United States in these fields. For example, while students from China are about 15.5 percent of all chemistry Ph.D.s, they are only 5.3 percent of degree recipients from top-five programs. At the other extreme, student from Canada and European countries tend to be represented in the top programs in shares in excess of their overall representation among Ph.D. recipients from U.S. universities. Countries that send a relatively high fraction of students to unranked or low-ranked Ph.D. programs are those where opportunities for graduate study in the sciences are quite limited. Put somewhat differently, these data are indicative of the quality of “home country” Ph.D. programs; for coun18. Empirical verification of this point is provided by consideration of the correlation between measures of U.S. Ph.D. production and home-country BA degrees awarded. Using available data, there is a negative (–.2) correlation between the ratio of Ph.D.s awarded in the United States and the ratio of BA degrees to population, indicating that countries with relatively well-developed university systems rely less on U.S. institutions for Ph.D. production. 19. We use the rankings at the discipline level assembled by the National Academy of Sciences at a point in time in the early 1990s. While there have been some changes over time in rankings, there have been few large movements (mobility from unranked to top five) over the last three decades. See National Research Council (1995). It is, of course, true that there are changes in the relative rank of Ph.D. programs over time; yet these changes tend to be modest relative to the overall correlation between rankings done in various years.

Table 2.2

Distribution of Ph.D.s awarded by country, field, and program quality, 1994–2003 Physics

Country Canada China Former Soviet Union France Germany India Italy Japan Korea Mexico Taiwan UK U.S.

Chemistry

% Country (nj/n) (1)

% Country Top 5 (n5j/n5) (2)

Country % Top 5 (n5j/nj) (3)

Country % Low (nBj/nj) (4)

% Country (nj/n) (1)

% Country Top 5 (n5j/n5) (2)

Country % Top 5 (n5j/nj) (3)

Country % Low (nBj/nj) (4)

1.3 12.4

3.8 8.3

31.4 7.0

11.8 51.3

0.8 15.5

2.0 5.3

20.6 2.7

33.9 58.6

4.0 0.3 1.9 3.3 0.7 0.6 3.7 0.5 2.8 0.4 49.6

3.6 0.4 1.1 1.7 0.6 0.4 1.9 0.2 1.7 0.9 56.1

9.5 14.7 5.9 5.6 9.8 6.7 5.5 4.9 6.6 27.7 11.9

40.5 41.2 40.0 44.9 52.2 33.3 44.7 49.2 40.7 31.9 34.5

1.5 0.7 0.7 3.4 0.2 0.4 2.8 0.3 2.4 0.6 56.1

0.8 0.4 0.8 1.3 0.3 0.3 1.9 0.1 1.7 0.4 71.2

4.0 4.2 9.4 3.0 11.6 6.5 5.3 3.3 5.6 5.8 10.1

50.2 43.1 43.9 66.1 39.5 40.3 50.9 63.3 46.4 56.7 39.0

Economics Canada China Former Soviet Union France Germany India Italy Japan Korea Mexico Taiwan UK U.S.

Biochemistry

1.3 6.0

2.5 3.8

27.5 8.4

17.5 37.8

1.0 16.6

3.3 6.3

22.7 2.5

40.0 63.2

1.1 0.7 1.4 5.0 2.0 2.2 7.1 1.1 2.9 0.8 39.0

0.8 1.5 1.8 2.0 3.8 2.7 3.4 2.3 1.0 1.8 40.7

10.3 30.8 17.3 5.5 25.5 16.5 6.5 27.3 4.7 29.1 14.1

30.8 21.5 26.3 40.9 7.3 14.6 31.2 23.6 35.4 17.7 33.8

1.0 0.3 0.5 3.4 0.1 0.3 2.6 0.4 2.8 0.4 58.5

0.2 0.4 0.6 1.2 0.0 0.2 1.0 0.0 2.5 0.2 73.1

1.2 10.0 7.7 2.2 0.0 4.5 2.4 0.0 6.0 3.6 8.3

65.9 60.0 51.3 75.6 45.5 63.6 57.8 53.1 52.5 64.3 47.8

Source: S&E Ph.D.s: NSF, Survey of Earned Doctorates microdata (authors’ tabulations) Ph.D. program rankings: Research-Doctorate Programs in the United States: Continuity and Change (1995) http://books.nap.edu/html/researchdoc/ researchdoc_intexp.html. Notes: National origin is defined by the country in which an individual went to high school. Fields defined using NSF classification, from survey of earned doctorates (SED) annual reports. In the column heading, j subscript is country. Col. (1) indicates the percent of degrees in the indicated field awarded to those from country j. Col. (2) is the ratio of degrees from top 5 institutions in country j relative to all degrees from top 5 institutions for the indicated field. Col. (3) presents the share of degrees awarded to individuals from country j that were from top 5 institutions. Col. (4) presents the share of degrees awarded to individuals from country j that were from institutions that were unranked or ranked below. Countries not specifically enumerated are in an “other” category that is included in totals but does not appear in the table.

Internationalization of U.S. Doctorate Education

73

tries like Canada, where Ph.D. recipients from U.S. institutions are concentrated in relatively high-quality institutions, the quality of home country Ph.D. options is relatively high. A related explanation is that the countries with the highest relative representation among top programs in the United States are those countries where there is considerable existing research exchange among scholars in the United States and abroad, providing a natural network linking students from foreign universities to graduate study in the United States. 2.4 Growth in Foreign Share Over Time The growth in the representation of foreign students among doctorate recipients from U.S. universities captures changes on both sides of the market for graduate education. In particular, the growth reflects some combination of the following circumstances: (a) shifts in demand for graduate study among foreign-born arising from changes in the sending country; (b) shifts in demand arising from changes in institutions that affect the “costs” of matching students with U.S. graduate programs, including the development of international networks; and (c) adjustments in the supply-side or offerings of U.S. universities that differentially affect foreign students. The forces affecting the representation of foreign students in U.S. doctorate education are presented through a basic supply-demand framework. Demand shocks generated by increases in the number of undergraduates (potentially) prepared for graduate study from abroad are one dimension of change. Those countries with relatively high BA growth might be expected to expand in the share of Ph.D. received from U.S. institutions. Growth in the size of cohorts prepared for graduate study (for simplicity, those with the BA) is the most obvious type of demand shift varying across countries. Such shifts may include growth in the fraction of college graduates or shifts in cohort size, varying in magnitude and timing across countries. Over the course of the last half century, a number of political transformations such as the fall of communism in the Soviet bloc or the normalization of relations with China have dramatically altered the demand for graduate study in the United States among foreign students. Beyond changes in the number of students prepared for graduate work in a country, a related change in demand comes from the development of networks that reduce the costs of foreign study. Following dynamic models similar to the Carrington, Detragiache, and Vishwanath (1996) of the South-North migration of blacks in the first half of the twentieth century, successful experiences of initial migrants lead to dramatic reductions in information costs among those in later cohorts. Students from specific regions or foreign universities may establish links with U.S. programs; in turn, U.S. universities may use past experience in recruiting and selecting students. Such network effects have the long term result of increasing the

74

John Bound, Sarah Turner, and Patrick Walsh

relative benefits of pursuing doctorate study and the share of students from abroad pursuing graduate study in the United States. By lowering the costs and increasing the value of graduate education in the United States, such networks serve to shift the demand for graduate education in the United States. The supply-side of the U.S. market for graduate education is by no means fixed over time. Because doctorate-level students do not pay full tuition for their studies, the availability of opportunities is likely to be determined by research funding and other institutional sources of support, including state funding and demand for teaching assistants. These sources of support have varied over time, with federal funding for science stagnant from the 1970s through the mid-1980s. Then, beginning in the mid-1980s, there were quite substantial increases in federal research funding to colleges and universities in both the physical sciences and the health sciences. As a result, we expect supply shocks affect the doctorate education market. Increases in research funding or direct public support for graduate programs in the United States have the effect of increasing the number of opportunities for study in U.S. graduate programs. If the elasticity of demand for graduate study among those from abroad is greater than for the United States (perhaps because the opportunity cost is study in another country rather than a different career), funding shifts will yield relatively larger responses in degree attainment among foreign students, resulting in increasing share with positive shocks and decreasing share with adverse shocks. When the fraction of a country’s potential doctorate students choosing to study in the United States is initially small (or when there is excess demand among foreign students for U.S. programs), expansions in U.S. opportunities could plausibly have proportionately larger effects on the number of individuals pursuing a degree in the United States than when the share pursuing degrees is already quite large. A second explanation is that when foreigners considering studying in the United States have alternatives that are close substitutes (e.g., studying in Australia) elasticity of demand will be much higher. For those from the United States, the alternative to pursuing a Ph.D. at a U.S. university is unlikely to be a close substitute, demand will likely be more inelastic, and the change in graduate study in response to a supply shock somewhat more limited. 2.4.1 Evidence on Changes in the Share of Foreign Ph.D.s A starting point for understanding the dynamic in the variation in the representation of foreign students among U.S. doctorate recipients is to examine how country and field specific patterns differ from overall trends, which are presumably a function of secular changes. Table 2.3 illustrates country and field of Ph.D. degree receipt relative to total degrees awarded by U.S. universities at the start of each decade and during the peak 1994 to 1996 interval. In terms of growth rates, Ph.D. receipt for U.S. residents has

7 65 91 16 17 25 112 10 36 4 34 10 2 6 43 2,320

3,201

Total

Engineering

6,792

6 265 41 36 3 37 247 18 26 11 32 17 4 2 48 5,300

8,617

14 216 76 21 21 60 221 11 22 10 78 10 6 4 61 7,102 10,635

40 135 162 97 53 76 654 100 84 13 109 18 0 9 75 7,228 14,492

27 282 78 96 8 35 460 67 41 10 42 39 0 7 102 11,309

Life science

Engineering

Life science

Physical science

1969–1971

1958–1961

17,033

25 260 216 58 46 63 423 55 64 27 98 25 0 12 114 13,634

Physical science

7,587

92 54 16 139 30 23 685 298 95 9 101 52 9 20 41 3,375

Engineering

16,092

162 139 9 80 13 28 194 123 36 9 42 76 3 21 66 12,668

Life science

1979–1981

Ph.D. degrees awarded over time by U.S. universities by field and country of origin

Brazil Canada China Egypt France Germany India Iran Israel Italy Japan Mexico Russia/USSR Spain UK U.S.

Table 2.3

12,645

66 119 9 48 29 42 452 163 71 28 66 43 24 18 89 9,040

Physical science

18,278

163 144 2,259 154 85 80 1,718 229 59 31 117 97 61 26 47 6,620

Engineering

23,881

193 314 2,752 47 58 155 720 82 58 36 92 187 42 73 114 13,787

Life science

1994–1996

(continued )

20,213

106 228 2,882 30 83 246 912 85 82 97 79 107 219 40 102 9,880

Physical science

0.120

Total

Engineering

0.076

0.150 0.006 0.064 0.098 0.098 0.006 0.062 0.131 0.046 0.010 0.027 0.083 — 0.125 0.075 0.076

Life science

0.068

0.058 0.019 0.104 0.102 0.078 0.005 0.065 0.161 0.107 0.099 0.023 0.092 — 0.110 0.063 0.065

Physical science

Annual Change 1958–1961 to 1969–1971

0.174 0.073 0.058 0.180 0.114 0.111 0.176 0.230 0.085 0.118 0.116 0.059 — 0.041 0.056 0.114

(continued)

Brazil Canada China Egypt France Germany India Iran Israel Italy Japan Mexico Russia/USSR Spain UK U.S.

Table 2.3

0.179 0.071 0.216 0.018 0.049 0.022 0.086 0.061 0.013 0.011 0.000 0.067 — 0.110 0.044 0.011 0.010

0.034

Life science

0.083 0.092 0.232 0.036 0.057 0.120 0.005 0.109 0.012 0.037 0.008 0.106 — 0.080 0.060 0.076

Engineering

0.030

0.097 0.078 0.318 0.019 0.046 0.041 0.007 0.109 0.010 0.004 0.040 0.054 — 0.041 0.025 0.041

Physical science

Annual Change 1969–1971 to 1979–1981

0.059

0.038 0.065 0.330 0.007 0.069 0.083 0.061 0.018 0.032 0.082 0.010 0.042 0.128 0.017 0.009 0.045

Engineering

0.026

0.012 0.054 0.382 0.035 0.100 0.114 0.087 0.027 0.032 0.092 0.052 0.060 0.176 0.083 0.036 0.006

Life science

0.031

0.032 0.043 0.385 0.031 0.070 0.118 0.047 0.043 0.010 0.083 0.012 0.061 0.147 0.053 0.009 0.006

Physical science

Annual Change 1979–1981 to 1994–1996

Internationalization of U.S. Doctorate Education

77

lagged overall university doctorate production, particularly since about 1980 (refer back to figures 2.1 and 2.3). For foreign countries, several regimes are apparent. Canada, as well as the UK, present one case where degrees awarded by U.S. institutions largely echo the overall trend. South Korea, India (except in engineering), China, and Germany are cases where Ph.D.s awarded by U.S. universities to students from these countries far outstrip the secular trend through much of the 1980s. 2.4.2 Doctorate Program Quality Expansion in doctorate attainment at U.S. institutions among foreign students is not uniform among differently ranked graduate programs and, indeed, much of the growth recorded from the mid-1980s to the mid-1990s occurred at Ph.D. programs outside the most highly ranked. Figure 2.4 shows doctorates awarded to foreign students by rank of program. In physics, biochemistry, and chemistry much of the expansion in doctorate receipt to foreign students occurs at unranked programs or those ranked outside the top fifty; while the growth in foreign students in engineering is distributed more evenly among programs. Among students from China, Taiwan, and South Korea growth has been particularly concentrated outside the most highly ranked institutions. 2.4.3 Demand Changes at the Country Level A basic proposition is that growth in undergraduate degree attainment is likely to translate to increases in the overall demand for doctorate-level training and, specifically, growth in the number of students pursuing Ph.D.s at U.S. institutions. Figure 2.5 illustrates the time-trend in BA degree attainment in the sciences by country relative to the United States. The top two panels of figure 2.5 show the growth of undergraduate degrees relative to the base year of 1975 for European countries (and Canada) and Asian countries, respectively, and the final panel shows the number of degrees awarded by year in China. The top panel, which illustrates relative BA attainment for North American and European countries, shows near parity among countries from the 1976 BA year through 1985. Then, there is retrenchment in the number of BA degrees awarded in the United States (and to a lesser extent the United Kingdom) as birth cohorts shrank markedly in subsequent years. In the other countries in this panel, degree attainment continues to rise into the 1990s, reflecting somewhat different demographic trends and net increases in collegiate attainment within cohorts. The trends in North American and European countries are quite modest when seen in comparison to changes BA degree receipt among Asian countries (second panel). Most dramatically, BA degrees awarded in Korea grew by about 150 percent over the period shown in the graph. Both India and Taiwan witness considerable growth in BA degree attainment during the years in which BA degrees awarded in the United States were stagnant. The

Ph.D. degrees awarded by country and program quality (physics and economics)

Source: NSF, Survey of Earned Doctorates restricted-use microdata; authors’ tabulations. Program rankings are from National Research Council (1995).

Fig. 2.4

Internationalization of U.S. Doctorate Education

Fig. 2.5

79

Changes in BA attainment relative to the U.S., selected countries

Source: Data for India and Taiwan are from NSF (1993, NSF 93-303, table A-9) and include degrees in natural sciences and engineering; data for France, Germany, and the U.K. are from NSF (1996, NSF 96-316, table A-16) and include degrees in the natural sciences, math, agriculture, and engineering; data for China, Korea, and Japan are from NSF (2004); data for Canada are from UNESCO (annual series, 1963–1999).

most dramatic story, however, is the case of China, with the number of science and engineering degrees shown in the bottom panel. Although consistent data on the number of science and engineering degrees are difficult to piece together for China until the mid-1980s, undergraduate degree attainment has risen meteriorically over the past quarter century in China, rising from near zero in the mid-1970s to more than 330,000 BA degrees in science and engineering fields at the start of the twenty-first century. What the trends in undergraduate degrees by country suggest is that those countries with growth relative to the United States at the undergraduate level may translate to increased demand for doctorate education from U.S. institutions. Figure 2.6 illustrates this point in a general sense, with the annual rate of growth in BA degrees on the x-axis and the annual rate of

80

John Bound, Sarah Turner, and Patrick Walsh

Fig. 2.6 Changes in BA degrees and Ph.D. degrees conferred from U.S. institutions, by country Source: The figure shows the average annual percentage change in BA degrees awarded in a country (x-axis) for 1975–1992 relative to the average annual change in S&E Ph.D. degrees awarded by U.S. universities in the 1982–1999 interval, calculated from regressions of the log of degrees awarded on a time trend. BA data for the United States, Austria, Belgium, Denmark, Finland, France, Germany, Greece, Ireland, Italy, Netherlands, Portugal, Spain, Spain, Sweden, UK: NSF (1996, appendix table 5), while BA data for India, Japan, Singapore, S. Korea, China, Taiwan: (1993, table A-9). The Ph.D. data are from the authors’ calculations using the restricted access Survey of Earned Doctorates microdata.

growth in Ph.D.s awarded by U.S. universities (seven years later) on the y-axis. While the link is by no means exact—with some countries well above and below unity—the relationship is clearly positive. What is more, the figure makes clear the variation in the expansion of undergraduate degree attainment across countries. At one extreme, the United States, United Kingdom, and Japan hover at growth below 2 percent while South Korea evidences growth in BA attainment over 10 percent. China—as we discuss shortly—is a case that is literally off the chart in terms of the growth in Ph.D.s awarded between 1982 and 1992. That there are a number of countries such as Germany, Italy, and India where the growth in Ph.D.s awarded by U.S. institutions outstrips the home country growth in BA degree receipt suggests that growth in undergraduate degree production is but one factor determining the rise in the representation of foreign students among doctorate recipients from U.S. uni-

Internationalization of U.S. Doctorate Education

81

versities. We note that the growth in the presence of students from Germany and Italy in U.S. Ph.D. programs reflects catching up to other European countries in U.s. doctorate receipt. We also suspect that dramatic growth in Ph.D. programs in Korea and Taiwan over the last decade (see table 2.1) may have recently begun to have an effect on the relative attractiveness of U.S. Ph.D. programs for students from these countries. While it would be hard, if not impossible, to quantify the importance of the growth of networks for explaining the growing representation for foreigners in U.S. Ph.D. programs, anecdotal evidence points to their importance. Repeatedly we have been told of cases where someone from, for example, Italy was encouraged to seek graduate education outside of Italy by a professor who, himself, had been trained in the United States. We also find the students from particular countries tend to be overrepresented in particular Ph.D. programs. Thus, for example, in economics, Italians are overrepresented at MIT, Columbia, and NYU, while students from India are underrepresented at Harvard and overrepresented at Rochester, Columbia, Boston University, and Cornell. In contrast, in physics, Italians are again overrepresented at MIT, while students from India are overrepresented at Ohio, Stony Brook, Maryland, Rochester, and Texas. Such patterns are consistent with the importance of department and institutionspecific networks. 2.4.4 Country-Specific Shocks Beyond gradual changes in the demand for U.S. doctorate training generated by expansion in home country BA production among countries with long-standing diplomatic and trade ties with the United States, political shifts produce sharp changes to foreign students’ access to the U.S. education market. Two of the most dramatic examples include the entry to the United States of Ph.D. students from China in the early 1980s and the dramatic decline in the flow of Ph.D. students from Iran in the late 1970s. Figure 2.7 illustrates these transformations for China and Iran. Representing the data by year of birth (in the right panels of figure 2.7) shows clearly the cohort-specific effects which tend to be somewhat attenuated when the data are arranged by year of Ph.D. given the natural variation in time to degree. China China represents the most extreme case. In the first part of the twentieth century there were relatively extensive exchanges between U.S. and Chinese universities, with many of China’s leading scientists trained in the United States. Exchange with western universities changed dramatically at midcentury. During Mao’s Cultural Revolution (from 1966 to 1976) university activity was largely disrupted. The establishment of diplomatic relations with the United States in 1979 dramatically changed the level of educational

Case studies of countries with large changes in graduate student flows to the U.S.

Source: NSF, Survey of Earned Doctorates restricted-use microdata; authors’ tabulations. Note: Doctorates awarded to students from China (panel A) or Iran (panel B) on left axis; doctorates awarded to students from United States shown on right axis.

Fig. 2.7

Internationalization of U.S. Doctorate Education

83

exchange with China. China sought to jump-start its development process through access to science and engineering technology through U.S. university education and, at least initially, most students and scholars visiting the United States from China came on J-1 exchange visas. A disproportionate share of the first wave of exchange students coming to the United States were related to high-level Chinese officials, including the son of Deng Xiaoping and the son of the Foreign Minister (Wong 1981), though there was also considerable competition among U.S. universities to identify the most talented among the Chinese students. The establishment of networks early on was particularly important in opening doctorate education. One important example was the ChinaUnited States Physics Examination and Application (CUSPEA) program initiated in the fall of 1979 by the Chinese-American Nobel Laureate physicist T. D. Lee of Columbia University. The intent of the initiative was to identify gifted graduate students through examination in China and to place these students at U.S. universities. During the course of the program, CUSPEA placed more than 900 students in physics programs at U.S. universities.20 When we look at the data for China organized by year of birth or year of college entry, the cohorts born in 1962 to 1963 and entering college in 1978 are extraordinary in representation among U.S. Ph.D. recipients in the sciences. These cohorts captured considerable pent-up demand for undergraduate education and represented the first full class of students admitted to Chinese universities y competitive examination in the aftermath of the Cultural Revolution. Add to this strong encouragement from the government to study abroad combined with relatively few domestic opportunities, and many students from this cohort received Ph.D.s from U.S. universities. To illustrate the unusual impact of this single cohort, we note that of the Ph.D. degrees awarded to students from China in the decade between 1985 and 1994, 46.6 percent o the 11,197 Ph.D.s awarded to students from China had started college in 1978. What is more, if one eliminates this cohort the downturn in degrees awarded to students from China after 1995 virtually disappears. Iran While the case of students from China over the course of the last two decades is one of increased involvement with U.S. universities, Iran represents a counterexample. In the late 1960s and early 1970s, Iranian doctorate attainment—particularly in engineering—rose rapidly, reflecting move20. To put these numbers in perspective, the total number of Ph.D. degree recipients from China receiving degrees in physics between 1980 and 1992 was 1,062. Of course, there were other channels through which Chinese students could study physics in the United States, but the CUSPEA program clearly had a substantial impact in generating a network or link between leading U.S. and Chinese universities.

84

John Bound, Sarah Turner, and Patrick Walsh

ments of the country to modernize and improve technological infrastructure, particularly in relation to the petroleum industry.21 The political upheaval associated with the fall of the Shah in January 1979 and the hostage crisis at the American embassy in Tehran in 1979 brought an abrupt halt to the substantial participation of Iranian students in U.S. graduate education. While entry of graduate students stopped largely after 1979, it is plain that many students of Iranian origin chose to stay in the United States to finish their graduate studies. What is apparent in figure 2.7 is the sharp drop-off in degree attainment by birth cohort and the more gradual decline by year of degree receipt. Eastern Europe and Former Soviet Union In the years before 1989, barely a trickle of students from the Soviet Union completed doctorate degrees in the United States, with most of those students likely related to political émigrés. Then, perestroika in the Gorbachev years initiated modest exchange of graduate students and scholars (Raymond 1989). But the collapse of the former Soviet Union also led to significant declines among the traditional Soviet universities, which had long standing strengths in the physical sciences and had been generously supported by the government during the Cold War. By one estimate, funding for science in Russia decline 44.2 percent between 1989 and 1991 (Shkolnikov 1995). The result was an exodus of scholars and graduate students to the United States and universities in Europe and Israel. In the Eastern European countries of Bulgaria, Czechoslovakia, Romania, Hungary, and Poland, there are similar shifts in the flow of doctorates students to the United States corresponding to political transitions of the late 1980s and early 1990s. In summarizing the country-specific trends in doctorate attainment at U.S. universities, it is clear that both secular growth in home country undergraduate education and the sharp changes produced by political transformations in countries like China affect the representation of foreign students at U.S. universities. It is also the case—if somewhat more difficult to measure directly—that the establishment of networks providing informa21. In Iran, the oil boom of the early seventies brought a half a dozen new universities and an increased premium on western-trained academics (Pace 1976). In addition, many relatively affluent Iranian families paid to send their children to U.S. universities and, by one estimate, as many as 50,000 Iranian students were attending educational institutions in the United States before the fall of the Shah, accounting for one-fifth of the foreign student population in 1979. With the crisis following the Iranian revolution in the United States, Iranian students were severely limited in their capacity to finance studies in the United States and student visas were unattainable as diplomatic relations ceased. At the extreme, institutions like the University of Southern California had as many as 1,000 students from Iran. While many universities were able to make accommodations for Iranian students, it was the small colleges that suffered financial setbacks with the political shock. For example, the small Windham College in Vermont depended on Iranians for 30 percent of its enrollment and went out of business when these students were unable to make tuition payments (Hechinger 1979).

Internationalization of U.S. Doctorate Education

85

tion about U.S. universities and opportunities builds the floor of foreign students to U.S. universities. Yet the flow of students to U.S. doctorate programs need not reflect a permanent exodus of the highly skilled from the sending country to the United States (what is sometimes described as “brain drain”); there is clear evidence that the initial flows following political transitions capture considerable pent-up demand that subsides, particularly with investment in home country universities (Blanchard, Bound, and Turner 2008).22 2.5 Stagnation in Degree Receipt Among U.S. Students While funding for science at U.S. universities has increased in the last two decades, the number of Ph.D.s in the sciences awarded to students from the United States has been largely stagnant over the last two decades, falling somewhat in economics and rising only modestly in the life sciences. The number of Ph.D.s awarded to those from the United States in 2003 in the physical sciences, engineering, and economics remains below corresponding numbers from 1970. How do we explain the relatively anemic participation of students from the United States in doctorate-level science? 2.5.1 Undergraduate Degree Attainment in the United States As we indicated before (fig. 2.5), the growth in the number of individuals receiving undergraduate degrees in the sciences has been quite muted in the United States. Change in cohort size plays a central role in these trends. In the United States, the size of the college-age population (and, by extension the broad pool of potential Ph.D. recipients) grew rapidly with the college entry of the baby-boom cohorts, peaked in the mid-1970s, and then declined through the early 1990s. Thus, despite the fact that the fraction of cohorts obtaining undergraduate degrees in science and engineering during the 1980s and early 1990s rose at an average rate of 2 percent per year, the number of science and engineering BA’s hardly rose at all. As figure 2.6 suggests, the growth in the number of science and engineering Ph.D.s being granted to U.S. residents is in line with the growth in the number of BAs awarded in science and engineering fields in the United States. Indeed, figure 2.6 would suggest that the slow growth in the number of science and engineering BAs being awarded in the United States relative to the growth in other countries, can go a long way toward explaining the drop in the U.S. share of science and engineering Ph.D.s in the country. Beyond overall changes in undergraduate degree attainment, the progression from baccalaureate attainment to Ph.D. completion has varied appreciably over the last four decades. The ratio of Ph.D. receipt to BA re22. See also Regets (2001) for a thoughtful discussion of determinants and timing of migration decisions of the highly-skilled.

86

John Bound, Sarah Turner, and Patrick Walsh

ceipt organized by year of BA rose during the early 1960s, fell through the 1970s, and has subsequently maintained a plateau. This ratio peaked at 0.056 in 1964 and had fallen to about 0.025 percent by 1974. Figure 2.8 makes this presentation at the level of field of study, aligning Ph.D.s by the year in which individuals received BA degrees in relation to the number of BA degrees awarded in a given year. While the number of Ph.D.s awarded in these sciences and engineering fields rose over the 1980s and early 1990s (fig. 2.3), the ratio of Ph.D.s to BAs did not change appreciably for those completing their undergraduate work (and potentially considering graduate study) in the 1980s. The growth of foreign students among overall Ph.D. recipients and Ph.D. recipients from U.S. institutions affects the flow of potential U.S. doctorate students through two potential channels. First, U.S. students may face increased competition for slots or admission to the most highly ranked programs, which typically have considerable excess demand. Second, beyond potential crowd-out effects in higher education, the overall growth in the number of foreign doctorates (both those who obtained their degrees in the United States and those who migrated after receiving their

Fig. 2.8

Ph.D.s to BA degrees for U.S. Residents by BA year

Source: The Ph.D. data are for doctorate recipients completing high school in the United States and organized by year of BA degree. These data are from authors’ tabulations of the Survey of Earned Doctorates microdata. The BA data are based on compilations of national data from the Earned Degrees Conferred Survey assembled in Goldin (1999).

Internationalization of U.S. Doctorate Education

87

degrees) is likely to have had a substantial effect on the labor market returns to Ph.D. awards in science (Bound and Turner 2006). 2.5.2 Direct Crowd-Out of U.S. Students by Foreign Students An important question is how changes in demand for U.S. doctorate education from foreign students affect the level and distribution of doctorate attainment among U.S. students. Changes in the rate at which U.S. students complete Ph.D. programs may reflect both student demand and the availability of opportunities in graduate programs. It is surely possible that, with a limited supply of places in graduate programs, the presence of foreign students may change opportunities for U.S.-born students, potentially initiating crowd-out at the doctorate level.23 It is hard to estimate the counter-factual of how large the growth of Ph.D. programs would have been in the absence of this substantial inflow of foreign students. Some crowd-out—with foreign students lowering degree attainment among U.S. residents—is likely to follow as U.S. students become less likely to receive admission offers from the top programs and expansion in the total number of degrees awarded reduces expected wages. Yet estimates of crowd-out are inherently difficult to estimate because it is necessary to separate increases in demand among foreign students from other factors such as funding shocks, which would lead to increases in scale of graduate programs. The magnitude of crowd-out effects ultimately depends on the elasticity of supply in U.S. doctorate programs. We suspect that at least in the short run, additional foreign students reduce the number of U.S. students 1:1 in the most highly-ranked programs where nearly all students enter with full funding and class size is essentially fixed. Somewhat further down the distribution of program quality, programs appear to be much more elastic in scale. Indeed, for the programs that are unranked or ranked very modestly, the period of growth in the 1960s and early 1970s represented both expansion in scale and the entry of new programs; the entry of new programs in this category was extraordinary, with a threefold increase in the primary science fields. As the market contracted in the 1970s and then expanded in the 1980s, the adjustment came in terms of the scale of programs, with apparently few programs either exiting or entering the market. 23. Some previous research attempts to estimate the extent to which foreign graduate students tend to crowd-out U.S. students. In general, there is little conclusive evidence to support substantial crowd-out effects. Using data from the Survey of Graduate Students and Postdocs and variation within academic departments, Regets (2001) finds a largely positive association between enrollment of U.S. students and foreign students. Borjas (2007) uses within institution variation in graduate student enrollment measured in the Integrated Postsecondary Education Data System (IPEDS) surveys and finds a negative effect of foreign enrollment on the level of enrollment of white men, though little effect on domestic enrollment in aggregate. This previous research is limited to the extent that increases in the representation of foreign students in U.S. graduate programs may well be endogenously related to other factors, such as the availability of funding simultaneously affecting the demand for graduate students.

88

John Bound, Sarah Turner, and Patrick Walsh

The case of the sharp increase in demand among Chinese graduate students beginning in the early 1980s presents a clear opportunity to assess the adjustment of the U.S. market to a sharp demand shock. Figure 2.9 illustrates using the example of the field of physics with the data on doctorates awarded by year of graduate school entry, which makes the magnitude of the change among the Chinese students all the more striking. At top-ranked programs, the number of additional students from China is small and there is little discernable change in the overall number of Ph.D.s awarded. At the other extreme (bottom right panel), the number of Chinese students receiving Ph.D.s from universities outside the top fifty increased from seven to 202 between the 1980 year of graduate entry and the 1985 year of graduate school entry. Notably, this large shock produced no

Fig. 2.9 Supply-shock case study of physics by rank and year of graduate school entry Source: NSF, Survey of Earned Doctorates microdata. Notes: National origin is defined by the country in which an individual went to high school. Year of graduate school entry is adjusted to reflect year of MA completion for those students with an MA degree received from an institution from outside the United States.

Internationalization of U.S. Doctorate Education

89

notable decline in Ph.D.s awarded to U.S. students at these institutions, with this number actually rising slightly from 164 to 199, while the number of students from other countries receiving Ph.D.s also rose over this interval of graduate school entry. Data for other fields show similar patterns. Remarkably, this large cohort of Chinese students had no discernable impact on the number of U.S., or for that matter, other foreign students receiving Ph.D.s in the sciences. We found this evidence that the large influx of Chinese students in the early 1980s seemed to have no noticeable crowd-out effects surprising. We suspect a combination of factors may have been at work. This was a period of time in which funding for the sciences in general, and the physical sciences in particular, was expanding rapidly. One senior physicist described how the influx of Chinese students at his research university met a need and allowed the department to expand, as funding for physics remained glowing in the 1980s with the persistence of Cold War federal funding. At the same time the number of undergraduates from the United States obtaining degrees in the physical and life sciences was stagnant or declining and the size of college-age population in the United States was declining.24 Thus, the capacity for U.S. graduate programs to expand rapidly had they relied on U.S. students might have been quite limited. Thus, it is unclear whether under different circumstances a similar demand shock would work in the same way (i.e., have no impact on other groups).25 With this in mind we did similar analyses using data from the former Soviet Union and Iran. Unfortunately, these shocks were not large enough for them to provide useful information with respect to crowd-out. Our analysis suggests that substantial changes in the doctorate study of foreign students have not led to direct crowd-out of the best and brightest U.S. students in top programs, as much of the expansion in study among students from abroad has come in less highly-ranked programs. 2.5.3 The Opportunity Cost of Ph.D. Attainment for U.S. Students Examination of the trends in the labor market rewards for Ph.D. scientists in the United States relative to other high-skill workers over the last quarter century suggests that the relative returns to advanced study in the 24. Note that the decline in the number of science and engineering BA degree recipients is largely a reflection of the decline in cohort size from its peak in the late 1970s. 25. It is also worth considering whether the absence of crowd that we see among doctorate recipients would also be apparent if we were able to examine first-year enrollment by country of origin over a long horizon. One hypothesis is that the new foreign students—particularly in the case of the Chinese—replaced relatively weak domestic students, many of whom might have been expected to bow out of doctorate programs with MA degrees after experiencing difficulties at the stage or qualifying or preliminary exams. Because the entering Chinese students were often extremely well technically prepared, the examinations in early stages of graduate study were less likely to be substantial hurdles to completion.

90

John Bound, Sarah Turner, and Patrick Walsh

sciences have not increased.26 Earnings of those early-career advanced degrees in the sciences have actually decreased in relation to the earnings of other college-educated workers, with the latter having risen overall in the last two decades as is well-known. Before turning to the earnings evidence, an important structural shift in the expected career paths of advanced degree recipients in science and engineering fields is the increasing reliance and expectation for postdoc appointments. Freeman et al. (2001) note that the time between graduate school entry and completion of training in the life sciences has increased from less than ten years in the 1970s to over 11.8 years in the 1990s, largely reflecting the increased expectation of postdoctorate appointments and the extended duration of these appointments. Similarly, National Science Foundation data show a dramatic increase in the number of Ph.D.s holding postdoctorate appointments in university departments of science and engineering between 1981 and 1998, rising from approximately 18,000 to 39,000. Relatively low wages associated with postdoc appointments combined with the increased uncertainty about permanent employment prospects detract from some of the attractive features of investment in doctorate-level training and careers in the sciences. Further, it may be that beyond the decline in relative earnings associated with science and engineering, these careers may be particularly unattractive given the long hours and difficulties in accommodating two-career families in university labor markets. The labor market provides considerable clues in understanding why there has not been a larger response among U.S. students to opportunities for doctorate study. Figure 2.10 shows the trends in salaries by field for those within ten years of doctorate receipt, from 1973 to the present. The dashed line represents the corresponding trend for BA recipients (ages twenty-five to thirty-four) more generally, calculated from the current population survey (CPS). To be sure, real earnings of doctorate-level scientists have increased over the last decade. Yet, as indicated by the dashed line to the field-specific series, the increases in the earnings of scientists have risen less rapidly than BA recipients more generally, with the exception of the physical sciences where the changes are near equal. Focusing just on faculty labor markets, the rate of growth for young academics in the sciences has lagged behind of that college-educated workers. Examining faculty salaries at public institutions by rank and field in all 26. Our data on the earnings of early career advanced degree recipients in the sciences come from two sources that tell similar stories. First, the Survey of Doctorate Recipients is a stratified random sample of Ph.D. recipients from U.S. universities across potential cohorts and provides earnings observations in odd years from 1973 to 2001, with data from the precursor National Register of Scientific and Technical Personnel providing evidence from 1958 to 1970. In addition, the Faculty Salary Survey series provides salaries by field and rank for public universities.

Internationalization of U.S. Doctorate Education

91

Fig. 2.10 Trends in earnings of science and engineering Ph.D.s relative to all BA recipients Source: Field-specific annual earnings are from the Survey of Doctorate Recipients and limited to those within nine-years of Ph.D. receipt. Notes: The “BA” trend is calculated from the March CPS and limited to those ages 25–34 and indexed to correspond to the mean within the indicated field in 1973. All data are limited to men.

of the broad science fields, the average annual rate of growth in academic salaries is less than 2 percent in both the 1980s and the 1990s, based on data from the Faculty Salary Survey (Oklahoma State University, various years). In comparison, the rate of growth in real earnings across all young workers with a BA degree was about 2.6 percent from 1994 to 2003.27 In effect, scientists employed in academics have done less well than college-educated workers more generally in the last decade. Moreover, there is some evidence that there have been changes in the profile of academic salaries by rank over the course of the last three 27. This rate of growth in earnings is yet larger if the comparison group is advanced degree recipients in the CPS.

92

John Bound, Sarah Turner, and Patrick Walsh

decades. The ratio of earnings of junior faculty to senior faculty has decreased over time, with the ratio of assistant to full salaries in the physical sciences falling from .62 in 1974 to 1975 to .57 in 2003 to 2004, and the ratio of assistant to full salaries in the physical sciences falling from .69 in 1974 to 1975 to .59 in 2003 to 2004 in the life sciences based on data from the Faculty Salary Survey (Oklahoma State University). (Economics is an exception, presumably as the nonacademic market remains strong in economics, with the ratio of junior to senior faculty salaries holding roughly constant over the interval.) Because reaching full professor is not guaranteed, this shift works to reduce incentives to enter science as rewards appear to have become more concentrated toward the senior level. What is striking is that growth in the earnings of new advanced degree recipients in science and engineering fields in the last two decades is muted relative to the overall market for college-educated workers. In contrast, during the scientific boom years of the late 1950s and the early 1960s, the increases in the salaries of scientists with advanced degrees tended to outstrip overall changes in earnings of college graduates (see table 2.4 as well as Freeman [1975]). It seems likely that other factors in addition to the rise in relative salaries were increasing the demand for graduate education during the 1950s and 1960s. Academic jobs were relatively plentiful in this period, owing to the expansion of undergraduate education through the early 1970s and that the federal government continued to provide substantial resources for the funding of scientific research. Moreover, the availability of student deferments provided an incentive for men to enter graduate school and persist toward the Ph.D. as a means to avoid military service in the late 1960s.28 Suggested by this comparison is a case that there has been a structural change in the labor market for those with advanced degrees in science and engineering. Where there were once only a modest number of potential students from foreign countries there are now, literally, thousands of potential students from countries like China. The resulting shift in the demand for U.S. doctorate programs over the last quarter century is surely central to the rising representation of foreign students among doctorate recipients from U.S. institutions. In turn, funding shocks (refer back to figure 2.2) in 28. For cohorts graduating from college in the early 1960s, the availability of 2-S deferments from military service for graduate study encouraged many students to seek out doctorate programs as a refuge from the risk of the draft. Then, in 1967, the provision allowing exemption for graduate study was eliminated. Under the Selective Service Act of 1967 (which became effective June 30, 1967) and Executive Order 11360, there would be a one-year grace period through the end of academic year 1967 to 1968 and then no more 2-S deferments would be granted to graduate students (except as specifically written into the law). Support for the proposition that the incentive to avoid military service inflated doctorate enrollment and attainment during this period is provided by the much larger relative decline in the progression of men relative to women into graduate education (Bowen, Turner, and Witte 1992).

Internationalization of U.S. Doctorate Education Table 2.4

93

Median salaries of Ph.D.-level scientists in the 1960s Median Ph.D. salaries (2000$) 1958

1960

1962

1964

1966

1968

1970

Biological sciences Physics Chemistry Economics

44,093 56,135 57,636 —

52,358 63,993 63,993 —

57,020 74,126 68,424 —

62,214 74,990 72,213 67,214

66,435 76,533 74,407 71,750

70,266 78,678 77,193 78,183

71,010 76,780 77,224 77,224

BA/CPS

40,911

—

—

48,910

52,297

55,703

56,279

Sources: American Science Manpower, various years. Current Population Reports, Consumer Income, Series P60, #’s 33, 48, 53, 66, and 80. Note: CPS numbers represent median money income for men aged 25 and older.

the sciences appear to be accommodated by foreign students as well as U.S. students, leading to much smaller changes in the wages of scientists and engineers in recent years, relative to the 1950s and 1960s. Analysis of the science and engineering labor market of the 1960s and 1970s found considerable empirical support for cobweb cycles in the labor market, with changes in labor market demand resulting in sharp fluctuations in wages. Boom periods led to substantial increases in the returns to science, where declines in funding brought about sharp downturns, resulting from the relative inelastic supply of scientific labor associated with the long time lag to doctorate production (see, e.g., Blank and Stigler [1957] and Freeman [1975]). Yet funding changes for science in the 1980s and 1990s did not lead to sharp increases in wages for scientists. One explanation is that the labor market drew in lots of foreign-trained scientists—in addition to retaining a number of foreign students educated in the United States—resulting in few incentives for U.S. students to change investments in scientific training. Many foreign-born workers among the highly skilled enter the United States having completed graduate study abroad. Indeed, a substantial number of foreigners first enter the United States as postdoctoral scholars (National Academy of Sciences 2005).29 According to the 2000 Census, close to 20 percent of foreign-born Ph.D.s in the United States had immigrated within the last four years, too short a time to obtain a Ph.D. Where the supply of those trained at the highest level in science and engineering disciplines in the United States might have accurately been described as inelastic in the short run during the 1960s, this structural feature 29. Data presented in a recent National Academy of Sciences report showed that of the 60 percent of academic postdocs who hold temporary visas, about 80 percent have non-U.S. doctorates (National Academy of Sciences 2005, 35), implying that about one-half of all U.S. postdocs in academic institutions have Ph.D. degrees from abroad.

94

John Bound, Sarah Turner, and Patrick Walsh

of the science and engineering labor market appears to have eroded.30 As such, our hypothesis is that the science and engineering labor market is much more internationally integrated now than three decades ago. The result is that changes in the labor demand for scientists are much more likely to be accommodated in the near term. The decision to stay in the United States by those from other countries receiving Ph.D.s from U.S. institutions is one mechanism for adjustment. In addition, the United States remains a net importer of doctorate degree recipients from other countries, which can be seen in the comparison of the number of doctorates awarded by U.S. institutions by country and year of birth and the representation of doctorate recipients by country in the Decennial Census files. 2.6 Conclusions and Discussion An undisputed empirical point is that there has been a dramatic rise in the share of doctorate degrees awarded by U.S. institutions to students from other countries. How do we explain the determinants of this change and the resulting variation in the countries of origin of these doctorate recipients in science and engineering fields? There is no single explanation for this quite dramatic change. And, perhaps more significantly, there is considerable variation across countries in the magnitude of the change in U.S. doctorate receipt and the underlying causal forces. A substantial part of the increased representation in foreign students can be explained in terms of growth in the demand for U.S. Ph.D. programs generated by the expansion of undergraduate degree attainment in countries with relatively modest university systems (particularly as they existed two decades ago) like South Korea. Changes in political circumstances— as with the cases of China and the former Soviet Union—also produce sharp changes in the flow of doctorate students to the U.S. university system. With substantial differences in home country opportunities, it is natural that students from countries where options are more limited will be distributed at a broader range of institutions (and less concentrated at the highest quality programs) than students from countries where opportunities are closer to those found in the United States. Still, increases in demand for doctorate study among foreign students cannot account for the full expansion of foreign doctorate attainment or the relative stagnation in attainment among U.S. students, particularly in 30. The detailed evidence on the earnings of those in faculty position in the sciences and Ph.D. recipients employed in the United States makes clear that this group of workers did not capture the rents to increases in federal funding for sciences in the last two decades as suggested by Goolsbee (1998). Instead, we believe the measures employed in this analysis capture essentially the secular changes in wages to the college-educated in this period (which, in turn, are correlated with science funding) rather than the effect of federal funding on the earnings of advanced degree recipients in science and engineering fields relative to other collegeeducated workers.

Internationalization of U.S. Doctorate Education

95

the 1990s. Substantial increases in public support for science and engineering research fueled supply-side expansion in many fields. It is quite plausible that elasticity of demand and associated response to such shocks among foreign students may be somewhat larger than for U.S.-based students. Beyond direct supply-side shifts, the role of international networks and the process by which they have expanded over the last quarter century surely contributes to the internationalization of U.S. doctorate education. That growth in Ph.D. receipt among U.S. students in the sciences has not kept pace with the outcomes for foreign students is also likely a response to the labor market for advanced degree recipients in these fields. Despite what is perceived as a relative boom period for scientific fields in the 1990s, the earnings gains for advanced degree recipients in the sciences actually trailed those of college-educated workers more generally. To this end, the educational choices of U.S. students should be no surprise. A change that is remarkable, nevertheless, is the increased internationalization of the labor market for advanced degree recipients in science and engineering fields. One immediate effect of this transformation is the reduction of the large swings in the earnings of scientists associated with changes in federal support for research. Much more work is yet to be done before one can present a full analysis of the welfare effects of the internationalization of both doctorate education and the science and engineering labor market. We suspect that the resources of U.S. research universities are a lure for the best and the brightest across the world. If there are benefits to concentrating talent (agglomeration effects), then international output is expanded. With some foreigners trained in the United States returning to their home countries, there are surely home country benefits if these scientists are able to spur development of science, while also engaging in the exchange of ideas internationally through networks developed in the United States. Benefits also accrue in the United States, as the influx of scientists—trained in both the United States and abroad—reduces labor costs and increases the flexibility in the supply of science and engineering workers. Yet all these benefits come with some costs and it seems clear that some individuals would have pursued advanced degrees in science and engineering in the absence of the substantial foreign flow into graduate education and the labor market.

References Berelson, B. 1960. Graduate Education in the United States. New York: McGrawHill. Blanchard, E., J. Bound, and S. Turner. 2008. Opening (and closing) doors: Country-specific shocks in U.S. doctorate education. In Doctoral Education and

96

John Bound, Sarah Turner, and Patrick Walsh

the Faculty of the Future, ed. R. Ehrenberg and C. V. Kuh, forthcoming. Ithaca, NY: Cornell University Press. Blank, D., and G. Stigler. 1957. The demand and supply of scientific personnel, no. 62. General Series, National Bureau of Economic Research. New York: National Bureau of Economic Research. Borjas, G. 2007. Do foreign students crowd out native students from graduate programs? In Science and the University, ed. P. Stephan and R. Ehrenberg, 134–49. Madison, WI: University of Wisconsin Press. Bound, J., and S. Turner. 2006. International flows of college-educated workers: Estimates of the effect on skilled workers in the United States. Unpublished Manuscript. University of Michigan and the University of Virginia. Bowen, W., M. Kurzweil, and E. Tobin. 2005. Equity and excellence in American higher education. Charlottesville, VA: University of Virginia Press. Bowen, W., and N. Rudenstine. 1992. In pursuit of the PhD. Princeton, NJ: Princeton University Press. Bowen, W., S. Turner, and M. Witte. 1992. The B.A.-Ph.D. nexus. Journal of Higher Education 63 (1): 65–86. Carrington, W., E. Detragiache, and T. Vishwanath. 1996. Migration with endogenous moving costs. The American Economic Review 86 (4): 909–30. Finn, M. 2003. Stay rates of foreign doctorate recipients from U.S. universities. Oak Ridge, TN: Oak Ridge Institute for Science and Engineering. Freeman, R., 1975. Supply and salary adjustments to the changing science manpower market: Physics, 1948–1973. American Economic Review 65: (1) 27–39. Freeman, R., E. Weinstein, E. Marincola, J. Rosenbaum, and F. Solomon. 2001. Competition and careers in biosciences. Science 294 (5550): 2293–94. Gereffi, G., V. Wadhwa, B. Rissing, and R. Ong. 2008. Getting the numbers right: International engineering education in the United States, China, and India. Journal of Engineering Education 97 (1): 13–25. Goolsbee, A. 1998. Tax and human-capital policy: Does government R&D policy mainly benefit scientists and engineers? The American Economic Review 88 (2), Papers and Proceedings of the 110th Annual Meeting of the American Economic Association: 298–302. Hechinger, F. 1979. Iranian plight puts a spotlight on U.S. colleges. New York Times, February 20. Lowell, L. 2000. H-1B temporary workers: Estimating the population. University of California, San Diego. The Center for Comparative Immigration Studies, Working Paper no. 12, May. Maeroff, G. 1979. Colleges facing decline, advised to lure students. New York Times, January 1. National Academy of Sciences. 2005. Rising above the gathering storm: Energizing and employing america for a brighter economic future. Committee on Science, Engineering, and Public Policy. Washington, D.C.: National Academy of Sciences. National Academy of Sciences-National Research Council. 1958. Doctorate production in United States universities 1936–1956 with baccalaureate origins of doctorates in sciences, arts and humanities. Publication 582. Washington, D.C.: National Academy of Sciences. National Research Council. 1995. Research-doctorate programs in the United States: Continuity and change. Washington, D.C.: National Academy of Sciences. National Science Board. 1996. Science and engineering indicators. National Science Foundation. Available at http://nsf.gov/statistics/seind96/. ———. 2000. Science and engineering indicators. National Science Foundation. Available at http://nsf.gov/statistics/seind00/.

Internationalization of U.S. Doctorate Education

97

———. 2004. Science and engineering indicators. National Science Foundation. Available at http://nsf.gov/statistics/seind04/. National Science Foundation. 1993. Human resources for science and technology: The Asian region. NSF 93-303. Washington, D.C.: National Science Foundation. ———. 1996. Human resources for science and technology: The European region. NSF 96-31. Washington, D.C.: National Science Foundation. National Science Foundation, Reports of the National Register of Scientific and Technical Manpower, American Science Manpower. Washington 1954–1955, 1956–1958, 1960, 1962, 1964, 1966, 1968, 1970. Oklahoma State University, Office of Institutional Research and Information Management. Faculty salary survey by discipline. Stillwater, OK. Pace, E. 1976. A new Middle East boom. New York Times, April 25. Raymond, C. 1989. Seventeen Soviets start graduate work at U.S. institutions. The Chronicle of Higher Education (October 11). Available at http://chronicle.com/ che-data/articles.dir/articles-36.dir/issue-06.dir/06a04302.htm. Regets, M. 2001. Research and policy issues in high-skilled international migration: A perspective with data from the United States. Institute for the Study of Labor (IZA) Discussion Paper no. 366. Roy, A. D. 1951. Some thoughts on the distribution of earnings. Oxford Economic Papers 3 (2): 135–46. Shanghai Jiao Tong University, Institute of Higher Education. 2003. Academic ranking of world universities. Available at http://ed.sjtu.edu.cn/ranking.htm. Shkolnikov, V. 1995. Potential energy: Emergent emigration of highly qualified manpower from the former Soviet Union. Ph.D. diss. RAND Graduate School. Santa Monica, CA. Stevens, R. 2005. University to uni: The politics of higher education in England since 1944. London: Politico’s Publishing. United Nations Educational, Scientific, and Cultural Organization. (UNESCO). (annual series) 1963–1999. Statistical yearbook. Paris: UNESCO. Wong, J. 1981. China’s leap to American campuses. New York Times, November 15.

3 Improving the Postdoctoral Experience An Empirical Approach Geoff Davis

3.1 Introduction The population of postdoctoral researchers (“postdocs”) in the sciences and engineering has undergone a large expansion, nearly tripling over the last thirty years (National Science Foundation 1983–2003). While these scientists have produced tremendous quantities of new research, the relatively rapid growth in their ranks has been accompanied by two problems. First, the increase in the supply of postdocs has not been accompanied by a commensurate increase in the demand for them, at least in the academic sector. Second, large postdoctoral populations on campuses have strained institutions’ capacities for providing these researchers with basic administrative oversight. To address these concerns, leaders in the scientific community have called for changes in the postdoctoral experience, most notably improved compensation, augmented professional development opportunities, and increased administrative oversight. Each of these recommended measures comes at a cost, so assessing their relative benefits is important if institutions are to allocate their resources efficiently. In this chapter we will develop such an assessment. The current absence of standards for the postdoctoral experience means that even within a single department there can be considerable variation in Geoff Davis is a senior quantitative analyst in the User Experience Research Group at Google. The Sigma Xi Postdoc Survey was funded by a grant from the Alfred P. Sloan Foundation. Analysis of the survey data was funded in part by The Wertheim Fund at the Labor and Worklife Program, Harvard Law School. The author would like to thank Richard Freeman for insightful comments and suggestions, Jenny Zilaro for her work managing the survey, and Sigma Xi, the Scientific Research Society, for hosting the project.

99

100

Geoff Davis

working conditions and compensation packages for postdocs. We can use this variability to gauge the impact of proposed changes to the postdoctoral experience. We use linear models to isolate the effects of specific measures on outcomes using data from a large-scale survey of postdocs. The results are striking: a handful of straightforward and relatively inexpensive measures appear to make a large difference in postdoc productivity and in the overall quality of the postdoctoral experience. 3.2 Background A postdoctoral appointment is a short-term apprenticeship immediately following the completion of doctoral work that is designed to further prepare new Ph.D.s to become independent researchers. When postdoctoral positions were first instituted a century ago, they represented rare opportunities for some of the most promising young scholars to enhance their skills. In recent years, however, postdoctoral scholars have become increasingly common. As of 2003 there were 46,807 postdocs employed at academic institutions (NSF 1983–2003) and roughly 11,000 to 12,000 in other sectors (primarily government labs and industry) (National Academy of Sciences [NAS] 2000). Postdocs perform a substantial fraction of the skilled work in research labs and are responsible for a disproportionate share of new discoveries. A 1999 study found that 43 percent of first authors of research articles in Science were postdocs (Vogel 1999) (in science and engineering journals, the primary contributor to a paper is usually listed first). The recent growth in the postdoctoral ranks is less a planned expansion than the result of a combination of economic and political factors. A substantial increase in the graduate student population in the late 1980s, fueled by increased National Science Foundation spending, a doubling of the budget of the National Institutes of Health over the latter half of the 1990s, and the increased ability of young researchers from the former Soviet Union, Eastern Europe, and China to come to the United States, have all increased the supply of postdocs. Over the same time period, university faculties—historically postdocs’ primary employment destination—have grown much more slowly. Many scientific and academic leaders have raised concerns about the side effects of this postdoc expansion. The first set of concerns has to do with structural changes in the labor market. In many fields, particularly in the life sciences, a postdoctoral appointment has evolved from an optional educational enhancement to a de facto prerequisite for a faculty position (Comission on Professionals in Science and Technology [CPST] 1998). The result has been a substantial lengthening of the time spent training: recent cohorts of Ph.D.s will not begin fully independent research until their

Improving the Postdoctoral Experience

101

early forties (National Research Council [NRC] 2005). Tenure-track faculty positions have become more difficult to come by, and as a result many scholars spend increasing amounts of time in a frustrating postdoctoral “holding pattern” waiting for an academic job (NRC 1998). The diminishing probability of obtaining a faculty position has engendered fierce competition for relative advantage among researchers (Freeman et al. 2001a). Because universities are now able to draw upon a large pool of able would-be postdocs from less-developed countries, declining career opportunities have not resulted in a corresponding reduction in the supply of postdocs as in the past (Freeman 1990). The second set of concerns is related to administrative matters: many institutions have been slow to address the needs of the postdoc population in a systematic fashion. “Postdoctoral education today is almost exactly where Ph.D. education was in the 1890s—very ad hoc,” declares Steven B. Sample, president of the University of Southern California and chair of the Association of American Universities (AAU) Committee on Postdoctoral Education (NAS 2000). At some institutions postdocs are not classified either as students or as faculty/staff and, as a result, receive the benefits and protections of neither. Postdocs are in some cases poorly remunerated, retirement benefits are the exception rather than the rule, and nonmonetary aspects of work are in some cases only addressed on an improvised basis. There are no standard expectations for the supervision and mentorship of postdocs. Grievance resolution procedures are often ill-defined. Campus career services are usually geared exclusively toward undergraduates, occasionally graduate students, and only rarely postdocs. Educational leaders, funding agencies, and postdocs all agree on the need for improvements in postdoctoral working conditions and have advocated five broad classes of practices be implemented by those employing and funding postdocs (Association of American Universities 1998; NRC 2005; National Postdoctoral Association 2005): 1. Fellowships: A larger fraction of postdocs should be funded individually (i.e., funded via a fellowship/traineeship as opposed to a grant made to a senior faculty member). 2. Salary: Postdoc stipends/salaries should be increased. 3. Benefits: Postdocs should receive basic benefits, particularly health and retirement benefits. 4. Professional Development: Employers of postdocs should provide professional development opportunities to prepare postdocs for a variety of careers. 5. Structured Oversight: Institutions employing postdocs should develop postdoc-specific policies and should require (or strongly encourage) such practices as individual development plans, regular reviews, and so on.

102

Geoff Davis

3.3 Improving Postdoctoral Training Implementation of the practices recommended previously is underway on many campuses. More than forty institutions have created postdoctoral offices tasked with ensuring the well-being of their postdocs. Postdocs have started forming institution-level organizations to advocate for improvements in their working environments, often with support from their institution’s administrations. At present there are roughly fifty such postdoctoral organizations in the United States, and the National Postdoctoral Association has been created with the goal of coordinating local efforts and sharing resources. Disciplinary societies have started postdoctoral initiatives to enhance the postdoctoral experience, one of the largest being the Postdoc Network at Science’s Next Wave, formed in November 2000. The National Science Foundation has sponsored two recent workshops intended to inform specific programmatic and policy initiatives that it might undertake (Merrimack Consultants, LLC 2003; Westat Inc. and Merrimack Consultants, LLC 2004). An important question for all stakeholders in postdoctoral training is determining which, if any, of the advocated measures have the greatest impact. Because administrative responsibility for a postdoctoral appointment is typically held by a postdoc’s advisor rather than by a department-level or university-level administrator, and because the implementation of recommended measures is just beginning, there is considerable variation in working conditions for postdocs even within individual departments. We can use this diversity of working environments to good effect: by comparing postdocs working with different recommended measures in place, we can estimate the effects of specific measures on the overall postdoctoral experience. We analyze data from the Sigma Xi Postdoc Survey, a multi-campus survey of postdoctoral scholars carried out between December 2003 and April 2005. Sigma Xi conducted the survey at forty-seven institutions, including eighteen of the twenty largest academic employers of postdocs and the largest government employer. Over the course of the survey, Sigma Xi contacted some 22,400 postdocs, roughly 40 percent of the U.S. postdoc population. The survey’s overall response rate was 38 percent (Sigma Xi 2005). We tested the data set for nonresponse biases in two ways. First, we compared demographics of survey responders to known postdoc demographics at an institution that had detailed records of the sex, citizenship, and underrepresented minority status of its postdoctoral employees. The observed differences were within the range that would be expected due to sampling. Second, we looked for differences between early and late responders by regressing citizenship, sex, underrepresented minority status, and reported levels of overall satisfaction on the time between the start of the survey at a given institution and the time at which the respondent be-

Improving the Postdoctoral Experience

103

gan the survey. Such differences, if present, suggest differences between responders and nonresponders. Our analysis suggested an underrepresentation of African American postdocs in the survey respondents as well as small underrepresentation of non-U.S.-citizen postdocs. No significant variation over time was found for other underrepresented minorities, for sex, or for levels of overall satisfaction. Further details of the nonresponse analysis may be found at http://postdoc.sigmaxi.org/results/tech_reports. 3.4 Outcome Measures How do we measure the quality of postdoctoral experiences? One possibility is to follow the example of private foundations in evaluating the impact of their investments in young scientists. Recognizing that research careers span decades and that events during postdoctoral study can have an impact that unfolds over long periods of time, many foundations assess their impact by measuring publication rates and awards for those they fund some five to ten years afterwards (Pion and Ionescu-Pioggia 2003). These longitudinal studies have the advantage of allowing time for long-term investments to pay off, but they are expensive and labor-intensive. Bibliometric measures are useful in evaluating the success of postdocs who end up in tenure-track academic positions, but historically only about a third of postdocs have ended up in such positions (Regets 1998) and this fraction is likely shrinking (Davis 2005). We need a measure of success that is both easily obtainable and applicable to people with a broad range of career trajectories. We construct four different measures of success metrics based on Sigma Xi survey data, two of which are subjective and two of which are objective: 1. Subjective Success: This measure reflects a postdoc’s overall assessment of the current appointment. How satisfied is the postdoc with her current position? Is the current appointment doing a good job at preparing the postdoc to be an independent researcher? Is the appointment providing preparation for key aspects of the postdoc’s future career? Postdocs’ opinions about the success of their appointments are one useful measure of success. Given that postdocs have typically completed more than ten years of undergraduate and graduate education, and a third have already done at least one previous postdoc, they should have some sense of what constitutes effective training. They also best know their own career goals and should have an idea of how well their current experiences are preparing them to meet those goals. Ensuring that postdocs view their experiences as positive and successful can help institutions in hiring new postdocs, since satisfied postdocs are much more likely to recommend their current institution to others than dissatisfied postdocs (84 percent versus 30 percent). Information from post-

104

Geoff Davis

docs may influence undergraduates’ career decisions as well (Freeman et al. 2001b), so preventing dissatisfaction in the postdoctoral ranks may be important in convincing younger students to pursue science careers. 2. Advisor Relations: This measure gauges the quality of the postdoc’s relationship with his advisor. How would the postdoc rate his advisor’s overall performance? How does the postdoc think his advisor would rate his overall performance? Does the postdoc consider his advisor to be a mentor? In the idealized postdoctoral appointment, a postdoc’s advisor serves as a mentor, and he and the postdoc have a close working relationship. Positive relationships are important because much of the training that takes place happens through the postdoc’s interaction with his advisor. 3. Absence of Conflict/Misconduct: Has the postdoc had a conflict with her advisor? Has she seen misconduct in her work group? The absence of conflict and misconduct is a more objective complement to the subjective measure of advisor relations as described previously. The scores are related: postdocs who reported conflict/misconduct had an average advisor relations score that is 0.4 standard deviations lower than those who did not. Keeping conflicts rare is particularly important because of the power disparity in the advisor/postdoc relationship: a serious conflict can end a postdoc’s career. A recent survey (Martinson, Anderson, and de Vries 2005) shows relatively high rates of minor misconduct in science. In this context, a conflict and misconduct-free postdoc is one form of success. 4. Productivity: Postdoctoral appointments are training experiences, but they are also a source of new research. An appointment that is scientifically productive, as measured by papers and grant proposals submitted, can be considered successful. To measure research productivity, we compute the rate at which postdocs submit papers to peer-reviewed journals per year. We also look at the rate of submission of papers for which the postdoc was the primary author as well as the rate of grant proposals submission. The Sigma Xi survey questions asked about the total number of papers and grants submitted as a postdoc, so our measures show productivity over a respondent’s entire time as postdoc, not just for the current appointment. The details and summary statistics for these success measures are shown in the appendix. Distributions of the measures are shown in figure 3.1. The subjective success and advisor relations have roughly normal distributions with a positive skew. The success distribution decays more slowly than a Gaussian, however, indicating the presence of more unsuccessful experiences than would be expected if the components of the measure were wellmodeled as jointly Gaussian. Productivity, as measured by the number of papers submitted per year (excluding first-year postdocs to avoid small-

Improving the Postdoctoral Experience

Fig. 3.1

105

Distributions of success measures

Table 3.1

Success Advisor No conflict Productivity

Correlations between success measures Success

Advisor

No conflict

Productivity

1.000 0.448 0.194 0.117

0.448 1.000 0.142 0.094

0.194 0.142 1.000 0.032

0.117 0.094 0.032 1.000

denominator problems), is roughly exponentially distributed; the distribution of the log of the rate, excluding zeros, is roughly normal. Pairwise correlations between the success measures are shown in table 3.1. There is a modest correlation between the subjective success and advisor relations measure, which is not surprising given the importance of the advisor-postdoc relationship in the overall success of the endeavor. Corre-

106

Geoff Davis

lations between other pairs of measures are all low, indicating that we are measuring disparate aspects of the experience. 3.5 Measures of Recommended Practices We next define a set of measures of the implementation of the recommended practices. The first is straightforward: the individual funding measure is an indicator variable that set to 1 if the postdoc is funded individually. For the responses we use in our analysis (those from nonclinical-fellow postdocs working full-time), 20 percent report that their funding was from “a grant, contract, or fellowship that was awarded directly to [the postdoc].” The primary sources of these fellowships are private foundations/ associations/disciplinary societies (37 percent) and NIH National Research Service Awards (22 percent). The salary measure is simply the postdoc’s annual salary. For the responses we analyze below, the mean salary was $39,305 and the standard deviation was $7,194. In our regressions we use the natural log of the salary, normalized to have zero mean and unit variance. The other measures are normalized counts of features of the postdoctoral experience. The structured oversight measure counts the number of recommended practices such as research plans, formal reviews, and so on, that are implemented in the current appointment. The professional development measure counts the types of training postdocs reported receiving, either via formal coursework or on-the-job experience, in their current appointments. The health insurance measure is an indicator variable set to 1 if the postdoc has health insurance; the benefits measure counts the other types of benefits available in the current appointment. The summary statistics and individual items counted for each measure are detailed in the appendix. On average, postdocs indicated that six of the sixteen forms of structured oversight were implemented, they received professional development in six of the twelve areas counted, and reported that eleven of the eighteen forms of benefits were available; 98 percent received health insurance. 3.6 Distribution of Practices Distributions of the measures of recommended practices are shown in figure 3.2. As with the outcome measures, most of the distributions resemble skewed normals, with heavy tails in some cases. Table 3.2 shows the pairwise correlations between the measures. There is a weak correlation between the structured oversight and professional development measures that likely arises from some institutions devoting more resources to postdocs via both oversight and formal training offerings. The other pairwise correlations are all very low.

Improving the Postdoctoral Experience

Fig. 3.2

107

Distributions of recommended practices measures

Table 3.2

Correlations between measures of recommended practices Structured oversight

Structured Oversight Professional Development Benefits log(Salary) Funding

Professional development

Benefits

log(Salary)

Funding

1.000

0.302

0.130

0.022

0.074

0.302 0.130 0.022 0.074

1.000 0.106 0.024 0.045

0.106 1.000 0.016 0.099

0.024 0.016 1.000 0.114

0.045 0.099 0.114 1.000

Table 3.3 shows regressions of the recommended practices measures on institution, field, duration variables, and postdoc demographics. There are citizenship-related differences in pay and likelihood of independent funding (citizens and permanent residents earn 4.7 percent more than temporary visa holders, about $1,850/year, and have 82 percent higher odds of being independently funded), and postdocs with medical degrees report

108

Geoff Davis

Table 3.3

(Intercept) Male Citizen_or_pr Underrepresented Married Children Age Medical_degree Years_total Years_current Previous_postdocs N

Recommended practices measures regressed on demographic variables Structure

Prof dev.

Health

Benefits

Log(Salary)

Funding

0.379** (0.136) 0.111** (0.034) 0.065. (0.035) 0.037 (0.093) 0.017 (0.038) 0.076. (0.041) 0.054** (0.019) 0.103. (0.057) 0.026 (0.018) 0.099*** (0.021) 0.045 (0.034)

0.369* (0.149) 0.04 (0.037) 0.299*** (0.038) 0.044 (0.102) 0.072. (0.042) 0.029 (0.044) 0.029 (0.021) 0.180** (0.062) 0.056** (0.020) 0.077*** (0.023) 0.011 (0.037)

13.532 (271.983) 0.381* (0.180) 0.282 (0.180) 0.327 (0.495) 0.142 (0.206) 0.248 (0.207) 0.049 (0.092) 0.884** (0.276) 0.271* (0.121) 0.206 (0.136) 0.118 (0.180)

0.387*** (0.113) 0.03 (0.028) 0.021 (0.029) 0.004 (0.077) 0.046 (0.032) 0.060. (0.034) 0.048** (0.016) 0.005 (0.047) 0.046** (0.015) 0.061*** (0.018) 0.067* (0.028)

0.532*** (0.095) 0.015 (0.024) 0.260*** (0.024) 0.089 (0.065) 0.032 (0.027) 0.078** (0.028) 0.028* (0.014) 0.076. (0.040) 0.145*** (0.013) 0.024 (0.015) 0.02 (0.024)

2.723 (7.192) 0.062 (0.057) 0.604*** (0.059) 0.280. (0.145) 0.057 (0.065) 0.046 (0.069) 0.166*** (0.035) 0.108 (0.098) 0.05 (0.037) 0.220*** (0.041) 0.184** (0.067)

3,552

3,552

3,477

3,552

3,552

3,463

Notes: Standard errors are shown in parentheses. Robust regression with an M-estimator was used for the structure, professional development, benefits, and salary measures. Logistic regression was used for health insurance and independent funding. All regressions also included 46 dummy variables for institution and 95 for field of research; these have been omitted to conserve space. For this and subsequent regressions, ***designates a p-value of  0.001, **designates a p-value of  0.01, *designates a p-value of  0.05, and . designates a p-value of  0.10. Boldface indicates that coefficients maintain their sign and statistical significance (p-value  0.10 for the smaller data sets) in regressions on subsets of the data consisting of (1) all postdocs in their first appointment and (2) all postdocs in their second appointment.

greater levels of professional development. Apart from these, there are few other demographically linked differences in the best practices measures. Related analyses of variance for the continuously valued practices measures support our claim of heterogeneity in working conditions within institutions. Field and institution together explain just 11 percent of the variation in structured oversight, 5 percent of the variation in professional development, 19 percent of the variation in benefits, and 28 percent of the variation in log(salary). Many of the structured oversight questions ask about events that occurred at the start of a postdoc’s current appointment, so the negative “years_current” coefficient in the structured oversight model may result from postdocs forgetting details of the start of their appointment over time.

Improving the Postdoctoral Experience

109

The positive “years_current” coefficient for professional development is consistent with the notion that training accumulates over the course of one’s appointment. The positive “years_current” term in funding probably results from postdocs extending existing appointments upon the receipt of a fellowship. One thing we must be careful of in these regressions is that we are combining responses from postdocs in their first appointments with those from postdocs who have had multiple appointments. This creates potential endogeneity problems, since several of our model variables are likely to be linked to the reasons postdocs choose to pursue or not pursue further appointments. For example, those with high subjective success in a first appointment may be more likely to pursue a second, while those experiencing conflicts in a first appointment may be less likely to do so. We can reduce this problem by performing separate regressions for those in their first appointment, those in their second and so on, so that we have more uniform pools of responses. However, such a disaggregation comes at the price of smaller data sets to work with and reduced test power. In table 3.3 we introduce a convention that we will use for the remainder of our regressions: we will report results for the full data set, and then, as a confirmatory measure, we will perform separate regressions for postdocs in their first appointment and those in their second. We show in boldface coefficients that maintain their sign and statistical significance in the pooled data as well as the two disaggregated sets (a p-value  0.10 on the smaller data sets) and will focus our discussion on those coefficients. 3.7 Impact of Recommended Practices Do recommended practices for the postdoctoral experience have any measurable benefits? Table 3.4 provides a rough answer: each pair of columns compares components of our outcome measures for postdocs reporting the highest and lowest levels of our measures of recommended practices. Postdocs reporting the highest levels of oversight and professional development are more satisfied, give their advisors higher ratings, report fewer conflicts with their advisors, and are more productive than those reporting the lowest levels. High levels of benefits are associated with similar but smaller differences in three of the four categories. Those with individual funding show little difference from those without. Health insurance is accompanied by higher rates of satisfaction and better advisor grades, but lower productivity (likely because an absence of health insurance is most commonly the result of a selective fellowship with inadequate provisions for benefits). Salary appears to be associated with only minimal differences in the measures.

110

Geoff Davis

Table 3.4

Components of success measures as a function of funding mechanism, levels of structured oversight, professional development, benefits, and salary Direct funding

Satisfied Advisor grade Conflict Papers/year

Structured oversight

Yes

No

Top 25%

Bottom 25%

Top 25%

Bottom 25%

74% 3.0 14% 1.1

69% 3.1 14% 1.2

78% 3.4 11% 1.3

63% 2.8 18% 1.0

82% 3.4 11% 1.3

56% 2.7 17% 1.1

Health insurance

Satisfied Advisor grade Conflict Papers/year

Professional development

Salary Benefits

Yes

No

Top 25%

Bottom 25%

Top 25% ≥ $42,000

Bottom 25%  $35,000

71% 3.1 14% 1.1

61% 2.9 14% 1.5

75% 3.1 12% 1.2

65% 3.0 15% 1.2

71% 3.0 16% 1.2

67% 3.1 13% 1.2

Notes: Each pair of columns compares those with and without direct funding/health insurance and those in the top quartile and bottom quartile of structured oversight, professional development, and so forth. “Satisfied” is the percent reporting that they are satisfied overall with their position. “Advisor grade” is the average “grade” (on a 4 point scale) that postdocs give their advisors. “Conflict” is the percent reporting that they have experienced a conflict with their advisor. “Publications/year” is the average number of peer-reviewed publications submitted per year for those who have been postdocs for at least twelve months.

While table 3.4 is a useful start, we need to be much more careful if we are to obtain a reliable estimate of impact. There are important differences in the postdoctoral experience across research fields and institutions, and there is interplay between the contributing factors. Special populations may have different experiences. To test the hypothesis that the recommended practices impact our success measures while controlling for these various potentially confounding factors, we perform a set of multivariate regressions. We regress each of the measures of success on log salary, dummy variables for independent funding and health insurance, and our composite measures of structured oversight, professional development, and benefits. We add variables to control for sex, underrepresented minority status, citizenship, age, marital status, children, field of research (ninety-six fields), and employing institution (forty-seven institutions). For those reporting research in multiple fields, we weight the field dummy variables so that they sum to 1. We control for years spent in the current postdoctoral appointment, years spent in all postdoctoral appointments taken together, and the total number of postdoctoral appointments. To compensate for

Improving the Postdoctoral Experience

111

differing response rates at surveyed institutions, we give each sample a weight inversely proportional to the response rate at the institution. We use a robust regression (an M-estimator with a Huber influence function) for the subjective success and advisor relations, logistic regression for the binary-valued absence of conflict measure, and Poisson regression with log(total_years) as an offset for the rates of production of papers and grants. The results, shown in table 3.5, confirm what we observed in our initial comparison: professional development is positively associated with all of our success measures, and structured oversight with five or six measures. The structured oversight relationship is the most robust in that the effects are seen in the full data set as well as in the subsets, consisting of those in their first appointment and those in their second. Professional development may have a smaller impact after a postdoc’s first appointment, since skills learned in a previous appointment do not need to be relearned, or its effects may be too small to see in the smaller set of postdocs in their second appointment (860 of the 3,552 postdocs). Professional development is the strongest predictor of subjective success and of good advisor relations, and structured oversight the strongest predictor of an absence of conflicts. Those with independent funding submitted grant proposals at a 66 percent greater rate than those without (not surprising, since one must request funding in order to receive it), and reported greater levels of subjective success, but there appear to be few other measurable benefits. Salaries are weakly linked with subjective success and positive advisor relations, but the association is not significant for those in their second appointments. Both salary and structured oversight are positively correlated with the rate of paper production, both for all peer-reviewed papers as well as for first-authored papers. One standard deviation in each (for salary, a 19 percent difference, or roughly $7,600) corresponds to 6.5 to 7 percent increase in the rate of paper production. The simplest explanation for the salary relationship is that the most productive postdocs are better able to land higher-paying appointments. For the structured oversight/productivity relationship, in contrast, there is reason to believe that there is causality in the opposite direction. 3.8 Correlates of Success To better understand the reasons for the observed associations, we perform another set of regressions, this time replacing the composite measures of structured oversight and professional development with their individual components. The results are shown in table 3.6. As before, we will take the conservative approach of focusing on relationships that appear in not only the full data set, but also in separate regressions for those in first and second appointments.

Table 3.5

(Intercept) Structure Professional_ development Health Benefits Log(salary) Funding Sex Citizen_or_pr Underrepresented Married Children Age Medical_degree Total_years Current_years Previous_postdocs N

Success measures regressed on recommended practices measures and other descriptive variables Subjective success

Advisor relations

Absence of conflict

Papers submitted

1st authored papers

Grants submitted

0.696*** (0.161) 0.157*** (0.016) 0.453*** (0.015) 0.240* (0.109) 0.102*** (0.017) 0.049** (0.018) 0.158*** (0.037) 0.090** (0.030) 0.073* (0.031) 0.041 (0.082) 0.026 (0.034) 0.018 (0.036) 0.031. (0.017) 0.155** (0.050) 0.003 (0.016) 0.043* (0.019) 0.067* (0.030)

0.162 (0.156) 0.159*** (0.015) 0.242*** (0.015) 0.084 (0.106) 0.006 (0.016) 0.069*** (0.018) 0.031 (0.036) 0.014 (0.029) 0.030 (0.030) 0.036 (0.079) 0.054. (0.033) 0.034 (0.035) 0.059*** (0.017) 0.093. (0.049) 0.040* (0.016) 0.025 (0.018) 0.012 (0.029)

3.621 (11.031) 0.239*** (0.037) 0.151*** (0.034) 0.632* (0.255) 0.152*** (0.035) 0.053 (0.041) 0.160* (0.081) 0.147* (0.066) 0.097 (0.069) 0.065 (0.188) 0.012 (0.075) 0.184* (0.078) 0.084* (0.037) 0.466*** (0.101) 0.034 (0.032) 0.324*** (0.037) 0.134* (0.057)

0.651*** (0.076) 0.065*** (0.007) 0.058*** (0.007) 0.193*** (0.049) 0.018* (0.008) 0.070*** (0.008) 0.006 (0.017) 0.137*** (0.014) 0.111*** (0.014) 0.048 (0.040) 0.044** (0.016) 0.014 (0.016) 0.038*** (0.008) 0.040. (0.024) 0.058*** (0.005) 0.043*** (0.006) 0.087*** (0.011)

0.111 (0.101) 0.063*** (0.010) 0.062*** (0.009) 0.102 (0.067) 0.000 (0.010) 0.069*** (0.011) 0.037 (0.023) 0.147*** (0.020) 0.129*** (0.020) 0.110. (0.057) 0.058** (0.022) 0.105*** (0.022) 0.000 (0.011) 0.022 (0.033) 0.090*** (0.008) 0.047*** (0.009) 0.121*** (0.014)

0.341 (2.170) 0.021. (0.011) 0.127*** (0.011) 0.112 (0.074) 0.046*** (0.012) 0.033* (0.013) 0.506*** (0.022) 0.084*** (0.021) 0.276*** (0.022) 0.033 (0.052) 0.099*** (0.023) 0.018 (0.024) 0.051*** (0.012) 0.237*** (0.040) 0.143*** (0.011) 0.101*** (0.013) 0.106*** (0.023)

3,552

3,552

3,552

3,348

3,348

3,348

Notes: Standard errors are shown in parentheses. Robust regression with an M-estimator was used for the subjective success and advisor relations measures. Logistic regression was used for the absence of conflict measure. Poisson regression with a log(total_years) offset was used for the measures of productivity. All regressions also included 46 dummy variables for institution and 95 for field of research; these have been omitted to conserve space. ***Designates a p-value of  0.001. **Designates a p-value of  0.01. *Designates a p-value of  0.05. .Designates a p-value of  0.10.

Table 3.6

(Intercept) Plan_oral Plan_written Advisor_plan Evaluations Contract_ compensation Contract_benefits Contract_ responsibilities Contract_advisor Contract_term Policy_authorship Policy_misconduct Policy_grievance Policy_ip Placement_services Career_counseling Ethics Writing Public_speaking Teaching Proposal_writing Lab_management Project_management

Success measures regressed on individual components of structured oversight and professional development Subjective success

Advisor relations

Absence of conflict

Papers submitted

First author

Grants submitted

0.557*** (0.165) 0.066. (0.036) 0.099. (0.056) 0.254*** (0.038) 0.113** (0.040) 0.024 (0.035) 0.019 (0.034) 0.069. (0.038) 0.062 (0.051) 0.035 (0.038) 0.062 (0.048) 0.099* (0.046) 0.053 (0.048) 0.016 (0.045) 0.083 (0.054) 0.226*** (0.055) 0.129*** (0.034) 0.102** (0.039) 0.094* (0.039) 0.475*** (0.033) 0.166*** (0.033) 0.231*** (0.035) 0.128*** (0.035)

0.709*** (0.166) 0.142*** (0.036) 0.191*** (0.056) 0.242*** (0.038) 0.129** (0.040) 0.019 (0.035) 0.014 (0.034) 0.007 (0.038) 0.075 (0.051) 0.072. (0.038) 0.038 (0.048) 0.057 (0.046) 0.060 (0.048) 0.037 (0.045) 0.034 (0.054) 0.081 (0.055) 0.100** (0.034) 0.174*** (0.039) 0.064 (0.039) 0.096** (0.033) 0.120*** (0.033) 0.183*** (0.035) 0.080* (0.035)

2.898 (10.899) 0.023 (0.080) 0.116 (0.138) 0.421*** (0.095) 0.124 (0.095) 0.177* (0.081) 0.262** (0.083) 0.184* (0.090) 0.088 (0.132) 0.001 (0.089) 0.429*** (0.117) 0.169 (0.108) 0.257* (0.121) 0.572*** (0.112) 0.234. (0.123) 0.347** (0.120) 0.231** (0.077) 0.115 (0.089) 0.037 (0.089) 0.197* (0.080) 0.221** (0.076) 0.274*** (0.083) 0.324*** (0.082)

0.427*** (0.082) 0.024 (0.017) 0.204*** (0.026) 0.038* (0.019) 0.086*** (0.019) 0.033* (0.017) 0.124*** (0.016) 0.083*** (0.018) 0.040 (0.024) 0.001 (0.019) 0.071** (0.022) 0.129*** (0.022) 0.045. (0.023) 0.010 (0.022) 0.143*** (0.025) 0.081** (0.026) 0.018 (0.016) 0.079*** (0.019) 0.105*** (0.019) 0.108*** (0.016) 0.031* (0.016) 0.046** (0.017) 0.072*** (0.017)

0.100 (0.110) 0.037 (0.024) 0.260*** (0.035) 0.067** (0.026) 0.131*** (0.026) 0.000 (0.023) 0.154*** (0.022) 0.048* (0.024) 0.009 (0.033) 0.089*** (0.026) 0.151*** (0.031) 0.156*** (0.030) 0.119*** (0.032) 0.079** (0.030) 0.067. (0.034) 0.071. (0.036) 0.025 (0.022) 0.098*** (0.027) 0.103*** (0.027) 0.128*** (0.022) 0.035 (0.022) 0.052* (0.024) 0.091*** (0.024)

0.822 (2.146) 0.052* (0.025) 0.226*** (0.039) 0.025 (0.028) 0.039 (0.030) 0.114*** (0.026) 0.073** (0.025) 0.076** (0.028) 0.034 (0.038) 0.024 (0.028) 0.094* (0.037) 0.036 (0.032) 0.091** (0.034) 0.053. (0.032) 0.006 (0.039) 0.021 (0.040) 0.007 (0.024) 0.093** (0.030) 0.001 (0.031) 0.097*** (0.023) 0.867*** (0.030) 0.035 (0.026) 0.030 (0.026) (continued )

Table 3.6

Negotiating Ip Conflict_resolution English Non_academic Health Benefits Log(salary) Funding Male Citizen_or_pr Underrepresented Married Children Age Medical_degree Total_years Current_years Previous_postdocs N

(continued) Subjective success

Advisor relations

Absence of conflict

Papers submitted

First author

Grants submitted

0.073* (0.037) 0.053 (0.035) 0.058 (0.036) 0.081* (0.032) 0.228*** (0.030) 0.216* (0.105) 0.081*** (0.016) 0.043* (0.018) 0.118*** (0.035) 0.093** (0.029) 0.004 (0.033) 0.069 (0.079) 0.022 (0.032) 0.005 (0.034) 0.034* (0.016) 0.164*** (0.048) 0.003 (0.016) 0.043* (0.018) 0.086** (0.028)

0.019 (0.037) 0.001 (0.035) 0.014 (0.036) 0.050 (0.032) 0.106*** (0.030) 0.072 (0.105) 0.008 (0.016) 0.071*** (0.018) 0.003 (0.035) 0.019 (0.029) 0.013 (0.033) 0.044 (0.079) 0.053 (0.032) 0.040 (0.034) 0.057*** (0.016) 0.090. (0.048) 0.038* (0.016) 0.025 (0.018) 0.008 (0.028)

0.359*** (0.089) 0.015 (0.083) 0.225** (0.087) 0.192* (0.076) 0.388*** (0.071) 0.621* (0.263) 0.135*** (0.036) 0.056 (0.042) 0.100 (0.083) 0.168* (0.067) 0.002 (0.076) 0.029 (0.192) 0.017 (0.077) 0.205* (0.080) 0.092* (0.038) 0.494*** (0.104) 0.056. (0.033) 0.292*** (0.038) 0.105. (0.059)

0.172*** (0.018) 0.006 (0.017) 0.002 (0.017) 0.023 (0.015) 0.000 (0.014) 0.167*** (0.049) 0.020* (0.008) 0.070*** (0.008) 0.008 (0.017) 0.139*** (0.014) 0.078*** (0.016) 0.113** (0.041) 0.037* (0.016) 0.004 (0.016) 0.032*** (0.008) 0.044. (0.024) 0.055*** (0.005) 0.044*** (0.007) 0.082*** (0.011)

0.209*** (0.024) 0.026 (0.023) 0.064** (0.024) 0.019 (0.021) 0.013 (0.020) 0.054 (0.068) 0.006 (0.011) 0.074*** (0.011) 0.020 (0.024) 0.146*** (0.020) 0.085*** (0.022) 0.217*** (0.058) 0.048* (0.022) 0.088*** (0.022) 0.008 (0.011) 0.012 (0.034) 0.085*** (0.008) 0.047*** (0.009) 0.118*** (0.014)

0.118*** (0.026) 0.021 (0.025) 0.042 (0.026) 0.097*** (0.024) 0.022 (0.021) 0.091 (0.074) 0.051*** (0.012) 0.045*** (0.013) 0.431*** (0.023) 0.092*** (0.021) 0.191*** (0.024) 0.061 (0.053) 0.085*** (0.023) 0.036 (0.025) 0.047*** (0.013) 0.210*** (0.040) 0.143*** (0.012) 0.100*** (0.013) 0.095*** (0.024)

3,552

3,552

3,552

3,348

3,348

3,348

Notes: Standard errors are shown in parentheses. Robust regression with an M-estimator was used for the subjective success and advisor relations measures. Logistic regression was used for the absence of conflict measure. Poisson regression with a log(total_years) offset was used for the measures of productivity. All regressions also included 46 dummy variables for institution and 95 for field of research; these have been omitted to conserve space. ***Designates a p-value of  0.001. **Designates a p-value of  0.01. *Designates a p-value of  0.05. .Designates a p-value of  0.10.

Improving the Postdoctoral Experience

115

3.8.1 Research/Career Plans The most interesting observation is that postdocs who plan their experience with their advisors at the outset of their appointments fare substantially better than those who do not. Postdocs with a written plan submit papers to peer-reviewed journals at a 23 percent higher rate, first-authored papers at a 30 percent higher rate, and grant proposals at a 25 percent higher rate than those without a written plan. These findings are in keeping with Drucker’s assertion that knowledge workers’ productivity requires that they have a role in shaping their responsibilities (Drucker 1999). Postdocs with plans that discuss what their advisors will do as well as what they will do score 0.25 standard deviations higher on the subjective success measure and 0.24 standard deviations higher on the advisor relations measure than those with no plan or a plan that includes no advisor component. Correlation does not necessarily mean causation, of course, but there are a number of mechanisms by which the process of planning might give rise to the positive outcomes we observe. A trivial explanation is that those with the greatest propensity to write are more likely to write both plans and papers. It is unlikely that this is the only mechanism, however. Contracts play a key role in labor exchanges. Without a contract guaranteeing compensation or credit for investments such as training or extra hours in the lab, postdocs may forego these investments even when they would benefit all parties involved (the hold-up problem) (Jacobsen and Skillman 2004). By serving as contracts, plans can foster greater levels of investment, leading to greater productivity. Satisfaction in some cases is a function more of how one’s circumstances compare to one’s expectations than of one’s absolute circumstances (Kahneman, Diener, and Schwarz 1999). Plans may improve satisfaction levels and relations with advisors by serving as an effective expectationsetting mechanism. Indeed, while 20 percent of postdocs who made no plan reported that their advisor was not meeting their initial expectations, only 5 percent of postdocs with written plans that addressed their advisors’ obligations as well as their own reported similar disappointment. Plans can help postdocs clarify their career goals early on. Postdocs with plans then have more time to pursue training opportunities appropriate for their goals. As a result they may judge their appointments as providing better preparation than those with no plans. When postdocs and their advisors craft a plan together, they are making an explicit commitment to each other. Studies have shown that even when promises are nonbinding, people who make them in writing are more likely to follow through (Cialdini 1993). Thus, well-crafted plans can promote success by helping to ensure that both advisors and postdocs live up to their obligations.

116

Geoff Davis

Planning is widely used as an important tool for managing time and resources efficiently, and they may increase postdocs’ productivity by focusing their efforts. Additionally, a number of studies have found positive associations between job satisfaction and worker performance, particularly among professional and managerial workers (Iaffaldano and Muchinsky 1985; Petty, McGee, and Cavender 1984), so plans’ expectation setting function may have additional productivity benefits. 3.8.2 Professional Development Exposure to nonacademic careers and training in teaching skills, proposal writing, project management, and ethics are all associated with greater levels of subjective success. Exposure to nonacademic careers and training in proposal writing are further correlated with better advisor relations and lower rates of conflict. These associations make sense, since training in these skills helps postdocs perform their jobs more effectively and prepares them for their future careers. On-the-job training has been linked to increased rates of worker productivity in other sectors (Bartel 1994), and postdoc productivity appears to benefit from some forms of training. Those reporting training in proposal writing reported submitting grant proposals at a 138 percent higher rate than those reporting no training. The direction of causality probably goes both ways here: formal training in proposal writing likely helps postdocs with the grant writing process, but also those who write grants may consider the act itself a form of experiential training. Training in negotiation skills is associated with a 19 percent increase in the rate of paper submissions. Negotiation skills may help postdocs to obtain resources needed for their research, as 50 percent of those reporting negotiation skills training are completely satisfied with the funds available for research and travel, compared with 39 percent of those without such training. 3.8.3 Salary and Benefits Compensation levels have been linked to workplace satisfaction among doctorate holders (Bender and Heywood 2004; Moguerou 2002), and our findings are consistent. Benefits, another form of compensation, have a similar relationship to satisfaction. Both of these factors have a much smaller effect than intrinsic features of employment such as levels of structure and training, however a finding that is in keeping with past studies (Iaffaldano and Muchinsky 1985). The weak relationship between compensation and satisfaction fits in with the notion of the academic labor market as a tournament (Lazear and Rosen 1981; Freeman 2001b) in which incentives for postdocs are provided by the prospect of future, more lucrative employment as tenured faculty members, rather than current salaries.

Improving the Postdoctoral Experience

117

3.8.4 Contracts and Policies Ambiguity in the ownership of intellectual property is a potential source of contention. Clear policies can help prevent problems from arising, and indeed, we see that such policies are associated with 77 percent lower odds of conflict between postdoc and advisor. Contracts/letters of appointment that spell out a postdoc’s benefits are also associated with lower rates of conflict. This association may arise because such contracts are proxies for well-organized central administration of postdocs. Under the tournament interpretation of the academic labor market, the prospect of future employment motivates postdocs more than current compensation levels, and credit for work done is important for gaining access to those future opportunities. Enforcement of authorship rights should increase paper writing by increasing the likelihood that effort will be rewarded. Authorship policies are in fact associated with an increase in publications, but also, interestingly, a decrease in grant writing (both only for postdocs in their first appointments). One interpretation is that at institutions where authorship rights are less secure, postdocs shift their efforts into activities for which credit is more assured, such as applying for fellowships. Somewhat surprisingly, postdocs who report a local authorship policy report 54 percent higher odds of conflict. One recent survey suggests that authorship problems are fairly common (Tarnow, Cohen, and de Young 2007), but authorship policies, in contrast, are relatively rare—only 23 percent of postdocs report knowing about such a policy. In this light, two explanations present themselves. One possibility is that authorship policies may simply encourage greater rates of reporting of a common but underreported problem. Alternatively, the individuals most likely to be aware of authorship policies are those who have experienced problems, or, similarly, the institutions most likely to have authorship policies in place may be the ones with the highest rates of authorship problems. 3.8.5 Time All of the productivity metrics worsen over time. For every year spent in a postdoctoral appointment, postdocs submit papers at a 6 percent lower rate. This decline is offset in part by an age-linked productivity increase of 0.7 percent per year—perhaps maturity brings with it better judgment about research directions to pursue or better time management skills. For each previous appointment a postdoc has held, there is an 8 percent increase in overall paper production and a 13 percent increase in the rate of first authored papers. This finding is particularly striking given that changing appointments can be a disruptive process involving relocating and even changing fields.

118

Geoff Davis

One explanation is that the selection process for subsequent appointments is linked to productivity. The least productive people are less able to obtain subsequent appointments, and some fraction of the most productive people find better opportunities. The positive “previous_postdocs” term suggests that on the balance, low productivity is the more likely reason for leaving the postdoc pool, and hence multiple appointments are a sign of fitness with respect to the selection function. 3.8.6 Demographics Men have higher levels of subjective success than women, at least in the first appointment, which agrees with previous findings (Moguerou 2002). Studies have found that male scientists publish at a higher rate than female scientists, and our findings are consistent. Xie and Shauman (2003) report that these sex-linked productivity differences for more senior scientists disappear when the type of institution and available resources are taken into account. However, the sex-linked productivity differences we observe for postdocs persist after controlling for institution, family structure, and levels of supervision and training. Interestingly, women submit grant proposals at a higher rate, which suggests that some of the difference in publication rates may be the result of different resource allocation strategies. Citizens of the United States submit more grant proposals but fewer papers than those on temporary visas, again suggesting different allocations of time and other resources. Underrepresented minority postdocs submit first authored papers at a lower rate than majority postdocs. Those with medical degrees report lower levels of subjective success and have 64 percent greater odds of reporting a conflict with their advisors than postdocs with other types of degrees. The reasons may have to do with differing cultures and workplace environments in medical fields. 3.9 Causality Structured oversight and professional development are correlated with our success measures, but we have not proved a causal relationship. The links we observe could arise from several possible mechanisms: 1. Structured oversight and professional development may directly cause greater levels of success via the previously discussed mechanisms. 2. Structured oversight and professional development might be associated with a common unobserved underlying cause. For example, these practices might be indicators of a particularly well-managed lab, or of a principal investigator with ample resources. 3. Positions that offer professional development and oversight might attract intrinsically successful postdocs.

Improving the Postdoctoral Experience

119

4. Successful postdocs might be more likely to seek out professional development opportunities and to initiate such things as research plans. In the first two scenarios, some aspect or aspects of the current appointment cause success; in the second two, success is linked to the individual postdoc and is not affected by local circumstances. It may well be that more than one or even all of these mechanisms play some role; the interesting question is whether any predominate. One way to establish causation would be to conduct an experiment: a major funder of postdocs such as the National Institutes of Health could test the effects of practices in a manner similar to a clinical trial for a new drug. Funded postdocs could be randomly assigned to one of two variants of a funding program, one with a requirement, say, of a written plan, and one without. Absent such an experiment, we must rely on more indirect means. One approach is to test whether there is a relationship between an exogenous indicator of success and the amount of structure and professional development present in appointments. A positive relationship would suggest that intrinsically successful people seek out or create structure and professional development as in scenarios 3 and 4 in the previous list. Conversely, the absence of a relationship would suggest that structure and professional development play a causal role or are indicators of some other causal factor as in scenarios 1 and 2. One crude indicator of a postdoc’s ability is the quality of her doctorategranting program. We obtained a National Research Council (NRC) quality rating (NRC 1995) for the doctorate-granting department of 38 percent of the surveyed postdocs (some did not earn their doctorate in the United States, some earned their degrees in departments that were not rated, and some did not provide their Ph.D.-granting department). Table 3.7 shows a regression of each of our measures of recommended practices on demographic characteristics, field, and the normalized NRC rating for the postdoc’s doctorate-granting department. We do not control for institution, as doing so would hide a tendency for those from more prestigious Ph.D. programs gravitating to institutions with greater overall structure or training. We see in table 3.7 that NRC rating does have some effect: each standard deviation increase in the rating (0.86 points on a 5 point scale) is associated with 71 percent greater odds of independent funding and a 1.4 percent higher salary. There is no indication, however, that those from higher rated doctorate-granting programs either seek out or create more structure or professional development opportunities for themselves. A regression like the one in table 3.4 minus the institutional controls shows that the “fitter” postdocs with multiple appointments do not do so either. These findings cast doubt on scenarios 3 and 4 in the list and suggest that structured over-

120

Geoff Davis

Table 3.7

(Intercept) Male Citizen_or_pr Underrepresented Married Children Age Medical_degree Nrc Total_years Current_years Previous_postdocs N

Regression of recommended practices measures on demographic variables, field, and the National Research Council rating for the postdoc’s doctorategranting department Structure

Prof dev.

Health

Benefits

log(Salary)

Funding

0.216 (0.212) 0.061 (0.055) 0.087 (0.065) 0.13 (0.129) 0.025 (0.060) 0.082 (0.064) 0.070* (0.033) 0.063 (0.191) 0.054* (0.026) 0.027 (0.041) 0.094* (0.045) 0.09 (0.079)

0.152 (0.230) 0.036 (0.059) 0.163* (0.070) 0.098 (0.139) 0.096 (0.065) 0.085 (0.070) 0.031 (0.036) 0.222 (0.207) 0.042 (0.028) 0.088* (0.045) 0.119* (0.048) 0.011 (0.086)

11.172. (6.478) 0.293 (0.289) 1.935** (0.716) 0.943. (0.511) 0.352 (0.378) 0.547. (0.326) 0.068 (0.199) 14.794 (848.151) 0.188 (0.137) 0.168 (0.264) 0.086 (0.271) 0.944 (0.614)

0.540** (0.206) 0.087 (0.053) 0.02 (0.063) 0.131 (0.125) 0.059 (0.058) 0.061 (0.063) 0.068* (0.032) 0.023 (0.185) 0.035 (0.025) 0.014 (0.040) 0.015 (0.043) 0.079 (0.077)

0.420* (0.164) 0.041 (0.042) 0.225*** (0.050) 0.032 (0.100) 0.001 (0.046) 0.092. (0.050) 0.001 (0.026) 0.111 (0.148) 0.082*** (0.020) 0.116*** (0.032) 0.101** (0.035) 0.193** (0.061)

2.593*** (0.362) 0 (0.089) 0.885*** (0.120) 0.441* (0.196) 0.102 (0.100) 0.041 (0.104) 0.057 (0.058) 0.391 (0.297) 0.537*** (0.049) 0.015 (0.073) 0.202** (0.077) 0.336* (0.143)

1,375

1,375

1,364

1,375

1,375

1,343

Notes: Standard errors are shown in parentheses. A robust regression with an M-estimator was used for the structure, professional development, benefits, and salary measures. Logistic regression was used for health insurance and funding. The regression also included 95 dummy variables for field of research; these have been omitted to conserve space. ***Designates a p-value of  0.001. **Designates a p-value of  0.01. *Designates a p-value of  0.05. .Designates a p-value of  0.10.

sight and professional development either cause the observed benefits themselves or are markers for some other underlying cause. 3.10 Conclusion Of the five major classes of practices that have been recommended for postdoctoral appointments, structured oversight and professional development appear to have the greatest impact. In particular, written research

Improving the Postdoctoral Experience

121

plans that lay out the obligations of both postdoc and advisor are correlated with broad-ranging and substantial benefits. Exposure to nonacademic careers and training in teaching skills, proposal writing, and project management are also associated with multiple positive outcomes. There are plausible causal mechanisms that can explain these correlations and indirect evidence against noncausal alternative explanations. Because structured oversight measures are sufficiently simple and commonsensical, many are advocating their implementation without waiting for irrefutable evidence of their efficacy. Recent reports from the National Science Foundation (Merrimack Consultants, LLC 2003) and the National Academies (NRC 2005) require written plans detailing advisor and postdoc contributions as part of the grant application process. Given the potential benefits of plans together with their relative rarity at present (11 percent of postdocs reported having a written plan; 34 percent had a plan that detailed their advisor’s obligations as well as their own), such a requirement has the potential to improve the postdoctoral experience considerably. If a universal requirement for written research/career plans were to bring about the same productivity increase that we see with existing, voluntary plans (an outcome that is by no means assured), the resulting increase in paper production would be the equivalent of having more than 10,000 additional postdocs working in the United States. Regardless, there is much to be gained from a more systematic investigation of the process of scientific training and research.

Appendix The following describes components of the measures of success, measures of recommended practices measures, and other descriptive variables. Component abbreviations are shown in italic. Summary statistics are in parentheses. The statistics are for responses from nonclinical-fellow postdocs working full-time. Missing values were imputed with mean values where appropriate. To improve readability, standard deviations are not shown for binary-valued data. Subjective Success Measure The measure is the normalized sum of the following items, scored as described below (before normalization   2.02,   5.19, N  3,719): • sat_overall (  0.71,   1.24, N  3,669)  overall satisfaction with current position; –2 points for very dissatisfied, –1 point for somewhat dissatisfied, 0 for neither satisfied nor dissatisfied, 1 point for somewhat satisfied, 2 points for very satisfied.

122

Geoff Davis

• Extent to which respondent is being prepared for his/her future career in the following areas: prep_research (  1.14,   1.08, N  3,701)  research skills, prep_teaching (  –0.87,   1.33, N  3,667)  teaching skills, prep_management (  –0.27,   1.39, N  3,669)  management skills, prep_communications (  0.39,   1.31, N  3,688)  communications skills; 2 points for “excellent,” 1 point for “good,” 0 points for “fair,” –1 points for “poor.” • independent (  0.92,   0.98, N  3,675)  2 points for “strongly agree” that position is preparing respondent to be an independent researcher, 1 point for “agree,” 0 points for “neither agree nor disagree,”–1 point for “disagree,”–2 points for “strongly disagree.” Advisor Relations Measure The measure is the normalized sum of the following items, scored as described below (  6.33,   2.37, N  3,719 before normalization): • postdoc_grade (  3.29,   0.72, N  3,228)  estimated letter grade advisor would give respondent for overall performance (A  4 points, . . . F  0 points). • advisor_grade (  3.06,   0.96, N  3,463)  grade respondent would give advisor for overall performance (A  4 points, . . . F  0 points). • mentor (  0.73, N  3,190)  Does postdoc consider advisor to be a mentor? (1 point for “yes,” 0 points for “no”). Absence of Conflict Measure (  0.86, N  3,719)  0 if the respondent has experienced one of the following with/from his/her advisor: a dispute over authorship or author precedence, a dispute over intellectual property ownership, a dispute over research ethics, discrimination or harassment, or other research misconduct; 1 if not. Productivity Measures • papers (  2.89,   4.11, N  3,478)  the number of papers submitted to peer-reviewed journals while a postdoc. Includes papers submitted during all postdoctoral appointments, not just the current one. • first_authored (  1.55,   2.48, N  3,478)  the number of papers for which the postdoc is the primary author submitted to peerreviewed journals while a postdoc. • grants (  1.25,   1.94, N  3,478)  the number of grant proposals submitted while a postdoc.

Improving the Postdoctoral Experience

123

Structured Oversight Measure The measure is the normalized sum of the following items, scored as described below (  6.34,   2.43, N  3,719 before normalization): • plan_oral (  0.62, N  3,632)  Did the respondent and postdoc advisor set expectations orally at the beginning of the appointment for what postdoc would do and learn? • plan_written (  0.10, N  3,632)  Did the respondent and postdoc advisor set expectations in writing at the beginning of the appointment for what postdoc would do and learn? • advisor_plan (  0.35, N  3,082)  (For those who made a plan/ plans) Did the plan set expectations for what advisor would contribute to the experience? • evaluations (  0.22, N  3,086)  Does advisor provide formal performance evaluations? • (For those with a letter of appointment or contract) 1 point for each of the following pieces of information included in the contract: contract_compensation (  0.65, N  3,123)  Compensation, contract_benefits (  0.43, N  3,123)  Benefits, contract_responsibilities (  0.37, N  3,123)  Your responsibilities, contract_advisor (  0.14, N  3,123)  Advisor’s responsibilities, contract_term (  0.77, N  3,123)  Term of appointment. • 1 point for each of the following policies at institution: policy_authorship (  0.233, N  2,500)  determining paper authorship and author precedence, policy_misconduct (  0.47, N  2,545)  defining misconduct, policy_ grievance (  0.34, N  2,314)  resolving grievances, policy_ip (  0.40, N  2,375)  determining ownership of intellectual property. • placement_services (  0.56, N  1,422)  Are job placement services available at institution? • career_counseling (  0.68, N  1,638)  Is career counseling available at organization? Professional Development Measure The measure is the normalized sum of the following items, scored as described below (  6.10,   3.19, N  3,719 before normalization): • Source of respondent’s primary training in current position. 1 point for each of the following answered “workshop/seminar/formal coursework,” 0 points for “informal, on-the-job training,”–1 points for “no training”: ethics (  0.68, N  3,669)  Research ethics, writing (  0.71, N  3,678)  Writing skills, public_speaking (  0.72, N  3,675)  Public speaking skills, teaching (  0.34, N  3,656)  Teaching skills, proposal_writing (  0.64, N  3,671)  Grant or

124

Geoff Davis

proposal writing, lab_management (  0.51, N  3,658)  Group or lab management, project_management (  0.57, N  3,649)  Project management, negotiating (  0.32, N  3,644)  Negotiating skills, ip (  0.35, N  3,636)  Intellectual property, conflict_resolution (  0.38, N  3,644)  Conflict resolution skills, english (  0.42, N  3,629)  English language skills. • non_academic (  0.48, N  3,609)  How much has current position exposed respondents to opportunities outside of academia? One point for “A lot” or “Some,” 0 for “Not at all.” Health Insurance Measure • (health) (  0.98, N  3,635)  1 if health insurance is available at the postdoc’s institution, 0 if not. Benefits Measure The measure is the normalized sum of the following items, scored as described below (  11.2,   3.2, N  3,719 before normalization): • A measure of benefits available to the respondent at his/her institution. Scoring: 1 point for each of the following: health_family (  0.91, N  3,235)  Health insurance for your family, dental (  0.80, N  3,484)  Dental insurance, vision (  0.59, N  2,744)  Vision insurance, disability (  0.71, N  2,032)  Disability insurance, life (  0.73, N  2,603)  Life insurance, mental_health (  0.82, N  2,108)  Counseling/mental health services, retirement (  0.50, N  2,840)  Retirement plan, child_care (  0.46, N  1,949)  Child care, family_leave (  0.71, N  1,765)  Family leave, gym (  0.78, N  3,118)  Athletic facilities, parking (  0.79, N  3,296)  Parking, tuition (  0.67, N  2,024)  Tuition/fees for courses at institution, flex_spending (  0.57, N  1,941)  Flexible spending account/medical savings account, credit_union (  0.80, N  2,454)  Credit union, 401k (  0.53, N  2,065)  Voluntary, tax-deferred savings plan, housing (  0.22, N  2,268)  Subsidized housing, transportation (  0.62, N  2,644)  Public transportation subsidies. Independent Funding Measure • funding (  0.21, N  3,620)  1 if the postdoc receives independent funding (e.g., a fellowship), 0 if not. Salary Measure • log_salary (  10.6,   0.18, N  3,225 before normalization)  The postdoc’s annual salary normalized to zero mean, unit variance.

Improving the Postdoctoral Experience

125

Other Descriptive Variables • male (  0.57, N  3,684)  1 if the postdoc is male, 0 if not. • citizen_or_pr (  0.46, N  3,683)  1 if the postdoc is a citizen or a permanent resident of the United States, 0 if not. • underrepresented (  0.03, N  3,622)  1 if the postdoc is a citizen or permanent resident who is a member of an underrepresented minority group, 0 if not. • married (  0.69, N  3,638)  1 if the postdoc is married or partnered, 0 if not. • children (  0.34, N  3,636)  1 if the postdoc has children, 0 if not. • age (  33.4,   4.41, N  3,571 before normalization)  The postdoc’s age. Normalized to mean 0, variance 1 in regressions. • total_years (  2.42,   1.82, N  3,589)  The total number of years spent as a postdoc in all postdoctoral positions taken together. • current_years (  1.82,   1.34, N  3,621)  The total number of years spent in the current postdoctoral position. • previous_postdocs (  0.41,   0.73, N  3,704)  The number of previous postdoctoral appointments the postdoc has held. • medical_degree (  0.12, N  3,719)  1 if the postdoc has a medical degree (an MD, DDS, or DVM), 0 if not. • institution (Not shown)  A set of 46 dummy variables (coded with deviation coding) for the postdoc’s institution. • field (Not shown)  A set of 95 dummy variables used for the postdoc’s field(s) of research. If a postdoc specifies more than one field, the field variables are normalized so that they sum to 1. • nrc (  3.56,   0.86, N  1,405 before normalization)  The National Research Council’s quality rating for the postdoc’s doctorategranting department.

References Association of American Universities, Committee on Postdoctoral Education. 1998. Report and recommendations. Washington, D.C. Available at http:// www.aau.edu/reports/PostdocRpt.html. Bartel, A. P. 1994. Productivity gains from the implementation of employee training programs. Industrial Relations 33 (4): 411–25. Bender, K. A., and J. S. Heywood. 2004. Job satisfaction of the highly educated: The role of gender, academic tenure, and comparison income. Science and Engineering Workforce Project (SEWP). Working Paper. Cialdini, R. 1993. Influence: The psychology of persuasion. New York: William Morrow. Commission on Professionals in Science and Technology (CPST). 1998. Employment of recent doctoral graduates in S&E: Results of professional society sur-

126

Geoff Davis

veys. Washington, D.C. Available at http://www.cpst.org/web/site/pages/pubs/ PDF-pubs/98data-1.pdf. Davis, G. 2005. Doctors Without orders. American Scientist 93 (Supplement). Available at http://postdoc.sigmaxi.org/results/. Drucker, P. F. 1999. Management challenges for the 21st century. New York: HarperCollins. Freeman, R. 1990. Labor markets in action: Essays in empirical economics. Cambridge, MA: Harvard University Press. Freeman, R., E. Weinstein, E. Marincola, J. Rosenbaum, and F. Solomon. 2001a. Competition and careers in biosciences. Science 294 (5550): 2293–94. ———. 2001b. Careers and rewards in bio sciences: The disconnect between scientific progress and career progression. The American Society for Cell Biology. Available at http://www.ascb.org/newsfiles/careers_rewards.pdf. Iaffaldano, M. T., and P. M. Muchinsky. 1985. Job satisfaction and job performance: A meta-analysis. Psychological Bulletin 97:251–73. Jacobsen, J. P., and G. L. Skillman. 2004. Labor markets and employment relationships: A comprehensive approach. Malden, MA: Blackwell Publishing. Kahneman, D., E. Diener, and N. Schwarz. 1999. Well-being: The foundations of hedonic psychology. New York: Russell Sage Foundation. Lazear, E. P., and S. Rosen. 1981. Rank-order tournaments as optimum labor contracts. Journal of Political Economy 89 (5): 841–64. Martinson, B. C., M. S. Anderson, and R. de Vries. 2005. Scientists behaving badly. Nature 435:737–38. Merrimack Consultants, LLC. 2003. Postdoctoral appointments: Roles and opportunities. A Report on a National Science Foundation (NSF) Workshop. Atlanta, GA. Available at http://www.MerrimackLLC.com/2003/postdoc-workshop.html. Moguerou, P. 2002. Job satisfaction among U.S. Ph.D. graduates: The effects of gender and employment sector. Working Paper, Institut de Recherche sur l’Education (IREDU). CNRS-University of Burgundy, France. National Academy of Sciences, Committee on Science, Engineering, and Public Policy. 2000. Enhancing the postdoctoral experience for scientists and engineers: A guide for postdoctoral scholars, advisers, institutions, funding organizations, and disciplinary societies. Washington, D.C.: National Academies Press. National Postdoctoral Association. 2005. Recommendations for postdoctoral policies and practices. Washington, D.C. Available at http://www.nationalpostdoc .org/policy/Recommended_Practices.pdf. National Research Council. 1995. Research-doctorate programs in the United States: Continuity and change. Washington, D.C.: National Academies Press. National Research Council, Committee on Bridges to Independence: Identifying Opportunities for and Challenges to Fostering the Independence of Young Investigators in the Life Sciences. 2005. Bridges to independence: Fostering the independence of new investigators in biomedical research. Washington, D.C.: National Academies Press. National Research Council, Committee on Dimensions, Causes, and Implications of Recent Trends in the Career of Life Scientists. 1998. Trends in the early careers of life scientists. Washington, D.C.: National Academies Press. National Science Foundation. (1983–2003). Survey of graduate students and postdoctorates in science and engineering, 1983–2003. Arlington, VA: National Science Foundation. Petty, M., G. McGee, and J. Cavender. 1984. A meta-analysis of the relationship between individual job satisfaction and individual performance. Academy of Management Review 9:712–21.

Improving the Postdoctoral Experience

127

Pion, G., and M. Ionescu-Pioggia. 2003. Bridging postdoctoral training and a faculty position: Initial outcomes of the Burroughs Wellcome Fund CareerAwards in the biomedical sciences. Academic Medicine 78 (2): 177–86. Regets, M. C. 1998. What follows the postdoctorate experience? Exmployment patterns of 1993 postdocs in 1995. National Science Foundation Division of Science Resources Studies Issue Brief (NSF 99-307). Available at http:// www.nsf.gov/sbe/srs/issuebrf/sib99307.html. Sigma Xi. 2005. Sigma Xi postdoc survey methods (technical report #2). Sigma Xi Postdoc Survey. Research Triangle Park, NC. Available at http://postdoc.sigmaxi.org/results/tech_reports/. Tarnow, E., M. B. Cohen, and B. R. de Young. 2004. Coauthorship in pathology, a comparison with physics and a survey-generated and member-preferred authorship guideline. Medscape General Medicine. Available at http://www.medscape.com/viewarticle/477492. Vogel, G. 1999. A day in the life of a topflight lab. Science 285 (September): 1531– 32. Westat, Inc., and Merrimack Consultants, LLC. 2004. Postdoctoral appointments: Policies and practices. Report on a Workshop. Rockville, MD and Atlanta, GA. Available at http://www.MerrimackLLC.com/2004/postdoc-workshop.html. Xie, Y. and K. A. Shauman. 2003. Women in science: Career processes and outcomes. Cambridge, MA: Harvard University Press.

4 Immigration in High-Skill Labor Markets The Impact of Foreign Students on the Earnings of Doctorates George J. Borjas

4.1 Introduction The rapid growth in the number of foreign students enrolled in U.S. universities has transformed the higher education system, particularly at the graduate level. In 1976, 72.4 thousand foreign students were enrolled in graduate programs, making up 5.5 percent of total enrollment. By 2000, 232.3 thousand foreign students were enrolled, or 12.6 percent of enrollment. The impact is even greater at the doctoral level. For example, the fraction of doctoral degrees awarded to foreign students rose from 11.3 to 24.4 percent during the same period, with nonresident aliens receiving a remarkably high share of the doctoral degrees awarded in the physical sciences (36.5 percent of all doctorates awarded in 2000), engineering (50.7 percent), and the life sciences (25.7 percent).1 Many of these newly-minted doctorates remain in the United States after receiving their doctoral degrees, so that the foreign student influx can have a significant impact in the labor market for high-skill workers.2 Despite the large size of the supply shock and despite the importance of the labor market for doctorates in determining technological change and ecoGeorge J. Borjas is the Robert W. Scrivner Professor of Economics and Social Policy at the John F. Kennedy School of Government, Harvard University, and a research associate of the National Bureau of Economic Research. I am grateful to Alberto Abadie, Ronald Ehrenberg, Richard Freeman, Rachel Friedberg, and Paula Stephan for helpful suggestions, and to the Sloan Foundation for research support. 1. Snyder and Hoffman (2002, tables 207, 270, 272). 2. Finn (2003) calculates the stay rate of foreign-born doctoral recipients. The proportion of foreign-born doctorates who remain in the United States after receiving their degree increased from 49 percent for the 1989 cohort to 71 percent for the 2001 cohort.

131

132

George J. Borjas

nomic growth, there has not been any study of how the foreign student program affects labor market conditions for high-skill workers.3 This chapter provides an initial attempt to address a question that inevitably lies at the core of any evaluation of the costs and benefits of the foreign student program: Has the foreign student influx into doctoral programs harmed the economic opportunities of competing native workers? There already exists a large literature in labor economics that attempts to analyze the labor market impact of immigration. This literature, however, has been in a state of flux and confusion for many years. The simplest supply-demand framework implies that “limitation of the supply of any grade of labor relative to all other productive factors can be expected to raise its wage rate; an increase in supply will, other things being equal, tend to depress wage rates” (Samuelson 1964, 552). Despite the intuitive appeal of these theoretical implications, and despite the large number of careful studies in the literature, it has proved surprisingly difficult to demonstrate empirically that immigration has a sizable and significant adverse effect on competing workers. For example, a widely cited survey by Friedberg and Hunt (1995, 42) concludes that “the effect of immigration on the labor market outcomes of natives is small.”4 This conclusion is difficult to reconcile with the textbook model because the immigrant supply shock in recent decades has been very large, and most studies of labor demand (outside of the immigration context) conclude that the labor demand curve is not perfectly elastic (Hamermesh 1993). Much of the existing literature exploits the fact that immigrants in the United States cluster in a small number of geographic areas and uses the geographic variation in the supply shock to identify the labor market impact of immigration.5 The stereotypical study defines a metropolitan area as the labor market that is being penetrated by immigrants. The study then goes on to measure the relation between the native wage in the locality and the relative number of immigrants in that locality. Although there is a great deal of dispersion across studies, the estimated correlations tend to cluster around zero, and this finding is often interpreted as saying that immigrants have little impact on the labor market opportunities of native workers. Recent research raises two questions about the validity of this interpre3. Freeman (1975, 1976) used a cobweb model to analyze how wages adjust to supply shifts in high-skill labor markets. Because Freeman studied the supply shifts that occurred between the late 1940s and the early 1970s, he did not address the question of how these markets responded to immigration-induced supply shifts. 4. Borjas (1999) and Smith and Edmonston (1997) also survey the literature and reach the same conclusion. 5. Representative studies include Altonji and Card (1991), Card (1990), Grossman (1982), LaLonde and Topel (1991), and Schoeni (1997). Friedberg (2001) presents a rare study that uses the supply shock in an occupation to identify the labor market impact of immigration in the Israeli labor market. Card (2001) uses data on occupation and metropolitan area to define the relevant labor markets and estimates a slight adverse impact of an immigration-induced supply increase.

Immigration in High-Skill Labor Markets

133

tation of the evidence. First, immigrants may not be randomly distributed across local labor markets. If immigrants tend to endogenously cluster in cities with thriving economies, there would be a spurious positive correlation between immigration and local outcomes. Second, natives may respond to the immigrant supply shock in a local labor market by moving their labor or capital to other cities. These flows of internal migrants or capital would reequilibrate the national labor market and spread out the impact of immigration over the entire economy. A comparison of the economic opportunities facing native workers in different cities would show little or no difference because, in the end, immigration affected every city, not just the ones that actually received immigrants.6 Because of the strong likelihood that the local labor market adjusts to immigration—through the internal migration of workers or jobs—recent studies have proposed changing the unit of analysis to the national level. Borjas (2003), for example, examines the evolution of the national wage structure for skill groups defined in terms of educational attainment and work experience.7 The use of work experience to classify workers across skill groups takes advantage of the notion that similarly educated workers with similar levels of experience are more likely to be substitutable with each other than similarly educated workers with very different levels of experience (Welch 1979; Card and Lemieux 2001). The empirical analysis reported in Borjas (2003) used Census data from 1960 through 2000 and indicated that immigration indeed harmed the earnings opportunities of competing native workers. An immigrant influx that increases the size of a particular skill group by 10 percent lowers the wage of native workers in that group by about 3 to 4 percent. This chapter uses data drawn from the Survey of Earned Doctorates and the Survey of Doctoral Recipients to analyze the impact of the influx of foreign students on the earnings of doctorates.8 These data provide detailed information on the size of the immigrant supply shock and the labor market experiences of doctorates in science and engineering. The data also contain information on doctoral fields and year of graduation, so that it is possible to construct specific cohorts of doctorates and examine how a particular supply shock affects the earnings of doctorates in that cohort. It turns out that the foreign student influx has differentially affected different 6. Borjas, Freeman, and Katz (1997) and Card (2001) provide the first attempts to jointly analyze labor market outcomes and native migration decisions. 7. See also Borjas, Freeman, and Katz (1997). 8. The labor market impact of the foreign student influx has not been examined in the existing literature even though the wage of doctorates is a crucial indicator of conditions in highskill labor markets and is a major part of the costs of running universities or firms engaged in research and development (Ehrenberg 2000). A related study by Levin et al. (2004) uses a “shift-share” methodology to analyze employment patterns of native- and foreign-born doctorates in science and engineering and finds that native-born doctorates are underrepresented in those fields most heavily penetrated by foreign students.

134

George J. Borjas

fields at different times. I exploit this variation in the supply shock to identify the impact of immigration on high-skill labor markets. In an important sense, the foreign student influx into the labor market for doctorates provides a near-ideal research framework for measuring the impact of immigration. The labor market for these high-skill workers is certainly national (and perhaps even international) in scope. It is also unlikely that the internal migration of doctorates across fields can help the high-skill labor market adjust to the supply shocks. A doctoral education in science and engineering is a highly specialized endeavor, requiring the investment of a great deal of time and effort, and the training is very specific.9 An exogenous supply increase in a particular field at a particular time may affect the education decisions of future generations of students, but there is relatively little that current doctorates can do about the situation except to absorb the supply shock—presumably through lower wages. The empirical analysis reported in this chapter clearly shows that a foreign student influx into a particular field at a particular time has a significant and adverse effect on the earnings of competing doctorates in that field who graduated at roughly the same time. A 10 percent immigrationinduced increase in the supply of doctorates lowers the wage of competing workers by about 3 to 4 percent—remarkably similar to the elasticity estimates reported in Borjas (2003) for the typical worker in the national labor market. About half of this adverse wage effect can be attributed to the increased prevalence of low-pay postdoctoral appointments in fields where immigration has softened labor market conditions. Because the magnitude of the immigrant supply shock in particular fields has been sizable, this elasticity implies that many doctorates employed in the United States, whether native-born or foreign-born, have experienced a substantial wage loss. 4.2 Data The analysis uses data drawn from the Survey of Earned Doctorates (SED) and the Survey of Doctoral Recipients (SDR). These data files, designed to provide detailed information on trends in the number of doctorates awarded and in labor market conditions for these high-skill workers, are maintained by the National Science Foundation.10 The SED provides a population census of all persons who receive doc9. The notion that labor supply is inelastic in the short run was a core assumption of the cobweb model used by Freeman (1975, 1976) to interpret wage and employment adjustments in high-skill labor markets. 10. The National Science Foundation has two websites that provide detailed descriptions of the SED and SDR data sets. The SED website is http://www.nsf.gov/sbe/srs/ssed/ sedmeth.htm; and the SDR website is http://sestat.nsf.gov. The data analyzed in this paper are available from the NSF under a licensing agreement designed to guard the confidentiality of the survey participants.

Immigration in High-Skill Labor Markets

135

torates from a U.S. institution in a particular calendar year, with a response rate of around 92 percent. I will use the SED to calculate the magnitude of the immigrant supply shock by field and year of degree. The SDR is a biennial longitudinal file that provides a 7 percent sample of persons who obtained their doctoral degrees in the United States in science or engineering, and contains detailed information on a worker’s employment and earnings. A sample of newly granted doctorates is added to the sample every two years and a “maintenance cut” of older doctorates is conducted so as to keep sample size relatively constant at around 30,000 per wave. The existing panel consists of five waves, beginning in 1993.11 The analysis reported following will use data from all of the five panels conducted between 1993 and 2001. By linking the two data sets, it is possible to ascertain if immigrant supply shocks specific to a particular cohort defined by field and year of graduation affected the labor market performance of competing workers. I restrict the analysis to persons who received their doctoral degree between 1968 and 2000. The SED did not collect data that identified a person’s detailed immigration status (such as the difference between a naturalized citizen or a native-born citizen) prior to 1967. After 1967, the “citizenship status” variable reports if the newly-minted doctorate was a native-born citizen, a naturalized citizen, a noncitizen with a permanent visa, or a noncitizen with a temporary visa at the time the degree was awarded. Throughout the analysis, I define an “immigrant” to be a person who is either a naturalized citizen or a noncitizen; all other persons are classified as “natives.” Because the SDR data contains information on labor market characteristics of doctorates only in science and engineering, I restrict the analysis of the SED data to those persons who received doctoral degrees in those fields. Consider the population of persons who are granted a doctorate in field f in calendar year c. The foreign-born share in this particular field-cohort cell is given by: (1)

Mfc pfc   , Mfc  Nfc

where Mfc gives the number of immigrants in cell ( f, c) and Nfc gives the corresponding number of natives. The top panel of figure 4.1 shows the trend in the number of doctorates granted each year to native-born and foreign-born students (aggregated across all fields), while the bottom panel of the figure shows the trend in the 11. The SDR actually dates back before 1993, but there was a major redesign of the sample in the early 1990s that makes it extremely difficult to longitudinally track persons before and after 1993. The sample redesign was prompted by the fact that the response rate had fallen to around 50 percent by the late 1980s, probably making the data collected by the SDR prior to 1993 quite unrepresentative of the underlying population.

136

George J. Borjas

A

B

Fig. 4.1 Doctorates awarded, 1968–2000: A, doctorates awarded each year; B, immigrant share Source: The data reported in panel A is drawn from the Survey of Earned Doctorates; the data reported in panel B is drawn from both the Survey of Earned Doctorates and the Survey of Doctoral Recipients. Note: The “immigrant share, stayers (SED)” series in the bottom panel gives the fraction of workers who are foreign-born when the foreign-born population includes only those newlyminted doctorates who intend to stay in the United States after graduation.

aggregate immigrant share. The annual number of doctorates granted to native students in science and engineering declined from about 16,000 in 1970 to about 14,000 in 1980. It then began a slow steady rise that lasted through the late 1990s. By the late 1990s, around 18,000 native persons were being granted doctorates in science and engineering each year.

Immigration in High-Skill Labor Markets

137

The figure also shows that there was an even steeper rise in the number of doctorates granted to immigrants. Between 1980 and 1995, the number of doctorates granted to persons that I have classified as immigrants rose from about 4,000 to almost 11,000. As a result, the immigrant share in the number of doctorates awarded each year rose rapidly over the period. It was 17.5 percent in 1968, peaked at 39.7 percent in 1994, and then fell to 34.8 percent by 2000. As previously noted, the SED reports the person’s citizenship and visa status at the time the doctorate was awarded. The timing of this information makes it impossible to ascertain exactly if the foreign-born doctorate entered the United States using a foreign student (temporary) visa. Nevertheless, it is likely that the overwhelming majority of these foreign-born doctorates entered the country using a student visa. Table 4.1 shows that 76.7 percent of all doctorates granted between 1968 and 2000 to foreign-born students were granted to students who had temporary visas at the time the doctoral degree was awarded. Moreover, it is possible that many of the students who had permanent status at the time the doctorate was awarded entered the country with a student visa but then adjusted their status to get a green card (e.g., through marriage to a U.S. citizen) or became naturalized citizens. As table 4.1 also shows, the fraction of foreign-born students who received their high school diploma abroad is over 95 percent, both for foreign students with permanent status and with temporary visas. Put differently, it seems very likely that the bulk of the foreign-born population receiving their doctoral degrees from a U.S. university initially entered the country using a foreign student visa. It is important to stress that not all of the immigrants granted doctorates by U.S. universities will influence conditions in the U.S. labor market (at Table 4.1

Doctorates awarded to foreign-born persons, 1968–2000 Type of visa

Number of doctorates Percent with high school diploma from abroad Percent with a bachelor’s diploma from abroad Percent who expect to remain in the United States

Total

Citizen or permanent visa

Temporary visa

203,791

45,356

154,193

97.9%

94.9%

98.9%

89.7%

80.5%

92.6%

70.9%

92.5%

64.3%

Source: Survey of Earned Doctorates. Notes: A total of 511,741 doctorates were granted to native-born persons during the 1968– 2000 period. The type of visa refers to the visa or citizenship status of the person at the time the doctorate was granted.

138

George J. Borjas

least directly). Many of these newly-minted doctorates will instead return to their home countries. It turns out, however, that the vast majority of foreign-born students—regardless of whether they have permanent status or a temporary visa at the time they receive their doctoral degree—intend to stay in the United States. The SED asks the newly-minted doctorates if they intend “to live, work or study in the United States or a foreign country after receiving the doctorate.” The bottom row of table 4.1 shows that 64.3 percent of the foreign students with a temporary visa intend to remain in the United States and that over 90 percent who are citizens or have a permanent residence visa will also stay.12 In short, the foreign student program is an important conduit for supply shocks that permanently increase the number of doctoral workers in the United States. I calculated an alternative measure of the immigrant supply shock for each field-cohort cell by using the information on whether the foreign student intends to stay in the United States after graduation. If these expectations are actually realized, the immigrant share that would be observed (and would determine conditions) in the U.S. labor market is given by: M∗fc (2) p∗fc   , M∗fc  Nfc where M∗fc is the number of foreign-born doctorates that intend to stay in the United States. The two panels of figure 4.1 also illustrate the trend in the number of foreign-born doctorates and the immigrant share that includes only the “stayers.” The supply shock to the U.S. labor market is sizable: the number of foreign-born doctorates who intend to stay in the United States after graduation rose from 2,000 in 1968 to over 7,000 by the late 1990s. This supply shock increased the immigrant share in the flow of doctorates to the U.S. labor market from about 15 percent in the early 1970s to around 30 percent in the late 1990s. One potential problem with the calculation of the immigrant supply shock using the “intend to stay” information in the SED is that intentions to remain in the United States do not necessarily coincide with the actual ability to stay in the country. There is, after all, the relatively nontrivial matter of obtaining some type of work permit or permanent visa after graduation. As I will show momentarily, however, the available information indicates that the immigrant share calculated in equation (2) tracks the actual immigrant share of doctorates in the U.S. labor market very closely. Because the SDR provides a sample of the foreign-born doctorates who actually stayed in the United States, I can use these data to validate the “intend to stay” question in the SED. In particular, I used the SDR data to cal12. Finn (2003) provides a detailed analysis of the trends in the stay rate for foreign-born doctorates.

Immigration in High-Skill Labor Markets

139

culate the immigrant share for each year-of-graduation cohort. This trend is illustrated as the broken line in the bottom panel of figure 4.1. It is clear that the immigrant share calculated from the SDR almost perfectly tracks the immigrant share calculated from the SED sample of intended stayers until about 1992. In other words, throughout much of the sample period, the actual immigrant share observed in the U.S. labor market is almost identical to the immigrant share that would be predicted from the SED based on the “intend-to-stay” question that is asked of all foreign-born doctoral recipients at the time they receive their degrees. Beginning in 1992, however, the two data series begin to diverge. The SDR went through a major redesign in the early 1990s, and part of the divergence may be due to this redesign (or perhaps to incorrectly defined sampling weights for the subsample of foreign-born doctorates). In fact, figure 4.1 suggests that the SDR prediction of the immigrant share in the post-1992 period is contaminated by measurement error. In particular, the immigrant share calculated in the SDR in the late 1990s is actually higher than the immigrant share calculated in the SED that includes all foreign-born doctorates, regardless of whether they intend to stay or not.13 The sampling error—and the divergence of the SDRcalculated immigrant share from the true immigrant share—is even larger when I calculate the immigrant share for each year-of-graduation cohort by field. As a result, I will use the SED counts of doctorates to measure the immigrant supply shock throughout the chapter. Figure 4.2 continues the analysis by calculating the immigrant share in the SED by cohort and field for the largest five fields of doctorates. It is evident that the nature of the immigrant supply shock differs substantially across fields, not only in terms of the size of the shock but also in terms of the timing. Consider, for example, the supply shock in electrical engineering. The immigrant share in this field rose rapidly in the 1970s, from about 19 percent in 1970 to about 40 percent in 1985, and then remained stable at that level through 1998, when it began to rise again. In contrast, the immigrant share in biological sciences actually declined throughout the 1970s, from 10.6 percent in 1970 to 7.8 percent in 1982, rose rapidly until 1996 to 31.4 percent, and then began to decline again. Finally, the immigrant share in psychology has hovered between 3 and 5 percent throughout the entire sample period. I exploit these differences in the size and timing of the immigrant supply shock to estimate the impact of immigration on the earnings of native-born doctorates. The chapter focuses on twenty-two distinct doctoral fields that can be identified in both data sets. Table 4.2 reports summary statistics on degrees granted, salaries, and the trend in the immigrant share for each of 13. This anomaly could also be explained by the unlikely possibility that a large (and growing) share of native doctorates choose to migrate abroad after receiving their degrees.

140

Fig. 4.2

George J. Borjas

The immigrant supply shock, selected fields

Source: Survey of Earned Doctorates. Note: The five fields in the figure are the fields that produced the largest number of doctorates between 1968 and 2000.

these fields. There is a great deal of dispersion not only in the immigrant share and the timing of the immigrant supply shock, but also in the average salary in the various fields. In economics, for instance, the average annual salary during the 1990s was $91.6 thousand; in the biological sciences it was $74.4 thousand, and in chemistry it was $83.1 thousand. Finally, it is worth noting that a study of the impact of immigration on the earnings of doctorates based on the SED and SDR data could potentially miss an important part of the story. Both the SED and the SDR sample only those persons who received their doctorates in U.S. institutions, and the vast majority of these persons entered the country through the foreign student program. There may also be a sizable number of foreign-born persons in the U.S. labor market who received their doctorates abroad and who migrated to the United States after their education was completed. The immigrant supply shock calculated in this chapter would then understate the size of the relevant migration flow. I suspect, however, that the size of the population of science-andengineering doctorates who received their degrees abroad and then migrated to the United States is relatively small. The 1999 wave of the SDR reported there are 114.6 thousand foreign-born doctorates in the sciences and engineering employed in the United States. The 2001 wave enumerated 123.3 thousand such persons. The 2000 Census specifically indicates if the person has a doctoral degree. I used the 5 percent sample of the 2000 Census to count how many foreign-born doctorates were enumerated and are employed in the mathe-

Immigration in High-Skill Labor Markets Table 4.2

Doctorates awarded in 1968–2000, by field

Field Computer and information sciences Mathematical sciences Agricultural and food sciences Biological sciences Environmental life sciences Health and related sciences Chemistry, except biochemistry Earth sciences, geology, and oceanography Physics and astronomy Other physical sciences Economics Political science Sociology and anthropology Other social sciences Psychology Aerospace and related engineering Chemical engineering Civil and architectural engineering Electrical, electronic engineering Industrial engineering Mechanical engineering Other engineering All fields

141

(%) foreign-born (includes only foreign students intending to stay)

Ph.D.s granted (1,000s)

Average salary ($1,000)

1970s

1980s

1990s

14.0 32.5 34.8 140.2 2.8 26.5 64.2

88.0 76.3 68.9 74.4 70.2 75.9 83.1

19.6 16.1 20.0 10.1 10.2 11.5 15.8

33.9 33.7 21.6 11.3 10.5 11.1 21.1

41.6 42.6 34.6 27.5 24.2 16.7 34.0

19.8 45.1 3.0 28.8 23.4 29.8 16.7 100.7

73.5 82.6 66.0 91.6 72.6 61.7 69.6 70.1

11.8 18.0 18.2 17.2 9.4 6.8 12.2 3.2

13.7 28.1 24.2 28.7 15.9 9.6 18.5 3.4

23.5 37.5 39.1 36.7 14.4 13.0 22.2 4.9

5.6 15.7

91.1 93.1

29.7 37.1

44.1 40.9

35.1 43.6

13.6

83.3

42.3

51.8

54.2

35.4 12.2 18.3 32.0

99.7 87.1 86.2 89.3

30.0 34.9 31.0 28.2

47.0 45.0 50.7 40.8

49.2 46.0 49.1 43.9

715.3

78.2

19.7

27.5

33.4

Source: Survey of Earned Doctorates (except for the average salary data, which is drawn from the Survey of Doctoral Recipients). Note: The salary statistic gives the mean salary (in 2001 dollars) calculated over all workers in each doctoral field throughout the 1993–2001 sampling period.

matical sciences, other sciences, or social sciences.14 The 2000 Census enumerated a total of 133.1 thousand such doctorates. In short, almost 90 percent of all foreign-born doctorates employed in the United States in 2000 received their degrees in the United States, are enumerated in the SED, and are sampled by the SDR.15 The joint study of the SED and SDR surveys 14. The relevant occupation codes in the 2000 Census range from 100 through 196. 15. Of course, the estimate of the undercount is imprecise because it depends on the worker’s reported occupation in the 2000 Census, rather than on the field of doctoral degree. Some science-and-engineering doctorates may be employed outside these fields; and some persons with other types of degrees may be employed in science-and-engineering jobs.

142

George J. Borjas

that is the foundation of the empirical analysis reported in this chapter, therefore, should provide a comprehensive account of how immigration in high-skill labor markets—primarily through the foreign student program—affects economic opportunities for high-skill workers. 4.3 Regression Analysis As previously noted, the SDR gives a panel of recipients of doctoral degrees in sciences and engineering. The empirical analysis reported in this chapter uses all five waves of the SDR. Let wifc(t) denote the annual earnings of worker i, who has a doctorate in field f, received his doctoral degree in year c, and is observed at time t. Most studies of the labor market impact of immigration typically estimate regressions that relate the worker’s earnings to some measure of immigrant penetration in the relevant labor market. Consider the following generic model: (3)

log wifc(t)  pfc  xifc(t)  df  yc  t  (df  t )  εifc(t),

where xifc(t) is a vector indicating the number of years that the worker has been in the labor market; df is a vector of fixed effects indicating the worker’s field of doctoral study; yc is a vector of fixed effects indicating the worker’s year-of-graduation cohort; and t gives a vector of period fixed effects indicating the calendar year in which the worker’s earnings are observed. The worker’s experience is defined as the number of years elapsed between the time the worker is observed in a particular SDR wave and the time the worker received the doctoral degree. The vector xifc(t) then contains as many fixed effects as there are values for the experience variable (i.e., a dummy variable indicating if the worker has one year of experience, two years, and so on). To avoid contamination by composition effects, the sample used to estimate equation (3) includes only native-born doctorates. The linear terms of the fixed effects included in equation (3) adjust for differences in earnings across different doctoral fields, experience cells, and over time. The regression model also includes a set of interactions between the field and period fixed effects. These interactions account for the possibility that the economic returns to particular fields has been changing over time. Note that the regression cannot contain additional vectors of interactions among the various fixed effects because they would be either perfectly collinear with the variables already included in the regression or they would make it impossible to identify the parameter . For instance, interactions between the cohort fixed effects and the period fixed effects would be perfectly collinear with the xifc(t). Similarly, the inclusion of an additional vector of interactions between the worker’s experience and the field fixed effects would make it impossible to identify the parameter . The application of ordinary least squares to the regression model in equation (3) leads to incorrect standard errors for two distinct reasons.

Immigration in High-Skill Labor Markets

143

First, the same worker can be observed up to five times during the duration of the SDR panel, so that the estimation technique must adjust for withinworker correlation in the error term. Second, the immigrant share for a particular cohort-field combination is constant within the subset of workers who graduated at the same time with a doctoral degree in the same field. I use a two-stage estimation approach to adjust the standard errors both for the correlations in errors across the observations belonging to a particular individual and for the impact of the clustering of the key independent variable along the cohort-field dimension. In the first stage, I stack all workers across all panels and estimate the fixed effect for worker i in field f and cohort c. In particular, consider the regression model: (4)

log wifc(t)  vifc  xifc(t)  t  (df  t)  εifc(t),

where vifc is the fixed effect that measures the individual’s earnings potential after controlling for the worker’s experience, for any period-specific labor market effects on earnings, and for the possibility that there are secular trends in the wages paid in different doctoral fields. This regression yields an estimate of the person fixed effect, or vˆifc. In the second stage, I aggregate the estimated individual fixed effects within each field-cohort group—that is, within each ( f, c) cell. Let vˆfc be the mean value of the individual fixed effects within each of these groups. The second-stage regression model is then given by: (5)

vˆfc  pfc  df  yc  fc.

Note that the second-stage regression has one observation per field-cohort cell. I use the total of the sampling weights assigned to each person in the SDR (i.e., added across all the waves that a particular person appears in the survey) to calculate the average vˆfc. The standard errors of the second-stage regression are adjusted using a standard Huber-White correction to account for the heteroscedasticity introduced by the sampling error in the dependent variable.16 I use two alternative measures of a native worker’s earnings as the dependent variable in the regression analysis. The first gives the adjusted annual salary as constructed by the NSF from information on a worker’s income per pay period. The second is the total annual (earned) income of the worker in the calendar year prior to the survey. Although the total annual income would seem to be a preferable measure of earnings, it is not available for the 1993 survey (cutting down the size of the first-stage regression by approximately 20 percent).17 16. All second-stage regressions reported in this chapter also include a variable indicating the fraction of the ( f, c) cell that is male. This variable is typically not very important and its exclusion would not alter the quantitative nature of the results in any appreciable way. 17. The first stage regression has 105,921 observations when the dependent variable is the log of adjusted annual salary and 84,036 observations when it is the log of annual income.

144

George J. Borjas

As suggested by Welch’s (1979) study of the impact of cohort size on the earnings of baby boomers, workers who received their doctoral degree in the same field at roughly the same time are more likely to influence each other’s labor market opportunities than workers who are in the same field but graduated at very different times. I initially capture the (within field) similarity across workers who share the same years of experience by aggregating the flow data from the SED into three-year cohort intervals, indicating if the worker earned his doctorate between 1968 and 1970, 1971 and 1973, 1974 and 1976, and so on. There are a total of eleven three-year cohorts in the data (for each field). I then calculated the immigrant share for each of these cohorts and this is the key independent variable pfc in the second-stage regression model. The first two rows of table 4.3 report the estimates of the coefficient . The first row uses the immigrant share defined by (1); in other words, it uses all foreign-born persons who received a doctoral degree in a particular field and year. The second row estimates the regression models using the immigrant share defined by equation (2), using only those immigrants who intend to stay in the United States. Column (1) of the table reports the coefficient of the simplest specification, a regression model that does not

Table 4.3

Basic estimates of wage impact of immigration (Coefficient of immigrant share)

Measure of immigrant share 1. Three-year cohort 2. Three-year cohort, including only intended stayers 3. Five-year moving average 4. Five-year moving average, including only intended stayers Controls: (Field  period) interactions State of residence fixed effects

Adjusted annual salary

Income earned last year

(1)

(2)

(3)

(1)

(2)

(3)

.313 (.141)

.370 (.155)

.378 (.155)

.415 (.163)

.481 (.175)

.487 (.176)

.417 (.151) .286 (.097)

.489 (.166) .351 (.102)

.496 (.166) .354 (.101)

.536 (.174) .371 (.113)

.618 (.187) .426 (.117)

.623 (.187) .430 (.117)

.382 (.102)

.461 (.108)

.464 (.108)

.486 (.119)

.553 (.123)

.554 (.123)

No

Yes

Yes

No

Yes

Yes

No

No

Yes

No

No

Yes

Notes: The standard errors are reported in parentheses. The regressions have 240 observations when using the three-year cohort groups and 714 observations when using the five-year moving average. All regressions are weighted by the total sampling weight for the field-cohort cell. The standard errors are adjusted for heteroscedasticity by using the Huber-White correction.

Immigration in High-Skill Labor Markets

145

include the (df  t ) interactions. Column (2) includes the interactions and column (3) adds a vector of fixed effects indicating the worker’s state of residence. Regardless of the specification of the regression model, the definition of the immigrant supply shock, and the dependent variable chosen for the model, the evidence consistently reveals a numerically and statistically significant negative relation between the average earnings of doctorates in a particular field-cohort cell and the immigrant supply shock as measured by the immigrant share. Because the measure of the immigrant share that includes only the intended stayers is a conceptually better indicator of the supply shock actually affecting the U.S. labor market, the remainder of this chapter exclusively uses the supply variable that includes only the foreignborn intended stayers. In the most general specification of the regression model (columns 3 in row 2), the coefficient of the supply shock variable is –.496 (with a standard error of .166) in the adjusted salary equation and –.623 (.187) in the annual income equation. It is easier to interpret these coefficients by converting them to an elasticity that gives the percent change in earnings associated with a percent change in labor supply. Let mfc  Mfc/Nfc, or the percentage increase in the labor supply of group ( f, c) attributable to immigration. The implied factor price elasticity is then given by: (6)

∂ log wfc   (1 pfc )2. ∂mfc

By 2000, immigration had increased the immigrant share in the stock of doctorates in the United States to 23.6 percent. Equation (6) then implies that the factor price elasticity—evaluated at the mean value of the supply increase—can be obtained by multiplying  by approximately 0.6. The implied elasticity in the adjusted salary regression is then –0.30 (or –0.496  0.6), while the implied elasticity in the annual income regression is –.37. Put differently, a 10 percent supply shock (i.e., an immigrant flow that increases the number of doctorates in a particular field-cohort group by 10 percent) reduces the annual earnings of native-born doctorates by about 3 to 4 percent.18 As previously noted, I aggregated the supply measures into three-year cohorts to capture the notion that workers who share the same field and graduate at roughly the same time are perfect substitutes. An alternative approach, introduced by Welch (1979), uses some type of moving average of the supply shock. In other words, the type of supply shock encountered 18. The results that would be obtained by using the immigrant share that can be calculated from within the SDR are qualitatively similar, but not as large. For instance, the coefficient that would be analogous to that reported in the last column of row 2 in table 4.3 is –.281 (.131), implying a factor price elasticity of –.17. The smaller size of the coefficient is consistent with the conjecture that the SDR-implied immigrant share for specific field-cohort cells contains more measurement error than the comparable statistic in the SED.

146

George J. Borjas

by a native worker who received a doctoral degree in 1980 will be affected by the supply shock that occurred around 1980. The operational difficulty, of course, is the choice of the subset of years over which workers in a particular field are relatively substitutable. Suppose that workers in k adjacent cells around degree-granting date c are similar (with k odd). The relevant supply shock facing a worker in cell ( f, c) is then given by:

∑

(k 1)/2

(7)

Mf,c  (k 1)/2 pfc   , (k 1)/2 ∑ (k 1)/2(Mf,c  Nf,c)

so that the immigrant supply shock is approximately given by a k-year moving average of the immigrant share series. I used equation (7) to calculate a five-year moving average of the supply shock for each doctoral field.19 Note that there are now more observations in the second-stage regressions since each year-of-graduation cohort provides independent information about the immigrant supply shock to the regression model. Rows 3 and 4 of table 4.3 report the regression coefficients obtained by using the five-year moving average measure of the immigrant share. The coefficients are very similar to those obtained when using the three-year grouping. For example, the coefficient  in the annual income equation is –.554 (with a standard error of .123), implying a factor price elasticity of –.33. It is worth noting that the adverse wage effects reported in table 4.3 are likely to be underestimates of the true wage impact. After all, the flow of foreign immigrants into particular fields will likely be greater when the market in those fields is tight. For instance, foreign students will have a greater likelihood of remaining in the United States in those fields (and in those years) where they expect a high demand (and relatively high rewards) for their labor. This behavioral response would build in a positive correlation between immigration and native wages, attenuating the potential adverse wage impact of immigration.20 In sum, the coefficients reported in table 4.3 indicate that the immigration of doctorates (mainly through the foreign student program) had a sizable adverse impact on the earnings of competing native workers. Moreover, as table 4.4 shows, the results are roughly similar even when the regression model is subjected to a variety of major specification changes. For simplicity, the coefficients reported in table 4.4 are calculated from the most general specification of the regression model (which includes the field-period interactions and the state of residence fixed effects). 19. The moving average is calculated over all available data, even at the truncated endpoints of the time series of year-of-graduation cohorts. As a result, there are no missing values for the immigrant share defined by equation (7). 20. This argument might explain why the inclusion of field fixed effects in table 4.3 tends to increase (in absolute value) the negative correlation between immigration and wages.

Immigration in High-Skill Labor Markets Table 4.4

Sensitivity analysis (Coefficient of immigrant share) Annual adjusted salary

Sample 1. Baseline, all natives 2. Male 3. Female 4. Academic employer 5. Nonacademic employer 6. Received degree in 1971–1979 7. Received degree in 1981–1989 8. Received degree in 1991–1999

147

Income earned last year

Three-year cohort

Five-year moving average

Three-year cohort

Five-year moving average

.496 (.166) .452 (.129) .821 (.258) .388 (.169) .366 (.162) .641 (.339) .459 (.227) .803 (.412)

.464 (.108) .435 (.098) .854 (.218) .382 (.122) .331 (.129) .622 (.340) .605 (.228) 1.249 (.430)

.623 (.187) .515 (.150) .778 (.272) .475 (.189) .529 (.180) .522 (.521) .373 (.255) 1.309 (.396)

.554 (.123) .490 (.118) .844 (.222) .476 (.137) .479 (.144) .305 (.444) .537 (.286) 1.884 (.424)

Notes: The standard errors are reported in parentheses. The number of observations in each of the regressions using the three-year cohort groups and five-year moving average is: baseline: 240, 714; male: 240, 714; female: 217, 590; academic employer: 238, 694; nonacademic employer: 239, 708; the 1970s cohort: 65, 192; the 1980s cohort: 88, 197; and the 1990s cohort: 88, 198. The reported regression coefficients come from the specification of the model that includes both field-period interactions and state-of-residence fixed effects. All regressions are weighted by the total sampling weight for the field-cohort cell. The standard errors are adjusted for heteroscedasticity by using the Huber-White correction.

Rows 2 and 3 of table 4.4 reestimate the regression models in the samples of male and female native doctorates, respectively. The estimated coefficients are negative and significant for both groups, with the point estimate of the effect being larger for women. The next two rows report the regression coefficients by type of employer: academic or nonacademic. The two coefficients hover around –.5 in the annual income equations, implying that the adverse impact of an immigrant supply shock on one segment of the market completely spills over into the other segment. Finally, the last three rows of the table report the coefficients when the model is estimated separately in the sets of cohorts that received their degrees in the 1970s, the 1980s, or the 1990s, respectively. Although there is a lot of dispersion in the estimated coefficients, the coefficients are always negative and often significant. In sum, the evidence suggests a remarkable consistency in the negative relation between the earnings of native-born doctorates who received their degrees in the same field at roughly the same time and the immigrant supply shock affecting that specific group.

148

George J. Borjas

4.3.1 Estimating a Marginal Productivity Model Although the studies in the immigration literature often estimate the generic regression model giving the relation between the wage of a particular worker (or group of workers) and the immigrant share, there is an alternative approach, more closely linked with economic theory, that can be used to directly estimate the relevant factor price elasticity. Consider the following specification of a marginal productivity equation: (8)

vˆfc  log Lfc  df  yc  ξfc ,

where Lfc gives the total number of doctorates in field f and cohort c; and ξfc is the error term. The parameter gives the factor price elasticity, the percent change in the wage associated with a 1 percent increase in labor supply. Ordinary least squares estimation of equation (8) would obviously lead to biased estimates of because the supply of workers to the various cohort-field groups is likely to be endogenous over the thirty-three-year period spanned by the data. The economic question at the core of this chapter, however, suggests an instrument for the size of the workforce in each field-cohort group: the number of immigrants in the ( f, c) cell. In other words, the influx of foreign students into particular doctoral fields at particular times provides the supply shifter required to identify the labor demand function. This instrument would be valid if the foreign student influx into particular doctoral fields were independent of the relative wages offered in the various fields. Since most foreign students intend to remain in the United States, however, the number of immigrants in a field will likely respond to shifts in the wage structure. Income-maximizing behavior on the part of potential foreign students would generate larger flows into those fields that have relatively high wages. This behavioral response would build in a positive correlation between the size of the workforce in a particular cohort-field cell and wages. It can be shown that the independent variable (IV) regression coefficients would then understate the adverse wage impact of a relative supply increase. I estimate the marginal productivity model in (8) by using the mean fixed effects computed from the first stage regression in equation (4), with log Mfc as the instrument.21 The top two rows of table 4.5 report the regression coefficients estimated in the sample of native doctoral recipients. The values of the factor price elasticities reported in these two rows are almost identical to those calculated earlier using the immigrant share specification in equation (3). For example, the factor price elasticities reported in 21. The R-squared of the first-stage regression in the IV regression model corresponding to column (3) of table 4.5 is .976 when the regression uses the three-year cohort groupings and .978 when the regression uses the five-year moving average. The coefficient of log M in these regressions is .452 (.079) and .455 (.045), respectively.

Immigration in High-Skill Labor Markets Table 4.5

149

Factor price elasticities (IV estimates) Adjusted annual salary

Sample / measure of supply

1. Three-year cohort 2. Five-year moving average

1. Three-year cohort 2. Five-year moving average

1. Three-year cohort 2. Five-year moving average

(Field  period) interactions State of residence fixed effects

(1)

(3)

(1)

(2)

(3)

.260 (.126) .289 (.081)

.275 (.129) .312 (.083)

.311 (.142) .341 (.088)

.306 (.141) .337 (.088)

.348 (.197) .373 (.138)

B. Immigrants .405 .423 (.220) (.223) .435 .454 (.150) (.150)

.382 (.230) .451 (.163)

.424 (.238) .497 (.168)

.432 (.235) .504 (.166)

.244 (.125) .267 (.080)

C. All workers .277 .285 (.139) (.140) .306 .313 (.087) (.088)

.302 (.148) .330 (.091)

.328 (.157) .361 (.096)

.329 (.158) .362 (.096)

No No

Yes No

Yes Yes

.227 (.112) .252 (.073)

No No

(2)

Income earned last year

A. Natives .259 (.125) .288 (.081)

Controls: Yes No

Yes Yes

Notes: The standard errors are reported in parentheses. The instrument is the log of the number of doctoral degrees awarded to foreign-born persons in a particular field-cohort group. The regressions in the native sample have 240 observations when using the three-year cohort groups and 714 observations when using the five-year moving average; the respective numbers in the immigrant sample are 235 and 684; and in the “all workers” sample, 240 and 717. All regressions are weighted by the total sampling weight for the field-cohort cell. The standard errors are adjusted for heteroscedasticity by using the Huber-White correction.

table 4.5 for the annual income equations are –.31 (.14) and –.34 (.09), depending on whether I use the three-year cohort groups or the five-year moving average. The implied elasticities reported earlier when I used the immigrant share as the independent variable were –.30 and –.36, respectively. In short, the evidence strongly suggests that an immigration-induced 10 percent increase in the supply of a narrowly defined high-skill group lowers the wage of that group by between 3 and 4 percent. It is worth noting that these factor price elasticities are slightly higher than those estimated by Freeman (1975, 1976) in his series of cobweb-based studies of high-skill science labor markets. For example, Freeman’s estimates of the factor price elasticity in engineering, based on (nonimmigrant induced) supply shocks in the 1950s and 1960s, lie between –.1 and –.2.22 22. I also estimated the IV regressions by doctoral field. These within-field regressions obviously cannot include cohort fixed effects or interactions between work experience and survey year, so that the measured impact of an immigration-induced supply increase is contaminated by important omitted factors (e.g., increases in immigration may be correlated with

150

George J. Borjas

It is important to investigate if the adverse wage impact of the immigrant supply shock also applies to the immigrants themselves. There is, in fact, very little difference between the average earnings of native- and foreignborn doctorates in the United States. In 2001, the average annual income of the typical native-born doctorate was $90.5 thousand, while that of the typical foreign-born doctorate was $88.5 thousand. The adjusted log wage gap between immigrant and native doctorates—after controlling for the worker’s gender, a vector of field fixed effects, and a vector of year-ofgraduation fixed effects—was  .013 (.010), a wage gap that is both numerically and statistically trivial. The middle panel of table 4.5 reestimates the marginal productivity model using the sample of foreign-born doctorates to examine if the immigrant supply shock also affects their earnings opportunities. More precisely, I run the first-stage earnings function using only the sample of foreign-born doctorates, obtain the mean vˆfc for each ( f, c) cell, and estimate the labor demand function in (8). Although the factor price elasticities estimated in the sample of immigrants tend to be slightly more negative than those estimated in the sample of native-born doctorates, the difference between the two sets of estimates is not statistically significant.23 The similarity between the two sets of elasticities is not surprising because the two groups have almost identical incomes (within field-cohort cells). Therefore, it seems that foreign and native doctorates who belong to the same field-cohort cell are close to being perfect substitutes.24 The bottom panel of table 4.5 uses this insight and estimates the labor demand function using the sample of all doctorates, regardless of whether they are native-born or foreign-born. Not surprisingly, the factor price elasticity for annual income lies between –.3 and –.4, indicating that immiimprovements in labor market conditions for a particular field). There is a great deal of interfield dispersion in the estimated coefficients, but the estimated elasticities tend to be negative. The estimated factor price elasticities using adjusted annual salary as the dependent variable and the three-year cohort groupings are: –.140 (.035) in computer science; –.388 (.322) in mathematics; –1.570 (1.280) in agricultural sciences; –.565 (.204) in biology; –.054 (.081) in environmental science; –.119 (.049) in health sciences; –.250 (.205) in chemistry; –.892 (.388) in earth sciences; –.629 (.543) in physics; –.003 (.278) in other physical sciences; 3.503 (1.357) in economics; –.291 (.392) in political science; –1.080 (1.550) in sociology; –1.475 (.688) in other social sciences; .586 (.411) in psychology; –1.444 (1.328) in aerospace engineering; –1.436 (2.132) in chemical engineering; –.642 (.264) in civil engineering; –.104 (.273) in electrical engineering; –.901 (.588) in industrial engineering; –.240 (.129) in mechanical engineering; and –.236 (.485) in other engineering. 23. For example, in the most general specification of the annual income regression, the elasticities are –.306 (.141) and –.432 (.235) in the sample of native and foreign doctorates, respectively. The t-ratio testing for the difference between these two statistics is 0.46. 24. Although the estimated factor price elasticities for immigrants and natives are similar, it is likely that the secular increase in the supply of foreign-born doctorates would be associated with a reduction in average quality. Unless more structure is imposed on the data, however, it is unclear how (or if) this quality decline biases the estimated elasticities.

Immigration in High-Skill Labor Markets

151

gration into a particular field-cohort group adversely affects all workers in that group by a numerically important amount.25 4.4 Postdoctoral Appointments In the 1980s and 1990s it became relatively common for many newlyminted doctorates in some fields to work in postdoctoral appointments for a number of years after graduation.26 These postdoc positions tend to offer relatively low wages when compared to the salary that would be offered, for instance, in a tenure-track academic job. In fact, the postdoc appointments are low-paying even when compared to the salary opportunities offered to new college graduates with little labor market experience. The various waves of the SDR report if the respondent is working at a postdoctoral appointment during the survey week.27 Table 4.6 summarizes the data on the propensity of native doctorates to be employed as postdocs, as well as the average salary of workers employed in such jobs. To more clearly show the importance of postdoctoral appointments in some doctoral career tracks, the table focuses on the sample of doctorates under the age of forty. The data indicate that postdoctoral appointments are very common in some fields.28 The proportion of young workers in postdoctoral positions is 28.7 percent in the biological sciences, 17.4 percent in physics, and 9.3 percent in chemistry. In contrast, postdoctoral appointments are relatively rare in economics and computer sciences, where only 1 to 3 percent of the 25. It would be of interest to determine if the adverse wage impact of the foreign student influx depends on the quality of the universities attended by the native-born doctorates. Unfortunately, the SDR does not identify the degree-granting institution, so that it is not possible to link the data with detailed information on institutional quality. The only qualityrelated variable available in the SDR is the Carnegie classification. However, due to confidentiality considerations, the Carnegie ranking is not reported for many of the persons who received their degree after 1992. Moreover, even in the sample of doctorates who graduated prior to 1991, the Carnegie classification is not a very discriminating measure of school quality: 70.8 percent of native-born and 69.1 percent of foreign-born doctorates received their degree from the top tier in the Carnegie ranking (Research University I). 26. Freeman et al. (2001) report that the career path for the typical doctorate in bioscience changed in the 1980s so that it is not uncommon for newly-minted doctorates to go through a series of postdocs before they start their first “real job” sometime in their midthirties. A National Academy of Sciences (2000) report on postdoctoral appointments summarizes many of the key issues and reports some of the relevant data. 27. The postdoctoral information provided by the 1995 wave differs slightly from that of the other waves. The information on postdoctoral appointments is typically obtained from questions relating to the respondent’s current job. In 1995, however, the information refers to the respondent’s “principal” job. All waves are included in the empirical analysis reported in the following. The results are only slightly different if the 1995 wave is excluded from the analysis. 28. The percent of workers employed as postdocs and the average salaries reported in table 4.6 are obtained by pooling all persons across all the available waves of the SDR between 1993 and 2001 and treating each person-year observation as an independent observation.

152 Table 4.6

George J. Borjas Summary statistics on postdoctorate appointments for native-born doctorates, 1993–2001

Percent employed as postdocs

Field Computer and information sciences Mathematical sciences Agricultural and food sciences Biological sciences Environmental life sciences Health and related sciences Chemistry, except biochemistry Earth sciences, geology, and oceanography Physics and astronomy Other physical sciences Economics Political science Sociology and anthropology Other social sciences Psychology Aerospace and related engineering Chemical engineering Civil and architectural engineering Electrical, electronic engineering Industrial engineering Mechanical engineering Other engineering All fields

Mean annual salary, workers aged 40 or less (in $1,000s)

All persons

Aged 40 or less

Postdoctoral appointment

Not a postdoc

1.6 2.2 2.8 10.1 2.0 2.5 3.5

2.6 7.7 7.9 28.7 6.2 8.9 9.3

58.2 42.8 34.8 34.2 34.3 35.8 35.7

85.7 61.6 63.4 64.0 56.3 62.8 69.8

4.2 5.7 5.9 0.6 1.4 1.6 1.1 2.5 2.1 1.1 1.7 1.1 2.5 1.9 1.4

12.6 17.4 14.2 1.1 3.3 5.3 3.3 6.8 5.2 2.5 6.0 2.1 5.1 3.7 4.6

40.9 41.3 41.4 47.0 40.2 34.3 44.3 32.4 40.7 45.5 45.4 45.2 47.2 47.8 39.5

59.6 69.6 61.3 72.4 54.5 46.5 52.2 56.8 74.5 80.8 67.4 85.7 77.6 74.9 73.4

4.2

12.4

36.0

65.9

Source: Survey of Doctoral Recipients, 1993–2001 waves.

doctorates hold such jobs. The last two columns of the table show that postdoctoral appointments typically pay a great deal less than regular jobs. On average, a doctorate under the age of forty working in a postdoctoral appointment earns $36,000 as compared to $65,900 for a doctorate working in a regular appointment. This wage gap is equally large within fields: postdocs in biology, for example, earn $34,200 as compared to $64,000 for biologists with regular appointment. It is insightful to contrast these salaries with the annual earnings reported by college graduates in the 2000 Census. Male workers who have only a college diploma, work full time, and are between twenty-five and twenty-nine years old earned $33,000, while those who were thirty to thirty-four years old earned $42,300. In sum, the salary opportunities provided by postdoctoral appointments fall far short

Immigration in High-Skill Labor Markets

153

of what even younger persons with less education could earn in the U.S. labor market. The prevalence (and growth) of low-pay postdoctoral appointments as part of the career path in some doctoral fields suggests that one possible channel through which immigration affects the wage structure is by increasing the probability that newly-minted doctorates must first serve an internship in a series of postdoctoral appointments. Put differently, the sizable immigration-induced increase in the supply of doctorates in some fields and for some cohorts may soften labor market conditions sufficiently that scientific labs, for example, can attract many newly-minted doctorates to work in low-pay postdoctoral positions for a relatively long period of time.29 I used a variation of the marginal productivity model presented earlier to determine if the immigrant supply shock indeed increases the probability that native doctorates end up in postdoctoral positions. In particular, I estimated the regression model summarized by equations (4) and (8) using the probability that a particular worker in field f and cohort c is employed in a postdoctoral appointment at time t as the dependent variable. Note that by including interactions between the field and period fixed effects, the regression model controls for the possibility that the demand for postdocs is driven partly by such factors as increased National Institutes of Health (NIH) funding by field; such funding is typically the financial constraint faced by Principal Investigators in university labs before they can staff postdoctoral positions. Table 4.7 summarizes the relevant IV coefficients from the second-stage regression. Using the three-year cohort grouping, the elasticity estimated in the sample of native doctorates is .406 (.153). An immigration-induced 10 percent increase in supply, therefore, raises the probability of being employed in a postdoctoral appointment by about 4.0 percent. The response is even larger when the model is estimated in the sample of younger native workers: a 10 percent immigration-induced increase in supply increases the probability of postdoctoral employment by about 21.6 percentage points. The bottom two panels of table 4.7 reestimate the postdoctoral propensity model in the sample of foreign-born doctorates (panel B) and in the pooled sample of doctorates (panel C). The estimated elasticities are roughly similar across the various samples. Among younger workers, for 29. Of course, the increasing prevalence of postdoctoral appointments in some fields may also reflect structural changes in the training process—for example, it may now take longer to acquire the skills expected of doctorates in some sciences and postdocs arise as a way of filling the need for the longer apprenticeship period. The empirical analysis reported, however, shows that there is a strong correlation between the prevalence of postdoctoral appointments and the size of the immigrant influx in a particular field-cohort cell. It seems unlikely that the increase in the frequency of postdoctoral appointments mandated by educational needs would be so strongly correlated with the field-cohort variation in the number of foreign students.

154

George J. Borjas

Table 4.7

The impact of immigration on the probability of being employed as a postdoctoral fellow (IV estimates) All persons

Sample / measure of supply

1. Three-year cohort 2. Five-year moving average 1. Three-year cohort 2. Five-year moving average 1. Three-year cohort 2. Five-year moving average

(1)

(3)

(1)

(2)

(3)

.406 (.153) .449 (.093)

1.086 (.302) 1.182 (.178)

2.174 (.578) 2.333 (.337)

2.160 (.569) 2.316 (.331)

.640 (.277) .667 (.170)

B. Immigrants .736 .727 (.321) (.321) .768 .760 (.197) (.198)

1.454 (.665) 1.705 (.445)

2.653 (1.184) 3.184 (.798)

2.690 (1.198) 3.227 (.806)

.441 (.168) .484 (.102)

C. All workers .475 .470 (.185) (.183) .522 .517 (.112) (.111)

1.269 (.397) 1.404 (.235)

2.403 (.731) 2.672 (.434)

2.373 (.718) 2.642 (.427)

.383 (.142) .423 (.086)

(2)

40 years old or younger

A. Natives .410 (.155) .453 (.093)

Controls: (Field  period) interactions State of residence fixed effects

No

Yes

Yes

No

Yes

Yes

No

No

Yes

No

No

Yes

Notes: The standard errors are reported in parentheses. The instrument is the log of the number of doctoral degrees awarded to foreign-born persons in a particular field-cohort group. The regressions in the native sample (over all age groups) have 240 observations when using the three-year cohort groups and 714 observations when using the five-year moving average; the respective numbers in the immigrant sample are 236 and 688; and in the “all workers” sample, 240 and 717. The regressions in the native sample (for workers under the age of forty) have 170 observations when using the three-year cohort groups and 478 observations when using the five-year moving average; the respective numbers in the immigrant sample are 158 and 434; the respective numbers in the “all workers” sample are 172 and 484. All regressions are weighted by the total sampling weight for the field-cohort cell. The standard errors are adjusted for heteroscedasticity by using the Huber-White correction.

example, a 10 percent immigration-induced increase in supply increases the probability of being employed in a postdoctoral appointment by 20 to 30 percentage points, regardless of whether the affected doctorates are native-born or foreign-born. Finally, because postdocs earn about 50 percent less than comparable workers in regular jobs, the results in table 4.7 suggest that an important part of the wage impact of immigration may be taking place through the crowding of workers in immigrant-penetrated fields into postdoctoral appointments. To measure the extent to which postdoctoral appointments provide a channel for the labor market to reduce the wages of the affected

Immigration in High-Skill Labor Markets

155

workers, I reestimated the labor demand model in the sample of workers who are not employed in postdoctoral appointments. If the sole impact of immigration on labor market opportunities for doctorates was through the increased placing of workers in low-pay postdoctoral appointments, this regression specification should generate zero factor price elasticities. In fact, as table 4.8 shows, the estimated elasticities are still negative, but only about half the size of the elasticities reported earlier in the chapter (see the analogous table 4.5). For example, the factor price elasticity estimated in the sample of native workers using the three-year cohort group is –.306 (.141) when using all native workers, and –.125 (.084) when using the sample of native doctorates not employed as postdocs. It seems, therefore, that roughly half of the adverse wage impact of immigration on high-skill

Table 4.8

Factor price elasticities for workers not in postdoctoral appointments (IV estimates) Adjusted annual salary

Sample / measure of supply

1. Three-year cohort 2. Five-year moving average 1. Three-year cohort 2. Five-year moving average 1. Three-year cohort 2. Five-year moving average

(1)

(2)

(3)

Income earned last year (1)

(2)

(3)

.053 (.057) .063 (.047)

A. Native .074 .074 (.064) (.064) .088 .087 (.051) (.051)

.111 (.077) .127 (.058)

.132 (.085) .152 (.062)

.125 (.084) .145 (.062)

.003 (.080) .011 (.080)

B. Immigrant .032 .055 (.088) (.086) .046 .075 (.083) (.080)

.042 (.121) .107 (.104)

.089 (.129) .160 (.109)

.125 (.128) .198 (.108)

.034 (.056) .040 (.044)

C. All workers .055 .058 (.063) (.063) .065 .069 (.047) (.047)

.089 (.081) .109 (.058)

.113 (.089) .137 (.062)

.110 (.089) .135 (.062)

Controls: (Field  period) interactions State of residence fixed effects

No

Yes

Yes

No

Yes

Yes

No

No

Yes

No

No

Yes

Notes: The standard errors are reported in parentheses. The instrument is the log of the number of doctoral degrees awarded to foreign-born persons in a particular field-cohort group. The regressions in the native sample have 240 observations when using the three-year cohort groups and 714 observations when using the five-year moving average; the respective numbers in the immigrant sample are 236 and 685; and in the “all workers” sample, 240 and 716. All regressions are weighted by the total sampling weight for the field-cohort cell. The standard errors are adjusted for heteroscedasticity by using the Huber-White correction.

156

George J. Borjas

labor markets can be attributed directly to the increased use of low-pay postdoctoral appointments as a way of adjusting to the increase in supply. 4.5 Simulating the Wage Effects of the Foreign Student Program I now use the factor price elasticity estimated in this chapter to simulate the wage impact of the foreign student influx that entered the United States between 1993 and 2001. Assuming that capital is constant and ignoring spillover effects across fields, the total impact of immigration on the log wage of native workers in field f is:30 (9)

 log wf  mf,

where mf gives the percentage change in labor supply due to immigration in field f. I define mf as: (10)

˜ M f,1993–2001 mf   , ˜ ˜ (N˜f,2001  M f,2001) Mf,1993–2001

where M˜f,1993–2001 gives the change in the stock of foreign-born doctorates in field f between 1993 and 2001 and is calculated from the SED as the total number of doctorates awarded in field f to foreign-born persons (who intend to stay in the United States) during that period; and N˜f,2001 and M˜f,2001 give the stock of native and foreign-born doctorates in field f as of 2001 and are estimated from the 2001 wave of the SDR. The definition of the supply shock in (10) treats all foreign-born persons who obtained their degree prior to 1993 as part of the native baseline. In effect, the predicted wage effect (multiplied by minus one) gives the additional log wage that native doctorates in 2001 would have earned had the foreign student program been halted in 1993. The simulation uses the log adjusted annual salary as the dependent variable because this measure of earnings is available in both 1993 and 2001. Table 4.9 summarizes the results of the simulation using the –0.260 estimate of the factor price first reported in table 4.5. On average, the 1993 to 2001 influx increased the supply of doctorates by 13.9 percent. This supply shock reduced the wage of the average worker with a doctorate in science and engineering by approximately 3.6 percent. The predicted losses are sometimes very large because the supply shock in particular fields has been substantial. In computer science and mechanical engineering, for example, immigration increased the supply of doctorates by over 36 percent. This supply shock resulted in predicted wage losses of nearly 10 percent. 30. The assumption of a constant capital stock implies that the resulting wage consequences should be interpreted as short-run impacts. Over time, the changes in factor prices will fuel adjustments in the capital stock that attenuate the wage effects. The simulation also ignores the cross-effects of supply shocks in a particular field on the earnings of doctorates in other fields.

Immigration in High-Skill Labor Markets Table 4.9

157

Predicted wage impact of the 1993–2001 immigrant influx, by field

Field

Immigrant supply shock

Predicted impact on log salary

Actual change in log salary

Computer and information sciences Mathematical sciences Agricultural and food sciences Biological sciences Environmental life sciences Health and related sciences Chemistry, except biochemistry Earth sciences, geology, and oceanography Physics and astronomy Other physical sciences Economics Political science Sociology and anthropology Other social sciences Psychology Aerospace and related engineering Chemical engineering Civil and architectural engineering Electrical, electronic engineering Industrial engineering Mechanical engineering Other engineering

0.364 0.173 0.206 0.130 0.022 0.095 0.137

0.095 0.045 0.054 0.034 0.006 0.025 0.036

0.216 0.049 0.081 0.058 0.041 0.084 0.091

0.109 0.150 0.339 0.131 0.054 0.045 0.071 0.018 0.203 0.213 0.289 0.335 0.235 0.369 0.223

0.028 0.039 0.088 0.034 0.014 0.012 0.018 0.005 0.053 0.055 0.075 0.087 0.061 0.096 0.058

0.038 0.073 0.169 0.131 0.012 0.127 0.050 0.121 0.110 0.036 0.073 0.188 0.137 0.115 0.078

All fields

0.139

0.036

0.052

Notes: The simulation uses the factor price elasticity reported in the third column of row 1 of table 4.5, or .260. The immigrant supply shock (within field) gives the ratio of the number of doctorates granted between 1993 and 2001 to the native stock in 1993 (where the native stock in 1993 is defined as the sum of the total number of doctorates granted to natives and the number of doctorates granted to foreign-born persons prior to 1993).

The last column of the table reports what actually happened to the logadjusted annual salary between 1993 and 2001. The typical doctorate experienced a 5.2 percent increase in real wages. The foreign student influx, therefore, reduced wage growth by about 40 percent of what it would have been in its absence. Note, however, that there is a great deal of dispersion across fields in the relative impact of foreign students. In earth sciences, for example, the foreign student influx explains most of the 3.8 percent drop in real wages experienced by doctorates in that field. In contrast, the real wage of economists would have risen by 20 percent more had there been no immigrant influx during the period. It is important to point out a number of conceptual problems and interpretation difficulties inherent with this type of simulation. Any simulation of the wage impact of immigration must be based on a particular set of assumptions describing how the economy adjusted to the immigrant influx.

158

George J. Borjas

Needless to say, different counterfactuals inevitably lead to different simulated impacts. The simulation summarized in table 4.9 explicitly holds all other factors constant, so that neither native workers nor firms adjust to the increased number of foreigners who sought doctorates in American universities and then chose whether or not to remain in the U.S. labor market.31 There are many ways in which such adjustments could take place, and the resulting estimates of the wage impact of immigration could be correspondingly lower or higher, depending on the assumed counterfactuals. Suppose, for example, that native students would have taken the place of the foreign students admitted to the various graduate programs if there had been an enforceable prohibition on the entry of foreign students. In this extreme case, the total supply of doctorates in particular field-cohort groups would have been the same regardless of whether foreign students had been admitted to U.S. universities. This counterfactual implies that the wage structure in the doctoral labor market today would be exactly what we now observe, despite the fact that not a single foreign student entered the country. Alternatively, suppose that native students responded to the immigrant influx in particular fields and in particular years by moving to other departments in the university, or perhaps by going to law or business school. This spillover effect of immigration would then tend to lower wages throughout the entire high-skill sector, not just in the fields penetrated by immigrants. These across-field migration flows suggest that the labor market impact of immigration estimated in this chapter is numerically smaller than the actual impact, since the movement of native students across fields would tend to arbitrage wage differences. The simulation exercise reported in table 4.9 is best seen as an attempt to calculate the short-run impact of immigration, before any adjustments take place. Neither the supply and career decisions of native students nor the level of demand for doctorates in particular fields is affected by immigration. It would be interesting, of course, to simulate the impact of immigration in the market for high-skill workers under alternative scenarios. 4.6 Summary This chapter analyzed the impact of immigration on high-skill labor markets. The analysis used data drawn from the Survey of Earned Doctorates, a population enumeration of all doctoral degrees awarded by U.S. universities, and the Survey of Doctoral Recipients, a biennial longitudinal 31. In other words, table 4.9 compares the “actual” world where foreign students come to the United States to obtain their doctoral degrees and then choose to remain and enter the U.S. labor market, to a counterfactual where foreign students come to the United States to obtain their degrees, but then choose to return to their home countries (or move elsewhere) after completion of their studies.

Immigration in High-Skill Labor Markets

159

data set that provides detailed information on labor market conditions for a sample of doctorates in science and engineering. The empirical study is based on the intuitively appealing notion that shifts in labor supply in a finely-detailed skill group should affect the earnings and employment opportunities of that skill group. Put differently, immigration-induced shifts in the supply of students entering particular doctoral fields at particular times can be used to identify the impact of immigration on the earnings of doctorates. The analysis indicates that increases in the number of foreign-born doctorates, primarily through the foreign student program, have a significant adverse effect on the earnings of competing workers, regardless of whether the competing workers are native-born or foreign-born. An immigrationinduced 10 percent increase in the supply of doctorates in a particular field at a particular time reduces the earnings of that cohort of doctorates by about 3 to 4 percent. About half of this adverse wage effect can be attributed to the increased prevalence of low-pay postdoctoral appointments in fields that have softer labor market conditions because of large-scale immigration. These results have implications in a number of different policy contexts. For instance, there has been a long-standing debate about whether immigration affects labor market conditions for native workers at all. This study, along with other recent empirical work, seems to suggest that the supplydemand textbook model is correct after all: increases in labor supply do move the labor market along the demand curve and lead to lower wages for competing workers. It is also the case that economic opportunities in high-skill labor markets are among the key determinants of the career decisions made by the nativeborn student population. The increase in the number of foreign doctorates has clearly reduced economic opportunities in some fields relative to others, and may be an important factor driving native students to enter particular occupations and avoid others. For example, the wage that could be earned by native postdoctoral workers employed in research biology labs is much lower than it would have been in the absence of the immigrant influx, perhaps motivating bright U.S.-born undergraduates to pursue professional occupations that have not been targeted by immigration. The low wage paid to postdoctoral workers in these biology labs, however, still offers a very attractive opportunity when contrasted to the compensation available in other countries, so that the incentives for even more foreign students to enter the United States are not greatly reduced. In a sense, there is a potential vicious cycle where the incentives of research labs to offer low wages to their workers barely affect the supply of foreign doctorates, but have a substantial impact on the career decisions of native workers. In the resulting equilibrium, research labs find that they must keep recruiting from abroad because of the

160

George J. Borjas

assumption that natives do not want to do the type of work that immigrants do. Although we do not yet know the magnitude of the supply elasticities that determine inter-field migration flows, the wage effects of largescale immigration into some doctoral fields are very large and would be expected to be a crucial factor in labor supply decisions. Finally, although the foreign student program grew rapidly in the past three decades, this growth occurred without any systematic study of the costs and benefits that such a program entails for the native-born population. This chapter addressed an important component in such a costbenefit analysis—the cost borne by doctorates in the U.S. labor market. There is an equally important component that has not yet been analyzed carefully, namely the benefits of the program, such as the possibility that the sizable increase in the skill endowment of the workforce accelerates the rate of scientific discovery. These benefits could be very large and accrue to particular parts of the population, so that high-skill immigration may have significant efficiency and distributional effects that have yet to be analyzed.

References Altonji, J. G., and D. Card. 1991. The effects of immigration on the labor market outcomes of less-skilled natives. In Immigration, trade, and the labor market, ed. J. M. Abowd and R. B. Freeman, 201–34. Chicago: University of Chicago Press. Borjas, G. J. 1999. The economic analysis of immigration. In Handbook of labor economics, vol. 3A, ed. O. C. Ashenfelter and D. Card, 1697–1760. Amsterdam: Elsevier. ———. 2003. The labor demand curve is downward sloping: Reexamining the impact of immigration on the labor market. Quarterly Journal of Economics 118 (4): 1335–74. Borjas, G. J., R. B. Freeman, L. F. Katz, J. DiNardo, and J. M. Abowd. 1997. How much do immigration and trade affect labor market outcomes? Brookings Papers on Economic Activity, Issue no. 1:1–67. Washington, D.C.: Brookings Institution. Card, D. 1990. The impact of the Mariel Boatlift on the Miami labor market. Industrial and Labor Relations Review 43 (2): 245–57. ———. 2001. Immigrant inflows, native outflows, and the local labor market impacts of higher immigration. Journal of Labor Economics 19 (1): 22–64. Card, D., and T. Lemieux. 2001. Can falling supply explain the rising return to college for younger men? A cohort-based analysis. Quarterly Journal of Economics 116 (2): 705–46. Ehrenberg, R. 2000. Tuition rising: Why college costs so much. Cambridge, MA: Harvard University Press. Finn, M. G. 2003. Stay rates of foreign doctorate recipients from U.S. universities, 2001. Oak Ridge Institute for Science and Education, Oak Ridge, Tennessee. Working Paper. Freeman, R. B. 1975. Supply and salary adjustments to the changing science manpower market: Physics, 1948–1975. American Economic Review 65 (1): 27–39.

Immigration in High-Skill Labor Markets

161

———. 1976. A cobweb model of the supply and starting salary of new engineers. Industrial and Labor Relations Review 29 (2): 236–48. Freeman, R. B., E. Weinstein, E. Marincola, J. Rosenbaum, and F. Solomon. 2001. Competition and careers in bioscience. Science 294 (5550): 2293–94. Friedberg, R. M. 2001. The impact of mass migration on the Israeli labor market. Quarterly Journal of Economics 116 (4): 1373–1408. Friedberg, R. M., and J. Hunt. 1995. The impact of immigration on host country wages, employment and growth. Journal of Economic Perspectives 9 (Spring): 23–44. Grossman, J. B. 1982. The substitutability of natives and immigrants in production. Review of Economics and Statistics 64 (4): 596–603. Hamermesh, D. 1993. Labor demand. Princeton, NJ: Princeton University Press. LaLonde, R. J., and R. H. Topel. 1991. Labor market adjustments to increased immigration. In Immigration, trade, and the labor market, ed. J. M. Abowd and R. B. Freeman, 167–99. Chicago: University of Chicago Press. Levin, S. G., G. C. Black, A. E. Winkler, and P. E. Stephan. 2004. Differential employment patterns for citizens and non-citizens in science and engineering in the United States: Minting and competitive effects. Growth and Change 35 (4): 456–75. National Academy of Sciences, Committee on Science, Engineering, and Public Policy. 2000. Enhancing the postdoctoral experience for scientists and engineers: A guide for postdoctoral scholars, advisers, institutions, funding organizations, and disciplinary societies. Washington, D.C.: National Academies Press. Samuelson, P. A. 1964. Economics, 9th Edition. New York: McGraw-Hill. Schoeni, R. F. 1997. The effect of immigrants on the employment and wages of native workers: Evidence from the 1970s and 1980s. The RAND Corporation. Unpublished Manuscript, March. Smith, J. P., and B. Edmonston, eds. 1997. The new Americans: Economic, demographic, and fiscal effects of immigration. Washington, D.C.: National Academies Press. Snyder, T. D., and C. M. Hoffman. 2002. Digest of education statistics, 2002. Washington, D.C.: U.S. Department of Education. Welch, F. 1979. Effects of cohort size on earnings: The baby boom babies’ financial bust. Journal of Political Economy 87 (October, part 2): S65–S97.

5 Does Science Promote Women? Evidence from Academia 1973–2001 Donna K. Ginther and Shulamit Kahn

Fewer women are present in science academe than in the workforce as a whole, and this is particularly true in the higher levels of academe, such as tenured jobs and full professorships at major research universities. This chapter begins from the point when scientists receive their Ph.D.s and investigates gender differences as they move up the academic career ladder through the stages of getting tenure-track jobs, being granted tenure, and being promoted to full professorships. There is a large body of literature about women and science, particularly since 1982 when Congress instructed the National Science Foundation (NSF) to report biennially on the status of women and minorities in science. The NSF reports have consistently shown that since 1982 and through the most recent report (NSF 2004a), women continue to be less likely than their male colleagues to be full professors and more likely to be assistant professors. Congress established its own committee, the Congressional Committee on the Advancement of Women and Minorities in Science, Engineering, and Technological Developments (CAWMSET), to review the status of women in science. This committee (CAWMSET 2000) also found that women in SET (Science, Engineering, and Technology) acDonna K. Ginther is an associate professor of economics at the University of Kansas. Shulamit Kahn is an associate professor of finance and economics at the School of Management, Boston University. We thank Al and Judy Erickson for making this chapter possible. We also thank the National Science Foundation for granting a site license to use the data and Kelly Kang of the NSF for providing technical documentation. Dylan Rassier and Ronnie Mukherjee provided research assistance. Ginther acknowledges financial support from NSF grant SES-0353703. Finally, we thank Anne Preston and Richard Freeman for their useful comments on an earlier draft. The use of NSF data does not imply NSF endorsement of the research, research methods, or conclusions contained in this report. Any errors are our own responsibility.

163

164

Donna K. Ginther and Shulamit Kahn

ademia are less likely to be tenured (29 percent of women versus 58 percent of men among full-time ranked academics at four year colleges) or hold full professorships (23 percent of women compared to 50 percent of men). More recently, the Government Accountability Office (GAO) reported to Congress that women scientists lag behind men in terms of salary and rank (GAO 2004). In site visits, some women report that tenure-track positions at research universities create difficulty in balancing work and family. Others report that a hostile climate makes academic employment unattractive. Another recent study by Donna Nelson and Diana Rogers (2005) found that smaller percentages of women than men who receive Ph.D.s proceed to become assistant professors in top fifty SET departments. These important sources represent only a few of the many studies on women in science. Even though women are underrepresented in upper echelons of academic science, one cannot conclude from the NSF, CAWMSET, or Nelson reports that unfair treatment in the promotion process is the underlying cause of the gender gap in academic promotion. Two alternative possibilities include that women choose careers that do not have the rigid academic timetable or that women are less productive, particularly in terms of research, than men. Of course, research productivity itself may result from the absence of an environment and the resources that foster research, as demonstrated at MIT (Goldberg 1999). In contrast to these negative findings, Long (2001) studies the careers of women in science from 1973 to 1995 and concludes that women have been successful in moving “from scarcity to visibility.” They find that the impact of marriage and children on women’s careers had largely been eliminated by 1995, although men were still 4 percent more likely to receive tenure. On the other hand, Xie and Shauman (2003) find that marriage and children exacerbate gender differences in promotion in nonacademic science. In addition, they find the gender publication gap is smaller than in previous studies and declining over time, suggesting a convergence in women’s and men’s academic productivity. A recent report by the NSF (NSF 2004b) is the most comprehensive study to date of the factors contributing to promotion in academic careers of scientists and engineers. This work, carried out contemporaneously to ours and also using NSF’s longitudinal Survey of Doctorate Recipients (SDR), finds that controlling for human capital, personal characteristics, and institutional factors, there remains a significant female disadvantage in the likelihood of being in a tenure-track job, of receiving tenure, and of being promoted to full professor. However, in most of their specifications, they find that these gender differences become statistically insignificant when family characteristics are allowed to affect men and women differently. Our findings are quite different qualitatively from theirs, for reasons we discuss in the conclusion. We find that in science, single women actually have an advantage over single men in obtaining tenure-track jobs and in

Does Science Promote Women?

165

being granted tenure after controlling for covariates, and that married men and women without children are quite similar at these two stages. Children lower the likelihood that women in science will advance up the academic job ladder beyond their early postdoctorate years. In contrast, children have a positive or zero effect on men’s career success in academic science. We also find that science is not homogeneous. There are particularly large gender differences in obtaining tenure-track jobs, getting tenure, and being promoted to full in the life sciences, the area that graduates the most women. The remainder of the chapter is organized as follows: we first describe the data and methodology. We then discuss the entry into tenure-track jobs, describe and model the tenure decision, and then describe and model promotion to full professor. The final section concludes. 5.1 Data and Empirical Methodology Our analysis of promotion uses data from the 1973 to 2001 waves of the Survey of Doctorate Recipients (SDR). The SDR is a biennial, longitudinal survey of doctorate recipients from U.S. institutions conducted by the National Research Council. The SDR collects detailed information on doctorate recipients including demographic characteristics, educational background, employer characteristics, academic rank, government support, primary work activity, productivity, and salary. The SDR has undergone substantial changes in the sampling frame and survey content between the 1973 and 1993 waves (Mitchell, Moonesinge, and Cox 1998). Technical reports provided by the National Science Foundation have allowed us to construct a longitudinal data set with consistent variable definitions over time.1 We have selected a longitudinal extract of doctorate recipients in the sciences who received their Ph.D. between the years of 1972 and 1991 and remain in the survey ten years after the Ph.D. Individuals are excluded if they are not observed more than once or if they skip more than three surveys. We estimate three career milestones. First, we examine the probability of obtaining a tenure-track job within nine years of the Ph.D. Then we restrict the analysis to those who have ever held a tenure-track job to estimate two promotion milestones, the first award of tenure and the first award of full professorship. 1. Many longitudinal surveys are plagued by nonresponse and attrition. However, the NSF does a remarkably good job at keeping SDR response rates high. The response rate for each survey is in the range around 78 to 80 percent. Many of these nonresponders do respond to the following survey. Only about 5 percent of the sample either do not respond for three consecutive surveys or cannot be found. Note that people are dropped from the SDR when (a) they die (b) they pass seventy-five years of age (c) they are non-U.S. citizens out of the United States for two surveys in a row and (d) on a random basis in order to maintain the target sample size while incorporating new Ph.D.s.

166

Donna K. Ginther and Shulamit Kahn

From the 1973 through 1991 surveys, respondents provided the exact year that they received tenure, which adds some accuracy given the biennial nature of the survey. For later surveys, tenure year is imputed as the first year a person is observed with tenure in the sample. We impute the year a person receives full professorship as the first year a person is observed as a tenured full professor in the sample. Given the biennial nature of the survey, years until tenure and years until full professor may be measured with one-year error. Our following analyses include both time-varying and nontime varying independent variables. Nontime varying variables include gender, race, whether foreign-born, field, and aspects of the person’s Ph.D. institution. Time-varying independent variables include marital status, children, employer characteristics, primary and secondary work activities, government support, and limited productivity measures (discussed following). These covariates are suggested by previous studies of academic promotion (Long, Allison, and McGinnis 1993; Ginther and Kahn 2004). Table 5.1 gives descriptive statistics about both dependent and covariate variables at different stages of academic careers of scientists. Measures of academic productivity are largely missing from the SDR data, but the SDR does ask questions about publications in the 1983, 1995, and 2001 surveys. The 1983 question refers to publications between 1980 and 1983 whereas the 1995 and 2001 questions refer to numbers of publications in the previous five years. We use these data to create rough measures of cumulative papers presented and publications per year past Ph.D. If productivity data are missing for a particular year (as they are prior to 1980), average observed productivity is used to impute total productivity—an admittedly rough correction that nevertheless seems preferable to omitting the information altogether. Research by Ginther and Hayes (1999, 2003), Ginther (2001, 2003, 2004), and Ginther and Kahn (2004) demonstrates that employment outcomes differ by academic field. Thus, promotion is analyzed for all scientific fields together and broken down into three major scientific fields—biological and life sciences, physical sciences, and engineering. It is particularly important to differentiate between fields for gender differences in academic careers, in that the combined science statistics on women are more likely to be picking up trends in the life sciences, where most of the women are, while the statistics on men are quite likely to pick up engineering, which is heavily male. Accordingly, we point out when major facts differ across these broad areas. We evaluate gender differences in academic careers using both probit and hazard methodologies. In our probit analyses, first we estimate whether significant gender differences exist in the probability of a tenure-track job within nine years of the Ph.D. for all individuals with valid surveys. Second, for those who hold a tenure-track job at some point in their careers, we estimate probit models of the probability of having tenure eleven years

0.544*** — — 31.942*** 0.051** 0.004 0.122 0.002 0.168 79.672*** 0.732*** 0.108 0.077*** 0.036*** 0.672*** 0.756*** 0.320*** — — — — — — —

Female 0.582*** — — 30.674*** 0.042** 0.004 0.130 0.002 0.181 80.701*** 0.766*** 0.114 0.039*** 0.029*** 0.804*** 1.163*** 0.470*** — — — — — — —

Male

All doctoratesa

Gender differences in mean characteristics

Tenure-track within 9 years of Ph.D. Promotion to tenure within 11 years of Ph.D. Promotion to full within 15 years of Ph.D. Age at Ph.D. African American Native American Asian Other race Foreign-born Year of Ph.D. Ph.D. from Research I Ph.D. from Research II Ph.D. from Doctorate I Ph.D. from Doctorate II Married Total children Children  6 Cumulative employers Private university Research I Liberal arts I Medical school Primary work research Primary work teach

Variables

Table 5.1

— 0.516 — 31.998*** 0.047 0.005 0.089** 0.003 0.150 79.265*** 0.739*** 0.102 0.082*** 0.035 0.648*** 0.752*** 0.261*** 1.718 0.274*** 0.267*** 0.200*** 0.216 0.352*** 0.456***

Female — 0.532 — 30.503*** 0.048 0.005 0.105** 0.002 0.157 80.254*** 0.775*** 0.113 0.039*** 0.027 0.846*** 1.361*** 0.378*** 1.686 0.230*** 0.306*** 0.166*** 0.221 0.456*** 0.377***

Male

Tenure-trackb

— — 0.257*** 31.763*** 0.044 0.002 0.076 0.001 0.130 77.263*** 0.740*** 0.101 0.082*** 0.035 0.645*** 0.767*** 0.169*** 2.058** 0.272*** 0.275** 0.201 0.214 0.331*** 0.440***

Female

Male — — 0.316*** 30.152*** 0.047 0.004 0.084 0.001 0.120 78.247*** 0.779*** 0.113 0.036*** 0.031 0.849*** 1.399*** 0.246*** 1.961** 0.222*** 0.310** 0.176 0.215 0.427*** 0.372*** (continued )

Tenuredc

— — — — — — — — — 0.110*** 0.431*** 0.083*** 0.060*** 0.094*** 0.162*** 8,141

0.103*** 0.360*** 0.246*** 0.127 0.135*** 0.454*** 1.987*** 5.587*** 6.839*** 0.159 0.559*** 0.101*** 0.037*** 0.047*** 0.056*** 2,218

Female 0.079*** 0.397*** 0.280*** 0.119 0.103*** 0.534*** 2.373*** 8.707*** 9.451*** 0.142 0.414*** 0.072*** 0.064*** 0.074*** 0.164*** 4,406

Male

Tenure-trackb

0.124 0.378 0.256 0.145 0.106 0.481*** 2.902*** 7.006*** 9.131*** 0.149 0.588*** 0.108*** 0.032*** 0.048*** 0.036*** 1,238

Female

Male 0.108 0.406 0.270 0.157 0.093 0.540*** 3.455*** 10.728*** 12.291*** 0.147 0.440*** 0.076*** 0.067*** 0.074*** 0.123*** 2,721

Tenuredc

a

Note: Tests are two-sided. Excluding doctoral recipients who were unlikely to have sought academic jobs as evidenced by their immediately entering nonacademic jobs upon or prior to receipt of their Ph.D. b Time-varying variables evaluated eleven years from Ph.D. c Time-varying variables evaluated fifteen years from Ph.D. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

— — — — — — — — — 0.126*** 0.566*** 0.118*** 0.038*** 0.052*** 0.051*** 4,604

Primary work manage Secondary work research Secondary work teach Secondary work manage Secondary work other Government support, current year Cumulative years of government support Cumulative papers Cumulative publications Computer science / mathematics Biology and Life sciences Chemistry Earth science Physics Engineering Observations

Male

All doctoratesa Female

(continued)

Variables

Table 5.1

Does Science Promote Women?

169

after receiving a Ph.D. In most scientific fields, this includes a period of one or more postdoctorates in addition to a period in tenure-track assistant professor jobs. In this analysis, time-varying covariates are evaluated eleven years after Ph.D. receipt. Third, for those who receive tenure (by fifteen years past Ph.D.), probit analyses model the probability of being a full professor fifteen years after receiving a Ph.D. In these latter probits, timevarying covariates are all evaluated at fifteen years past Ph.D. We also use proportional hazard models with time varying-covariates to estimate the likelihood each year of becoming tenured, given that the individual has survived untenured until that point, and to estimate the likelihood each year of being promoted to full, given that the individual has received tenure but has not yet been promoted to full.2 Although the hazard analysis allows time-varying covariates to change each year and this is an improvement over measuring characteristics at a single point in time, it is not the ideal way to evaluate the impact of these variables on an individual’s career. For example, the impact of children on one’s career accumulates in terms of time devoted to childcare instead of scientific productivity. However, it is very difficult to measure these impacts precisely and we are left to use crude proxies such as the number of children or presence of young children to account for their effect on career outcomes.3 5.2 Stepping Onto the Academic Career Ladder The NSF conducts a census of doctorates granted in the United States in its Survey of Earned Doctorates. Based on these data, figure 5.1 illustrates the continuous growth in the percentage of the Ph.D.s granted in science going to females in all fields between 1974 and 2004. Life sciences started this period with less than 20 percent of Ph.D.s granted to females, but by the year 2004 nearly 50 percent were. Engineering started this period as the field with fewest female doctorates, and remained so for the entire quarter century. Nevertheless, the percent of female engineering doctorate recipients rose from just over one percent in 1974 to almost 18 percent by 2004. Figure 5.2 uses weighted estimates from the SDR to show the changing 2. Hazard models are preferred to probit models because they account for censored observations. The proportional hazard model is given by: hi(t)  o(t) exp {1xi1  . . .  kxik} where the hazard of promotion hi(t) is a function of the baseline hazard o(t) and covariates x. The covariates in the hazard equation influence the scale of the hazard rate and are not a function of time. In addition, the hazard for any one individual is a fixed proportion to the hazard for any other person in the sample. 3. While the accumulated effect might be captured by a variable such as “the percentage of time since Ph.D. with children in the house,” construction of this variable would require loss of many survey years and much of our data.

170

Donna K. Ginther and Shulamit Kahn

Fig. 5.1 Percentage of doctorates granted to females, 1974–2004 Survey of Earned Doctorates Source: 1974–2004 Survey of Earned Doctorates.

percentage of females in each academic rank over the quarter century. The general upward trend in the percentage of females among assistant professors mirrors the trend in science Ph.D. awards from figure 5.1. Also similar to doctorates granted, life sciences have the highest percentage of females among assistant professors, with physical sciences at much lower levels and engineering at the very lowest. Other aspects of the time trends in assistant professorships (in fig. 5.2) compared to doctoral recipients (in fig. 5.1) differ by field. In life sciences, throughout the entire quarter century, fewer women than men proceed from Ph.D. receipt to a tenure-track assistant professorship, with the wedge during the past four years being especially large, averaging a difference of 6 percentage points. In fact, during these four years, the proportion of females among assistant professors in life sciences has actually fallen despite the fact that given increasing time trends in doctoral receipt, we would have expected them to have risen. In contrast, in physical sciences the percent of females among assistant professors has consistently kept pace with the percentage of female doctorates. In 2001, 25 percent of doctorates awarded to women in the physical sciences and 26 percent of assistant professors

Percentage female by academic rank, science disciplines

Source: 1973–2001 Survey of Doctorate Recipients.

Fig. 5.2

172

Donna K. Ginther and Shulamit Kahn

were women. Even more extreme, in engineering, from the late 1980s until very recently, the percentage of females among assistant professorships was greater than among doctorates. Between 1990 and 1997, on average 11 percent of doctorates in engineering were granted to women whereas on average 13 percent of assistant professors were women. However, since 1997, the percent of females among engineering doctoral recipients has pulled ahead of that of assistant professors. Among Ph.D. recipients immediately entering the academic sector (including those entering postdocs), fewer women than men have tenuretrack jobs within nine years of Ph.D. receipt. Table 5.1 indicates that 58 percent of men held tenure-track jobs within nine years of Ph.D., compared with 54 percent of women. This statistically significant although small gender difference in obtaining a tenure-track job may not represent different jobs among otherwise identical men and women, but instead may be picking up a large variety of factors such as field and Ph.D. quality that happen to be correlated with gender. Model 2 of table 5.2 shows the gender difference once we control for covariates. (Results of the entire probit specifications are presented in appendix table 5A.1.) Even after controlling for age at Ph.D., cohort, race, origin, Ph.D. quality tier,4 and field in model 2, there is a significant gender difference in the likelihood of being in a tenuretrack job nine years after Ph.D. receipt. Disaggregating by broad fields reveals that this gender difference is entirely due to life sciences, which has a significant and large gender difference in the likelihood of getting a tenuretrack job that increases to 7.7 percent once all of these covariates are added. In contrast, for engineering and physical sciences, the gender difference is small and insignificant. Marriage and fertility decisions may be affecting women’s tenure-track employment. To investigate this possibility, we include marital status, total number of children, an indicator for having children less than six years of age, and each of these interacted with gender (table 5.2, model 3). Several female interaction terms are statistically significant and suggest that family variables affect women and men quite differently. With family interactions entered as controls, the female coefficient in science as a whole, in life sciences, and in physical sciences becomes positive and statistically significant. This indicates that single, childless women are between 11 percent (in life science) to 21 percent (in physical science) more likely to get a tenure-track job within nine years of the Ph.D. than single, childless men. In engineering, there is no significant difference. Being married increases men’s probability of getting a tenure-track job by a whopping 22 percent, in science as a whole and for each field separately. Women’s chances are also helped somewhat by marriage, but only 4. Quality tiers are based on rankings from the Carnegie Foundation for the Advancement of Teaching and of Comprehensive and Liberal Arts Institutions.

0.033*** (0.010) — — — — — — — — — — — — Yes Yes Yes

0.038*** (0.009) — — — — — — — — — — — — No No No 0.156*** (0.018) 0.218*** (0.016) 0.029*** (0.007) 0.022 (0.016) 0.171*** (0.024) 0.029** (0.013) 0.059** (0.028) Yes Yes Yes

Model 3 0.041*** (0.012) — — — — — — — — — — — — No No No

Model 1

*Significant at the 10 percent level.

**Significant at the 5 percent level.

***Significant at the 1 percent level.

Notes: Coefficients report change in probability. Standard errors in parentheses.

Source: 1973–2001 Survey of Doctorate Recipients.

Demographics Degree characteristics Fields

Model 2

Model 1

Science

0.077*** (0.013) — — — — — — — — — — — — Yes Yes Yes

Model 2

Life science

0.108*** (0.025) 0.220*** (0.023) 0.036*** (0.010) 0.026 (0.023) 0.149*** (0.033) 0.022 (0.017) 0.068* (0.038) Yes Yes Yes

Model 3 0.002 (0.016) — — — — — — — — — — — — No No No

Model 1 0.015 (0.017) — — — — — — — — — — — — Yes Yes Yes

Model 2 0.206*** (0.029) 0.222*** (0.026) 0.021* (0.012) 0.001 (0.029) 0.236*** (0.041) 0.055*** (0.022) 0.021 (0.050) Yes Yes Yes

Model 3

Physical science

Probability of having tenure-track appointment within nine years of Ph.D. by scientific field

Female * Young children

Female * Total children

Female * Married

Children  6  1

Total children

Married

Female

Table 5.2

0.000 (0.033) — — — — — — — — — — — — No No No

Model 1

0.013 (0.035) — — — — — — — — — — — — Yes Yes Yes

Model 2

Model 3 0.072 (0.064) 0.221*** (0.045) 0.025* (0.014) 0.069* (0.037) 0.009 (0.092) 0.053 (0.045) 0.000 (0.100) Yes Yes Yes

Engineering

174

Donna K. Ginther and Shulamit Kahn

by 5 percent on average.5 As a result, in science as a whole a married man without children and a married woman without children are about equally likely to have a tenure-track job nine years after Ph.D. However, there are large differences between the scientific fields. At one extreme, in engineering both sexes have equally large positive impacts of marriage (21 to 22 percent). In life sciences, marriage increases women’s likelihood of entering a tenure-track job by a more modest 7.1 percent (again compared to 22 percent for men). Finally, in physical science, marriage does not affect women’s chances at all. Children create a marked divergence between men and women. For science as a whole, the presence of a prekindergarten-aged child nine years post-Ph.D. lowers women’s likelihood of having a tenure-track job by 8.1 percent. The presence of a grade school child has no significant effect, presumably because the demands of rearing very young children occurred before Ph.D. receipt rather than during these nine years post-Ph.D. In contrast to women, prekindergarten children have no effect on men’s likelihood of having a tenure-track job while each child above six years old increases a man’s probability of getting a tenure-track job by 2.9 percent. Disaggregating children’s impact by field, young children especially hurt the tenure-track prospects of women in life sciences (by –8.1 percent) and in physical sciences by (–5.6 percent). In engineering, while the point estimate is large (–9.8 percent), it is significant only at the 20 percent level (perhaps due to small numbers of females in engineering). Grade school children are negatively correlated with women having a tenure-track job for physical science only, where the impact is relatively small (–3.4 percent). The positive impacts of marriage and children on men’s prospects here recalls positive impacts on wages and promotion in the labor market as a whole, which has been attributed to three primary explanations. First, particularly with respect to marriage, it may be due to selection: “good catches” in the marriage market are correlated with “good catches” in the labor market. Second, it could be induced effort by men responsible for a family. Third, it could be paternalistic favoritism by employers who know that the man has a family to support. Neither the induced effort nor the paternalistic favoritism seem likely to apply to new job offers for Ph.D.s in academia. And they are unlikely to ever apply to women. We are thus left with selection as the key explanation for positive impacts of marriage in obtaining tenure-track jobs for both men and women and with the positive impacts of children for men. 5. This calculation adds the coefficient of “Married” to the coefficient of “Female∗Married.” Many other numbers later in this section similarly add several coefficients. For instance, the impact on men of one young child adds the coefficient on “Total Children” to the coefficient on “Children”  6, while the impact on women adds to this sum the coefficient of “Female∗Total Children” and the coefficient of “Female∗Children”  6. We note in the text when the sum is not significant.

Does Science Promote Women?

175

Gender differences in the likelihood of receiving a tenure-track job have changed over time. In additional specifications (available upon request), the gender difference between comparable men and single women (table 5.2, model 3) was allowed to differ by year of Ph.D. In pooled science, later cohorts of women did better relative to men. For instance, single women with 1972 Ph.D.s in science had a 12.1 percent higher likelihood of entering tenure-track jobs within nine years than single men of that cohort, and this gender difference widened to 24.4 percent for those with 1991 Ph.D.s. Disaggregating, life science and physical science fields actually saw even larger changes over cohorts, while engineering had no significant cohort differences between men and women. 5.3 Empirical Analysis of Moving Up the Career Ladder: Promotion of Academic Scientists 5.3.1 Estimates of the Probability of Promotion to Tenure Returning to figure 5.2, the dashed line shows the changing percentage of females among associate professors, while figure 5.3 shows the percentage of females among all tenured faculty. In science as a whole, the monotonically increasing trend in associate professorships mirrors trends in

Fig. 5.3

Percentage of tenured faculty who are female, by discipline

Source: 1973–2001 Survey of Doctorate Recipients.

176

Donna K. Ginther and Shulamit Kahn

assistant professorships five to ten years earlier, and the levels are comparable. For instance, 26 percent of females among associates in 1991 is the same as the percent of females among assistant professors six years earlier. Within broad fields, however, trends in percent of females among associate professorships are not at all smooth or monotonically increasing, with substantial drops in the percentage of females in 1996 in life sciences and in 1993 in engineering, and stagnation in the percentage of females in physical sciences between 1989 and 1995. The top panel of table 5.3 summarizes the impact of gender on tenure probabilities before we allow gender differences in the impact of family variables.6 The first row shows the probit analysis of gender differences in the probability of tenure by eleven years from the doctorate controlling for academic field, demographic, family and employer characteristics, primary and secondary work activity, government grant support, and productivity (but without interaction terms). These results show no significant gender differences in tenure; the point estimates of the impact of being female even vary in sign across fields. Hazard analyses are able to capture the entire year-by-year pattern of the likelihood of receiving tenure and thus in the duration until tenure. A particular strength of this analysis is that it takes into account those observed to not have received tenure by the last survey. The second row of table 5.3 presents the risk ratios from a proportional hazards model of promotion regressed on a dummy variable for gender. This risk ratio can be interpreted as the effect of being female rather than male on the probability of receiving tenure. (A number less than one indicates that on average the likelihood of tenure receipt in any given year for females is less than for males.) In the hazard analysis with no controls, there is no significant gender difference either for science as a whole or for any of the broad fields. However, after adding in controls, the risk ratios fall, indicating less tenure for women.7 With controls, the gender difference is only significant for life sciences, where the point estimate suggests an 8 percent lower likelihood of tenure for women (p  .07).8 Once again, adding in female-interaction terms for marriage and children changes the picture. Table 5.4 reports probit coefficients of these gender and family terms when included in addition to other covariates.9 We have done the same estimation with hazard analyses (details available on request). Those results are qualitatively similar except where noted. 6. Appendix table 5A.2 provides detailed parameter estimates for the probit model. Hazard estimates are available by request. 7. The fact that the gender differences arise after controlling for these covariates suggests that the women who obtain tenure-track jobs have better credentials (X variables) than the men. 8. The complete hazard analysis is available upon request. 9. Note that the family variables measure the status as of eleven years after Ph.D. receipt.

Does Science Promote Women? Table 5.3

177

Gender differences in the probability and hazard of promotion Full sample

Female probit coefficient Promoted eleven years past Ph.D. (Including all covariates)

Life science

Physical science

Engineering

Promotion to tenure 0.00 0.03 (0.88) (0.19)

0.01 (0.73)

0.02 (0.75)

Risk ratio estimate 0.97 1.02 (0.33) (0.60) 0.95 0.89** (0.14) (0.02) 0.97 0.92* (0.29) (0.07)

1.00 (0.96) 0.93 (0.22) 0.94 (0.28)

1.06 (0.56) 1.00 (0.97) 1.03 (0.82)

0.02 (0.51)

0.09 (0.37)

Female risk ratio (No covariates) Model 1 female risk ratio (Covariates ex. productivity) Model 2 female risk ratio (Including productivity covariates) Female probit coefficient Promoted fifteen years past Ph.D. (Including all covariates) Female risk ratio (No covariates) Model 1 female risk ratio (Covariates ex. productivity) Model 2 female risk ratio (Including productivity covariates)

Promotion to full 0.05** 0.09*** (0.02) (0.00) Risk ratio estimate 0.90*** 0.96 (0.01) (0.48) 0.95 0.93 (0.34) (0.37) 0.97 0.96 (0.54) (0.61)

0.79*** (0.00) 0.87 (0.11) 0.89 (0.19)

0.95 (0.74) 1.09 (0.89) 1.04 (0.82)

Source: 1973–2001 Survey of Doctorate Recipients. Notes: P-values in parentheses. Probit coefficient reports change in probability. Hazard coefficients are risk ratios—estimates the impact of female on the likelihood of promotion in each period. Promotion to tenure is estimated on those who receive a tenure-track status within this period. Promotion to full is estimated on those who have been given tenure. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

As was true with entrance into tenure-track jobs, single women are more likely than single men to receive tenure. For science as a whole, the difference is 6.4 percent. However, disaggregating by broad field, it is only significantly true in engineering, where there is a very large difference (20.2 percent). In the life and physical sciences, differences between single men and women are essentially zero. Marriage does not have as large an effect on men with regard to tenure receipt as it did for obtaining tenure-track jobs. Marriage only significantly increases men’s likelihood of tenure in engineering (by 12.3 percent) and in science as a whole (6.2 percent). Marriage does not have a statistically significant effect on women in any field, although in physical sciences there is

178

Donna K. Ginther and Shulamit Kahn

Table 5.4

Marriage and children in probit analysis of tenure and promotion to full professor Full sample

Life science

Physical science

Promotion to tenure (11 years past Ph.D.) 0.064** 0.017 0.021 (0.0325) (0.047) (0.054) Married 0.062** 0.055 0.058 (0.029) (0.042) (0.046) Total children 0.026*** 0.027** 0.015 (0.010) (0.014) (0.017) Children  6 0.028 0.032 0.013 (0.023) (0.032) (0.041) Female * Married 0.024 0.032 0.028 (0.042) (0.058) (0.070) Female * Total children 0.048** 0.015 0.045 (0.020) (0.026) (0.036) Female * Children  6 0.010 0.026 0.031 (0.045) (0.061) (0.079) Female

Female Married Total children Children  6 Female * Married Female * Total children Female * Children  6

Promotion to full (15 years past Ph.D.) 0.013 0.011 0.034 (0.040) (0.054) (0.067) 0.046 0.066 0.048 (0.032) (0.045) (0.052) 0.013 0.008 0.001 (0.011) (0.015) (0.018) 0.048* 0.036 0.042 (0.027) (0.038) (0.045) 0.071 0.136** 0.074 (0.048) (0.059) (0.088) 0.030 0.010 0.093** (0.023) (0.031) (0.039) 0.098 0.011 0.242** (0.063) (0.081) (0.102)

Engineering

0.202** (0.082) 0.123* (0.073) 0.039* (0.021) 0.047 (0.052) 0.005 (0.152) 0.264*** (0.064) 0.174* (0.080) 0.235 (0.262) 0.026 (0.103) 0.064** (0.033) 0.122 (0.088) 0.218 (0.330) 0.050 (0.114) 0.413 (0.104)

Source: 1973–2001 Survey of Doctorate Recipients. Notes: Coefficients report change in probability. Standard errors in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

a possibility it increases tenure (p  .12). Combining this with the direct gender effect, married (childless) women have similar likelihoods of tenure as married (childless) men in life sciences and physical sciences. In engineering, married (childless) women have a 20.7 percent higher likelihood of tenure than married (childless) men ceteris paribus, similar to the difference between single men and women in engineering. Each child of school age by eleven years past Ph.D. has a positive 2.6 per-

Does Science Promote Women?

179

cent impact on the promotion probability of men in science as a whole. This is statistically significant in engineering (where the impact is 3.9 percent) and life science (with an impact of 2.7 percent). In physical sciences, the impact is not significant in the probit but is marginally significant in the more powerful hazard analysis. As was true in the previous section for entering the tenure track, the impact of younger children on men’s tenure likelihood is zero in all fields. For women, preschool children also have no significant impact in any field. Having grade-school children at eleven years past Ph.D., implying that they had children for many of those eleven years, has a large negative (–22.5 percent) impact on women’s tenure in engineering only. This engineering result must be seen in the context of two other differences between engineering and other scientific fields. Engineering was the only field where young children did not significantly hurt a woman’s likelihood of being in a tenure-track job nine years after Ph.D., and engineering has the lowest incidence of postdocs (NSF 2006). Together, these three facts suggest that women with children in all three fields are more likely to drop off the tenure-track career path within several years post-Ph.D. However, in engineering, this occurs while they are assistant professors. Before proceeding, we note a seeming contradiction. In the hazard analysis of the life sciences without gender-specific family variables, women had an 8 percent disadvantage in achieving tenure than men, ceteris paribus. However, in the probits with gender-specific family variables, there was no family situation where a woman in the life sciences was less likely to get tenure than a comparable man. Hazard analyses with family interaction terms confirm the results of these probits. The only explanation consistent with these facts is compositional. For instance, there are few single men but many married men with older children, and these characteristics work to men’s advantage. We discuss the issue of family composition in the conclusion. In additional specifications, other variables were allowed to have different coefficients by gender. We found that in the life sciences (only), private universities were less likely to give tenure in general, but women were penalized less than men. For men, the likelihood of receiving tenure fell over the cohort years. For women, this effect was smaller, still negative in life sciences but marginally positive in engineering and physical sciences. 5.3.2 Estimates of the Probability of Promotion to Full Professorship Returning to figure 5.2, among full professors in science, the percentage of females has been steadily increasing over the quarter century shown, but by 2001 had only achieved the level of female representation that had been achieved in assistant professorships and Ph.D. recipients in the early 1970s. Disaggregated, the same can be said for physical sciences and engineering. Life sciences here is the “success story,” with the percentage of females

180

Donna K. Ginther and Shulamit Kahn

among full professorships continually rising to 20 percent by 2001, comparable to the level of assistant professorships in 1983. Trends and levels among tenured faculty in figure 5.3 combine the trends in associate professors with those in full professors. The bottom panels of tables 5.3 and 5.4 summarize the impact of gender on promotion to full professorship.10 The probits and hazards reported here use two different beginning time points. The probits estimate the likelihood that someone who has received tenure has a full professorship fifteen years after Ph.D. The hazards start with first tenure receipt and study the likelihood of becoming a full professor and the duration of time it takes to get there. The first row of the bottom panel of table 5.3 gives the probit coefficient on promotion to full professor within fifteen years of Ph.D. (with covariates). For all sciences pooled, there is a significant gender difference. Breaking this down by broad fields allows us to see that this is entirely due to life sciences, where women have a 9 percent lower likelihood of being promoted to a full professorship. In the other two fields, differences are not significantly different from zero.11 In the hazard analysis of table 5.3, as before, the second row of the panel includes no covariates. The risk ratio from the proportional hazard analysis (without covariates) indicates highly significant gender differences in promotion to full in science as a whole. On average, the likelihood of promotion to full in any given year for females is 90 percent that of males. Disaggregating by field, we see a significantly lower promotion rate in physical science only, where the likelihood of being promoted to full professor in any given year for females is only 79 percent that of males. Adding in a full set of controls in the last two rows of table 5.3, however, moves the risk ratio in both the full sample and in physical sciences closer to one and makes them insignificant at standard levels of significance. The gender promotion gap remains the largest in physical science after controlling for all covariates at 11 percent but has only a p-value of .19. Of course, a much larger sample than ours12 could very well identify statistically significant gender gaps. Details on the family interaction terms are reported in the bottom panel of table 5.4. Adding gender-family interaction terms, we see no significant differences in either the broad fields or in science overall between single 10. Appendix table 5A.3 provides detailed parameter estimates for the probit models. Hazard estimates are available on request. 11. In analysis not shown, we estimated the probability of promotion to full professor seventeen and nineteen years after Ph.D. for life scientists as well. While women are 5 percent less likely to be promoted to full professor by fifteen years after Ph.D. receipt, they were 6 percent less likely by seventeen years. There was no significant difference in the probability of being promoted to full professor by nineteen years after Ph.D. receipt. 12. There are 2,721 tenured females in science as a whole and only 990 in the physical sciences.

Does Science Promote Women?

181

men and women. Marriage does not have an impact on men’s promotion to full. For women, there is no impact of being married in the pooled sample, but disaggregating, a married woman in life sciences has a 7.0 percent lower chance of achieving full (p  .11) than a single woman. Consequently, married childless women have lower rates of promotion to full than married childless men in the life sciences. In contrast, in physical science a married woman is 12.2 percent more likely to have a full professorship than a single one (p  .07). Having school-aged children fifteen years post-Ph.D. has no effect on men’s promotion to full, except in engineering where each child makes promotion 6.4 percent more likely. Having school-aged children fifteen years post-Ph.D. has an effect on women’s promotion to full only in the physical sciences, where in this case it lowers the probability of becoming full by 9.4 percent. Finally, having young children has no clear effect on full professorship for either sex in any field, with one exception: women in engineering. In engineering, young children may raise the probability of a woman receiving full in engineering (30.5 percent, significant at the 11 percent level). A few other variables were shown to have different impacts on men and women. For life sciences and consequently for science as a whole, private universities significantly hurt women’s chances of being promoted to full in the hazard model specifications—opposite to the gender difference found at promotion to tenure. Finally, unlike the tenure decision, both women’s and men’s likelihood of promotion to full worsen for later cohorts. 5.4 Conclusions: Putting Gender Differences in Promotion into Perspective One conclusion we make from this research is that aggregate statistics on gender differences in academic science careers are often misleading. Within science as a whole, there seem to be only small (between 0 and 3 percent) and sometimes insignificant differences between men and women scientists’ probability of obtaining a tenure-track job within nine years of doctorate receipt, receiving tenure, or being promoted to full professorships, after controlling for demographic and employer characteristics, academic field, primary and secondary work activity, government grants, and publications.13 However, the broad fields of science are very dissimilar. There are particularly large gender differences in the life sciences, the area that graduates the most women. Within the life sciences, men are approximately 8 to 9 percent more likely than women to obtain a tenure-track job within nine years of their Ph.D., to receive tenure, and to be promoted to 13. The full set of controls is included in all specifications discussed in this conclusion unless otherwise noted.

182

Donna K. Ginther and Shulamit Kahn

full. In contrast, there are no appreciable or statistically significant differences within physical science or engineering with the exception of a large but statistically insignificant gender difference in promotion to full professorship in physical sciences. In addition, aggregate gender differences often mask much more substantial differences between men and women with particular kinds of family structures. In fact, single women are more successful in the early years of academic careers, which in life and physical sciences covers the transition into tenure-track jobs but in engineering—with its few short postdocs—covers the tenure decision. If they make it into a tenure-track job, single women and men in the life and physical sciences are equally likely to receive tenure and to be promoted to full. Marriage greatly increases the likelihood that men get tenure-track jobs (by 22 percent), but has smaller and generally less significant effects on men’s promotion at either level. Marriage tends not to hurt women’s likelihood of getting tenure-track jobs, being granted tenure, or becoming full.14 Indeed, marriage increases the likelihood of obtaining tenure-track jobs, although not as much as it helps men. The positive effects of marriage on obtaining tenure-track jobs for both men and women seems most likely to be due to selection, insofar as it is unlikely to be either induced effort or paternalistic favoritism. Combining gender differences of singles and gender differences in the impact of marriage, a married man without children and a married woman without children are typically similar in their academic progress, with some exceptions. It is striking that marriage does not hurt women in science. Dual career problems do not seem to deter women from getting a tenure-track job, from getting tenure, or from becoming a full professor, despite the fact that more than 60 percent of women scientists are married to scientists (Rosser 2004). The presence of children, however, does disadvantage women during the early post-Ph.D. years that coincide with the child-bearing window. In life sciences and physical sciences, young children make it less likely for women to make it through the postdoc hurdle and get a tenure-track job. In engineering, people tend to go directly from the doctorate receipt to jobs, bypassing the postdoc stage. Here, too, however, having had children for much of their early career (as indicated by the school-aged children at eleven years post-Ph.D.) lowers women’s likelihood of succeeding in academia (in this case, of receiving tenure), while the absence of children makes women in engineering more successful in getting tenure than similar men. These results indicate that to some extent, women in science must make an early choice between a family and an academic career. Opting out of academic career jobs because of children dovetails with some of 14. The single exception is a 7 percent lower chance of achieving full in life sciences.

Does Science Promote Women?

183

Preston’s (2004) results, which show a major reason that women leave science is because of childcare responsibilities. In contrast, for men the presence of grade-school children (but not young children) is positively correlated with their likelihood of receiving tenure-track jobs and receiving tenure. A gender difference is not surprising, given that men spend much less time than women in childcare, even in professional couples. Preston (2004) finds that those male scientists who do spend time in childcare have similar impacts on their academic careers. It is possible that the negative impact of children on women early in their academic careers is also partly selection. We cannot know whether the women who have children during the formative years of their careers would be less devoted to their careers even in the absence of children than those who do not (i.e., a selection story), or are being hampered by the children’s presence. However, to the extent that children do indeed hamper women’s early career progression, science departments and associations should not therefore conclude that gender differences in early academic careers are nothing to be concerned about. Our results indicate that women must face a choice between having children or succeeding in their scientific careers, while men do not face these same choices. While science departments are clearly not responsible for the cultural expectation that mothers are the primary parental caregivers, the findings here should encourage conversations on whether the present system within academic science of long postdocs requiring long hours, particularly in life sciences, are necessary or even desirable to good science.15 The estimated gender differences that we have found among women scientists entering academic jobs post-Ph.D. are different from the recent NSF report (NSF 2004a) using the same data set. Where we find that single women have greater rates entering tenure-track jobs and being promoted to tenure and full (ceteris paribus), the NSF found no gender differences for entering tenure-track jobs and lower rates of women promoted to full. Where the NSF found that marriage hurt women’s careers at various stages, we find that marriage in the absence of children does not hurt. While NSF found negative impacts of children at all levels, we find the negative effect of children at the point of entry into tenure-track jobs only. What accounts for these very different results? There are some small differences in our research that are not responsible for the large discrepancies in results. For instance, our analysis uses the most recent data available from the 2001 SDR. Also, other studies stopped their analysis in 1999 or earlier. The NSF (2004a) included a somewhat different set of controls and did not include any publication controls. Instead, the important explanation for differences between our results 15. See, for instance, Freeman et al. (2001).

184

Donna K. Ginther and Shulamit Kahn

and the others is that we are looking only at the life sciences, physical sciences, and engineering. In contrast, both Long (2001) and NSF define science as including social science. Indeed, there is a gender difference in academic promotion in social sciences that we have demonstrated in previous work. Ginther (2002) and Ginther and Kahn (2004) estimates the probability and duration to promotion for faculty in the social sciences and economics, respectively. Ginther (2002a) finds a gender promotion gap in the social sciences that ranges between 10 to 12 percent (through 1997), with only half of the gap being explained by observable characteristics. In the field of economics, Kahn (1993) and Ginther and Kahn (2004) both find large gender promotion differences. Ginther and Kahn (2004) use data from the SDR (as well as independently collected data) through 2001 and find a 21 percent gender promotion gap in economics with less than half of the gap explained by observable characteristics. That paper also estimates an 8 percent promotion gap in social sciences, excluding economics, through 2001. Our results on promotion in sciences also differ from findings by Ginther and Hayes (1999, 2003) for faculty in the humanities. Using the 1977 to 1995 waves of the SDR and performing similar estimates, Ginther and Hayes find a gender promotion gap ranging between 7 to 9 percent. Some of the promotion gap in the humanities is explained by fertility and the treatment of work experience. Taking all of this work together, women’s disadvantages in promotion to tenure not explained by any covariates are largest in economics and other social sciences, are smaller in the humanities (in part explained by marriage and family characteristics), and nonexistent in the physical or life sciences or in engineering once all variables are taken into account. This is not to say that there are no gender differences at all in academic science careers once scientists enter tenure-track jobs. We have shown that promotion to full professorships is substantially different for men and women. Other research also finds different salaries of academic men and women in science. Ginther (2001, 2003, 2004) shows a significant gender salary gap in academic science especially at the full professor rank, after controlling for similar covariates including productivity. In 2001, male full professors in science earned 12 percent more than female full professors and one-third of this salary gap is not explained by observable characteristics (Ginther 2004). Although there is no significant difference in the likelihood of being promoted to full professor, compensation is apparently not equivalent.

0.033*** (0.010) 0.002*** (0.001) 0.084*** (0.021) 0.041 (0.068) 0.113*** (0.017) 0.019 (0.093) 0.047*** (0.015) 0.000 (0.001) 0.052*** (0.021) 0.062*** (0.023) 0.100*** (0.026) 0.066*** (0.031) — — — —

0.038*** (0.009) — — — — — — — — — — — — — — — — — — — — — — — — — —

Female

Total children

Married

Ph.D. from Doctorate II

Ph.D. from Doctorate I

Ph.D. from Research II

Ph.D. from Research I

Year of Ph.D.

Foreign-born

Other race

Asian

Native American

African American

Age at Ph.D.

Model 2

Model 1

Science

0.156*** (0.018) 0.001 (0.001) 0.104*** (0.021) 0.025 (0.069) 0.111*** (0.017) 0.014 (0.093) 0.044*** (0.015) 0.000 (0.001) 0.052*** (0.021) 0.054** (0.023) 0.099*** (0.026) 0.062** (0.031) 0.218*** (0.016) 0.029*** (0.007)

Model 3 0.041*** (0.012) — — — — — — — — — — — — — — — — — — — — — — — — — —

Model 1 0.077*** (0.013) 0.003** (0.001) 0.102*** (0.028) 0.091 (0.086) 0.096*** (0.025) 0.001 (0.126) 0.087*** (0.022) 0.004*** (0.001) 0.043 (0.024) 0.023 (0.029) 0.095*** (0.036) 0.022 (0.043) — — — —

Model 2

Life science

0.108*** (0.025) 0.002 (0.001) 0.117*** (0.028) 0.072 (0.088) 0.090*** (0.025) 0.022 (0.126) 0.086*** (0.022) 0.005*** (0.001) 0.043* (0.024) 0.018 (0.029) 0.099*** (0.036) 0.026 (0.043) 0.220*** (0.023) 0.036*** (0.010)

Model 3 0.002 (0.016) — — — — — — — — — — — — — — — — — — — — — — — — — —

Model 1 0.015 (0.017) 0.003 (0.002) 0.044 (0.039) 0.008 (0.160) 0.089*** (0.028) 0.039 (0.139) 0.048** (0.025) 0.005*** (0.001) 0.112** (0.054) 0.155*** (0.050) 0.199*** (0.048) 0.197*** (0.051) — — — —

Model 2

Physical science

0.206*** (0.029) 0.004** (0.002) 0.070* (0.038) 0.012 (0.164) 0.088*** (0.029) 0.003 (0.139) 0.046* (0.025) 0.005*** (0.001) 0.122** (0.054) 0.156*** (0.050) 0.198*** (0.048) 0.200*** (0.051) 0.222*** (0.026) 0.021* (0.012)

Model 3 0.000 (0.033) — — — — — — — — — — — — — — — — — — — — — — — — — —

Model 1

Probability of having a tenure-track appointment within nine years of Ph.D.: 1973–2001 survey of doctorate recipients

Variables

Table 5A.1

Appendix

0.013 (0.035) 0.008*** (0.003) 0.081 (0.056) 0.103 (0.162) 0.182*** (0.041) — — 0.032 (0.033) 0.007*** (0.002) 0.034 (0.076) 0.085 (0.074) 0.092 (0.099) 0.018 (0.097) — — — —

Model 2

0.072 (0.064) 0.010*** (0.003) 0.103* (0.054) 0.129 (0.165) 0.192*** (0.042) — — 0.044 (0.033) 0.007*** (0.002) 0.003 (0.077) 0.048 (0.081) 0.119 (0.103) 0.025 (0.105) 0.221*** (0.045) 0.025* (0.014)

Model 3

(continued )

Engineering

Health sciences

Zoology

Microbiology

Biochemistry

Biology and life sciences

Earth science

Chemistry

Physics

Computer science / mathematics

Female * Young children

Female * Total children

Female * Married

Children  6  1

Variables

Table 5A.1

— — — — — — — — — — — — — — — — — —

— — — — — — — —

Model 1

(continued)

0.137*** (0.022) 0.256*** (0.023) 0.229*** (0.023) 0.109*** (0.028) 0.166*** (0.020) — — — — — — — —

— — — — — — — —

Model 2

Science

0.154*** (0.022) 0.244*** (0.024) 0.218*** (0.024) 0.102*** (0.028) 0.161*** (0.020) — — — — — — — —

0.022 (0.016) 0.171*** (0.024) 0.029** (0.013) 0.059** (0.028)

Model 3

— — — — — — — — — — — — — — — — — —

— — — — — — — —

Model 1 0.026 (0.023) 0.149*** (0.033) 0.022 (0.017) 0.068* (0.038)

Model 3

— — — — — — — — — — 0.126*** (0.018) 0.072*** (0.024) 0.034 (0.029) 0.111*** (0.017)

— — — — — — — — — — 0.126*** (0.018) 0.075*** (0.025) 0.034 (0.029) 0.111*** (0.017)

Academic Field

— — — — — — — —

Model 2

Life science

— — — — — — — — — — — — — — — — — —

— — — — — — — —

Model 1

0.232*** (0.022) 0.162*** (0.025) 0.137*** (0.025) — — — — — — — — — — — —

— — — — — — — —

Model 2

Physical science

0.245*** (0.022) 0.156*** (0.025) 0.130*** (0.025) — — — — — — — — — — — —

0.001 (0.029) 0.236*** (0.041) 0.055*** (0.022) 0.021 (0.050)

Model 3

— — — — — — — — — — — — — — — — — —

— — — — — — — —

Model 1

— — — — — — — — — — — — — — — — — —

— — — — — — — —

Model 2

Engineering

— — — — — — — — — — — — — — — — — —

0.069* (0.037) 0.009 (0.092) 0.053 (0.045) 0.000 (0.100)

Model 3

— — — — — — — — — — — — — — — — — — 12,745 17.42

— — — — 0.007 (0.024) — — — — — — — — — — — — 12,745 820.28

— — — — 0.005 (0.024) — — — — — — — — — — — — 12,745 1,156.36

— — — — — — — — — — — — — — — — — — 6,826 10.93

*Significant at the 10 percent level.

**Significant at the 5 percent level.

***Significant at the 1 percent level.

Notes: Coefficients report change in probability. Standard errors in parentheses.

Observations Likelihood ratio stat

Industrial

Mechanical

Electrical

Civil

Chemical

Aerospace

Engineering

Agriculture and food

Environmental science

0.052 (0.049) 0.158*** (0.020) — — — — — — — — — — — — — — 6,826 368.44

0.033 (0.050) 0.152*** (0.021) — — — — — — — — — — — — — — 6,826 545.49 — — — — — — — — — — — — — — — — — — 4,365 0.01

— — — — — — — — — — — — — — — — — — 4,365 555.55

— — — — — — — — — — — — — — — — — — 4,365 691.06

— — — — — — — — — — — — — — — — — — 1,554 0.00

— — — — — — 0.092 (0.059) 0.055 (0.042) 0.157*** (0.033) 0.091*** (0.029) 0.135*** (0.033) 0.273*** (0.030) 1,554 102.37

— — — — — — 0.092 (0.059) 0.067 (0.041) 0.160*** (0.033) 0.100*** (0.029) 0.148*** (0.033) 0.277*** (0.028) 1,554 147.38

Table 5A.2

Probit estimates of the probability of tenure within eleven years of Ph.D. by field Science

Female Age at Ph.D. African American Native American Asian Other race Foreign-born Year of Ph.D. Ph.D. from Research I Ph.D. from Research II Ph.D. from Doctorate I Ph.D. from Doctorate II

Married Total children Children  6 Cumulative employers Private university Research I Liberal arts I Medical school Primary work research Primary work teach Primary work manage

0.003 (0.017) 0.008*** (0.002) 0.048 (0.035) 0.057 (0.103) 0.001 (0.030) 0.250 (0.143) 0.062*** (0.025) 0.008*** (0.002) 0.054 (0.039) 0.038 (0.043) 0.058 (0.049) 0.077 0.056

Life science 0.032 (0.024) 0.008*** (0.003) 0.037 (0.050) 0.036 (0.134) 0.004 (0.047) — — 0.120*** (0.039) 0.017*** (0.003) 0.060 (0.044) 0.050 (0.055) 0.134** (0.066) 0.122 0.079

11 years after Ph.D. 0.046** 0.033 (0.021) (0.029) 0.015 0.023 (0.008) (0.012) 0.028 0.039 (0.020) (0.028) 0.161*** 0.137*** (0.009) (0.012) 0.108*** 0.155*** (0.017) (0.024) 0.044** 0.044* (0.018) (0.025) 0.097*** 0.105*** (0.021) (0.033) 0.100*** 0.095*** (0.020) (0.024) 0.090*** 0.137*** (0.035) (0.044) 0.435*** 0.442*** (0.030) (0.037) 0.210*** 0.213*** (0.035) (0.049)

Physical science

Engineering

0.010 (0.029) 0.001 (0.003) 0.048 (0.061) 0.070 (0.246) 0.028 (0.047) 0.143 (0.215) 0.060 (0.042) 0.001 (0.003) 0.015 (0.096) 0.023 (0.105) 0.044 (0.111) 0.030 0.115

0.017 (0.052) 0.002 (0.005) 0.066 (0.083) 0.498** (0.195) 0.043 (0.061) — — 0.016 (0.050) 0.005 (0.005) 0.029 (0.136) 0.023 (0.137) 0.039 (0.148) 0.020 0.185

0.067* (0.036) 0.005 (0.015) 0.009 (0.036) 0.211*** (0.016) 0.068** (0.029) 0.002 (0.035) 0.121*** (0.032) 0.047 (0.047) 0.087 (0.076) 0.498*** (0.065) 0.214*** (0.064)

0.124 (0.065) 0.005 (0.019) 0.016 (0.047) 0.140*** (0.025) 0.045 (0.046) 0.057 (0.045) 0.011 (0.058) 0.121** (0.063) 0.094 (0.105) 0.406*** (0.100) 0.146 (0.089)

Table 5A.2

(continued)

Science Secondary work research Secondary work teach Secondary work manage Secondary work other Government support in current year Cumulative years of government support Cumulative papers Cumulative publications

0.011 (0.030) 0.260*** (0.029) 0.077** (0.034) 0.071** (0.036)

Life science 0.045 (0.046) 0.138*** (0.047) 0.021 (0.051) 0.147*** (0.048)

0.003 (0.022)

0.007 (0.031)

0.004 (0.006) 0.002*** (0.001) 0.007*** (0.001)

0.006 (0.009) 0.001 (0.001) 0.009*** (0.001)

Physical science 0.026 (0.045) 0.388*** (0.036) 0.145*** (0.048) 0.038 (0.060) 0.023 (0.040)

Engineering 0.081 (0.077) 0.221*** (0.063) 0.193** (0.061) 0.064 (0.088) 0.075 (0.052)

0.002 (0.012) 0.001 (0.001) 0.006*** (0.002)

0.048*** (0.016) 0.005*** (0.002) 0.002 (0.002)

0.085*** (0.041) 0.146*** (0.047) 0.125*** (0.046) — — — — — — — — — — — — — — — — — — — — —

— — — — — — — — — — — — — — — — — — — — — — — — — 0.059 (0.109)

Academic fields Computer science / mathematics Physics Chemistry Earth science Biology and life sciences Biochemistry Microbiology Zoology Health sciences Environmental science Agriculture and food Engineering Engineering Aerospace

0.010 (0.039) 0.235*** (0.041) 0.198*** (0.041) 0.104** (0.047) 0.215*** (0.033) — — — — — — — — — — — — 0.047 (0.039) — — —

— — — — — — — — — — 0.100*** (0.033) 0.079* (0.043) 0.124*** (0.047) 0.142*** (0.030) 0.162** (0.080) 0.216*** (0.035) — — — — —

(continued )

Table 5A.2

(continued)

Chemical Civil Electrical Mechanical Industrial Observations Likelihood ratio stat

Science

Life science

Physical science

— — — — — — — — — — 5,187 1,393.41

— — — — — — — — — — 2,756 768.98

— — — — — — — — — — 1,757 570.18

Engineering 0.017 (0.076) 0.009 (0.063) 0.002 (0.052) 0.013 (0.065) 0.043 (0.100) 669 145.83

Source: 1973–2001 Survey of Doctorate Recipients. Notes: Coefficients report change in probability. Standard errors in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

Table 5A.3

Probit estimates of the probability of full professor within fifteen years of Ph.D. by field

Female Age at Ph.D. African American Native American Asian Other race Foreign-born Year of Ph.D. Ph.D. from Research I Ph.D. from Research II

Science

Life science

Physical science

Engineering

0.048** (0.021) 0.004** (0.002) 0.102** (0.040) 0.091 (0.138) 0.028 (0.038) 0.367 (0.217) 0.036 (0.032) 0.016*** (0.003) 0.049 (0.048) 0.034 (0.056)

0.085*** (0.028) 0.002 (0.003) 0.058 (0.057) 0.216 (0.161) 0.023 (0.055) — — 0.022 (0.049) 0.016*** (0.004) 0.062 (0.050) 0.076 (0.065)

0.024 (0.037) 0.004 (0.004) 0.132* (0.069) 0.129 (0.229) 0.122*** (0.063) — — 0.050 (0.053) 0.014*** (0.004) 0.127 (0.153) 0.071 (0.172)

0.088 (0.097) 0.011 (0.008) 0.224* (0.127) 0.221 (0.275) 0.001 (0.106) — — 0.114 (0.083) 0.030*** (0.009) 0.175 (0.245) 0.081 (0.281)

Table 5A.3

(continued)

Ph.D. from Doctorate I Ph.D. from Doctorate II

Married Total children Children  6  1 Cumulative employers Private university Research I Liberal arts I Medical school Primary work research Primary work teach Primary work manage Secondary work research Secondary work teach Secondary work manage Secondary work other Government support in current year Cumulative years of government support Cumulative papers Cumulative publications

Computer science / mathematics

Science

Life science

Physical science

Engineering

0.097 (0.065) 0.075 (0.070)

0.158** (0.084) 0.105 (0.090)

0.215 (0.172) 0.197 (0.180)

0.528** (0.071) 0.322 (0.240)

15 years after Ph.D. 0.014 0.005 (0.024) (0.033) 0.009 0.013 (0.009) (0.013) 0.029 0.039 (0.024) (0.033) 0.044*** 0.034*** (0.009) (0.011) 0.050** 0.055* (0.021) (0.030) 0.003 0.024 (0.022) (0.029) 0.108*** 0.099*** (0.025) (0.038) 0.107*** 0.085*** (0.024) (0.029) 0.055 0.042 (0.050) (0.058) 0.161*** 0.135** (0.049) (0.060) 0.231*** 0.203*** (0.054) (0.067) 0.002 0.063 (0.042) (0.066) 0.136*** 0.026 (0.046) (0.071) 0.101*** 0.024 (0.047) (0.073) 0.046 0.002 (0.051) (0.075) 0.010 0.004 (0.026) (0.034) 0.012** 0.007 (0.006) (0.007) 0.002** 0.000 (0.001) (0.001) 0.008*** 0.009*** (0.001) (0.001) Academic field 0.045 — (0.040) —

0.074 (0.040) 0.020 (0.016) 0.001 (0.041) 0.069*** (0.017) 0.047 (0.034) 0.005 (0.043) 0.150*** (0.038) 0.096* (0.056) 0.109 (0.116) 0.276*** (0.101) 0.291** (0.115) 0.019 (0.058) 0.264*** (0.072) 0.106 (0.069) 0.015 (0.079) 0.048 (0.045) 0.031*** (0.010) 0.002 (0.002) 0.008*** (0.002) 0.046 (0.046)

0.061 (0.097) 0.061** (0.032) 0.065 (0.083) 0.061* (0.037) 0.045 (0.070) 0.039 (0.074) 0.044 (0.092) 0.056 (0.100) 0.379** (0.168) 0.361* (0.190) 0.326* (0.153) 0.111 (0.164) 0.179 (0.170) 0.255 (0.148) 0.042 (0.197) 0.056 (0.087) 0.015 (0.019) 0.005** (0.003) 0.004 (0.003) — — (continued )

Table 5A.3

(continued)

Physics Chemistry Earth science Biology and life sciences Biochemistry Microbiology Zoology Health sciences Environmental science Agriculture and food Engineering Engineering Aerospace Chemical Civil Electrical Mechanical Industrial Observations Likelihood ratio stat

Science

Life science

0.115** (0.045) 0.139*** (0.041) 0.060 (0.047) 0.125*** (0.036) — — — — — — — — — — — — 0.022 (0.045) — — — — — — — — — — — — — 3,223 355.72

— — — — — — — — 0.081** (0.039) 0.005 (0.051) 0.121** (0.051) 0.105*** (0.038) 0.044 (0.091) 0.156*** (0.042) — — — — — — — — — — — — — — — 1,728 187.11

Physical science 0.073 (0.052) 0.083 (0.050) — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — 1,161 187.23

Source: 1973–2001 Survey of Doctorate Recipients. Notes: Coefficients report change in probability. Standard errors in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

Engineering — — — — — — — — — — — — — — — — — — — — — — — 0.112 (0.180) 0.129 (0.127) 0.028 (0.097) 0.029 (0.082) 0.181* (0.103) 0.001 (0.169) 330 66.27

Does Science Promote Women?

193

References Congressional Commission on the Advancement of Women and Minorities in Science, Engineering, and Technology (CAWMSET). 2000. Land of plenty: Diversity as America’s competitive edge in science, engineering, and technology. Report available at http://www.nsf.gov/publications/pub_summ.jsp?ods _key-cawmset0409. Freeman, R., E. Weinstein, E. Marincola, J. Rosenbaum, and F. Solomon. 2001. Competition and careers in biosciences. Science 294 (5550): 2293–4. Ginther, D. K. 2001. Does science discriminate against women? Evidence from academia 1973–1997. Federal Reserve Bank of Atlanta Working Paper 2001-02, February. ———. 2002. Gender differences in employment outcomes for academics in the social sciences. Working Paper. University of Kansas. ———. 2003. Is MIT the exception? Gender pay differentials in academic science. Bulletin of Science, Technology, and Society 23 (1): 21–26. ———. 2004. Why women earn less: Economic explanations for the gender salary gap in science. AWIS Magazine 33 (1): 6–10. Ginther, D. K. and K. J. Hayes. 1999. Salary and promotion differentials by gender for faculty in the humanities. American Economic Review, Papers and Proceedings 89 (2): 397–402. ———. 2003. Gender differences in salary and promotion for faculty in the humanities, 1977–1995. The Journal of Human Resources 38 (1): 34–73. Ginther, D. K. and S. Kahn. 2004. Women in economics: Moving up or falling off the academic career ladder? Journal of Economic Perspectives 18 (3): 193– 214. Goldberg, C. 1999. MIT acknowledges bias against female professors. The New York Times, March 23. Government Accountability Office (GAO). 2004. Women’s participation in the sciences has increased, but agencies need to do more to ensure compliance with title IX. Washington D.C.: GAO. Available at http://www.gao.gov/cgi-bin/ getrpt?-GAO-04-639. Kahn, S. 1993. Gender differences in academic career paths of economists. American Economic Review, Papers and Proceedings 83 (2): 52–56. Long, J. S., P. D. Allison, and R. McGinnis. 1993. Rank advancement in academic careers: Sex differences and the effects of productivity. American Sociological Review 58 (5): 703–22. Mitchell, S. B., R. Moonesinghe, and B. G. Cox. 1998. Using the survey of doctorate recipients in time-series analyses: 1989–1997. Working Paper. National Science Foundation, Washington, D.C. National Science Foundation (NSF). 2004a. Gender differences in the careers of academic scientists and engineers. NSF 04-323, Project Officer, Alan I. Rapoport. Arlington, VA: NSF. ———. 2004b. Women, minorities, and persons with disabilities in sciences and engineering: 2004. NSF 00-327. Arlington, VA: NSF. National Science Foundation (NSF), National Science Board. 2006. Science and engineering indicators (2006). Available at http://www.nsf.gov/statistics/seind06/ 2006. Nelson, D. J., and D.C. Rogers. 2005. A national analysis of diversity in science and engineering faculties at research universities. Available at http://cheminfo .chem.ou.edu/~djn/diversity/briefings/Diversity%20Report%20Final.pdf.

194

Donna K. Ginther and Shulamit Kahn

Preston, A. E. 2004. Leaving science: Occupational exit from scientific careers. New York: Russell Sage Foundation. Rosser, S. V. 2004. The science glass ceiling. New York: Routledge. Xie, Y., and K. A. Shauman. 2003. Women in science: Career processes and outcomes. Cambridge, MA: Harvard University Press.

6 Patterns of Male and Female Scientific Dissemination in Public and Private Science Kjersten Bunker Whittington

6.1 Introduction Information on the patenting and publishing activity of scientists and engineers has long been an interest among scholars of science and technology. Publishing transmits valuable knowledge and resources to other scientists, both in the academy and in industry, while patenting is thought to spur innovation through economic and proprietary incentives. Traditionally, scientists within academia have primarily published, shying away from pursuing economic ends through patenting or other marketable ventures, while industrial scientists have predominantly pursued commercial goals. Aided by federal and state promotion as well as university infrastructure, the organization of scientific research within universities and industrial firms has undergone a sea change in the past two decades. Academic scientists are now commonly involved in a variety of commercial activities, including patenting, licensing, start-up incubation, and firm founding, especially in the life sciences (Rosenburg and Nelson 1993; Cohen, Florida, and Goe 1994; Kleinman and Vallas 2001; Owen-Smith and Kjersten Bunker Whittington is an assistant professor of sociology at Reed College. This research is based upon work supported by a National Bureau for Economic Research (NBER) Dissertation Fellowship from the Science and Engineering Workforce Project, as well as an Association for Institutional Research (AIR) grant. Any opinions, findings, conclusions, or recommendations expressed in this publication are those of the author and do not necessarily reflect the views of NBER or AIR. Analyses in this work are conducted with restricted National Science Foundation SESTAT data, made available to researchers through the U.S. government (http://sestat.nsf.gov). The use of restricted data does not imply NSF endorsement of the research methods or conclusions contained in this report. I wish to thank Walter Powell, Jason Owen-Smith, Laurel Smith-Doerr, Michael Rosenfeld, Cecilia Ridgeway, Justine Tinkler, and Stefanie Mollborn for their helpful comments and feedback on this project. Any remaining errors are, of course, my own.

195

196

Kjersten Bunker Whittington

Powell 2001). At the same time, in some sectors there is much greater involvement in basic research by industry (Powell and Owen-Smith 1998). Research suggests that increases in academic-industry relations are shaping faculty careers in new directions, and altering the standards by which occupational success and reward are determined (Etzkowitz 1993; Packer and Webster 1995; Owen-Smith and Powell 2001). There is evidence that commercialization may be a new arena for disparities between men and women in scientific productivity. Recent work indicates that academic women are less likely to become involved in commercial activity than men (Whittington and Smith-Doerr 2005; Ding, Murray, and Stuart 2006; Whittington 2007; Whittington and SmithDoerr 2008). But past studies have yet to examine scientists from a variety of disciplines (most focus on the life sciences) or consider how men’s and women’s involvement in the joint activities of public and private science arise. In addition, much of the research on men and women in science leaves the actions and rationale of industry scientists unaddressed (but see Whittington [2007]). Given that the rate of women’s entrance in the industrial labor force has eclipsed that of men in the past few decades (Long 2001), now more than ever it is relevant to consider industrial scientists, and to compare the nature of sex disparities across sectors and work environments. I present a two-part analysis to address patterns of men’s and women’s dissemination in patenting and publishing activities (separately, together, or not at all) across sectors and disciplines. The first analysis uses loglinear modeling of a national sample of scientists and engineers to address the association between sex, discipline, employment sector, and involvement in scientific dissemination. I test the extent to which sex disparities in productivity are created and maintained by sorting mechanisms (i.e., the extent to which men and women scientists are differentially located in higher producing sectors or disciplines), as well as through organizational settings after controlling for sex distributions. In the second analysis I explore the ways in which various organizational contexts may differentially influence men and women scientists. I contrast two basic forms of organization—academic hierarchy, and what is sometimes called the “network form” of organization (Powell 1990; Podolny and Page 1998). I present network visualizations of coinventor collaborations between life science inventors working in the academy, public research organizations, and biotechnology firms, and address how the structure of science within each sector may contribute to sex disparities in productivity. Assessing the effects of organizational context on multiple forms of dissemination is of great importance as scholars begin to sort out the contemporary pushes, pulls, and constraints operating on women scientists in an era where commercial and academic science are much more closely linked.

Patterns of Male and Female Scientific Dissemination

197

6.2 Theoretical Background 6.2.1 Gender and Productivity Past research on scientific women documents their many structural and socialized constraints (for reviews of the literature, see Long and Fox [1995]; and Xie and Shauman [2003]). Women are less likely to participate in science, have less prestigious positions, and have received less recognition than men (Cole and Zuckerman 1984; Zuckerman 1991). Studies show that at all levels, there is a great disparity in the career attainments and opportunities of women scientists (Etzkowitz, Kemelgor, and Uzzi 2000). These disparities extend into accounts of research productivity as well, as women scientists have traditionally published less than men in the sciences (Cole 1979; Fox 1983; Cole and Zuckerman 1984; Long 2001; Xie and Shauman 2003). Given the current climate of science, it is important to consider sex disparities in commercial involvement as well as publishing. Research shows that women faculty do not sit on scientific advisory boards at the same rate as men scientists, and at the highest level of commercial involvement, they make up miniscule percentages of company founders (Ding, Murray, and Stuart 2006; Stuart and Ding 2006; Murray and Graham 2007). Sex disparities also exist in rates of academic patenting. Women faculty engage in patenting behavior at a decreased rate than male scientists, and produce less patents overall (Morgan, Kruytbosch, and Kannankutty 2001; Whittington and Smith-Doerr 2005; Whittington 2007; Whittington and SmithDoerr 2008). Patents are an increasingly available academic activity, and like publishing, can be an important signal of scientists’ research capabilities; thus a determinant of career outcomes. In addition, commercial involvement may bring academic scientists substantial increases in research funding, access to better research tools and equipment, potentially large gains in personal wealth, and an increased attractiveness to potential graduate students, postdocs, and other academic and industry collaborators. While in industry commercial output is often an expected productivity outcome, it remains a way for scientists to prove their company worth and value. Whether or not, and the degree to which, a scientist is active in publishing or patenting are particularly defining signals of his or her research and development goals and opportunities. Scientists are increasingly pursuing academic and industrial activities hand in hand. Stephan et al. (2007) find that publishing is positively related to patenting, thus one form of productivity does not seem to preclude the other (but see Owen-Smith and Powell [2001]; Bonaccorsi, Daraio, and Simar [2006]). In addition, if an academic life scientist’s colleagues, co-

198

Kjersten Bunker Whittington

authors, or department chair engage in commercial behavior, it significantly increases the chances that the scientist will as well (Stuart and Ding 2006; Bercovitz and Feldman Forthcoming). Research also suggests that patenting is strongly affected by the institutional environments where academic science takes place. Stephan et al. (2007) show that the number of patent applications is higher for those working in universities than house medical institutions and research institutes. Ding, Murray, and Stuart (2006) find evidence that university support for patenting plays an important role in scientists’ propensity to patent. Like publishing, the conclusion is that context matters. Researchers have proposed many explanations for the productivity differences in publishing between men and women, yet early research efforts that focused on individual status characteristics have been unable to fully account for variation in publication output (Zuckerman 1991; Long and Fox 1995; Ward and Grant 1995). Characterized most famously by Cole and Zuckerman (1987), this inequity has traditionally been referred to as the gender “productivity puzzle.” Much of the early research, however, fails to consider how resource distribution, job placement, and the structure of academic work is gendered. Organizational context likely plays an important role in gender equality, as successful scientific work relies on equal access to facilities and funds, available help, and a supportive research environment (Fox 1991, 2001). Indeed, in recent work, Xie and Shaumann (1998, 2003) are able to render much of the direct effects of sex on publication productivity insignificant by taking into account organizational positions and resources. They suggest that the traditional productivity puzzle should be replaced with a new puzzle to explain differences in resources and structural characteristics. The correlates and variations of the sex gap in patenting are not welldocumented. In addition, while numerous studies have contributed to our understanding of inequality among scientists, most have concentrated on either one type of science or another, combined similar sciences together (often in an attempt to increase the small percentages of women in their sample), or controlled for discipline or employment sector effects irrespective of their joint interaction with sex. Furthermore, few analyses have concentrated on how these contexts may affect men and women scientists differentially. Lastly, by focusing solely on faculty members, previous research neglects to address how the existing organization of academic life compares with that of other science and technology sectors. These issues are particularly important as there has been a slow but steady increase in the percentage of scientists working beyond the academic sector (Long 2001). 6.2.2 Gender, Scientific Dissemination, and Organizational Context Historically, industry has been seen as separate from and less prestigious than academia (Caplow and McGee 1961), yet it has also provided some

Patterns of Male and Female Scientific Dissemination

199

women with favorable workplace incentives (such as flexibility and higher mobility) not present in the academy (Aisenburg and Harrington 1988; Long and Fox 1995). In the past there has been speculation that women make a tradeoff in prestige for the slight advantages available in industry (Etzkowitz, Kemelgor, and Uzzi 2000). Recently, scholars are recognizing that careers outside the academy are growing in numbers and in prestige, as well as offering increased incentives for scientists, who choose to leave the ivory tower (Rabinow 1996). Whether this changing context has had a differential effect on men and women scientists is unclear, as measures of inequality among industrial scientists are lacking. However, Smith-Doerr (2004) shows that men and women tend to hold comparable management positions in industrial biotechnology firms. Past research by Long, Allison, and McGinnis (1993) shows a causal relationship between academic rank and dissemination, with higher-producing scientists receiving more returns to career advance. We might expect that similar processes may be acting upon industrial scientists, and as such, sex differences in involvement among them may be smaller than their academic counterparts. To the extent that sex differences in research productivity are the result of the different positions women hold rather than differences in capability or motivation, one can expect the sex gap in dissemination involvement to vary across sectors as well. The effects of location and context may also vary by the type of science, as resources and opportunities to publish or patent operate differently among disciplines. Commercialization in the physical sciences (comprised mostly of optics and solid-state applications) has more distant and less direct economic payoffs than the life sciences, where research results translate into new medicines with some urgency and for considerable profit. In addition, the public drive to finance and invest in physical science research is considerably less than that of the life sciences. Computer science faces a different issue. Its technology moves at a much faster pace in both discovery and development than other sciences, and the average three-year lag between filed and issued inventions can be problematic. Often computer science and similar fields resort to other methods of dissemination such as trade secrets, publishing, or copyright to preserve property rights and/or transmit new knowledge. Also relevant is the fact that the proportion of women across scientific fields varies. Women make up a much higher percentage of life and computer scientists than physical scientists and engineers. In this case, assessing the sorting of scientists across sectors and disciplines would provide useful information about the nature of sex disparities in productivity across sectors. In as much as patenting varies by sector and discipline, dissemination is also influenced by the norms and characteristics of the working environment. The broad contexts of academic and industrial work also likely have differing effects on men and women (Fox 1991, 2000). The analyses in this

200

Kjersten Bunker Whittington

chapter build upon my previous research, which finds sex disparities in commercialization among life scientists to vary across work settings (Whittington and Smith-Doerr 2005; Whittington 2007; Whittington and Smith-Doerr 2008). Using a combination of career history data and patenting information for a sample of life scientists across a period of two decades, this work finds that women engage in and produce less commercial work than men. The degree of disparity remains relatively constant across time. Importantly, sex differences in commercial involvement are greatest in academia, where the percentage of men involved in patenting more than doubles that of women. Although women patent less than men in industry as well, it is a much smaller difference.1 This previous research provides a starting place to address how dissemination varies across sectors, but has several implications for future work. First, the focus on a sample of life scientists leaves the broader context of discipline unaddressed. It would be useful to address the extent to which sector- and discipline-level sorting mechanisms account for sex disparities in productivity. Second, while the focus of this research puts emphasis on the degree to which academic scientists are patenting, it does not address the extent to which industrial scientists publish, as well as joint involvement in patenting and publishing. In both public and private science, men and women who are involved in these dual activities possess an ability to speak to multiple domains of science, and to apply complementary application to their research. In this work, I argue that it is important to know the extent to which men and women are participating in both patenting and publishing, across all sectors and disciplines. To this end, the first analysis addresses the relationship between organizational context and gender on a macro-level across three disciplines and three employment sectors. I present loglinear models of patenting and publishing activity using a nationally representative sample of doctoral recipients, for publishing as well as patenting. Specifically, I address the following questions: Does men and women scientists’ propensity to patent and publish vary by type of science and employment sector? Is there a sex gap after controlling for the distribution of men and women into lower- or higher-producing sciences and sectors? In sum, what is the association between sex, discipline, employment sector, and involvement in scientific dissemination? In the second analysis I explore the ways in which various organizational contexts may differentially influence men and women, highlighting the specific case of life scientists working in the Boston area, one of the most 1. However, among those involved in commercial activity, the quality and impact of women’s commercial work remains the same or better than that of male scientists (as measured by forward and backward patent citations).

Patterns of Male and Female Scientific Dissemination

201

active and fertile biotechnology clusters in the world (Owen-Smith and Powell 2004; Porter, Whittington, and Powell 2005). Using patenting data made available through the National Bureau of Economic Research (NBER) on U.S. inventors across time (Hall, Jaffe, and Trajtenberg 2001), I present network visualizations that depict inventor collaborations of university and industry life science inventors. I suggest how the structure of science within each sector may contribute to existing trends in sex differences in productivity. Combined, the two analyses have important implications for the effects of organizational context on men’s and women’s productivity. 6.3 Patterns of Publishing and Patenting Across Sectors and Disciplines 6.3.1 Data To examine the relationship between scientific dissemination and sex, type of science, and employment sector, I analyze data from the 1995 National Science Foundation’s Survey of Doctoral Recipients (SDR).2 The SDR incorporates a complex survey design that stratifies respondents by scientific discipline, employment sector, receipt of a doctoral degree, and certain demographic variables. When weighted, SDR data characterizes a nationally representative population of individuals trained and/or working as scientists or engineers between 1990 and 1995.3 The SDR includes scientists working across a variety of employment sectors and disciplines, making it a useful data set to address these research questions. For this analysis I focus only on scientists whose principal work responsibilities include research and development. The SDR asked respondents to indicate the primary and secondary work activities on which they “spend the most hours during a typical work week.”4 I restrict the sample of scientists to those who listed applied research, basic research, development, or design as their primary work activity. Limiting the sample to scientists who indicate that they spend the most time on research and development helps to reduce concern over unequal allocations of work activities

2. The SDR population consists of all individuals under the age of seventy-six who received a research doctorate in science or engineering from a United States institution prior to June 1994 and who resided in the United States as of April 1995. 3. All results and data presented here incorporate weighted sample statistics. 4. Scientists could choose from the following work activities: (a) accounting, (b) applied research, (c) basic research, (d) computer applications, programming, or system development, (e) development, (f) design of equipment, processes, or models, (g) employee relations including recruitment, training, or personal development, (h) managing and supervising, (i) production or operations, (j) professional services, (k) sales, purchasing, marketing, customer service, or public relations, (l) quality assurance, (m) teaching, or (n) other.

202

Kjersten Bunker Whittington

among women and men in similar positions (for example, gender variations in teaching loads or committees, management responsibilities, etc.).5 6.3.2 Measures Dissemination activity is measured by combining responses to two survey questions. Scientists were asked whether or not they had been named as an inventor on (a) a U.S. patent application and (b) as an author or coauthor on a peer-reviewed published paper, in the past five years. These two variables were combined to make one variable with four categories: (a) respondent neither published nor patented, (b) respondent published, but did not patent, (c) respondent patented, but did not publish, and (d) respondent both published and patented.6 Scientists are classified by what NSF terms “major employment sector,” that is, two-year colleges, four-year colleges, government, or business/industry. Institutions designated as four-year colleges include baccalaureate and master’s institutions, and Research I and II universities. The industrial sector includes private, for-profit companies, as well as scientists who are self-employed. Most scientists within the government sector are federal workers; however, state and local government scientists are included as well. For the purposes of this research, however, government is included largely as a control. In addition, scientists are placed in one of six disciplines according to the type of science they perform in their current job. These categories are computer and mathematical sciences, life sciences, physical sciences, social sciences, engineering, and nonscientific occupations. For this analysis, I exclude scientists who work at two-year colleges, as well as those located in computer sciences, the social sciences, or in nonscience or nonengineering disciplines. Theoretically, I am mainly concerned with scientific occupations that produce research that is potentially patentable as well as publishable. Although publications may be common, nonscientific occupations and the social sciences are not oriented toward commercialization in the same way other sciences are (for example, only 0.4 percent 5. In addition, only scientists who are working full time are included in the sample. Parttime scientists may not have an equal opportunity to publish and patent to the same degree as compared to their full-time counterparts. The percentage of scientists who work part time while engaging in research and development as a primary work activity is small (3.6 percent of the sample (N  636). Models run with and without part-time scientists do not differ qualitatively. All coefficients have the same sign and significance, and magnitudes are negligibly different. 6. Across and within sectors, disciplines vary in their propensity to publish and patent. Using a measure of publications and patents accrued over a period of five years may seem stringent for some fields and too lengthy for others. Because this analysis controls for distributions across discipline and sector, however, it is possible to compare scientists within fields and sectors without having to choose a global average across fields.

Patterns of Male and Female Scientific Dissemination

203

[N  7] of social scientists in the sample patent).7 In computer science, the rate by which innovation moves often means that the lag between filed and issued patents will outlive the novelty of the invention. As such, patenting is a rarity for scientists in this discipline and data set (approximately 2 percent [N  15] of the patenting academic sample). Once all variables and constraints are taken into account, the final sample for this analysis consists of 10,144 scientists, 16.5 percent of whom are women. There are no missing data. 6.3.3 Survey Statistics Table 6.1 presents weighted summary statistics on the sample’s participation in publishing and patenting activities, broken down by employment sector. The table shows that scientists vary in their propensity to publish and patent. Involvement in commercial activity remains an activity pursued by a minority of the scientific population. The majority of all scientists, women and men, tend to have only published, if at all, between 1990 and 1995. Approximately 25 percent of the national population of scientific doctorates indicated that they had patented between 1990 and 1995. This percentage is heavily skewed by the higher rates of industrial patenting, however. About 14 percent of university scientists had patented in the period leading up to 1995, as compared with 39 percent of industrial scientists.8 As table 6.1 shows, men and women have similar preferences for patenting and publishing, although women’s participation differs from that of men in a few notable areas. First, a higher percentage of women than men only publish. This is true in both academic and industrial settings, although industrial differences are larger. In addition, the 1995 SDR data echoes previous findings of commercial participation, which show that women participate less in patenting than men. Approximately 13 percent of women as compared with 21 percent of men patent. The proportion of academic men involved in patenting is about twice that of women academic scientists. In7. However, because women represent a greater percentage of social and behavioral scientists than they do in the natural, physical, and engineering sciences, it is important to know how the exclusion of this group may bias the analysis. The majority of social scientists publish only or disseminate nothing (54.5 percent and 44.7 percent, respectively), and differences between men and women are small (although higher in academia than industry). Because very few members of these groups have a propensity to patent, many cells in the contingency table are small in count or have no activity at all. This sparseness ultimately causes methodological problems in the stability of the loglinear models that cannot be resolved. While instable, models run with and without social scientists yield coefficients similar in sign, magnitude, and significance, and the substantive findings of this work remain the same. 8. Although not in the table, the three disciplines also vary somewhat in their propensity to publish and patent. Engineers are more likely than life or physical scientists to patent in the university (24 percent versus 13 percent and 12 percent, respectively), and physical scientists are slightly more likely to patent in industry than engineers and life scientists (46 percent versus 38 percent and 31 percent, respectively). Tables available from the author upon request.

7.5

15.1

100.0

56,554 3,624

Total

Weighted N Unweighted N

71,783 4,715

100.0

13.5

4.0 82.0 0.5

Percent total

— —

0.21

0.12

0.19 0.23 0.09

Proportion women

60,493 3,698

100.0

28.7

21.3 38.5 11.5

Percent men

8,263 620

100.0

24.5

18.6 50.4 6.5

Percent women

68,756 4,318

100.0

28.2

21.0 39.9 10.9

Percent total

Industry

— —

.12

.10

.11 .15 .07

Proportion women

131,475 8,225

100.0

21.3

12.7 60.3 5.6

Percent men

25,897 1,919

100.0

13.1

9.3 75.4 2.2

Percent women

Note: This table reflects statistics from the weighted sample. Total N(unweighted)  10,144

Total N(weighted)  157,372

157,372 10,144

100.0

20.0

12.2 62.8 5.1

Percent total

— —

.17

.11

.13 .20 .07

Proportion women

Total (includes government scientists)

Source: National Science Foundation/Division of Science Resources Statistics. Scientists and Engineers Statistical Data System (SESTAT), 1995.

15,229 1,091

100.0

3.8 88.5 0.2

Percent women

4.1 80.2 0.6

Percent men

Academia

Percent of women and men involved in dissemination activity, by employment sector (1990–1995)

Neither publish nor patent Publish only Patent only Both publish and patent

Dissemination activity

Table 6.1

Patterns of Male and Female Scientific Dissemination

205

dustrial men and women scientists, on the other hand, appear to be more similar regarding commercialization. Women are only slightly less likely to participate in both patenting and publishing activities (25 percent versus 29 percent, respectively). The sector-level disparity in patenting involvement is roughly 50 percent higher in academia than in industry. Similar sector-level differences are seen when conceptualized as probability differences. The academic difference in probabilities between men and women is .076, compared to an industrial probability difference of .042, which yields a 45 percent probability difference between the sectors. The previous descriptive statistics suggest the need to further investigate the effects of men’s and women’s locations on dissemination activity. While women are less likely to engage in commercial behavior, once sorted by sector the dissemination trends among men and women scientists diverge. What effect does location have on women’s dissemination activity once the distribution of men and women across disciplines is addressed? Moreover, how do documented sex disparities in dissemination change after accounting for location and organizational context? I turn to these questions in the following section. 6.3.4 Methodology To study publishing and patenting behavior by sex, discipline, and employment sector, I use loglinear models that identify the associations among these variables independent of the marginal distributions of men and women across sciences and sectors. The primary objective of loglinear analysis is to determine if the distribution of counts among the cells of a table have an underlying structure. With the case of sex and dissemination, we can make several predictions about how the distribution of women and men in sciences and sectors may shape the degree to which women participate in various scientific dissemination activities. This methodology is both useful and necessary for this type of analysis because it permits the modeling of relationships between two or more categorical variables. Rather than looking solely at the effects of each variable on a single outcome, this method is especially practical for this analysis because it allows for the complexities of association linkages among all of the variables. Thus, loglinear modeling accounts for the interrelationships between science, sector, sex, and dissemination. Importantly, this method is also able to estimate the effects of multiple-order interactions in conjunction with and controlling for the effects of other multiple-order interactions.9 To examine these relationships, the scientific dissemination table was 9. One possible reason for the lack of research addressing joint program and sector effects on gender and dissemination is the methodological difficulty of multiple third-order interactions. Loglinear analysis provides a way to account for these multi-level interactions simultaneously.

206

Kjersten Bunker Whittington

cross-classified by sex, employment sector, and discipline. There are 72 cells in the data set (4 [dissemination]  2 [sex]  3 [employment sector]  3 [discipline]). Table 6.2 presents the seven nested models in this analysis. In simplified hierarchical terms, they can be presented as follows: (1) Ln (U )  Constant  Error (2) Ln (U )  Model 1  patpub  sex  discipline  empsectr (3) Ln (U )  Model 2  patpub  discipline  patpub  empsectr  discipline  empsectr  patpub  discipline  empsectr (4) Ln (U )  Model 3  sex  patpub (5) Ln (U )  Model 4  sex  empsectr  discipline (6) Ln (U )  Model 5  sex  patpub  discipline  sex  patpub  empsectr (7) Ln (U )  sex  patpub  empsectr  discipline where only the highest-order terms are listed, and the lower-order terms are assumed. Here, U is the predicted number of scientists, discipline is the scientific discipline, empsectr is the employment sector, patpub is patenting and publishing activity, and sex is the whether the respondent is a man or a woman. While the loglinear model framework may seem foreign, logistic regression equivalency is possible for some loglinear models depending on the relationship between the dependent model and the predictor terms. For example, the multinomial equivalencies for the final two models above take the following form: (6) Multinomial Logit(patpub)  Constant  sex  discipline  empsectr (7) Multinomial Logit(patpub)  Constant  sex  empsectr  discipline where only the highest-order terms are listed, and the lower-order terms are assumed. Because loglinear methodology does not assume a specific dependent variable but rather predicts cross-classified cell counts, it gives more flexibility to explore the relationships between all and various combinations of variables in the model. 6.3.5 Gendered Dissemination across Sectors and Disciplines Table 6.2 presents the results from a series of loglinear models, in which the associations between gender, discipline, employment sector, and dissemination activity are estimated from the data. Models 1 and 2 in table 6.2 account for the constant-only model and the direct effects of the variables. The direct effects model accounts for dissemination, sex, type of science, and employment sector without making any assumptions about the relationships between the variables. The likelihood-ratio (LR) chi-square for model 2 is 4,903 with 63 degrees of freedom. As is to be expected, this model

Patterns of Male and Female Scientific Dissemination Table 6.2

207

Loglinear models of scientific dissemination, sex, discipline, and employment sector, 1990–1995

Model description 1. Constant only 2. Model 1  direct effects 3. Model 2  Pat/pub activity * Employment sector  Pat/pub activity * Scientific discipline  Scientific discipline * Employment sector  Pat/pub activity * Employment sector * Scientific discipline 4. Model 3  Sex * Pat/pub activity 5. Model 4  Sex * Scientific discipline * Employment sector 6. Model 5  Sex * Pat/pub activity * Scientific discipline  Sex * Pat/pub activity * Employment sector 7. Model 6  Sex * Pat/pub activity * Scientific discipline * Employment sector (saturated model)

Terms in the model

Residual df

Goodness of fit

1 9

71 63

22,686.8 4,903.2

0.0 0.0

22,031.8 4,322.1

37

35

1,085.4

0.0

762.5

40

32

891.3

0.0

596.1

48

24

43.7

0.001

177.7

60

12

24.8

0.02

85.9

72

0

n.a.

LR p-value

n.a.

BIC

n.a.

Source: National Science Foundation/Division of Science Resources Statistics. Scientists and Engineers Statistical Data System (SESTAT), 1995. Notes: df  degrees of freedom; LR  likelihood-ratio; BIC  Bayesian Information Criterion; n.a.  not applicable.

fits very poorly by both the likelihood ratio test and by the Bayesian Information Criterion (BIC).10 This is not surprising because model 2 makes the unlikely assumption that scientists are distributed evenly across sectors and programs, and disseminate equally without effects from these areas. The third model adds twenty-eight terms to account for the effects of differential dissemination rates across disciplines and employment sectors, and the distribution of scientists across combinations of fields and sectors. Model 3 considerably reduces the goodness-of-fit chi-square from 4,903 to 1,085, but it still does not pass the likelihood ratio test or meet the BIC criterion. Model 4 adds three more terms to the existing model to account for 10. The BIC calculates the goodness of fit in terms of how the data compares to the saturated model. When the BIC is negative, it is considered better than the saturated model (Raftery 1986). Although generally accepted in the literature as a satisfactory measure of a model’s goodness of fit, the BIC measure is considerably less stringent than the likelihood ratio test (Weakliem 1999). For this reason, the results of both tests are reported for each of the models.

208

Kjersten Bunker Whittington

the effects of gender on dissemination. This additional factor reduces the goodness-of-fit chi-square (from 1,085 to 891), and significantly improves the fit of the model, but it is not yet a well-fitting model by either LR or BIC criteria. This improvement indicates the presence of differential dissemination trends among the sexes. Model 5 includes eight additional terms for the three-way interaction between gender, employment sector, and discipline, which accounts for the distribution of women across sectors and sciences. The addition of these terms drastically reduces the goodness-of-fit chi-square, from 891 to 44. With a p-value of 0.001, this model’s significance gets closer to, but not above, the 5 percent probability threshold needed to reject the LR test. Model 5 is the first model that achieves a good fit by the BIC criterion (–177.7). The significant increase in fit indicates the importance of accounting for the different locations of men and women when analyzing dissemination. The significance of model 5 suggests that aside from the uneven distribution of scientists across sectors and disciplines (model 3), women scientists have their own unique pattern of location. The tremendous improvement in fit lends powerful support to the sorting mechanism as a primary way in which gender stratification takes place. Apart from sex distribution, however, disciplines and sectors may have differential effects on men’s and women’s dissemination. The next two models test whether or not sex disparities change across disciplines or sectors. Model 6 adds two three-way interactions between sex, dissemination, and discipline, and sex, dissemination, and employment sector. This model improves on the previous one with a goodness-of-fit chi-square of 24.8 and 12 degrees of freedom. The significant improvement of this model over the last suggests that there is a differential sex effect operating among disciplines and sectors (to be discussed in more detail later). The model’s p-value, however, is 0.02, slightly below the 0.05 threshold needed to reject the LR test. This result indicates that the saturated model—one that includes associations between all variables in the model—is most appropriate for this data.11 Model 7 incorporates the saturated model by including an interaction between dissemination, sex, discipline, and employment sector. Model 7 is the final and best-fitting model of this analysis, indicating that men and women disseminate differently across sectors and disciplines, other things being equal.12 Model 7 also documents the persistent presence 11. The significance of the saturated model indicates that a unique relationship exists between all four variables in the analysis, rather than an overarching simpler trend based on combinations of the four. In laymen’s terms, this means that sex and dissemination likely vary by discipline and employment sector, concurrently, rather than just discipline or sector alone. 12. Although model 5 is best-fitting by BIC standards, I choose the best-fitting model by the more stringent likelihood ratio test. I make this decision primarily because the BIC criterion, in which models with BIC less than zero are preferred to the saturated model, is far easier to satisfy than the more stringent likelihood ratio test of model fit. In addition, there have been some critiques of the BIC’s ability to correctly take into account the sample size of the hypotheses in question (Weakliem 1999).

Patterns of Male and Female Scientific Dissemination

209

of direct sex effects on dissemination. Interaction terms between sex and dissemination activity remain significant in the final model despite accounting for the distribution of women across sciences (model 5) and gendered employment sector and discipline effects on dissemination (models 6 and 7). Changes in the coefficients for the interaction between sex and dissemination across models are telling. Before controlling for the distribution of women across sectors and sciences, women are more likely to publish than men, and only slightly less likely to engage in dual dissemination. The final model portrays a very different story, one in which women are at a clear disadvantage. The fact that the coefficients become more negative with the addition of sex distribution controls suggests that women tend to be located in sciences and sectors that are less disadvantaging. When controlling for this distribution, however, we see that independent of sector and science, women experience disparities between themselves and their male counterparts. The final results indicate that the propensity to publish and patent varies with sex as well as discipline and employment sector. Including interactions between (a) sex and dissemination, (b) sex, discipline, and employment sector, and (c) sex, employment sector, and dissemination greatly improve the fit of the model. The models show that the sorting of women across disciplines and sectors is one primary way to account for overall sex disparity in dissemination. Despite accounting for the distribution of scientists across fields, however, the location of scientists alone does not explain patenting and publishing differences between men and women scientists. Furthermore, sex effects on dissemination operate uniquely within disciplines and sectors, and the best fitting model includes interactions between all four variables. The final model coefficients suggest substantial discipline, sector, and sex effects on dissemination, yet further analysis is needed to discern the relationship among all interactions combined. From these results, it is possible to construct comparisons of log odds across and within groups to investigate the implications of the final model. Table 6.3 presents the log odds and odds ratios of dissemination between men and women across academia and industry, broken down by discipline. 6.3.6 Odds Ratios The raw percentages in the earlier descriptive statistics suggest that gender disparities may vary by employment sector. In particular, industry settings appear to be more gender equal with regard to both patenting and publishing, and industrial women appear to engage in publishing behavior more than industrial men. The loglinear models echo this finding, yet with an important caveat—only for the life sciences. Due to significant, positive interactions between industry, sex, and dissemination, women industrial scientists in the life sciences are equally as likely to publish as men scien-

210

Kjersten Bunker Whittington

Table 6.3

Odds ratios for men and women scientists by employment sector, holding discipline constant Life Sciences Academia

Publishing onlya Men Women Relative log odds (M-F) Relative odds Patenting onlya Men Women Relative log odds (M-F) Relative odds Both publishing and patentinga Men Women Relative log odds (M-F) Relative odds

3.705 3.122

Industry

Engineering Academia

Industry

Physical sciences Academia

Industry

1.430 1.679

2.272 5.752

0.066 0.038

2.584 3.139

0.777 0.626

0.249 0.780

3.480 0.030

0.028 1.028

0.555 0.574

0.151 1.160

0.950 1.110

3.275 1.349

2.112 2.917

3.275 3.575

0.253 0.796

1.622** 5.063**

0.160 1.174

0.146 0.146

0.300 1.350

0.543 1.721

1.926 0.647

0.730 0.784

1.181 3.925

0.196 0.533

0.705 0.388

0.741 0.141

0.054 0.950

2.744 0.060

0.336 1.399

0.317 1.370

0.600** 1.820**

0.583*** 1.791*** 1.187 2.809

1.279*** 3.593***

0.805** 2.237**

Source: National Science Foundation/Division of Science Resources Statistics, Scientists and Engineers Statistical Data System, 1995. a “Publishing only,” “Patenting only,” and “Both publishing and patenting” have “Neither publishing nor patenting” as a comparison group. ***Significant at the 1 percent level (two-tailed). **Significant at the 5 percent level. *Significant at the 10 percent level.

tists. Table 6.3 shows that life science men in academia have approximately 3.6 times higher odds of publishing and patenting than life science women academics ( p  .001), but there is no disparity between men and women life scientists in industry. Thus, the models show significant gender disparities among life scientists in the academic sector, whereas men and women industrial life scientists disseminate equally. Table 6.3 also shows that uniform employment sector effects are not apparent in the other disciplines. Men and women engineering and physical scientists in the academic setting do not exhibit statistically significant sex differences in involvement. The results are similar for industrial engineers, but male physical scientists in industry are approximately 1.8 times more likely to patent and publish than female physical scientists in industry. Thus, it appears that organizational context plays a significant role in addressing the sex disparity seen in the descriptive statistics in these models.

Patterns of Male and Female Scientific Dissemination

211

In particular, the trends seen in the raw data stem from those exhibited by scientists in the life sciences. Engineers are statistically gender equal with regard to involvement in dissemination activities, and only in industry do men and women physical scientists differ with regard to dual dissemination outcomes. 6.3.7 Implications of the Survey Models The national sample sheds light on the degree to which the sector-level differences seen in my previous research extend beyond the life sciences to other disciplines and sectors. At least on a macro-scale, broad-based sector effects appear to be most prominent for scientists in the life sciences. Why do sex disparities in the life sciences vary so clearly across sector lines? Or conversely, what factors in the academic and industrial settings of engineering, for example, maintain a similar level of sex disparity across sectors? One way in which the life sciences differ from the physical sciences and engineering is in the proportion of women scientists working in the field. The life sciences have the highest female composition of the science and engineering disciplines in both industry and academia. Previous research has heralded the inroads women have been able to make in the life sciences, and presented it as one of the more women-friendly disciplines. In contrast, women continue to remain a very low proportion of physical scientists and engineers. Whereas the proportion of women in this sample in the life sciences is 27 percent, they make up only 10 percent and 5 percent of the physical sciences and engineering, respectively. The relative equality of men and women in sciences of extremely low female proportion may reflect that, either by choice or necessity, women tend to more closely resemble the men that dictate the norms of their working environments. I suggest that an additional factor may be important to consider when looking at large scale sector-level effects—organizational form. The structure of industrial firms differs from that of academic settings, and industry context may vary across disciplines as well. Organizational work settings within an industrial sector are not necessarily comparable; they may vary in terms of whether they are more hierarchical or of a “network form” (Powell 1990; Podolny and Page 1998; Smith-Doerr and Powell 2005). Hierarchical companies tend to focus in-house, while network firms engage in durable, yet flexible, ties with external partners. Even the largest network firms are highly relational organizations, embedded in a variety of interorganizational relationships with a diverse mixture of organizational forms—universities, public research institutes, dedicated science-based companies, and large, multinational organizations (Powell, Koput, and Smith-Doerr 1996). In addition, whereas the organization of work in hierarchal companies is bureaucratic and rigidly arranged, network firms exhibit a more horizontal relational organizational structure among scientists.

212

Kjersten Bunker Whittington

Of relevance to this issue is the difference between small firms across disciplines. The life science industry, in particular, is notable for its abundance of small, dedicated biotechnology firms (DBFs)—research-intensive organizations primarily concentrating on genetic engineering and molecular biology for human therapeutic and diagnostic applications (Powell, Koput, and Smith-Doerr 1996). Increasingly, life scientists in industry must decide whether to work for large diversified pharmaceutical corporations or smaller start-up DBF organizations. Biotechnology firms utilize small numbers of employees to leverage connections with other organizations, facilitate the transfer of basic science into new medicines and research tools, and build research connections and alliances with other organizations. For many scientists, the biotechnology setting allows its employees a considerable amount of freedom regarding publishing and basic science activities, and in doing so, facilitates an atmosphere that resembles academic autonomy (Carre and Rayman 1999). Small life science firms, in particular, depend on the development of basic research for new ventures. Smith-Doerr’s research (2004) speaks to the relevance of gender in these two very different industrial work settings. She finds equality in the types of management positions men and women hold in small DBF organizations, but finds greater sex disparities among those in corporate labs and in academia. My additional research with Smith-Doerr using the same sample of life scientists also shows evidence that women perform especially well in these science-based types of organizations (Whittington and SmithDoerr 2008). Women life scientists in industry are equally likely to become involved in patenting activity as industrial men, while male academics are over twice as likely to patent as female life science academics. Importantly, however, the models suggest that this sector difference is limited by type of industrial organization—that is, only in small, dedicated biotechnology firms. Life scientists working in industry for large, multinational companies are not privy to this industry advantage. Thus, the industry effect in this sample mimics that of this previous work. The commercial activity of women scientists located in dedicated life science startups may be driving the sector-level differences seen in the loglinear models. Clearly there is something unique about industrial DBF work settings. If the culture and organization of scientific work across these varied work settings matters for predicting sex disparities in productivity, the structure of science within smaller, dedicated biotechnology firms may operate as an opportunity structure for sex equality in dissemination. In the next section, I present network visualizations of academic and industrial inventors in biotechnology to examine the structure of science across these two work settings. This structure contains clues to the ways in which the network form may result in a more equitable environment for women scientists. I compile connections among coinventors through U.S. patent activity across a period of two decades, and present network graphs of the scien-

Patterns of Male and Female Scientific Dissemination

213

tific community in both sectors. Investigating collaboration networks provides visual strength to otherwise intangible differences in sector-level structure, and allows for inferences to be made about the degree to which women and men may be affected by critical differences in organizational context. In addition, collaboration networks can address how the structure of informal personal relationships among inventors speaks to the broader arrangement of the production of knowledge across biotechnology firms and the academy. 6.4 Networks of Collaborations in Public and Private Life Science 6.4.1 Data The network data consist of inventor-level information from United States patents filed between 1976 and 2002 from the academic and industrial sectors. A list of 482 public and private biotechnology firms from Powell et al. (2005) provided the industry sample of scientists. Powell and colleagues collected the data on firms and firm networks from Bioscan, an industry publication. Bioscan includes nearly the entire population of biotechnology firms in existence between 1988 and 2002 and thus provides a representative sample of industry activity. The academic sample was drawn from Research I universities in the United States. Academic and firm-level data are matched with patent information extracted from the U.S. Patent and Trademark Office (USPTO) database. I limit the collaboration network to firms and universities in the Boston region. I do so to provide a natural boundary on the scope of the network so as to maintain a manageable size for which to analyze the data. The Boston region is one of the top three areas of regional biotechnology development in the United States. Boston is unique in the sense that, in conjunction with biotechnology firms, its university activity and the activity of other public research organizations (for example, Dana Farber Cancer Institute and Massachusetts Eye and Ear) play a significant role in driving regional biotechnology innovation (Owen-Smith and Powell 2004; Porter, Whittington, and Powell 2005).13 For this reason it is useful to analyze the Boston area because of the jointly significant roles that universities and private firms play.14 To ensure patent comparability across sector samples, the academic 13. In addition to firm and university inventors, I include collaboration networks of those in public research institutions. Although not the focus of this chapter, the significant role such networks play in the Boston region suggests a more accurate picture of the joint firmuniversity network is achieved with their inclusion. 14. There is a noticeable absence of large multidivisional corporations, such as pharmaceutical companies, in all major biotechnology regions in the United States, including Boston, during this time period (Owen-Smith and Powell 2004). Their lack of presence in the Boston network makes it difficult to ascertain the structure of inventor collaborations in large,

214

Kjersten Bunker Whittington

database is limited to patents in the biotechnology sector. Suitable patent classes were chosen by limiting the university sample to the subclasses that account for the ninetieth percentile and below of life science firm patenting.15 In addition to including academic scientists that always patent in biotechnology classes (i.e., scientists for whom all of the patents in their portfolio fall within core biotechnology subclasses), I include all academic scientists that sometimes do (i.e., scientists for whom at least one of the patents in their portfolio is assigned to a core biotechnology subclass). This allows me to make comparisons across scientists who are located in a similar world of science, while limiting unnecessary truncation of scientists’ collaborative communities. I collected inventor names from all patents granted to Boston firms, universities, and public research organizations. Multiple inventions by the same person also involve the confirmation of similar names. Inventions are considered to be from the same person when two inventors match in first, middle, and last name (or part thereof, in the case of missing middle or first names). Importantly, however, two names are only considered a match if they have similar first, middle, and last names and a similar city and state, assignee name, or the same primary and secondary technology class.16 It is not possible to obtain from the USPTO information regarding any previous last names an inventor may have had (i.e., maiden names or names changed legally for other reasons), and background searching the list of more than 15,000 name records is not possible. The lack of this information may undercount the patents of women scientists, as they are more likely to change their names than men scientists. My name-matching algorithm addresses this by flagging records for hand coding that match on other criteria but not last name (i.e., combinations of first and/or middle name, assignee, patent class, and/or town). In addition, all hyphenated surnames have been hand coded to search for matches on either name. About 7 percent of the records are hand coded due to flag issues that fall under one of two types—those that are deemed a match yet may not be, or those that are not deemed a match yet may be (name changes fall in this latter type). Flagged records are then hand coded using available information from a variety of publicly searchable sources.17 hierarchically-oriented firms such as these. Novartis and Pfizer, however, both moved their R&D facilities into the Boston area since 2002, no doubt attempting to anticipate benefits from a closer location to such a dominant biotech regional economy. Although not included here, work currently in progress includes these organizations to enable a more fine-grained observation of the differing organizational contexts of industrial work settings. 15. It happens to be the case that these same classes hold the majority of the top patenting classes in university patents. 16. See Whittington (2007) for a more detailed description of the name-matching algorithm. 17. Most record decisions fall into agreement with the algorithm, and the remainder benefit from the hand-coding. This process presents a useful way to manage this data absent the presence of a unique identifier for individuals from the USPTO.

Patterns of Male and Female Scientific Dissemination

215

Inventor gender is coded through the assistance of an algorithm that determines gender based on first name comparisons with a list of names (broken out by gender) from the ninetieth percentile and below of the 1990 U.S. Census. This list contains not only 90 percent of the most common names in 1990, but the cumulative percentage of the U.S. population with each name. In the case where a name appears on both the male and the female list, the cumulative percent for the name in each sex is compared. Names in dispute that are above the seventieth percentile are considered rare, and sex is assigned to the more common sex.18 Decisions are not made about androgynous names (i.e., names found on both the male and female list) where a cumulative percentage is not given or jointly rare. There is a tendency in the name-matching algorithm to find matches for a greater percentage of scientists with a typical “American” name, as many foreignbased names are not present in the ninetieth percentile of the U.S. population. Much of the missing data (around 17 percent) lies in names of Indian and Asian descent, both of which are difficult to code for sex.19 The missing data is still included (labeled “Sex Unknown”) in the network to maintain its integrity. In the final sample there are sixty-three firms, four universities (Harvard, Tufts, Boston University, and MIT), and thirteen public research organizations (other than the universities) in the Boston area between 1976 and 2002. There are 4,994 inventors who have participated in a total of 5,598 patents from 1976 to 2002; 1,921 inventors have been granted 1,995 patents in biotechnology firms, and 1,174 inventors have been granted 1,246 patents in Boston universities. The Boston firm and university samples contain 26 percent men (N  411) and 20 percent women (N  199), respectively. The proportion of women biotechnology scientists is greater than that in academia or public research organizations (PROs) for each year in the network, although the gap is changing over time as the proportion of women grow in academic and PRO settings. The average yearly percentage female in biotechnology is 24 percent, versus 17 percent in both academia and PROs, respectively. 6.4.2 Constructing the Network Sample The data for this analysis represent two-mode affiliation data, where the inventors are the actors and each patent is the event. In this way, a connection between actors is assumed strictly by their collaboration activity. The 18. When possible, names not on the 1990 census are coded using secondary name data, found in books and websites with downloadable lists of ethnic and foreign names. 19. In their complexity, Indian names are hard to compile into lists of common names. Asian names tend to be androgynous when spelled in English, where gender is typically relayed in tone and written character in those languages. Presumably these missing names contain a similar percentage of men and women to the matched data, although that itself is not knowable.

216

Kjersten Bunker Whittington

affiliation network, when multiplied with its transpose, produces a onemode actor-by-actor network. I refer solely to this one-mode inventorinventor network in this analysis. I focus my analysis of the patent collaboration network on the largest, weakly connected component in a network, which follows the strategy of others in this area (White and Harary 2001; Moody and White 2003). This structure, the main component, represents the greatest concentration of coinventors, and the largest hub of patenting collaboration in the Boston region.20 I present network visualizations through the network program Pajek. Pajek produces network images by applying energy-based physical science algorithms that act to minimize node strain based on the inventor ties across the network. I create network images using Pajek’s Kamada and Kawai (1989) and Fruchterman and Reingold (1991) energy algorithms.21 This two-step process is done first to map and spread the structure of the network, and then to optimize that spread according to the constraints inherent in the node relationships.22 6.4.3 Academic and Industry Patenting Collaboration Figure 6.1 is a visualization of the main component of the copatenting network, with 2,371 inventors aggregated across time (1976 to 2002). Circles represent university inventors, squares are biotechnology firms, and triangles are public research organizations. Sex is coded by color— black nodes are women inventors, gray nodes are men inventors, and white nodes are scientists of unknown sex. The visualization of the combined network reflects the integration of academic and industrial collaboration networks in the Boston region. The expansive pump of activity in the center of the network represents the commercial side of a lab run by academic Dr. Robert Langer, an MIT professor in the Department of Chemical Engineering who does work at the interface of biotechnology and material science. Robert Langer is unique in the sense that he is world-renowned for his science, as well as his commercial engagement. Forbes, Bio World, Discover, and Time magazines have all independently ranked Langer as a top individual in biotechnology in the United States and the world. His close collaborators are from both university and industry settings, as well as a select few scientists who have moved from MIT to companies, and scientists in a handful of PROs (for more detail on Langer’s scientific advisory board and founding connections, see Porter, Whittington, and Powell [2005]). Spreading out from Langer are 20. Visualizations of the main component omit all inventors with no collaboration ties to others, and smaller clusters of collaborators that do not connect to the largest cluster. In this network, there are very few small clusters of collaborators besides the main component. 21. For more information on the algorithms or their use for visualization, see http:// vlado.fmf.unilj.si/pub/networks/pajek. 22. All networks are optimized to account for the number of coinventor ties between scientists.

Patterns of Male and Female Scientific Dissemination

Fig. 6.1

217

Boston patent coinventor network, main component (1976–2002)

Notes: Circles = University Inventors; Squares = Biotechnology Firms; Triangles = Public Research Organizations.

students, postdocs, and technicians in his lab, other clusters of university lab activity, and several firm and public research organization clusters connected through cross-sector inventors. Though not the immediate focus of this research, the combined network shows the important role that multivocality plays in linking the two sectors together (Padgett and Ansell 1993; Powell et al. 2005). A select group of 379 (16 percent of the network) scientists have moved across sectors over time or have patents jointly assigned to universities, industry, or public research organizations. These inventors are represented with diamonds. The main component would be vastly unconnected without these inventors. Upon their removal from the network, it essentially dissolves and breaks into almost 200 different clusters—the largest involving only 191 scientists. These scientists are translators in a dual sense—they are familiar with the mores of both university science and science-based companies, and their research translates from the laboratory bench to clinical treatment. In this analysis, I suggest that differences in the structure of collaborative relations in academia and industry have implications for the ways in which sector-level gender disparities in commercial activity arise. In particular, I focus on the degree of hierarchy in each sector, and discuss (a) how

218

Kjersten Bunker Whittington

the level of hierarchy in an organizational work setting may arise from sector-level goals and supply-side demands of academia and industry, and (b) how these goals and demands interact with men and women scientists embedded in the structure to influence dissemination patterns. Breaking the full network down by sector, there are distinct clusters of activity within academic and industrial life science. There are four main industrial and three main academic components in the Boston region. Figures 6.2 and 6.3 display the largest components in both the academic and industrial sectors, respectively. The network of collaborations in each of these components has implications for the structure of science in academic and industrial settings, the placement of men and women in these structures, and the level of gender disparity seen in both. 6.4.4 The Structure of Academic Science Figure 6.2 shows the structure of the largest academic component in the Boston university network. This network contains 14 percent women scientists, and is composed mainly of a cluster of MIT scientists, although inventors from Harvard, Boston University, and Tufts are also represented. Langer’s laboratory in the center and the activity extending out from the center of the graph are his students and other university professors (MIT and otherwise) and their lab activities. The structure of academic science resembles that of a bicycle wheel. “Star scientists” are located in the middle of the wheel structure, and collaboration networks extend out from the scientists like spokes. Connections may exist to other wheel-like structures of star scientists, in which collaboration networks again surround the one, central scientist in the center. The structure of academic science reflects the organizational goals of the university. Here, head scientists represent the most central locations, and the domain within which each operates is kept largely separate from other head scientist domains. Networks of academic collaborations are highly centralized, and only a few linkages extend to other centralized networks of collaborations. The bottom visual in figure 6.2 presents an alternative view of the same academic structure, but organized with the most connected individuals at the top. To scale the figure, I use degree centrality and array the nodes by standard deviation. Those at the bottom level represent scientists who have a degree centrality score that is at the mean of the group or below. Each subsequent level brings the threshold up one standard deviation of degree centrality. At the peak of the figure we find Dr. Langer, whose degree centrality is 18 standard deviations above the mean. When viewed this way, the network suggests that collaborative action among inventors is closely orchestrated through the actions of a few highly influential individuals. In academic science, star scientists reside at the top with limited paths and connections between individuals. The visualization provides insight into the ways in which scientists’

A

B

Fig. 6.2 Largest collaboration component, Boston universities (1976–2002): A, Network graph; B, Degree distribution, normalized. Notes: Circles = University Inventors; Triangles = Public Research Organizations.

220

Kjersten Bunker Whittington

patenting choices are constrained by and given opportunity through the structure of reward allocation and the normative prescriptions of their social system. In the realm of the academy, reputation, status, and prestige are the reward generators. Recognition for priority and discovery is payment for scientists’ labors (Merton 1973). The resulting status order is apparent in this visualization of the academy. This organization also likely reflects the need for academic scientists to facilitate large research projects within labs, and to maintain prestige, offset priority-loss, and maximize economies of scale through connecting to other labs that can provide useful collaboration. Figure 6.2 also lends insight in how supply-side demands may be shaped by the system. Consider one’s location in the Langer hierarchy. Moving or finding opportunities and other resources beyond the Langer cluster may be a difficult task given the single node collaboration connections from which the Langer cluster is linked to other clusters at MIT. In university settings, scientist’s ability to appropriate information and reach potential collaborators is limited structurally by the inherent linkages (or lack of them) between clusters. Whether a scientist wants to obtain other collaborators, knows about his or her resource limitations, or is only aware of the difficulty in learning of them yields similar hindrances. Visualizing how scientists are located in the structure of academic science suggests ways to conceptualize how men and women scientists’ decisions and actions may interact with the existing structure of academic science. The overall centralization of the network is a useful way to describe the hierarchy of node distribution in a given system. A simple star network, where everyone is connected to one person but to no one else, has the highest possible degree centralization score of 1. Centralization measures express the degree of variance in the distribution of central positions across a network as a percentage of a star network of the same size. A network with high degree centralization has few nodes with many ties and many nodes with few ties. In the case at hand, the academic centralization is .28, a high value for a large network. Thus the power of individual actors varies substantially, and positional advantages are unequally distributed in this network. Only a few nodes in academia act as bridges to other groups. 6.4.5 The Structure of Industrial Science Industrial scientists’ collaborative arrangements look quite different from that of academic science. Figure 6.3 presents visualizations of the largest connected component in the Boston firm network. The inventors in this component have patented technology developed from a handful of notable biotechnology firms, such as Genetics Institute, Genzyme Corp., Immulogic, and Biogen, among others. In comparison with the main academic component, which has 14 percent female scientists, the industrial component has 25 percent. In addition, we see a clear lack of “star

A

B

Fig. 6.3 Largest collaboration component, Boston firms (1976–2002): A, Network graph; B, Degree distribution, normalized. Notes: Squares = Biotechnology Firms; Triangles = Public Research Organizations.

222

Kjersten Bunker Whittington

scientist” activity, where many nodes are connected to one node but not to others, and an increase in nodes that are reachable through many paths. The industry visualization exemplifies how collaborative activity reflects commercial goals. Across firms, research collaborations with other companies suggest a flow of ideas and resources across many individuals. Whereas academic scientists are organized around specific labs and top scientists, industrial biotechnology scientists appear to be more uniformly organized, perhaps around research problems or the sharing of information, supplies, and human capital across firms. Like the academic network, figure 6.3 reflects the incentive structure of industrial science. Figure 6.3 presents the same industry component, but organized hierarchically by degree centrality. The hierarchy in the industry network is more dispersed as compared with the academic network. Instead of all nodes connected to one or two top scientists, in these firms multiple inventors are connected to many other inventors. The graphs suggest that network collaborations may be more fluid in small firm industrial science. The industrial degree centralization is much lower than that of academia, at .07. Thus, academic scientists are arranged in a structure with four times the degree centralization of industrial scientists. In biotech firms, the lack of a star scientist hierarchy speaks directly to the nature of the industrial science system, and has implications for how the supply side of sector activity may shape or be shaped by the structure of industrial science. Because the network is more dispersed hierarchically, scientists may have more collaborative resources to draw from, or more ability to form new research ties within companies. In addition, industrial networks may be less internally competitive than academic ones. In this way, the structure may help to shape the decisions scientists make, and their opportunities for involvement in dissemination. These figures represent networks of the sum total of ties across all years. By looking aggregately, we see how new and old ties fit into an existing organization of relations among scientists. Equally important, however, is how collaborative activity is organized in real-time slices of available activity—that is, the collaborative opportunities that are present year by year as new ties are made and old ties die out. My additional work on this topic looks at changes in the network structure over time, and finds that across time the academic hierarchy remains, in different capacities, as does the relatively fluid structure of biotechnology (Whittington 2007). This further work also uses the time dynamics of the network to provide evidence of causality relating structural influence to scientists’ dissemination outcomes. 6.4.6 Gender and Network Structure The network visualizations suggest that the structure of scientific collaboration closely follows that of the arrangements of work in the academic and industrial organizational forms. A key difference between aca-

Patterns of Male and Female Scientific Dissemination

223

demia and dedicated biotechnology firms is how the structure of knowledge production is organized within each. In the former, the university is arranged around the laboratories of tenured scientists. Collaborators, often graduate students and/or research scientists, are frequently responsible for small pieces of work within the laboratory’s focus. The structure of science in this academic setting is such that all research (commercial or otherwise) is tied through the head scientist, and collaborations with others in the lab and across labs with the same focus are often kept to a minimum. Like academic science, in biotech firms there may also be a comparable focus on a specific therapeutic arena. In this setting, however, scientists move from research project to research project within the firm, and are often characterized as “voting with their feet” because of the way they move across and between successful research projects in the organization.23 I suggest that the difference between the broadly distributed work of academic science and the more horizontal distribution of knowledge in biotech firms may be important in explaining differences in productivity between men and women within industry and academia. Previous research suggests that women and men create and exist in qualitatively different patterns of interaction within their work setting (Brass 1985; Ibarra 1992; Smith-Lovin and McPherson 1993). Women tend to exist in networks that have more strong relations, and potentially, have access to a fewer number of important bridging, influential ties in their networks. Because industrial settings may be less competitive within firms, or there may be more designated or directed positions of collaboration, women may be better able to access these types of beneficial ties in these settings. Additionally, scientists’ position at the confluence of reputation and collaborative networks may just not be as important in industrial settings as they are in academic ones, where scientists’ networks can matter greatly to help offset loss through priority-based competition. If this is the case, there may be less of an influence of network position and productivity for scientists in industrial jobs, and differences in men’s and women’s networks are not as likely to matter as much for research output as in the academy. Thus, gender differences in patenting may be amplified in public versus private settings due to variations in (a) the differing network positions of men and women, and (b) the varying importance of network position on productivity in industry and academia. My additional work on this topic explores these ideas more fully (Whittington 2007, 2008b). While the graphs show key visual differences, from the network data it is possible to construct standard measures of individual network position and overall network centralization. There are important distinctions between the centrality of men and women across sectors. 23. For an example of this, see the Harvard Business School Case on “planning the unplannable” in Amgen, a large, successful biotechnology firm in Los Angeles, California (Nohria and Berkley 1992).

224

Kjersten Bunker Whittington

My research shows that women in the academy are located in more peripheral and less central collaborative relationships than academic men, and the gender difference in positioning in the academy is significantly larger than in industry. The average yearly male to female ratio of betweenness centrality scores for industry scientists is .96; thus, industry men and women are almost positioned equally central in the network, with women slightly more central. In contrast, the corresponding male to female ratio for academic scientists is 3.1; on average men hold positions approximately three times more central than women in the academy. We might expect that variation between academic men and women in the level of betweenness centrality might lead to differentiation in their ability to appropriate further viable research partners and scientific information. In addition, fixed effects models show that women in biotechnology firms (like men) gain patenting advantages when moving to more central positions that optimize the number of reachable paths to influential others. Women in academia do not see a positive advantage from increasing such connections, but rather see patenting advantages only from increasing their percentage of close ties to others (as do women in industry, and men in both locations). Thus, not only are women achieving less central positions in the academy then men, but they are also benefiting less from those central positions than their counterpart women in other locations. Combined, this work suggests that the structure of academic and industrial science is related to the degree of sex disparity in patenting. One’s position in the network matters, but in addition, the structure of one’s network also plays a defining role. 6.5 Conclusion The loglinear models confirm that sector-level differences in sex disparities exist in the life sciences, while also showing this discipline to be relatively unique in the level of industrial equality among men and women. I link this industry effect with previous research that suggests the industry effect is located among scientists in the network form—science-based biotechnology firms (Whittington and Smith-Doerr 2008). The research in this chapter provides a structural account for the advantages of these particular locations. The network visualizations show how the structure of science varies across these diverse work settings, and illustrate how men’s and women’s commercial involvement may be related to their positions within this structure. Gender differences in involvement within the university may suggest that fewer women have exposure to the commercial process, or alternatively, foster a research focus that lends itself to becoming commercially involved. A more structural explanation for this may be that women lack institutional support for patenting (Murray and Graham 2007), or have

Patterns of Male and Female Scientific Dissemination

225

disadvantageous positions in the collaboration structure of academic science. One factor that has yet to be focused on in other research is the mentor effect in productivity networks. Mentors have been posited to have a significant effect on the commercial orientation of their students, and a strong influence on women scientists (Kram 1988; Fort 1995). In addition, students and their advisors provide many of the cross-sector linkages that knit together academic and industrial networks, thus acting as a key channel for information to diffuse across public and private science. On the individual level, other background factors are important as well, such as scientists’ graduate institution, personal abilities, or research focus, and the presence or absence of monetary and other nonpecuniary research support. In particular, my additional research suggests that motherhood is a particularly salient factor predicting patenting by women in the academy (Whittington 2007, 2008a). A final related and pertinent issue of the network graphs is how the organizational arrangements of large, broadly distributed corporations (such as pharmaceutical firms) are similar or dissimilar from that of dedicated biotechnology firms, as well as the academic life science setting. Elsewhere I have suggested that large corporations may have a similar structure and dynamic to that of university science, and indeed, there are no differences among life science sex disparities among scientists in hierarchical corporations and academic settings (Whittington 2007; Whittington and Smith-Doerr 2008). Including the collaborative activities of scientists in large pharmaceutical firms in future work will help to illuminate how this structure may also contribute to greater gender disparities in commercial involvement. At the heart of this research is the goal of gaining a better understanding of how work environments and changes in the context of science may make a difference for the known disparities between men and women scientists. With academic participation in both basic science and commercial endeavors on the rise, the fact that female scientists in the academy may lag behind their male counterparts (for whatever reason) and are more comparable to men in industrial settings has significant implications for the future labor market choices of women scientists. In addition, as commercialization becomes more common and has more repercussions for academic scientists, these trends have considerable implications not only for the scientific labor market, but the wider pursuit of knowledge as well.

References Aisenburg, N., and M. Harrington. 1988. Women of academe: Outsiders in the sacred grove. Amherst, MA: University of Massachusetts Press.

226

Kjersten Bunker Whittington

Bercovitz, J., and M. Feldman. Forthcoming. Academic entrepreneurs: Social learning and participation in university technology transfer. Organization Science. Bonaccorsi, A., C. Daraio, and L. Simar. 2006. Advanced indicators of productivity of universities: An application of robust nonparametric methods to Italian data. Scientometrics 66 (2): 389–410. Brass, D. J. 1985. Men’s and women’s networks: A study of interaction patterns and influence in an organization. Academy of Management Journal 28 (2): 327–43. Caplow, T., and R. McGee. 1961. The academic marketplace. New York: Science Editions. Carre, F., and P. Rayman. 1999. Professional pathways: Examining work, family, and community in the biotechnology industry. Cambridge, MA: Radcliffe Public Policy Institute. Cohen, W., R. Florida, and R. Goe. 1994. University industry research centers in the United States. Pittsburgh: Carnegie-Mellon University Press. Cole, J. R. 1979. Fair science: Women in the scientific community. New York: Free Press. Cole, J., and H. Zuckerman. 1984. The productivity puzzle: Persistance and change in patterns of publication among men and women scientists. In Women in science, ed. M. W. Steinkamp and M. L. Maehr. Greenwich, CT: Jai Press. ———. 1987. Marriage, motherhood, and research performance in science. Scientific American 256 (2): 119–25. Ding, W. W., F. Murray, and T. E. Stuart. 2006. Gender differences in patenting in the academic life sciences. Science 313 (5787): 665–67. Etzkowitz, H. 1993. Enterprises from science: The origins of science-based regional economic development and the venture capital firm. Minerva 31:326–60. Etzkowitz, H., C. Kemelgor, and B. Uzzi. 2000. Athena unbound: The advancement of women in science and technology. Cambridge: Cambridge University Press. Fort, D.C., ed. 1995. A hand up: Women mentoring women in science. Washington, D.C.: Association for Women in Science. Fox, M. F. 1983. Publication productivity among scientists: A critical review. Social Studies of Science 13 (2): 285–305. ———. ed. 1991. Gender, environmental milieu, and productivity in science. New York: W. W. Norton & Company. ———. 2000. Organizational environments and doctoral degrees awarded to women in science and engineering. Women’s Studies Quarterly 28:47–61. ———. 2001. Women, science, and academia: Graduate education and careers. Gender & Society 15 (5): 654–66. Fruchterman, T., and E. Reingold. 1991. Graph drawing by force-directed replacement. Software-Practice and Experience 21:1129–64. Hall, B. H., A. B. Jaffe, and M. Trajtenberg. 2001. The NBER patent citation data file: Lessons, insights and methodological tools. NBER Working Paper no. 8498. Cambridge, MA: National Bureau of Economic Research, October. Ibarra, H. 1992. Homophily and differential returns: Sex differences in network structure and access in an advertising firm. Administrative Science Quarterly 37 (3): 422–47. Kamada, T., and S. Kawai. 1989. An algorithm for drawing general undirected graphs. Information Processing Letters 31 (1): 7–15. Kleinman, D. L., and S. P. Vallas. 2001. Science, capitalism, and the rise of the “knowledge worker”: The changing structure of knowledge production in the United States. Theory and Society 30 (4): 451–92. Kram, K. 1988. Mentoring at work. New York: University Press of America.

Patterns of Male and Female Scientific Dissemination

227

Long, J. S. 2001. From scarcity to visibility: Gender differences in the careers of doctoral scientists and engineers. Washington, D.C.: National Academies Press. Long, J. S., P. D. Allison, and R. McGinnis. 1993. Rank advancement in academic careers: Sex differences and the effects of productivity. American Sociological Review 58 (5): 703–22. Long, J. S., and M. F. Fox. 1995. Scientific careers: Universalism and particularism. Annual Review of Sociology 21:45–71. Merton, R. K., ed. 1973. Sociology of science. Chicago: University of Chicago Press. Moody, J., and D. R. White. 2003. Social cohesion and embeddedness: A hierarchical conception of social groups. American Sociological Review 68:103–27. Morgan, R., C. Kruytbosch, and N. Kannankutty. 2001. Patenting and invention activity of U.S. scientists and engineers in the academic sector: Comparisons to industry. Journal of Technology Transfer 26 (1–2): 173–83. Murray, F., and L. Graham. 2007. Buying science and selling science: Gender differences in the market for commercial science. Industrial and Corporate Change 16 (4): 657–89. Nohria, N., and J. Berkley. 1992. Amgen, Inc: Planning the unplannable. Harvard Business School case. 9-492-052. Owen-Smith, J., and W. W. Powell. 2001. Careers and contradictions: Faculty responses to the transformation of knowledge and its uses in the life sciences. Research in the Sociology of Work 10:109–40. ———. 2004. Knowledge networks as channels and conduits: The effects of spillovers in the Boston biotechnology community. Organization Science 15 (1): 5–21. Packer, K., and A. Webster. 1995. Inventing boundaries: The prior art of the social world. Social Studies of Science 25 (1): 107–17. Padgett, J. F., and C. K. Ansell. 1993. Robust action and the rise of the Medici, 1400–1434. American Journal of Sociology 98 (6): 1259–1319. Podolny, J. M., and K. L. Page. 1998. Network forms of organizations. Annual Review of Sociology 24:57–76. Porter, K. A., K. B. Whittington, and W. W. Powell. 2005. The institutional embeddedness of high-tech regions: Relational foundations of the Boston biotechnology community. In Clusters, networks, and innovation, ed. S. Breschi and F. Malerba, 261–96. Oxford: Oxford University Press. Powell, W. W. 1990. Neither market nor hierarchy: Network forms of organization. Research in Organizational Behavior 12:295–336. Powell, W. W., and J. Owen-Smith. 1998. Universities and the market for intellectual property in the life sciences. Journal of Policy Analysis and Management 17 (2): 253–77. Powell, W. W., K. W. Koput, and L. Smith-Doerr. 1996. Interorganizational collaboration and the locus of innovation: Networks of learning in biotechnology. Administrative Science Quarterly 41 (1): 116–45. Powell, W. W., D. R. White, K. W. Koput, and J. Owen-Smith. 2005. Network dynamics and field evolution: The growth of interorganizational collaboration in the life sciences. American Journal of Sociology 110 (4): 1132–1205. Rabinow, P. 1996. Making PCR. Chicago: University of Chicago Press. Raftery, A. 1986. Choosing models for cross-classification. American Sociological Review 51:145–56. Rosenburg, N., and R. Nelson. 1994. American universities and technical advance in industry. Research Policy 23 (3): 323–48. Smith-Doerr, L. 2004. Women’s work: Gender equality vs. hierarchy in the life sciences. Boulder, CO: Lynne Rienner Publishers.

228

Kjersten Bunker Whittington

Smith-Doerr, L., and W. W. Powell. 2005. Networks and economic life. In The handbook of economic sociology, second edition, ed. N. Smelser and R. Swedberg, 379–402. Princeton: Russell Sage Foundation. Smith-Lovin, L., and M. McPherson, eds. 1993. You are who you know: A network perspective on gender. New York: Aldine de Gruyter. Stephan, P. E., S. Gurmu, A. J. Sumell, and G. Black. 2007. Who’s patenting in the university? Evidence from the survey of doctorate recipients. Economics of Innovation and New Technology 16 (2): 71–99. Stuart, T. E., and W. W. Ding. 2006. When do scientists become entrepreneurs? The social structural antecedents of commercial activity in the academic life sciences. American Journal of Sociology 112 (1): 97–144. Ward, K. B., and L. Grant. 1995. Gender and academic publishing. In Higher education: Handbook of theory and research, vol. 11, ed. A. E. Bayer and J. C. Smart, 172–212. New York: Agathon. Weakliem, D. L. 1999. A critique of the Bayesian information criterion for model selection. Sociological Methods and Research 27 (3): 359–97. White, D. R., and F. Harary. 2001. The cohesiveness of blocks in social networks: Node connectivity and conditional density. Social Methodology 31:305–59. Whittington, K. B. 2007. Employment sectors as opportunity structures: The effects of location on male and female scientific dissemination. Ph.D. diss. Department of Sociology, Stanford University. ———. 2008a. Gender and patenting in academia and industry: Interactions between individual and contextual characteristics. Working Paper. Reed College. ———. 2008b. The influence of network structure on men’s and women’s scientific collaboration. Working Paper. Reed College. Whittington, K. B., and L. Smith-Doerr. 2005. Gender and commercial science: Women’s patenting in the life sciences. Journal of Technology Transfer 30 (4): 355–70. ———. 2008. Women inventors in context: Disparities in patenting across academia and industry. Gender & Society 22 (2): 194–218. Xie, Y., and K. A. Shauman. 1998. Sex differences in research productivity: New evidence about an old puzzle. American Sociological Review 63 (6): 847–70. ———. 2003. Women in science: Career processes and outcomes. Cambridge, MA: Harvard University Press. Zuckerman, H. 1991. The careers of men and women scientists: A review of current research. In The outer circle: Women in the scientific community, ed. H. Zuckerman, J. Cole, and J. Bruer, 27–56. New York: W. W. Norton.

7 Educational Mismatch among Ph.D.s Determinants and Consequences Keith A. Bender and John S. Heywood

7.1 Introduction According to assignment theory, the returns to investments in human capital vary dramatically with the quality of the match between the worker and the job (see Sattinger 1993; Belman and Heywood 1997). Mismatches between worker skills and job requirements have substantial costs as workers are either unable to fulfill job requirements or have skills that go unused. Mismatches waste educational resources, resulting in dissatisfied workers (Tsang 1987) and higher turnover (McGoldrick and Robst 1996). Lower job satisfaction and higher turnover may reduce formal training and lower labor productivity, and so firms’ profits (Groot 1993; Sloane, Battu, and Seaman 1996). Finally, frustration over being mismatched may independently reduce worker effort (Belfield 2000). In this chapter we focus on workers with a Ph.D. in science and examine the predicted consequences of mismatch. First, we examine these workers because they play a key role in innovation and creating technological progress. As a consequence, economists have estimated the determinants of productivity for scientists (Levin and Stephan 1991), the adequacy of their supply (Stephan and Levin 1991; Stephan 1996; National Science Board Keith A. Bender is an associate professor of economics at the University of WisconsinMilwaukee. John S. Heywood is a professor of economics at the University of WisconsinMilwaukee. The authors thank the National Science Foundation (NSF) and William Rayburn for helping to secure access to the restricted data, and Richard Freeman, Dan Hamermesh, Shulamit Kahn, Kostas Mavromaras and the participants of seminars at the University of Aberdeen (Scotland), Lancaster University (England), the University of Wisconsin-Milwaukee, the National Bureau of Economic Research, and the NSF for helpful comments. The use of NSF data does not imply NSF endorsement of the research methods or conclusions contained in this chapter.

229

230

Keith A. Bender and John S. Heywood

2003; Teitelbaum 2004) and the rewards to their education (Stephan and Everhart 1998). Second, the homogeneity of this sample allows us to control for variables excluded from typical estimates in examining the consequences of mismatch. Third, managers concerned with maximizing the impact of their research and development staff need to understand the relationship between mismatch and job satisfaction, since it influences productivity (Kim and Oh 2002). Finally, substantial governmental resources are devoted to educating these workers and to improving their diversity. Yet concern continues as the growth in U.S. university students pursuing advanced degrees in science slows and as trained scientists increasingly abandon scientific careers (Preston 2004). We are the first to use this group of workers to examine the three major consequences that have been identified with mismatch: lower earnings, lower job satisfaction, and turnover. The findings are striking. Mismatch is associated with lower earnings, reduced job satisfaction, and greater turnover even after controlling for a wide range of other explanatory variables and even given the relative homogeneity of our sample. Several major categories of theoretical conjecture explain why mismatches between workers and jobs persist in equilibrium. First, government subsidization may result in “overeducation” in which the general level of educational attainment exceeds that demanded (Freeman 1976). While this may create a surplus, it is less transparent why the wage will not fall to clear the surplus. Second, there may be information problems. If productivity is not known or is revealed only over a lengthy period, workers may remain mismatched based on signals that need not reflect their true productivity (Tsang and Levin 1985). Similarly, search and information costs may be large enough that workers remain mismatched as a cheaper alternative to searching for a new job or to the firm searching for a new employee. Third, institutional theories of the labor market have long contended that internal labor market considerations force employers to base earnings on easily observable characteristics of the worker and job (Thurow 1975). Thus, institutional issues within the firm help determine pay and allow mismatches to remain, especially when specific human capital investments bind the worker and firm together across a wide range of pay and productivity relationships. These concerns may be particularly prominent in science occupations in which the skill vintage changes rapidly, hastening worker mismatch. Alternatively, these concerns may be less prominent if those in science occupations are more likely to have earnings that reflect performance. Finally, jobs in science occupations may represent tournament winnings. Only the best are able to be carefully matched to their training and pursue continued employment. The mismatched in this view may be less able and, indeed, mismatch may even be evidence of an efficient tournament. A large empirical literature attempts to measure the consequences of

Educational Mismatch among Ph.D.s

231

mismatch. While the measure by which workers have too much education varies (Belfield 2000, 35–37), Groot and Maasen van den Brink (2000) provide a meta-analysis showing that the overeducated suffer a 14 percent earnings penalty (also see Chevalier 2003; Dolton and Vignoles 2000; Battu, Belfield, and Sloane 1999). Yet mismatch goes beyond overeducation. Borghans, Bruinshoofd, and de Grip (2000) show that workers holding a job unrelated to their field of education suffer significantly diminished earnings. Allen and van der Velden (2001) measure the wage effects of skill mismatches, controlling for both educational levels, and apparent educational mismatches showing that those with unused skills earn significantly less. Psychological theories of expectation suggest that underutilized skills cause diminished job satisfaction. Those with the greatest education and skills have the highest expectations for their jobs and careers and are more easily disappointed (Tsang and Levin 1985; Clark and Oswald 1996). Solomon et al. (1981) and Allen and van der Velden (2001) confirm that both underutilized skills and skill deficits are associated with significantly diminished job satisfaction. Belfield and Harris (2002) and Moshavi and Terborg (2003) find that the overeducated suffer diminished job satisfaction. Yet the evidence remains mixed as Buchel (2002) presents German evidence that overqualified employees have the same job satisfaction as those properly matched. Much of the original reason for examining subjective job satisfaction is that it influences real economic variables including quit rates, shirking, and absenteeism (Freeman 1978; Clark and Oswald 1996). Thus, mismatched workers that have lower job satisfaction will be more likely to search for a new job. Moreover, reduced productivity associated with mismatch may encourage employers to seek alternative workers. The consequence is that the turnover rate among the mismatched should be higher (Wolbers 2003; Allen and van der Velden 2001). Unlike the literature reviewed previously, we go beyond simply identifying the consequences of mismatch to ask which reasons for being mismatched have the greatest consequences. We find that those mismatched because of the lack of jobs or family considerations suffer very large reductions in wages and job satisfaction. Women who are mismatched because of family considerations have particularly large reductions in earnings. We also estimate the causes of mismatch. The estimations suggest that the vintage of scientific skills is a critical determinant. In general, scientists are more likely to be mismatched, and the penalty for mismatch grows as they age beyond the time of their degree. There is evidence of the traditional movement from the lab to management for workers in rapidly changing disciplines and the indication that married men are least likely to be mismatched, all else equal.

232

Keith A. Bender and John S. Heywood

7.2 Data and Methodology We draw our primary data from the 1997 and 1999 Survey of Doctorate Recipients (SDR) conducted by the National Opinion Research Center for the National Science Foundation. The SDR is a nationally representative sample of all Ph.D. graduates in the hard and social sciences, prior to 1997, living in the United States. Collected in response to the National Research Council’s demand for data that allows the integration of occupational detail and academic training, the SDR is conducted every other year. The 1997 SDR asks about overall job satisfaction and is the only wave asking what we identify as the secondary mismatch indicators. We selected all currently employed scientists for which full information was available, yielding a sample of 31,845.1 The primary indicator of mismatch comes from responses to the question, “Thinking about the relationship between your work and your education, to what extent is your work related to your doctoral degree?” The possible responses are “closely related,” “somewhat related,” and “not related.” Those scientists working in jobs not related to their education are presumably using less of the knowledge, training, and skills learned in that education. In this critical sense they may be identified as mismatched because there is not a close fit between their education and job. As table 7.1 shows, only 7.3 percent of the sample report their education and job are not at all related, although another 23.4 percent report that their education and job are only somewhat related. Two secondary mismatch indicators probe related aspects of the link between the workers’ scientific education and their current job. The first asks, “Thinking back to when you completed your degree would you say your work during a typical week on your job is 1) very similar to what you expected to be doing, 2) somewhat similar to what you expected to be doing or 3) not very similar to what you expected to be doing?” To the extent that expectations upon completing the doctoral degree reflect the training and experience in their field, those who are far away from their expectations may also be far away from their training or field. Obviously, expectations could be imperfect and, if so, one might expect a weaker relationship between this question and underlying mismatches. The means indicate that 20.6 percent report that their job is not very similar to what they expected, with another 32.5 reporting that their job is only somewhat similar to what they expected. The other secondary question asks, “If you had the chance to do it over again, knowing what you do now, how likely is it that you would choose the 1. While much of the data from the SDR are publicly available, we add variables from the restricted use version. These variables include annual earnings, detailed job codes, race/ethnicity, age, and marital status. See the SDR website at http://sestat.nsf.gov/ for details on both versions of the SDR data.

Table 7.1

Means of the variables All

Academic

Nonacademic

0.693 0.234 0.073 0.469 0.325 0.206 0.549

0.835 0.141 0.024 0.617 0.289 0.094 0.601

0.564 0.319 0.118 0.333 0.359 0.308 0.503

0.302 0.149

0.281 0.118

0.320 0.177

Dependent variables $70,449 (48,905) 3.399 (0.741) 0.183

$59,881 (39,116) 3.426 (0.734) 0.136

$80,070 (54,608) 3.373 (0.747) 0.226

Demographic variables 0.130 0.097 0.186 0.587 0.818 0.133 0.023 0.023 0.004 14.1 (10.3) 304.9 (375.1) 0.078 0.163 0.137 0.063 0.184 0.043 0.079 0.069 0.182

0.127 0.114 0.214 0.546 0.832 0.108 0.027 0.029 0.004 14.0 (10.6) 309.7 (386.6) 0.086 0.155 0.155 0.079 0.162 0.058 0.087 0.067 0.149

0.132 0.080 0.160 0.628 0.805 0.155 0.019 0.017 0.003 14.2 (10.0) 300.3 (363.4) 0.072 0.170 0.121 0.048 0.205 0.029 0.071 0.070 0.213

Job variables 0.534 0.922 0.817 0.882

0.497 0.937 0.902 0.953

Mismatch variables Job and education closely related (excl) Job and education related Job and education not related Job very similar to expectations Job similar to expectations Job not very similar to expectations Very likely to choose similar field Somewhat likely to choose similar field Not likely to choose similar field Annual salary Satisfaction with job Changed jobs Single male (excl) Single female Married female Married male White, non-Hispanic (excl) Asian, non-Hispanic Hispanic Black, non-Hispanic Other race, non-Hispanic Experience Experience squared Northeast region Middle Atlantic region East North Central region West North Central region South Atlantic region East South Central region West South Central region Mountain region Pacific region Supervisor Full time employment Pension is available Profit sharing is available

0.568 0.909 0.740 0.818 (continued )

234 Table 7.1

Keith A. Bender and John S. Heywood (continued)

Employer size  1,000 (excl) Employer size  1,000 and  5,000 Employer size  5,000 Number of memberships in professional organization Academic sector Business sector (excl) Government sector Research is main work activity (excl) Teaching is main work activity Management is main work activity Computer work is main work activity Other main activity Economics (excl) Computer Math Hard science Social science (not economics) Engineering Management Health Teacher Other (nonscience) occupation

All

Academic

Nonacademic

0.261 0.044 0.695

0.117 0.051 0.831

0.392 0.038 0.571

2.472 (1.941) 0.477 0.103 0.420 0.407 0.218 0.161 0.048 0.166

2.943 (2.021) 1.000 0.000 0.000 0.371 0.450 0.092 0.013 0.073

2.042 (1.759) 0.000 0.197 0.803 0.439 0.007 0.225 0.080 0.250

0.040 0.006 0.081 0.392 0.198 0.093 0.077 0.023 0.081 0.010

0.020 0.071 0.017 0.269 0.146 0.172 0.192 0.032 0.003 0.077

Discipline indicators 0.030 0.040 0.047 0.328 0.171 0.135 0.137 0.028 0.040 0.045

Notes: All means are weighted using sample weights. Numbers in parentheses are standard deviations for continuous variables; “(excl)” indicates that this variable was a reference variable in the regressions.

same field of study for your highest degree?” The answers are “very likely,” “somewhat likely,” and “not likely at all.” While this question might simply be seen as identifying those who feel they made a bad career choice, it seems reasonable that those who are mismatched would especially regret their choice of field. Consistent with evidence that scientists leave their careers in large numbers (Preston 2004), only slightly more than half of the sample report being very likely to repeat their study for their highest degree. In sum, these three questions reflect slightly different aspects of the relationship between education and the workers’ current positions. Despite these differences, the three measures appear highly correlated, as we will show shortly. Table 7.2 presents cross-tabulations of the two secondary measures of mismatch with the primary measure. The top panel shows that the distribution of expectations of the job given the education is broadly similar to that of how closely the job and education relate. The diagonal terms are the

Educational Mismatch among Ph.D.s Table 7.2

235

Frequency distribution between three mismatch variables How related is job and education? Not closely related

Closely related

Very closely related

Job not very similar to expectations Job similar to expectations Job very similar to expectations

5.54 0.79 0.64

9.14 10.67 3.65

4.59 21.59 43.39

Not likely to choose same field Somewhat likely to choose same field Very likely to choose same field

2.78 2.10 2.09

4.75 8.51 10.21

7.39 19.95 42.23

largest within each column and comprise roughly 60 percent of the sample. The distribution of the likelihood of repeating the same education is a bit less similar to the primary indicator. The diagonal terms are not the largest in two columns and comprise a smaller share of the sample. The less than perfect correlation suggests the three measures capture somewhat different aspects of mismatch and that considering each may be valuable. Using these measures, we investigate the effect of mismatch on earnings, job satisfaction, and job change. The earnings measure is annual earnings in 1997 including all wages, salaries, bonuses, overtime, commissions, consulting fees, and net income from business and has an average of over $70,000. The critical question on job satisfaction asks, “How would you rate your overall satisfaction with your principal job during the week of April 15th?” The choices are “very dissatisfied,” “somewhat dissatisfied,” “somewhat satisfied,” and “very satisfied.” These responses are ordered values from 1 to 4 with an average of 3.4. Job change information comes from the 1999 survey, which asks workers to identify one of the following: (a) their current job is different than that held in the 1997 survey but with the same employer; (b) their current job is different than that they held in the 1997 but with a different employer; (c) they hold the same job and employer as in 1997; or (d) they hold the same job with a different employer. We view positive answers to either of the first two options as evidence of a job change. A relatively large share of sample, 18.3 percent, changed jobs within the two years. We also use only those who changed employers and jobs as true job changes, and this does not alter the basic results. The bottom portion of table 7.1 indicates the rich set of demographic and job dimension variables that we use as controls in an effort to isolate the influence of the indicators of mismatch. These include race, gender, marital status, regional variables, many job dimensions that might influence earnings, job satisfaction, turnover, and indicators of discipline (see Bender and Heywood [2006] for a detailed list of these indicators). Knowing the age of the respondent and the year of their Ph.D., we also derive an

236

Keith A. Bender and John S. Heywood

imputed measure of experience that we take to be an indicator of vintage of training. Table 7.3 collects the primary mismatch indicator, earnings, job satisfaction, and turnover. The data are presented separately for each discipline and divided between those holding academic and nonacademic jobs. Academics, in general, are much less likely to report any degree of mismatch (this is true across all three mismatch indicators). Academic economists report the lowest share of any degree mismatch at only 2.2 percent. Those working in computer science report a very high degree of mismatch, with two-thirds reporting a degree of mismatch. This may reflect a vintage effect in which the discipline changes very quickly, heightening the gap between skills learned in school and those needed on the job. Such a possibility emphasizes the need to control for the years since degree (experience) for the worker. The single highest degree of mismatch is among those nonacademics working in “other disciplines.” This follows, in part, from construction. This category includes disciplines other than those in the sciences. Thus, workers in this category are necessarily working in disciplines other than that in which they were trained.

Table 7.3

Discipline All Economics Social science Computer science Math Hard science Engineering Management Health Other disciplines

Educational mismatch, earnings, job satisfaction, and turnover by discipline and sector

Sector

% Closely related

% Very satisfied

Average salary ($USD)

% Change jobs

Academic Nonacademic Academic Nonacademic Academic Nonacademic Academic Nonacademic Academic Nonacademic Academic Nonacademic Academic Nonacademic Academic Nonacademic Academic Nonacademic Academic Nonacademic

83.5 56.4 97.8 76.5 93.7 88.7 38.4 28.6 86.9 68.7 85.9 66.9 87.7 57.2 55.7 41.9 50.5 39.8 35.0 17.4

54.8 50.9 56.6 55.1 55.8 55.7 59.9 43.5 51.3 50.2 54.1 48.9 53.8 44.7 63.9 57.4 55.6 58.1 45.0 48.6

59,881 80,070 62,911 88,308 51,986 63,918 63,172 78,025 56,565 76,758 56,067 72,572 68,200 80,444 90,202 104,885 80,098 101,106 39,462 65,627

13.6 22.6 10.6 19.5 12.0 14.6 26.9 26.1 10.2 18.2 14.4 22.5 11.6 22.3 18.4 29.3 19.9 13.7 22.9 23.6

Note: “% Change jobs” includes those who changed jobs regardless of whether or not they changed employers.

Educational Mismatch among Ph.D.s

237

In terms of the outcome variables, those groups with better matches—a closer fit between education and job—on average also have higher job satisfaction and lower job change. For example, academics, who say that they are better matched on average, report modestly higher job satisfaction and have much lower turnover. On the other hand, although more highly matched, academics earn less than nonacademics in every discipline. Further investigation, simultaneously controlling for other determinants of earnings, is warranted to see if this correlation holds. 7.3 Consequences of Mismatch Table 7.4 presents typical log-linear earnings equations revealing a number of anticipated results. As expected in a sample of the highly skilled, the racial differences are small or simply absent. Supervisors, married men, those receiving pensions and working full time each earn more. Married women earn less than single men in the academic sector, while single women earn less than single men in the nonacademic sector. Those working in the government earn less than those working in the private nonacademic sector. Experience shows the standard concave pattern with earnings. Thus, for the academics in the sample, the experience coefficients do not reflect Ransom’s (1993) finding that higher seniority for university professors is associated with lower salaries, all else equal. The results indicate that among academics, those who view their primary activity as teaching earn less and that only those in management science earn more than the excluded group, economists. Although not reported, those in hard science earn more than 20 percent less than economists, with the gap larger in percentage terms in the nonacademic sector than in the academic sector. Importantly, a large decrement in earnings is associated with the critical mismatch variables. If academic workers report that their education only relates somewhat, their earnings are 6.9 percent lower holding, all else equal.2 The comparable earnings penalty for nonacademic workers is 4.7 percent. If academic workers report that their education does not relate, their earnings are 13.8 percent lower. The comparable earnings penalty for nonacademic workers is 9.8 percent. The greater penalty in academia may reflect a greater importance of appropriate educational credentials or it may be that academic jobs not in a worker’s Ph.D. field reflect a greater degree of mismatch than is true for nonacademic jobs. Recalling that at least one consensus estimate of the penalty associated with overeducation was 14 percent, we are estimating a mismatch penalty of roughly the same size. This is surprising, as those estimates were taken to measure the influence of having an unnecessary degree. While that may be true for some of our 2. The coefficient  is transformed into a percentage change in earnings, e – 1.

Supervisor

Experience squared

Experience

Other race, non-Hispanic

Black, non-Hispanic

Hispanic

Asian, non-Hispanic

Married males

Married females

Single female

Nonacademic 0.056*** (4.72) 0.117*** (5.44) 0.116*** (4.01) 0.013 (0.57) 0.093*** (5.20) 0.006 (0.48) 0.034 (0.95) 0.018 (0.56) 0.007 (0.11) 0.040*** (19.44) 8.5E-4*** (13.59) 0.103*** (8.76)

Academic

0.073*** (5.70) 0.167*** (4.38) 0.026 (1.55) 0.048*** (3.40) 0.060*** (4.84) 0.034*** (2.57) 0.016 (0.97) 0.021 (1.24) 0.112*** (2.94) 0.036*** (22.22) 5.4E-4*** (11.75) 0.119*** (14.31)

Ln(Annual earnings)

Regressions estimating the consequences of mismatch

Job-Education somewhat related Job-Education not related

Table 7.4

0.332*** (9.46) 0.542*** (6.75) 0.057 (1.17) 0.054 (1.25) 0.160*** (4.21) 0.085** (2.48) 0.081 (1.32) 0.120** (2.13) 0.216 (1.55) 0.002 (0.43) 4.7E-4*** (4.47) 0.097*** (3.72)

Academic 0.357*** (13.72) 0.479*** (11.15) 0.038 (0.72) 0.180*** (4.20) 0.128*** (3.65) 0.116*** (3.91) 0.058 (0.86) 0.118* (1.68) 0.245 (1.33) 0.009** (2.35) 5.4E-4*** (5.10) 0.147*** (5.82)

Nonacademic

Job satisfaction

0.043*** (3.62) 0.087*** (3.10) 0.016 (1.00) 0.004 (0.27) 0.019 (1.52) 0.022* (1.81) 0.004 (0.19) 0.019 (1.21) 0.010 (0.24) 0.016*** (12.45) 4.1E-4*** (12.23) 0.034*** (3.80)

Academic

0.042*** (3.49) 0.063*** (3.30) 0.031 (1.30) 0.040** (2.07) 0.009 (0.60) 0.043*** (3.04) 0.065*** (2.64) 0.017 (0.53) 0.133 (1.50) 0.018*** (10.34) 4.3E-4*** (9.66) 0.030*** (2.63)

Nonacademic

Changed jobs?

0.133*** (14.65) 0.001 (0.08) 0.064 (1.52) 0.024 (1.16)

0.537*** (16.71) 0.284*** (14.42) 0.422*** (10.52) 0.035* (1.94) 0.031 (1.45) 0.043*** (3.09) 0.029*** (13.61) —

0.656*** (17.92) 0.125*** (7.13) 0.085*** (3.17) 0.096*** (8.76) 0.004 (0.16) 0.032** (2.28) 0.023*** (7.16) 0.17*** (12.36) 0.083 (0.89) 0.007 (0.41) 0.038** (1.96) 0.008 (0.42) 0.052* (1.79) 0.055 (0.98) 0.246** (2.27) 0.017 (0.33)

0.057 (1.15) 0.156*** (3.74) 0.049 (0.82) 0.114** (1.98) 0.041 (0.67) 0.082** (2.17) 0.025*** (3.88) — 0.099** (2.11) 0.120*** (3.15) 0.268*** (5.71) 0.189*** (6.95) 0.287*** (5.16) 0.197*** (6.52) 0.014* (1.87) 0.02 (0.54) 0.395*** (2.82) 0.103*** (2.88) 0.092* (1.92) 0.122*** (3.34) 0.038*** (4.08) 0.022 (1.29) 0.004 (0.10) 0.014 (0.89)

0.065*** (3.16) 0.054*** (3.51) 0.062*** (2.69) 0.012 (0.66) 0.021 (1.07) 0.021* (1.79) 0.010*** (4.97) —

0.080*** (3.77) 0.051*** (2.83) 0.059*** (3.02) 0.023** (1.96) 0.090*** (3.17) 0.067*** (4.90) 0.004 (1.24) 0.048*** (3.30) 0.040 (0.65) 0.046*** (2.81) 0.044** (2.05) 0.019 (1.15)

Notes: The numbers under the coefficient estimates are t-statistics. All estimates use sample weights and include controls for academic discipline and region of the country. Cut points and constants also estimated where appropriate. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

Other main activity

Computer is main activity

Managing is main activity

Teaching is main activity

Number of professional memberships Gov’t., nonacademic

Employer size  1,000 and  5,000 Employer size  5,000

Profit sharing

Health insurance

Pension

Full time

240

Keith A. Bender and John S. Heywood

mismatched workers, it need not be. Workers need only report that they are working in an area (a subject matter) outside their degree. Job satisfaction measures force workers to consider elements of the job in addition to earnings. These elements include fringe benefits, interactions with coworkers, the value of the work itself, relations with supervisors, hours, and intrinsic rewards to name only a few. As Hamermesh (2000) argues, job satisfaction measures, at their best, hope to capture the sum of utility derived from all aspects of the job. As the job satisfaction measure is an ordered response, the estimation follows a cumulative normal in an ordered probit. The results in table 7.4 confirm some expectations from past work on job satisfaction but present a few surprises. Women do not emerge as the routinely “contented workers” that estimations on general populations often report (Clark 1997). Both single and married women in academic jobs have the same job satisfaction as single men while only married women in nonacademic jobs report greater job satisfaction than single men.3 Married men routinely show greater job satisfaction than single men. Blacks report lower job satisfaction as do those working for larger employers. Many of the other results roughly follow the wage equations with supervisors, those working full time, and those with pensions all reporting higher job satisfaction. The pattern of discipline effects (suppressed to save space) also follows familiar lines, with those in engineering and the hard sciences reporting lower job satisfaction than the excluded group (economists), and those in management reporting higher job satisfaction. The experience results do not follow the wage equation, with job satisfaction declining with experience but at a decreasing rate in the nonacademic sector. Job satisfaction declines well into mid-career before starting to rebound for these workers. Ward and Sloane (2000) find that satisfaction with salary decreases with age for male academics while the more general studies confirm that job satisfaction is lowest for those in middle age. The mismatch variables are associated with substantially lower job satisfaction. For both academics and nonacademics, working in jobs not related to their education is associated with a highly significant and large reduction in overall job satisfaction. The magnitude of this influence is understood by making projections. If we assume all variables are held at their mean levels except the mismatch variables, we can use the cut points and project the probability of being in each satisfaction level. As an illustration, if we assume that academic workers have the mean characteristics and are in jobs closely related to their education, they have a 0.655 probability of reporting the highest level of satisfaction, which is very satisfied. If 3. Bender and Heywood (2006) show that while women academics report lower job satisfaction than men, women in government report the same satisfaction as men and women in business report greater satisfaction than men.

Educational Mismatch among Ph.D.s

241

they hold jobs only somewhat related to their education, the probability of reporting being very satisfied drops to 0.543. Finally, if they hold a job not at all related to their education, the probability of reporting being very satisfied is only 0.428. A complete set of projections is available from the authors, but it is apparent that the marginal influence of mismatch on job satisfaction is very large. As discussed, one advantage of the SDR is the ability to follow the workers two years after the 1997 survey to determine whether or not they have changed their job. This measure of turnover is a dichotomous measure and becomes the dependent variable in probit specifications as shown in the last two columns of table 7.4. These estimations repeat some familiar patterns. Those who are supervisors and have fringe benefits are less likely to have changed jobs, with the latter perhaps reflecting deferred compensation that binds workers to employers.4 Job change becomes less likely with experience but eventually turns around and becomes more likely later in life. Married women in nonacademic jobs are more likely to change jobs than single males. Black workers appear no more likely to change jobs than do white workers, while those in large firms are more likely to change jobs. The critical mismatch variables reveal that those working positions not related to their education have a higher probability of turnover than those in positions closely related to their education: 8.7 percentage points higher for academics and 6.3 percentage points higher for nonacademics. Given that the average turnover rate across the two subsamples is slightly above 18 percent, these are very large marginal effects. Again, there is a more muted effect for a worker being in a field only somewhat related to their education. Among academics, these workers are 4.3 percentage points more likely to change jobs and among nonacademics, these workers are 4.2 percentage points more likely to change jobs. The marginal effects for both of these degrees of mismatch are statistically significant for each subsample.5 Our examination of the consequences of mismatch presents a consistent picture. Using this primary indicator of mismatch, we find routine and robust partial correlations. Mismatch remains associated with lower wages, lower job satisfaction, and an increased probability of changing jobs. 7.3.1 Other Indicators of Mismatch The other survey measures related to mismatch present a similar, if slightly less dramatic, picture. Our estimates of the earnings equations 4. Such an implication makes sense to the extent that pensions are back-loaded (as in defined benefit plans—see Lazear [1979]) and to the extent that health insurance creates “job lock,” making mobility difficult (Adams 2004). 5. We altered our definition of job change, restricting it to include only those workers who simultaneously changed jobs and employers. This removes normal promotions from being classified as job changes but runs the risk of excluding real job changes within an employer. The mismatch coefficients remain statistically significant and of the same direction. Mismatch continues to increase the probability of job change.

242

Keith A. Bender and John S. Heywood

used precisely the same set of controls as reported in table 7.4 but merely replaced the mismatch variables with their alternatives. As the first panel in table 7.5 shows in specification 2, academics who find their work not very similar to their expectations at the time of degree completion have 9.6 percent lower earnings while nonacademics who find their work not very similar to their expectations have 8.5 percent lower earnings. Academics who report that they would not be very likely to repeat their Ph.D. degree in the same field have 8.0 percent lower earnings while those nonacademics who report the same thing have 8.8 percent lower earnings when compared to those very likely to repeat their degree. All of the coefficients from which these percentage measures are taken are statistically significant at the one percent level and serve to further confirm the suggestion that mismatch among the highly educated is associated with reduced earnings. The results on job satisfaction are more dramatic, with substantially larger declines than those associated with the original mismatch indicator. Academics whose work is not very similar to their expectations have a 31.1 percentage point reduction in the probability of being in the highest satisfaction category compared to those whose work is similar to their expectations (this is now a marginal effect). Nonacademics whose work is not very similar to their expectations have a 29.4 percentage point reduction in the probability of being in the highest satisfaction category. Those academics not very likely to repeat their degree have a 40.5 percentage point reduction in the probability of being in the highest satisfaction category and those nonacademics whose work is not similar to expectations have a 31.0 percentage point reduction in the same probability. While these secondary measures may simply be alternative satisfaction measures, it remains possible that they capture important elements of mismatch as each directs the respondent to compare their work to an aspect of their education. To the extent this is correct, the results serve to confirm those presented with the original indicator. The two secondary measures of mismatch also reinforce the results on job turnover. Those academics whose expectations are not very similar are 6.9 percentage points more likely to change jobs and those nonacademics whose expectations are not very similar are 11.5 percentage points more likely to change jobs. Academics who would not be likely to repeat their graduate education are 4.5 percentage points more likely to change jobs and nonacademics who are not very likely to repeat their graduate education are 7.0 percentage points more likely to change jobs compared to those who are very likely to repeat their graduate education. Thus, the secondary measures show a somewhat different pattern on job change that suggests that mismatch is more likely to result in job change for nonacademics than for academics. Nonetheless, the general patterns are similar across all three measures and across all three consequences. All of the

Table 7.5

Comparing consequences across different measures of mismatch Academic

Specification 1

Specification 2

Specification 3

Specification 1

Specification 2

Specification 3

Specification 1

Specification 2

Specification 3

Nonacademic

Log annual earnings regression Education relates 0.073*** (5.70) Education does not relate 0.167*** (4.38) Similar to expectations 0.032*** (3.77) Not very similar 0.107*** (5.87) Somewhat likely 0.047*** (5.48) Not very likely 0.083*** (6.49)

0.056*** (4.72) 0.117*** (5.44) 0.058*** (4.86) 0.089*** (6.25) 0.060*** (5.14) 0.092*** (6.02)

Job satisfaction ordered probit regression (marginal effects) Education relates 0.132*** (9.54) Education does not relate 0.212*** (7.16) Similar to expectations 0.237*** (23.89) Not very similar 0.311*** (20.20) Somewhat likely 0.238*** (23.93) Not very likely 0.405*** (37.57)

0.142*** (13.91) 0.187*** (11.71) 0.203*** (19.09) 0.294*** (23.80) 0.180*** (18.23) 0.310*** (28.62)

Change jobs probit regression (marginal effects) Education relates 0.043*** (3.62) Education does not relate 0.087*** (3.10) Similar to expectations 0.036*** (4.02) Not very similar 0.069*** (4.21) Somewhat likely 0.038*** (4.23) Not very likely 0.045*** (3.53)

0.042*** (3.49) 0.063*** (3.30) 0.072*** (5.50) 0.115*** (7.45) 0.027** (2.31) 0.070*** (4.77)

Notes: Each specification includes the covariates listed in table 7.4 through 7.6, except where the measures of educational mismatch are replaced by the mismatch measures listed in this table. Numbers in parentheses are t-statistics (for the earnings regression) or asymptotic z-statistics ( job satisfaction and change jobs regressions). The results for the job satisfaction ordered probits are marginal effects of the probability of being in the highest job satisfaction category, holding all other variables at their mean value. ***Significant at the 1 percent level.

244

Keith A. Bender and John S. Heywood

coefficients for both levels of mismatch are statistically significant at the one percent level. 7.3.2 The Role of Reasons for Mismatch In this section we expand the dimensions of mismatch, including in the analysis the reasons that workers see themselves as mismatched. The SDR asks each respondent who identifies themselves as in a job that does not at all relate to their education what is the most important reason why they have taken such a job. The reasons (with percentages giving that reason in parentheses) include: better pay and promotion opportunities (20.3 percent, better working conditions (4.2 percent), the job’s location (4.7 percent), family-related reasons (5.8 percent), a job is not available in their doctoral field (24.5 percent), a change in career/professional interests (28.6 percent), or other (11.9 percent). The vast majority of workers cite one of three responses: better pay and promotion, the lack of jobs, or changed interests. We transform these into a series of dummy variables to replace the previous general measure of job and education not at all related. Table 7.6 summarizes earnings regressions in which the reasons for mismatch replace simply being in a job not very closely related to one’s education. It is clear that a change in career interests is associated with the smallest decline in earnings, while being mismatched to improve pay and promotion opportunities shows up increasing earnings. Once these are controlled for, the other reasons for being mismatched are associated with remarkably large declines in earnings. Approximately one-quarter of those in jobs unrelated to their education identify no jobs in their field and 7.3 percent of all respondents indicate that their jobs are unrelated to their education. Thus, slightly less than 2 percent of all doctoral recipients report being mismatched because of a lack of jobs in their field. These recipients earn almost 20 percent less. The differences by gender are dramatic. In general, the extremes are more pronounced in the female sample. The gain associated with being mismatched because of a desire for better pay and promotion shows women earning nearly 21 percent more. At the same time, the loss associated with being mismatched because of family reasons is nearly 40 percent. Women who cite their family as a reason for their job and education not being related are clearly earning substantially less than what would otherwise be the case. The augmented job satisfaction estimates reveal that mismatch is never associated with greater job satisfaction even when it is associated with greater pay or promotion opportunities. All but one of the coefficients in the right-hand side of table 7.6 are negative, including that on pay and promotion opportunities. This pattern suggests that those who leave their doctoral discipline behind in order to earn more are not pleased with the decision (the coefficient being strongly significant for males). The need to meet

Educational Mismatch among Ph.D.s Table 7.6

Selected earnings and job satisfaction regressions results: Simple and augmented for reasons for mismatch Log annual earnings Male

Job-education somewhat related Job-education not at all related R2 or 2 statistic Job-education somewhat related Pay and promotion opportunities Working conditions Job location Family-related reasons Job not available in field Change in career/professional interests Other reason R2 or 2 statistic

245

Female

Job satisfaction Male

Female

Specification 1 0.064*** 0.096*** (6.24) (5.01) 0.130*** 0.173*** (6.20) (4.17) 0.392 0.382

0.143*** (15.00) 0.191*** (11.83) 981.1***

0.124*** (8.28) 0.177*** (6.86) 356.8***

Specification 2 0.064*** 0.095*** (6.26) (4.98) 0.041 0.187** (1.17) (2.19) 0.259*** 0.250 (2.96) (1.43) 0.340*** 0.253** (4.71) (2.16) 0.228** 0.481*** (2.50) (4.07) 0.249*** 0.219*** (6.02) (2.68) 0.086** 0.166** (2.36) (2.16) 0.177** 0.241** (2.49) (2.14) 0.395 0.368

0.144*** (15.08) 0.141*** (4.64) 0.006 (0.08) 0.253*** (4.16) 0.290*** (5.56) 0.338*** (16.85) 0.046 (1.56) 0.255*** (6.16) 1072.4***

0.124*** (8.24) 0.066 (1.05) 0.112 (1.16) 0.055 (0.51) 0.193*** (3.41) 0.358*** (11.63) 0.007 (0.14) 0.280*** (5.91) 404.8***

Notes: Numbers under coefficient estimates are t-statistics. Marginal effects from the job satisfaction ordered probit regressions are shown for the probability of being in the highest job satisfaction category. The probability for males and females to be in the highest job satisfaction category (based on average characteristics) is 0.536 and 0.502, respectively. All variables from table 7.4 are also included in the estimations. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

budget expenses or provide a better standard of living for one’s family may come at the cost of not working in an area the scientist would rather pursue.6 Indeed, Preston (2004) highlights just this kind of tradeoff in exploring those scientists who choose to work outside of science. While it need not be the case that all of those mismatched in our sample have left science, 6. While possible, we also note that these estimates are based on a cross-section and we do not observe the counterfactual of these same workers actually pursuing jobs closely related to their education.

246

Keith A. Bender and John S. Heywood

working outside one’s education in order to earn more comes with reduced job satisfaction. When looking within genders, it is clear that even for women who reported a more than 20 percent increase in earnings associated with working outside their educational field in order to improve pay and promotion opportunities, the move brought no improvement in job satisfaction. More generally, to the extent that job satisfaction can be taken as an indicator of utility, the fact that mismatch is never associated with increased utility (even when associated with increased income) further illustrates its high social costs.7 7.4 Using the Panel Data One of the important aspects of the SDR is its longitudinal design. While not all variables are available in all years, we undertake a series of panel estimates to confirm the largely cross-sectional results of the previous sections. In particular, we are concerned that workers who are less productive will naturally earn less and, perhaps, be less satisfied and subject to greater turnover. Yet the fact that they are less productive also makes such workers more likely to be mismatched. To the extent that our controls in the previous cross-sectional analysis do not capture differences in productivity, the associations we have shown run the risk of merely reflecting fixed worker effects rather than the true influence of an exogenous mismatch. This may be especially true if being matched, closely related to one’s education, is the result of winning a tournament. In order to examine this issue, we use five waves of the public-use SDR sample for 1993, 1995, 1997, 1999, and 2001. As mentioned, job satisfaction is only available in 1997 and the finest breakdown by discipline is unavailable in the public-use data. An additional limitation of the public-use data is that annual earnings are rounded to the nearest thousand and topcoded at $150,000. Recognizing these limitations, we use the public-use data to estimate an earnings equation across these five waves (in 2001 dollars). Accounting for individual fixed effects, the unbalanced panel estimation examines the role of mismatch on earnings. The variables that are constant across waves (such as race and gender) drop out of the estimation, while the coefficients on the mismatch variables reflect the consequences of individual workers changing in the degree of mismatch. However, these coefficients should be purged of the role of individual effects, such as lower productivity, that might simultaneously generate both mismatch and lower earnings. 7. The augmented estimation of the turnover equations were also estimated. There were fewer statistically significant reasons for mismatch. The importance of the absence of jobs and the change in career interests emerged as positive determinants of turnover. There were no significant negative determinants of turnover and there were few differences by gender.

Educational Mismatch among Ph.D.s

247

Table 7.7 presents the estimations and reveals that all of the coefficients on the mismatch variables remain negative and all but one statistically significant. This happens despite the errors in variables associated with the public-use sample. Academics no longer show a larger earnings penalty for mismatch. The penalty for education and job not relating at all is essentially the same for men and women. Most notably, and as anticipated, controlling for fixed effects causes the coefficients themselves to be smaller. Indeed, the percentage penalty associated with mismatch is roughly half the size, on average, as that estimated in the cross-section. We are quick to emphasize that this should be taken with care as the specification and construction of the variables differ between the two sets of estimations. Nonetheless, the fixed effect estimates emerge as smaller, but still negative and typically statistically significant. In addition to controlling for fixed effects, the panel data can help inform the extent to which mismatch reflects temporary disequilibrium. Rubb (2003) has estimated that in general samples, less than one in five overeducated workers moves to being matched within one year. Beginning with the group of workers that are in jobs either not related or only somewhat related to their education in 1997, 25.7 percent reported closely related jobs in 1999 and slightly more, 27.0 percent, reported closely related jobs four years later in 2001. Thus, the vast majority of those mismatched in the core year of our study remain so four years later, suggesting that mismatch and

Table 7.7

Mismatch earnings penalties from fixed effects regressions

Variable

Full

Academic

Nonacademic

Female

Male

Job-education somewhat related Job-education not at all related

0.014*** (3.71) 0.066*** (9.07)

0.011** (2.01) 0.026** (2.00)

0.015*** (2.63) 0.064*** (6.66)

0.009 (1.02) 0.056*** (3.29)

0.008* (1.82) 0.052*** (6.56)

R2: within R2: between R2: overall Variance of fixed effect Variance of the error term 

0.141 0.373 0.328 0.447 0.306 0.681

0.139 0.413 0.371 0.423 0.257 0.730

0.098 0.266 0.239 0.497 0.321 0.706

0.139 0.174 0.166 0.525 0.363 0.677

0.149 0.302 0.264 0.436 0.290 0.693

Notes: Variables are in comparison to those who report their job and education are closely matched. The number under the coefficient estimates are t-statistics. Data are from the 1993, 1995, 1997, 1999, and 2001 SDR public-use files. Earnings are in 2001 U.S. dollars. Other controls include: experience, experience squared, supervises individuals, full-time contract, U.S. citizenship, main activity is teaching, main activity is management, main activity is computer work, main activity is other, and principle job is mid/top level manager. The proportion of the variance due to the fixed effect is . ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

248

Keith A. Bender and John S. Heywood

its consequences are persistent.8 Yet, focusing on job-changers presents a rather different picture. First, of those who closely relate in 1997, 13.2 percent report a job change by 1999. This compares to 22.4 percent of those who did not relate in 1997 reporting a job change by 1999. Second, 48.7 percent of those job-changers who did not relate in 1997 report improving the quality of their match by 1999. Thus, there remains movement out of mismatch toward better matches. This movement, together with the fixed effect estimate, suggests that mismatch indicates something beyond merely a proxy for low ability and losing a tournament. 7.4.1 The Influence of Experience on the Penalty for Mismatch The complementarity between education and the experience gained working in the field of one’s education may generate a quality and quantity of human capital not generated for those who are mismatched. As this process of accumulating human capital and receiving a return on that investment takes time, it seems sensible that the penalty to being mismatched may be small in the early years of a career but grow larger in the later years of a career. To test this hypothesis we added interactions between the mismatch indicators, how closely one’s education and job relate, and the experience and experience squared variables. These interactions were added to the earnings specification. The results provide broad confirmation that the penalty associated with mismatch is larger for more experienced workers. The upper panel of figure 7.1 shows the predicted log earnings from a cross-sectional estimate for a hypothetical female worker with mean female characteristics while the lower panel repeats the exercise for a hypothetical male worker with mean male characteristics. The pattern generally shows the lower earnings for those in jobs that are only somewhat related and the still-lower earnings of those in jobs that do not relate at all. As anticipated, the size of the penalty associated with mismatch for older workers typically exceeds that for younger workers. Men face a small penalty early in their careers for being mismatched, but it grows dramatically. At thirty years of experience, the penalty for being in a job that does not relate at all to education is $14,267 (a penalty of 18.2 percent), compared with only $2,883 (5.6 percent) at five years of experience. Women face a penalty that starts larger than men, but the penalty does not grow much over time. Thus, at thirty years of experience, the penalty for being in a job that does not relate at all is $9,570 (16.2 percent), compared with $6,446 (15.1 percent) at five years of experience. The greatest penalty occurs at $10,161 (17.2 percent) at twenty-three years of experience. Relatedly, the penalty for women whose job only somewhat 8. The base group of workers in 1997 differs between these two years because of attrition that is substantial in the SDR. We did examine the determinants of those who dropped out between 1997 and 1999, finding that mismatch itself is not a significant determinant.

Educational Mismatch among Ph.D.s

249

Fig. 7.1 Cross-section based experience-earnings profiles by gender and degree of mismatch (top, female; bottom, male)

relates compared to those whose job closely relates to their education increases substantially as experience increases, from $837 (1.6 percent) at five years of experience to $8,089 (10.3 percent) at thirty years of experience. On the other hand, it might be argued that these cross-sectional estimates identify differences across cohorts but give no guide as to what happens when a single cohort ages. We attempted to repeat the unbalanced panel wage estimations with the experience-mismatch interactions for the

250

Keith A. Bender and John S. Heywood

five waves taken from the public-use sample. The coefficients on the interactions with experience come from the actual aging of workers as fixed effects hold worker specific determinants constant. Note, however, that the experience variables generally represent a trend across the waves and so represent change over a decade. The estimations revealed no statistically significant interactions. This might suggest that correlations with unmeasured ability generated the differences over the experience profile shown in the cross-section. On the other hand, the constructed measure of experience required in the panel, the limitations of the public-use data used in the panel, and the relatively short time of the panel all suggest that few firm conclusions should be drawn. 7.5 Determinants of Mismatch In an attempt to describe the mismatched we return to our broad mismatch measures and estimate their determinants. For each of these variables, we use the broad controls examined to date as potential determinants. As each of the mismatch variables has three ordered responses, the estimations follow an ordered probit specification measuring the degree of mismatch. In general, the results suggest that mismatch is likely to result from a dating of scientific skills. The first column of table 7.8 estimates the extent of mismatch as measured by the closeness of job and education. The results indicate that the likelihood of mismatch increases with experience but at a decreasing rate. However, the coefficients are such that the likelihood of mismatch increases (the net coefficient across both terms is positive) until thirty-six years of experience. Thus, throughout most of a scientist’s work life the chance of mismatch increases with experience. This pattern is reinforced by the specifications using the other measures of mismatch and tends to indicate that mismatch is associated with retaining an older vintage of scientific knowledge and skills. Married men, those working full time, and supervisors are less likely to report being mismatched. Similarly, those in academia and government are less likely to report being mismatched. Compared to those who primarily do research, those who have a primary activity of teaching are less likely to be mismatched, but those who manage are more likely to be mismatched. The latter correlation would tend to support the commonly observed career path within private industry where scientists move from the lab to the front office, or within academia where they leave the lab for administration. Again, this may reflect the vintage of scientific skills. This transition might be evidenced by the fact that those in the largest employers are more likely to be mismatched. These workplaces have the largest internal labor markets and longest career ladders and can more easily make such transitions available to scientists. Interestingly, those who have a primary

Table 7.8

Determinants of mismatch

Single female Married female Married male Asian, non-Hispanic Hispanic Black, non-Hispanic Other race, non-Hispanic Experience Experience squared Supervisor Full time Pension Health insurance Profit sharing Employer size  1,000 and  5,000 Employer size  5,000 Academic Government Teaching is main activity Management is main activity Computer is main activity Other main activity Computer discipline

Job and education not related

Expectations not similar

6.9E-4 (0.23) 0.004 (1.56) 0.009*** (3.91) 1.4E-3 (0.67) 0.002 (0.42) 0.004 (0.79) 0.006 (0.48) 1.2E-3*** (4.96) 1.8E-5*** (2.88) 0.012*** (7.51) 0.006* (1.87) 0.013*** (4.65) 0.003 (1.08) 0.004** (2.22) 0.006 (1.35) 0.007*** (3.66) 0.035*** (15.03) 0.009*** (4.12) 0.012*** (5.87) 0.039*** (10.07) 0.060*** (8.32) 0.023*** (6.80) 0.257*** (8.57)

0.016* (1.68) 0.007 (0.88) 0.021*** (3.11) 0.015** (2.48) 0.002 (0.16) 3.5E-4 (0.03) 0.003 (0.12) 0.008*** (10.87) 1.2E-4*** (6.65) 0.023*** (5.02) 0.030*** (3.00) 0.010 (1.32) 0.002 (0.20) 6.0E-4 (0.10) 0.010 (0.91) 0.004 (0.67) 0.112*** (17.39) 0.006 (0.77) 0.021*** (3.24) 0.154*** (15.79) 0.143*** (8.69) 0.072*** (8.26) 0.253*** (8.53)

Likelihood of not adopting same field 0.002 (0.30) 0.005 (0.82) 0.018*** (2.90) 0.071*** (10.84) 0.011 (1.05) 0.003 (0.31) 0.012 (0.44) 0.002*** (3.34) 1.1E-4*** (6.26) 0.030*** (7.07) 0.008 (0.98) 0.033*** (4.57) 0.025*** (3.19) 0.016*** (3.02) 0.014 (1.34) 0.016*** (3.19) 0.039*** (6.72) 0.009 (1.28) 0.016*** (2.80) 0.042*** (5.60) 0.069*** (5.16) 0.034*** (4.78) 0.021 (1.17) (continued )

252

Keith A. Bender and John S. Heywood

Table 7.8

(continued)

Mathematics discipline Hard science discipline Social science discipline Engineering discipline Management science discipline Health science discipline Teacher discipline Other (nonscience) discipline Chi Squared

Job and education not related

Expectations not similar

Likelihood of not adopting same field

0.051*** (3.79) 0.044*** (5.36) 0.015*** (3.21) 0.077*** (5.71) 0.146*** (7.74) 0.232*** (7.90) 0.148*** (6.44) 0.411*** (12.22) 4,894.7***

0.028 (1.51) 0.047*** (3.01) 0.068*** (5.35) 0.088*** (4.69) 0.208*** (9.12) 0.114*** (4.53) 0.100*** (4.37) 0.449*** (15.80) 4,957.5***

6.0E-4 (0.04) 0.027** (2.07) 0.011 (0.90) 0.007 (0.54) 0.013 (0.87) 0.003 (0.16) 0.027 (1.60) 0.092*** (4.52) 888.8***

Notes: The number under the coefficient estimates are t-statistics. The coefficients are the marginal effects of being in the most mismatched category from ordered probit estimations. All estimates use sample weights and include controls for the region of the country. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

activity of using the computer are also more likely to be mismatched. This may reflect that some of those who are mismatched have been forced to reduce the scope of their knowledge to a single activity, often in a supportive rather than lead research role. Finally, the pattern of fields may be seen as further evidence on the role of vintage. The fields that are least likely to be mismatched are economics (the base category) and other social science, while those among the most likely to be mismatched are hard science, computer science, and health science. Such a pattern might be expected if the vintage of knowledge is less crucial in the former fields than in the latter fields. Another way of putting this is that the speed of change in the former fields is much slower than in the latter fields, making mismatch less likely. 7.6 Conclusions This chapter has examined the consequences of job mismatch—lack of fit between education and jobs—among the most highly-educated workers in the economy. These workers of the knowledge economy are often thought

Educational Mismatch among Ph.D.s

253

to be critical for technological progress and growth. Understanding the consequences of mismatch is important when considering governmental policies toward scientific education, including issues of subsidizing students, supporting universities, and encouraging technology transfer. The evidence assembled here uses three related measures of mismatch from the Survey of Doctoral Recipients and estimates their influence on three job outcomes: earnings, job satisfaction, and turnover. Mismatch is associated with worse outcomes: lower wages, lower job satisfaction, and higher turnover. This persists across substantial variations in estimation and holds for academics and nonacademics and for men and women. The size of these influences is surprisingly large, including a double-digit reduction in earnings, a 20 percent increase in the likelihood of being dissatisfied, and a one-third increase in the turnover rate. The fixed effect panel estimates suggest a statistically significant earnings penalty of about half the size estimated in the cross-section. While this chapter has not tried to estimate rates of return (either public or private) on scientific education, one cannot help but be concerned about the implications of these findings. The primary mismatch variable indicates one in six academics report a degree of mismatch and nearly one in two nonacademics report a degree of mismatch. Given the large share of mismatched workers and the apparently severe consequences of mismatch, a thorough review of policy options would seem in order. Our attempts to estimate the determinants of mismatch suggest that there may be substantial vintage effects at work as the fields in which the knowledge base changes most quickly appear to be associated with a greater chance of being mismatched. Also, there appears to be confirmation of the typical pattern of moving from the first line of science research into more administrative positions as scientists age. Moreover, the influences of the reasons for mismatch are particularly interesting. Those who are mismatched in an attempt to earn more, do earn more but remain less satisfied. This is intriguing and suggests greater inquiry into exactly what these individuals do and whether they regret the decision to remain matched. If they do, it might be worth considering the options for recreating matches.

References Adams, S. J. 2004. Employer-provided health insurance and job change. Contemporary Economic Policy 22 (3): 357–69. Allen, J., and R. van der Velden. 2001. Education mismatches versus skill mismatches. Oxford Economic Papers 53:434–52. Battu, H., C. Belfield, and P. Sloane. 1999. Over-education among graduates: A cohort view. Education Economics 7 (1): 21–38.

254

Keith A. Bender and John S. Heywood

Belfield, C. R. 2000. Economic principles for education: Theory and evidence. Cheltenham, UK: Edward Elgar. Belfield, C. R., and R. D. F. Harris. 2002. How well do theories of job matching explain variations in job satisfaction across educational levels? Evidence for UK graduates. Applied Economics 34 (5): 535–48. Belman, D., and J. S. Heywood. 1997. Sheepskin effects by cohort: Implications of job-matching in a signaling model. Oxford Economic Papers 49 (4): 623–37. Bender, K. A., and J. S. Heywood. 2006. Job satisfaction of the highly educated: The role of gender, academic tenure and comparison income. Scottish Journal of Political Economy 53 (2): 253–79. Borghans, L., A. Bruinshoofd, and A. de Grip. 2000. Low wages, skills and the utilization of skills. In The Overeducated Worker, ed. L. Borghans and A. de Grip, 191–202. Cheltenham, UK: Edward Elgar. Buchel, F. 2002. The effects of overeducation on productivity in Germany: The firm’s viewpoint. Economics of Education Review 21:263–75. Chevalier, A. 2003. Measuring over-education. Economica 70 (3): 509–31. Clark, A. E. 1997. Job satisfaction and gender: Why are women so happy at work? Labour Economics 4 (4): 341–72. Clark, A. E., and A. J. Oswald. 1996. Satisfaction and comparison income. Journal of Public Economics 61:359–81. Dolton, P., and A. Vignoles. 2000. The incidence and effects of over-education in the UK graduate labor market. Economics of Education Review 19:179–98. Freeman, R. 1976. The over-educated American. New York: Academic Press. ———. 1978. Job satisfaction as an economic variable. American Economic Review 68 (2): 135–41. Groot, W. 1993. Overeducation: The returns to enterprise related schooling. Economics of Education Review 12:299–309. Groot, W., and H. Maasen van den Brink. 2000. Over-education in the labor market: A meta-analysis. Economics of Education Review 19 (2): 149–58. Hamermesh, D. 2000. The changing distribution of job satisfaction. Journal of Human Resources 36 (1): 1–40. Kim, B., and H. Oh. 2002. Economic compensation compositions preferred by R&D personnel of different R&D types and intrinsic values. R & D Management 32:47–59. Lazear, E. P. 1979. Why is there mandatory retirement? Journal of Political Economy 87 (6): 1261–84. Levin, S., and P. Stephan. 1991. Research productivity over the life cycle: Evidence for academic scientists. American Economic Review 81 (1): 114–32. McGoldrick, K., and J. Robst. 1996. Gender differences in overeducation: A test of the theory of differential overqualification. American Economic Review 86 (2): 280–84. Moshavi, D., and J. R. Terborg. 2002. The job satisfaction and performance of contingent and regular customer service representatives: A human capital approach. International Journal of Service Industry Management 13 (4): 333–47. National Science Board. 2003. The science and engineering workforce: Realizing America’s potential. Arlington, VA: National Science Foundation. Preston, A. E. 2004. Leaving science: Occupational exit from scientific careers. New York: Russell Sage Foundation. Ransom, M. 1993. Seniority and monopsony in the academic labor-market. American Economic Review 83 (1): 221–33. Rubb, S. 2003. Overeducation: A short or long run phenomenon for individuals? Economics of Education Review 22 (4): 389–94.

Educational Mismatch among Ph.D.s

255

Sattinger, M. 1993. Assignment models of the distribution of earnings. Journal of Economic Literature 31 (2): 831–80. Sloane, P. J., H. Battu, and P. Seaman. 1996. Over-education and the formal education/experience and training trade-off. Applied Economics Letters 3:511–15. Solomon, L. C., L. Kent, N. L. Ochsner, and M.-L. Hurwicz. 1981. Underemployed Ph.D.s. Lexington, MA: D.C. Heath and Company. Stephan, P. 1996. The economics of science. Journal of Economic Literature 34 (3): 1199–1235. Stephan, P., and S. S. Everhart. 1998. The changing rewards to science: The case of biotechnology. Small Business Economics 10 (2): 141–51. Stephan, P., and S. Levin. 1991. Ph.D. supply. Issues in Science and Technology 7:28–29. Teitelbaum, M. S. 2004. Do we need more scientists? In The U.S. scientific and technical workforce: Improving data for decision making, ed. T. K. Kelly, W. P. Buts, S. Carroll, D. M. Adamson, and G. Bloom, 11–20. Santa Monica, CA: RAND Corporation. Thurow, L. 1975. Generating inequality. New York, NY: Basic Books. Tsang, M. C. 1987. The impact of underutilization of education on productivity: A case study of the U.S. Bell companies. Economics of Education Review 6 (3): 239–54. Tsang, M. C., and H. Levin. 1985. The economics of overeducation. Economics of Education Review 4 (2): 93–104. Ward, M., and P. J. Sloane. 2000. Non-pecuniary advantages versus pecuniary disadvantages: Job satisfaction among male and female academics in Scottish universities. Scottish Journal of Political Economy 47 (3): 273–303. Wolbers, M. 2003. Job mismatches and their labor-market effects among schoolleavers in Europe. European Sociological Review 19:249–66.

8 Capturing Knowledge The Location Decision of New Ph.D.s Working in Industry Albert J. Sumell, Paula E. Stephan, and James D. Adams

8.1 Introduction The placement of newly-minted science and engineering Ph.D.s provides one means by which knowledge is transferred from the university to industry. The placement of Ph.D.s with industry can be especially important in facilitating the movement of tacit knowledge. Despite this role, we know very little about industrial placements. One dimension of ignorance involves the extent to which students stay where trained or leave the area/ state after receiving the degree. The policy relevance of this question is obvious. Creating a highly-skilled workforce is one of several ways universities contribute to economic growth (Stephan et al. 2004). The mobility of the highly educated affects the extent to which knowledge created in uniAlbert J. Sumell is an assistant professor of economics at Youngstown State University. Paula E. Stephan is a professor of economics at the Andrew Young School of Policy Studies at Georgia State University and a research associate of the National Bureau of Economic Research. James D. Adams is a professor of economics at the Rensselaer Polytechnic Institute and a research associate of the National Bureau of Economic Research. The authors wish to thank Grant Black for comments and the provision of certain MSA data. Financial support for this project was provided by the Andrew W. Mellon Foundation, the Science and Engineering Workforce Project, National Bureau of Economic Research, and the National Science Foundation, grant number 0244268. We have benefited from the comments of participants at the REER conference, Georgia Institute of Technology, November 2003, the NBER meeting on the Economics of Higher Education, fall 2003, and the NBER meeting of the Science and Engineering Workforce Project, fall 2005. We have also benefited from comments of seminar participants at Université Jean Monnet (St. Etienne, France), Université Pierre Mendes France (Grenoble, France), Katholieke Universiteit Leuven (Leuven, Belgium), The European Forum, Robert Shuman Center (San Domenico, Italy), and the Universitat Pompeu Fabra (Barcelona, Spain). Mary Beth Walker, Rene Belderbos, and Bill Amis made helpful comments on an earlier draft. The use of NSF data does not imply NSF endorsement of the research methods or conclusions contained in this chapter.

257

258

Albert J. Sumell, Paula E. Stephan, and James D. Adams

versities is absorbed by the local economy.1 Having graduates work for neighboring firms strengthens the interface between the university and firms at the local or state level, and makes it easier for future graduates to find jobs with employers near the university. Moreover, the availability of a highly-trained workforce attracts new businesses to the local area. To the extent that students “fly the coop,” one rationale for investing state and local resources in universities is weakened. This is especially the case in today’s environment when universities, in an effort to attract resources, herald the role they play in local economic development, mindful of Stanford’s role in the creation of Silicon Valley, MIT and Harvard’s role in Route 128, and Duke and the University of North Carolina’s role in the Research Triangle Park (Link 1995).2 The migration behavior of the highly educated thus not only has longterm implications for the economic health of a region, but also may affect the amount policymakers are willing to invest in higher education. The stakes are somewhat different for private institutions than for public institutions. Not beholden to the public sector for funding, it is less essential that private institutions demonstrate a local economic impact. Nonetheless, private institutions receive a number of benefits from the state and local area, not the least of which is tax-exempt status. This is not to say that universities are solely focused on keeping their graduates close at hand. Placements outside the local area are an indication of success, signaling that the university has the necessary connections and reputation to warrant more distant placements.3 Moreover, strong industrial placements, regardless of whether or not they are local, can enhance future funding opportunities with industry. They can also enrich the alumni base and thus potential donations to the university. The objective of this chapter is to examine factors that influence the probability that a highly skilled worker will remain local or stay in the state. Specifically, we measure how various individual, institutional, and geographic attributes affect the probability that new Ph.D.s going to industry 1. Ph.D.s working in industry clearly contribute more than knowledge transfer. Stern (1999) discusses industrial scientists’ interest in “Science,” which to continue Stern’s typology, leads to “Productivity” for the firm. The ability to engage in “Science” provides psychic rewards for the scientist. The productivity effects experienced by the firm result in part from the “ticket of admission” that the practice of “Science” provides the firm to the wider scientific community (Stern 1999, 11). We focus on the knowledge-transfer role here because of our interest in the interface between industry and academe and the geographical dimensions of this interface. 2. There is a culture in universities of expecting Ph.D.s going into academe to seek the best available positions, regardless of locale. Attitudes toward industrial placements are less clearcut. Stephan and Black (1999) find that in the field of bioinformatics, often faculty do not even know the name of the firms their students go to work for. 3. Mansfield’s work (1995) suggests that industry, when looking for academic consultants, is likely to use local talent for applied research, but focuses on getting the “best,” regardless of distance, when basic research is involved.

Capturing Knowledge

259

stay in the metropolitan area or state where they trained. Our study focuses on Ph.D.s who received their degree in one of ten fields in science and engineering (S&E) during the period 1997 to 1999. Data come from the Survey of Earned Doctorates, administered by Science Resources Statistics National Science Foundation. The chapter proceeds as follows. Section 8.2 provides a discussion of the role new Ph.D.s play in knowledge transfer. Section 8.3 briefly discusses the role of geographic proximity in promoting knowledge transfer. Section 8.4 offers a conceptual model of the individual decision to migrate. Section 8.5 discusses the data used for this study and provides some descriptive statistics on the migration of industrial Ph.D.s from metropolitan areas and states, focusing on the ability of metropolitan statistical areas (MSAs) and states to retain Ph.D.s produced in their region and/or import human capital from other regions. Section 8.6 gives the results from our empirical analyses and discusses the policy implications. Section 8.7 concludes by summarizing and discussing the key findings. 8.2 The Role of New Ph.D.s in Knowledge Transfer The transmission mechanism by which knowledge flows from universities to firms is varied, involving formal means, such as publications, as well as less formal mechanisms, such as discussions between faculty and industrial scientists at professional meetings. Graduate students are one component of the formal means by which knowledge is transferred. Much of graduate students’ training is of a tacit nature, acquired while working in their mentor’s lab. These new techniques, which cannot be codified, can be transmitted to industrial R&D labs through the hiring of recently-trained scientists and engineers. New hires also establish and reinforce existing networks between firms and university faculty whereby the firm can acquire more ready access to new knowledge being created in the university.4 The Carnegie Mellon Survey of R&D labs in manufacturing located in the United States asked respondents to rank the importance of ten possible sources of information concerning public knowledge for a recently completed major R&D project (Cohen, Nelson, and Walsh 2002). A four-point Likert scale was used. The ten sources included patents, publications/ reports, meetings or conferences, informal interaction, recently-hired graduates, licenses, cooperative/JVs, contract research, consulting, and personal exchange. The findings show that—across all industries—publications/ reports are the dominant means by which R&D facilities obtain knowledge from the public sector. Next in importance are informal information exchange, public meetings or conferences, and consulting. Recently hired 4. Networks have been found to relate to firm performance (Powell et al. 1998; Zucker and Darby 1997).

260

Albert J. Sumell, Paula E. Stephan, and James D. Adams

graduates show up in the second cluster, which, in the overall rankings, is lower than the first cluster of sources of public knowledge. In certain industries, however, 30 percent or more of the respondents to the Carnegie Mellon Survey indicate that recently hired graduates played at least a “moderately important” role in knowledge transfer. These industries are: drugs, mineral products, glass, concrete, cement, lime, computers, semiconductors and related equipment, and TV/radio. This finding likely relates to the relative importance of tacit knowledge in certain fields and the key role that graduate students play in the transmission of tacit knowledge.5 In a related study, Agrawal and Henderson (2002) interviewed sixtyeight engineering faculty at MIT, all of whom had patented and licensed at least one invention, asking them to “estimate the portion of the influence your research has had on industry activities, including research, development, and production” (53) that was transmitted through a number of channels. Consulting headed the list, with a weight of 25.1 percent, followed by publication at 18.5 percent. Placement of MIT graduates was a close third at 16.8 percent. 8.3 The Role of Geographic Proximity in Transmitting Knowledge Considerable research has focused on the role that geographic proximity plays in transmitting knowledge. Early work by Jaffe (1989), for example, used university research and development expenditures as a proxy for the availability of local knowledge spillovers as did work by Audretsch and Feldman (1996a, 1996b). More recent work by Feldman and Audretsch (1999), Anselin, Varga, and Acs (1997, 2000), and Black (2001) has followed suit, shifting the analysis from the state to the consolidated metropolitan statistical area (CMSA). In each study a significant relationship is found between the dependent variable, which is a measure of innovation, and the proxy measure for local knowledge. Zucker, Darby, and Brewer (1998) take a different path and examine the role that the presence of star scientists in a region play in determining the regional distribution of biotech-using firms. They find the number of active stars in the region to play an important role in determining firm activity. Moreover, the effect is in addition to the role played by general knowledge sources, as measured by a “top quality university” or number of faculty with federal support. Two recent studies use patent citations to examine the degree to which knowledge spillovers are geographically bounded. Thompson (2006) finds that inventor citations in the United States are 25 percent more likely to match the state or metropolitan area of their citing patent than are exam5. The second tier-ranking of graduates as a means of knowledge transfer reflects in part the fact that graduate students contribute indirectly through networking to several pathways of knowledge transfer (such as informal information exchange, public meetings or conferences, and consulting) that are listed separately on the questionnaire.

Capturing Knowledge

261

iner citations. Almeida and Kogut (1999) explore why patent citations are more regionally concentrated in certain areas than others, focusing on the semiconductor industry. They argue that the mobility of engineers plays a key role in explaining citation rates by region. Regions that have high interfirm mobility of inventors (as measured by inventor address) have higher rates of intraregional citation than regions with low interfirm migration. This suggests that “a driving force for local externalities in semiconductor design is the mobility of people” (Almeida and Kogut 1999, 906). These, and countless other studies, go a long way toward establishing that geographic proximity promotes the transmission of knowledge. They do not, however, address the extent to which knowledge spillovers are local. One of the few papers to examine this question was written by Audretsch and Stephan (1996) and examines academic scientists affiliated with biotech companies. Because the authors know the location of both the scientist and the firm, they are able to establish the geographic origins of spillovers embodied in this knowledge-transfer process. Their research shows that although proximity matters in establishing formal ties between university-based scientists and companies, its influence is anything but overwhelming. Approximately 70 percent of the links between biotech companies and university-based scientists in their study were nonlocal. Audretsch and Stephan also estimate the probability that the link is local. Here we extend the Audretsch–Stephan framework, examining the location decisions of recent graduates. We are particularly interested in knowing the degree to which available knowledge spillovers, as measured by the placement of Ph.D. students, are local and in knowing factors related to the “stickiness” of Ph.D.-embodied knowledge to the local area. 8.4 Determinants of Migration There is a vast literature examining factors that influence human migration, much of which owes its origin to the work of Sjaastad (1962), and that views migration as an investment decision. An individual will move if she or he perceives the present value of the stream of benefits resulting from the move, composed primarily of gains in real income, to be greater than the costs, composed of both pecuniary and psychic costs to moving. Here we are interested in modeling the decision of a Ph.D. headed to industry to locate outside the city (state) of training versus to stay in the city (state) of training. We assume that the new Ph.D. is interested in maximizing the present value of utility over the life cycle, where the utility function has arguments of both income and psychic attributes such as family wellbeing. The cost of moving involves psychic costs as well as monetary costs of relocation (some of which may be paid by the firm). We assume that the individual engages in search in an extensive way while in graduate school and thus does not forego actual income while looking for a job. Moreover,

262

Albert J. Sumell, Paula E. Stephan, and James D. Adams

we assume that capital markets are not perfect and thus individuals with little debt are more able to absorb the costs of moving than those with debt. We also assume that individuals with access to a wider network of information are more likely to move than are those with more limited access. Our model focuses on whether the Ph.D. leaves where she or he is trained. Three sets of explanatory variables are of interest: variables that reflect attributes of the state and local area, variables that reflect individual characteristics affecting the present value of the discounted stream of utility from moving compared to the present value of the discounted stream of utility from staying in the area, and variables that reflect field of training and institutional characteristics. From a policy perspective, we are also interested in knowing whether individuals trained at a private institution are more likely to leave than are individuals trained at a public institution. We are also interested in knowing whether in-state students, as measured by receiving one’s high school, college, and Ph.D. degrees in the same state, are more likely to stay. Attributes of the local area include the degree of innovative activity, job market prospects in industry for Ph.D.s, and the desirability of the location. Innovative activity is measured by such standard measures as patent counts, R&D expenditures, and so forth; desirability is measured by level of education and per capita income. Job market prospects for Ph.D.s in industry are measured by an index, explained later, that computes the employment absorptive capacity of the area. Personal characteristics affecting the net present value include age, marital status, and the presence of dependents. Variables that reflect wider access to networks include the rank of the department as well as whether or not the individual was supported on a fellowship during graduate school. We expect individuals who work full or part time during their last year in graduate school to be more connected to the local area and therefore more likely to stay. We also expect individuals who return to a job they held before coming to graduate school to be more likely to remain in the area. The assumption is that proximity plays a role in selecting the graduate program. Imperfect capital markets lead us to expect that individuals who leave graduate school with substantial debt face more constrained searches and thus are more likely to remain local. Preferences are also assumed to affect the decision to relocate. While difficult to measure, we make inferences concerning preferences based on the individual’s past pattern of mobility. 8.5 S&E Ph.D.s in Industry: Where They Come from and Where They Go Data for this chapter come from the Survey of Earned Doctorates (SED) administered by Science Resources Statistics (SRS) of the National Science Foundation (NSF). The survey is given to all doctorate recipients in

Capturing Knowledge

263

the United States, and has a response rate of approximately 92 percent. While the SED has always asked graduates whether they have definite plans to work with a firm, the identity and geographic location of the firm has only become available to researchers since 1997 and then only in verbatim form. We have recently used these verbatim files to code the identity of the firm for the period 1997 to 1999. The analysis is thus restricted to Ph.D.s in science and engineering who made a definite commitment to an employer in industry between 1997 and 1999. This undercounts Ph.D. placements in industry in two notable ways. First, many Ph.D.s who eventually end up working in industry initially take postdoctoral appointments, particularly Ph.D.s in the life sciences. Secondly, 37.1 percent of Ph.D.s who were immediately planning to work in industry did not list a specific firm or location because they had not made a definite commitment to an employer at the time the survey was administered.6 Our results are thus conditional on the acceptance of a position with industry at the time the survey was completed and do not apply to all Ph.D.s headed to industry. The fields of training of the 10,121 new Ph.D.s with definite plans to work in industry are given in table 8.1. Not surprisingly, the data is dominated by large fields having a tradition of working in industry as well as a tradition of not accepting a postdoc position prior to heading to industry. Engineers made up 53 percent of the sample; 12 percent of the sample is made up of chemists. For Ph.D.s who had made a definite commitment to an employer in industry and identified the specific name of the firm they plan to work for between 1997 and 1999, 36.7 percent had commitments with an employer that lay within the same state as their doctoral institution.7 The stay rate is low compared to that for bachelor’s and master’s degree recipients in science and engineering. The National Science Foundation reports that 62 percent of all recent bachelors in science and engineering in the United States stay in the state where they received their degree and 60.2 percent of all recent masters stay. The stay rate is highest for computer scientists (68.4 percent for bachelors and 70.8 percent for masters) and lowest for bachelors in engineering (55.1 percent and masters in the physical sciences (54.1 percent).8 The Ph.D. stay-rate of 36.7 percent is also low compared to recent law school graduates, for whom 57.0 percent with 6. Of the Ph.D.s awarded in the twelve broad S&E fields during this time period, 17,382 of the 75,243 had plans to work in industry. Of these, 10,932 (14.5 percent of all Ph.D.s in S&E during this time period) had made a definite commitment to an employer in industry and identified the specific name of the firm they planned to work for. Of these, 10,121 Ph.D.s were awarded by institutions in the continental United States in one of ten “exact” S&E fields. 7. The percent is based on the 10,932 referred to in footnote 6, which includes Ph.D.s trained in psychology and economics, as well as the ten fields listed in table 8.1. 8. The data are not strictly comparable since the NSF data include U.S. degree recipients who also received a high school diploma or equivalency certificate in the United States.

264

Albert J. Sumell, Paula E. Stephan, and James D. Adams

Table 8.1

Firm placements of new S&E Ph.D.s by field of training: 1997–1999

Field All S&E fields All engineering Agriculture Astronomy Biology Chemistry Computer science Earth science Math Medicine Physics

% of all Ph.D.s awarded that identified a firm

% in field of Ph.D.s that identified a firm

14.5 30.7 9.0 7.8 3.8 18.7 28.4 12.3 12.5 5.0 16.1

100 (n  10,121) 53.0 (n  5,364) 3.0 (n  308) 0.4 (n  44) 6.0 (n  609) 12.0 (n  1,216) 7.5 (n  762) 2.5 (n  252) 4.7 (n  477) 4.3 (n  435) 6.5 (n  654)

known employment status remain in the state of training (National Association for Law Placement 1998). The low stay-within-state rate does not necessarily indicate that the production of new Ph.D.s is entirely a poor investment from the perspective of state policymakers. In the majority of states (twenty-six), one-third or more of all newly employed Ph.D.s hired by in-state firms graduated from an institution within the state, and in eight states, institutions within the state supplied the majority of new Ph.D.s to firms within the state. Table 8.2 displays interstate and interregional migration data.9 Several notable patterns become evident. Pacific states are major net importers of new Ph.D.s; approximately 40 percent more Ph.D.s have definite plans to work in California, Oregon, and Washington than are produced there. California dominates in several respects. More Ph.D.s going to industry are produced in California than in any other state, the state retains a higher percent of the Ph.D.s it produces than does any other state, and more Ph.D.s produced in other states head to California than to any other state. The strong presence of IT firms in Pacific states, especially during the period of study—as well as the heavy proportion of engineers in the database—no doubt contribute to this finding. New England and Middle Atlantic states train approximately the same number of Ph.D.s that they hire. If it were not for New Jersey, however, the Middle Atlantic region would be a net exporter. New Jersey’s remarkable gain is in large part due to its ability to attract new Ph.D.s from neighboring New York and Pennsylvania. New York provides other states or countries with 591 new industrial Ph.D.s, sending 115 of those to New Jersey 9. Six states (Alaska, Nevada, Hawaii, North Dakota, South Dakota, and Wyoming) either produced or received too few Ph.D.s to report their interstate migration numbers.

1,890 311 898 681

2,102 611 376 430 445 240

698a 168 106 270 97

East North Central Illinois Indiana Michigan Ohio Wisconsin

West North Central Iowa Kansas Minnesota Missouri

958 145 8 713 30 54 8

Number of new Ph.D.s trained in state/region

504a 47 47 266 109

1,346 441 166 308 314 117

1,998 766 801 431

885a 220 7 594 39 25 s

Number of new Ph.D.s working in state/region

794 179 46 142 147 45 244 27 24 99 27

36.0 27.8 55.9 28.4 29.4 51.3 27.8 72.0 55.7 1.5 12.4

7.6 51.7 12.5 16.7 30.0 53.7 s 923 142 307 163

415 43 s 259 9 8 s

Percentage gain or loss

5.7 146.3 10.8 36.7

Number of new Ph.D.s produced that stay in state/region

Interstate and Interregional migration patterns of new industrial Ph.D.s 1997–1999

Mid-Atlantic New Jersey New York Pennsylvania

New England Connecticut Maine Massachusetts New Hampshire Rhode Island Vermont

State/Region

Table 8.2

35.0 16.1 22.6 36.7 27.8

37.8 29.3 12.2 33.0 33.0 18.8

48.8 45.7 34.2 23.9

43.3 29.7 s 36.3 30.0 14.8 s

Percent of new Ph.D.s produced that stay in state/region

51.6 42.6 48.9 62.8 75.2 (continued )

41.0 59.4 72.3 53.9 53.2 61.5

53.8 81.5 61.7 62.2

53.1 80.5 s 56.4 76.9 68.0 s

Percent of new Ph.D.s imported from other states/regions

s

1,692 64 271 324 266 321 91 269 23 63

297 102 46 49 100

896 22

South Atlantic Delaware Florida Georgia Maryland North Carolina South Carolina Virginia West Virginia Washington, D.C.

East South Central Alabama Kentucky Mississippi Tennessee

West South Central Arkansas

37 20

Number of new Ph.D.s trained in state/region

(continued)

Nebraska North Dakota South Dakota

State/Region

Table 8.2

7

28

1,050 15

193 56 37 12 88

1,195a s 173 171 233 197 69 233 35 84

s

Number of new Ph.D.s working in state/region

712 s 93 91 63 90 19 81 s 7 97 28 s s 40

24.3 s s 29.4 s 36.2 47.2 12.4 38.6 24.2 13.4 52.2 33.3 35.0 45.1 19.6 75.5 12.0

491 8

12 s s

Percentage gain or loss

17.2 31.8

Number of new Ph.D.s produced that stay in state/region

54.8 36.4

32.7 27.5 s s 40.0

42.1 s 34.3 28.1 23.7 28.0 20.9 30.1 s 11.1

32.4 s s

Percent of new Ph.D.s produced that stay in state/region

53.2 46.7

49.7 50.0 s s 54.5

40.4 s 46.2 46.8 73.0 54.3 72.5 65.2 s 91.7

57.1 s s

Percent of new Ph.D.s imported from other states/regions

1,831a s 1,539 99 161 15

17

Pacific Alaska California Oregon Washington Hawaii

Other Puerto Rico 10,303

18

2,534 s 2,126 s 187 s

474a 181 154 29 9 38 47 14 s

78 49 908

n.a.

5.6 n.a.

13

1,270 s 1043 40 57 s

228 79 73 s s 16 27 s s

14.9 8.1 21.4 141.7 40.0 7.3 44.7 s s 39.7 s 38.1 s 16.1 s

26 27 366

18.8 49.0 33.1

n.a.

76.5

69.4 s 67.8 s 35.4 s

40.9 40.1 37.2 s s 39.0 31.8 s s

27.1 28.1 53.7

n.a.

27.8

50.2 s 50.9 s 69.5 s

51.9 56.4 52.6 s s 57.9 42.6 s s

66.7 44.9 59.7

Notes: s  suppressed. At the request of Science Resources Statistics, National Science Foundation (2005), counts not reported if 6 or less or if a specific firm contributes half or more of the count in a cell. Counts include Ph.D.s trained in economics and psychology. n.a.  not applicable. a Suppressed cells not included in sums to prevent identification of cells.

10,932

557a 197 196 12 15 41 85 s 11

Mountain Arizona Colorado Idaho Montana New Mexico Utah Nevada Wyoming

Sum/means U.S.

96 96 682

Louisiana Oklahoma Texas

268

Albert J. Sumell, Paula E. Stephan, and James D. Adams

alone. Pennsylvania is not far behind, losing 518 new industrial Ph.D.s to other areas—seventy-seven to New Jersey. States in the Midwest (East North Central and West North Central) are net exporters, hiring approximately one-third fewer Ph.D.s than they train. The brain drain is substantial. As a region, the Midwest retains slightly more than one-third of those trained, but retention within Midwestern states (as opposed to within the region) is considerably lower, averaging less than 28 percent. Indiana Ph.D.s are the most likely to find employment in other states. Of the 376 new industrial Ph.D.s graduating from Indiana universities in the three-year period, forty-six, a meager 12.2 percent, had definite plans to work for a firm in Indiana. Iowa is not far behind. A state’s ability to retain its highly-trained workers is largely contingent upon the strength of its metropolitan areas. More than 67 percent of new industrial Ph.D.s who remain in-state work in the same CMSA in which they were trained. Table 8.3 takes a closer look at the ability of metropolitan areas to retain new industrial Ph.D.s by examining the top twenty-five destinations and the top twenty-five producing metropolitan areas.10 Overall, slightly more than 70 percent of those trained in a CMSA were trained in a top twenty-five CMSA, while approximately 80 percent of those going to work in a metropolitan area go to a top twenty-five destination city. It is evident from table 8.3 that areas that produce more industrial Ph.D.s generally hire more Ph.D.s in industry. This is accomplished by both retaining Ph.D.s produced in the city and attracting Ph.D.s from other cities. Eighteen metropolitan areas are in the top twenty-five in terms of both producing and employing new Ph.D.s going to industry. Furthermore, slightly more than one out of every three Ph.D.s trained in a top twenty-five metropolitan area stays in the area of training, whereas only about one in five produced in all other metropolitan areas stays where trained. This suggests that a dynamic is at work: Cities that produce more highly-skilled workers foster the development of new firms and attract firms wanting access to a highly-skilled workforce. This in turn attracts more highly skilled workers from other areas and encourages retention of those trained in the area. Particularly interesting is the role of New York/Northern New Jersey, San Francisco/San Jose, Boston, Los Angeles, and the District of Columbia/ Baltimore. These five metropolitan areas (although not in the same order) represent the top five metropolitan areas, both in terms of destination and in terms of the production of Ph.D.s heading to industry. Slightly over one in four of all new S&E Ph.D.s headed to industry was trained in one of

10. Here we focus on Ph.D.s awarded in a CMSA; 1,027 of the new Ph.D.s headed to industry were trained outside a CMSA. Note also that the number of Ph.D.s produced in CMSAs is not equal to the number hired by a CMSA for three reasons: some work outside CMSAs in the United States, others leave the United States for industrial employment abroad, and others are trained outside a CMSA but work in a CMSA.

2,427a 564

34.1 20.3

32.7 19.4 3.3 7.7 28.7 42.0 13.6 37.5 s 17.4 30.4 18.9 45.1

38.8 44.4 48.9 3.2 33.6 42.1 25.9 23.8 2.9 32.3

58.9

57.8

% that stay

Sum top 25 metropolitan areas All other metropolitan areas

San Francisco-Oakland San Jose, CA New York-No. New Jersey-Long Island, NY-NJ-CT-PA Boston-Worcester-Lawrence-LowellBrockton, MA-NH NE Los Angeles-Riverside-Orange County, CA Washington-Baltimore, D.C.-MD-VA-WV Houston-Galveston-Brazoria, TX Chicago-Gary-Kenosha, IL-IN-WI Portland-Seattle-Tacoma, OR-WA Philadelphia-Wilmington-Atlantic City, PA-NJ-DE-MD Dallas-Fort Worth, TX Detroit-Ann Arbor-Flint, MI Minneapolis-St. Paul, MN-WI Austin-San Marcos, TX San Diego, CA Atlanta, GA Raleigh-Durham-Chapel Hill, NC Phoenix-Mesa, AZ Denver-Boulder-Greeley, CO Cincinnati-Hamilton, OH-KY-IN Albany-Schenectady-Troy, NY Pittsburgh, PA Cleveland-Akron, OH Indianapolis, IN St. Louis, MO-IL Rochester, NY MSA

Consolidated metropolitan area

2,540 453

86 46 102 86 67 55 73 51 35 54 27 24 42 42 0 25 17

296 273 241 233 182 159 150 144 121 120 109 105 101 96 81 81 63 7,750 1,812

238 233 160 48 122 68

423

416

# local

588 484 443 340 339 339

1,293

1,369

N

32.8 25.0

29.1 16.8 42.3 36.9 36.8 34.6 48.7 35.4 28.9 45.0 24.8 22.9 41.6 43.8 0.0 30.9 27.0

40.5 48.1 36.1 14.1 36.0 20.1

32.7

30.4

% local

Notes: s  suppressed. Counts of 6 or less not reported at the request of Science Resources Statistics, National Science Foundation (2005). Counts include Ph.D.s trained in economics and psychology. a Suppressed count not included in total to prevent identification of the suppressed count.

7,122 2,783

86 42 7 16 51 68 21 54

263 217 209 208 178 162 154 144 142 138 138 127 122

Sum top 25 metropolitan areas All other metropolitan areas

238 233 160 10 102 122 73 67 8 86

614 525 327 313 304 290 282 282 279 266

24 42 24 55

416

706

s

423

732

New York-No. New Jersey-Long Island, NY-NJ-CT-PA San Francisco-Oakland-San Jose, CA Boston-Worcester-Lawrence-LowellBrockton, MA-NH NE Los Angeles-Riverside-Orange County, CA Washington-Baltimore, D.C.-MD-VA-WV Champaign-Urbana, IL Detroit-Ann Arbor-Flint, MI Chicago-Gary-Kenosha, IL-IN-WI Atlanta, GA Austin-San Marcos, TX Lafayette, IN Minneapolis-St. Paul, MN-WI Philadelphia-Wilmington-Atlantic City, PA-NJ-DE-MD Pittsburgh, PA State College, PA Madison, WI Raleigh-Durham-Chapel Hill, NC Portland-Seattle-Tacoma, OR-WA Columbus, OH Denver-Boulder-Greeley, CO Greensboro-Winston-Salem-High Point, NC Albany-Schenectady-Troy, NY Cleveland-Akron, OH Tucson, AZ San Diego, CA

# that stay

N

Top 25 destination consolidated metropolitan areas

Top twenty-five producing and destination consolidated metropolitan areas: 1997–1999

Top 25 producing consolidated metropolitan areas

Consolidated metropolitan area

Table 8.3

270

Albert J. Sumell, Paula E. Stephan, and James D. Adams

these five metropolitan areas, while approximately three out of eight were headed to one of these five metropolitan areas.11 Table 8.3 also shows that striking disparity exists in the ability of metropolitan areas to retain new industrial placements. The New York and San Francisco areas top the list; each employs about 58 percent of new industrial placements trained in their area. On the other hand, areas like Urbana-Champaign, Illinois; Lafayette, Indiana; and State College, Pennsylvania, all of which have a long tradition of training scientists and engineers, retain only about 3 percent of their new Ph.D.s headed to industry. This high attrition rate demonstrates that the presence of a large university does not guarantee sufficient job opportunities in the industrial sector to retain S&E Ph.D.s trained locally. Certainly, other factors necessary for economic development, such as transportation nodes, nearby amenities, access to venture capital, and so forth, present in cities like San Jose, are lacking in cities like Urbana-Champaign.12 While the universities like Illinois-Urbana/Champaign, Purdue, and Pennsylvania State appear to have a low return on their investment in terms of the fact that new Ph.D.s leave the city upon graduating, they do supply new talent to the state and nearby metropolitan areas. The University of Illinois-Urbana/Champaign supplies Chicago with about 10 percent of its new industrial hires, Purdue University is far and away the top supplier to Indianapolis, accounting for 21 percent of that city’s industrial hires, and firms in Pennsylvania recruit 8 percent of their new Ph.D. talent from Pennsylvania State University. Table 8.4 shows how migration behavior differs by a Ph.D.’s field of training. While 36 percent of engineers, who constitute about half of all industrial S&E hires in our sample, stay in state, 26 percent have plans to stay in the same metropolitan area; both are close to the mean of all S&E industrial hires. Doctorates in agriculture have the lowest stay rates of all S&E fields, with about one in four staying in state, and less than one in ten with plans to work in the same metropolitan area they were trained in. This reflects in part the fact that Ph.D.s in agriculture on temporary visas are the most likely of any group of S&E Ph.D.s to leave the United States upon graduation (Black and Stephan 2003). By way of contrast, astronomers are 11. The extreme geographic concentration displayed in table 8.3 has been found using several other measures of innovation. For example, Black (2001) examined the geographic concentration of innovation using Small Business Innovation Research (SBIR) awards and patent counts. There is significant overlap with the Ph.D. metropolitan areas: the top five metropolitan areas in terms of SBIR phase II awards are the same as the top five areas in terms of industrial Ph.D.s produced and hired. Four of the five metropolitan areas are also in the top five in terms of utility patents issued (Chicago is fourth on the list, while the District of Columbia is eleventh). 12. The lack of a booming industrial sector could prove an asset in the long run. That is, “college towns” may indirectly use their small city size as a tool to attract niche industries as well as a highly-trained workforce, marketing the lack of disamenities that are present in cities with large industrial sectors, such as high crime rates, congestion, and air pollution.

Capturing Knowledge

271

Table 8.4 Percent of firm placements staying in state and consolidated metropolitan areas by field of training: 1997–1999 Field All engineering Agriculture Astronomy Biology Chemistry Computer science Earth science Math Medicine Physics All fields

% staying in state

% staying in CMSA

36.3 26.0 56.8 45.0 28.6 36.4 28.6 35.0 46.0 45.0 36.4

26.2 9.7 54.5 34.6 19.7 30.6 17.9 29.4 35.2 35.0 26.6

the most likely to work in the state and metropolitan area in which they trained. More than 56 percent of astronomers have employment plans to work in the state of training and about 55 percent have plans to work in the metropolitan area of their doctoral institution. 8.6 Empirical Results In order to investigate specific factors affecting the decision to stay in the area of training, we estimate two equations, using two definitions of staying. These equations are shown in table 8.5. In equation (1) we estimate the probability that a new Ph.D. has made a definite commitment to an industrial employer in the same state as their doctoral institution; the dependent variable in equation (2) is whether or not the new Ph.D. stays in the same primary metropolitan area.13 Both equations are estimated using a logit model. Table 8A.1 presents the definitions, means, and standard deviations for all variables included in the regressions. Table 8.5 provides the coefficients and z-statistics for the two equations. We restrict the analysis to Ph.D.s trained in the continental United States, excluding those trained in Alaska, Hawaii, and Puerto Rico. Table 8.5 also reports the marginal effects of a change in an independent variable, evaluated at the mean. For a dummy variable these marginal effects show by how much the probability will change with a change in status; in the case of a continuous variable, they show how much the probability will change with a one-unit change in the value of the variable. All Ph.D.s who did not report their postdoctoral state 13. The difference between CMSA and PMSA is one of size. Thus, while San Jose is a PMSA, the larger CMSA includes San Francisco and Oakland as well as San Jose. Because of issues related to confidentiality, we are not able to display the data at the PMSA level; however, we are able to analyze the data at this level.

Intercept age agesq female asian nonwhite_asian permres tempres married female_married wchild singlepar samece_phd samehs_phd sameb_phd return debtlevel preftemp preptemp supp_fellow supp_teachasst supp_RA_trainee supp_employer astr agri alleng chem math comp

Variable

Table 8.5

3.4812*** 0.0634 0.0004 0.0875 0.1498** 0.2188** 0.1335 0.2913*** 0.0671 0.2413* 0.0019 0.1479 0.4742*** 0.2609* 0.0747 0.4428*** 0.0057** 0.4087*** 0.8163*** 0.2600*** 0.0325 0.1125 0.0550 0.2647 0.8708** 0.3713** 0.6905*** 0.2930 0.5299**

Estimate 17.71 2.58 0.63 0.83 5.11 4.77 2.28 16.82 1.15 3.82 0.01 1.09 21.01 3.18 0.31 37.99 6.01 46.49 68.47 8.32 0.14 2.54 0.23 0.21 5.62 4.43 12.12 1.67 5.97

z-stata n.a. 0.0142 n.a. 0.0196 0.0336 0.0478 0.0306 0.0647 0.0151 0.0559 0.0004 0.0326 0.1112 0.0605 0.0170 0.1036 0.0013 0.0941 0.1974 0.0567 0.0074 0.0254 0.0125 0.0619 0.1660 0.0839 0.1407 0.0631 0.1099

Marginal effect

Equation (1): Dependent variable  sameSTATE (N  10,000)

Empirical results Sample  placements trained in the continental United States

3.2185*** 0.0637 0.0004 0.0785 0.2897*** 0.2385** 0.0296 0.4297*** 0.0952 0.0947 0.0034 0.1113 0.3410*** 0.1956 0.2966** 0.3455*** 0.0078*** 0.3443*** 0.8029*** 0.1616 0.0393 0.0570 0.0274 0.2034 0.6840 0.0348 0.2954 0.1751 0.1990

Estimate 12.43 1.93 0.41 0.45 13.84 4.02 0.08 25.55 1.61 0.41 0.01 0.44 8.63 1.41 3.89 17.63 7.55 22.57 55.42 2.33 0.14 0.47 0.05 0.09 0.99 0.03 1.65 0.45 0.66

z-stata

n.a. 0.0091 n.a. 0.0112 0.0412 0.0323 0.0043 0.0597 0.0137 0.0141 0.0005 0.0156 0.0530 0.0270 0.0465 0.0537 0.0011 0.0521 0.1432 0.0225 0.0057 0.0083 0.0040 0.0276 0.0796 0.0051 0.0398 0.0267 0.0273

Marginal effect

Equation (2): Dependent variable  SamePMSA (N  8,838)

1.1897*** 0.2376 0.2280 0.1078 0.0107 0.2423*** 0.4438** 0.3724** 0.2738 0.0394 0.4171* 0.5861*** 0.1874 0.0445 0.00041 0.000020 0.000026*** 0.000058*** 0.00012 0.0098 0.0413** 0.2286*** n.a. n.a. n.a. n.a. n.a. n.a. n.a.

13,117.0

Note: n.a.  not applicable. a z-stats are based on chi-square distribution. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

–2 Log-likelihood

earth medi phys topsastr topsagri topsalleng topsbiol topschem topscomp topsearth topsmath topsmedi topsphys private STpats STacadRD STindRD STsize STpop STperhe STpcinc ABPhDST pmsapats milkenind pmsapop pmsasize pmsapcinc pmsaperhe ABPhDMSA

12.94 1.04 1.09 0.02 0.01 10.88 4.98 6.52 2.41 0.01 3.82 6.72 1.08 0.60 0.54 0.30 11.85 68.53 0.45 0.63 4.37 7.54 n.a. n.a. n.a. n.a. n.a. n.a. n.a.

0.2093 0.0516 0.0497 0.0239 0.0024 0.0541 0.0929 0.0794 0.0592 0.0088 0.0875 0.1187 0.0433 0.0101 0.000092 0.000004 0.000006 0.000013 0.00003 0.0022 0.00933 0.0516 n.a. n.a. n.a. n.a. n.a. n.a. n.a.

1.2719*** 0.1371 0.0895 0.4210 0.1003 0.3268*** 0.2406 0.4651** 0.1882 0.0297 0.1820 0.5087** 0.1474 0.1814** n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. 0.00295*** 0.3645*** 0.00009*** 0.0333** 0.0030 0.0084 0.0966*** 9,496.5

8.67 0.28 0.13 0.29 0.02 12.89 1.15 6.49 0.87 0.00 0.55 3.97 0.49 6.00 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. 21.45 33.59 33.53 5.36 0.11 1.74 47.63

0.1226 0.0191 0.0133 0.0695 0.0141 0.0464 0.0325 0.0592 0.0258 0.0044 0.0249 0.0627 0.0223 0.0258 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. 0.00043 0.0529 0.000014 0.0048 0.00043 0.0012 0.0140

274

Albert J. Sumell, Paula E. Stephan, and James D. Adams

of location or age are excluded from equation (1); Ph.D.s whose doctoral institution does not lie in a U.S. primary metropolitan statistical area (PMSA), as well as those who did not report a readable city name or age are excluded from equation (2). Table 8.5 shows that, other things being equal, the market for Ph.D.s trained in certain fields is significantly less local than for other fields. Specifically, relative to the benchmark of biology, we find individuals trained in agriculture, engineering, chemistry, computer science, and earth science to be significantly more likely to leave the state of training. The effects, in many instances, are substantial, as can be seen by examining the marginal effects. With the exception of earth science, there are no significant differences at the PMSA level. Few of the demographic variables play a significant role in determining whether the new Ph.D.s stay in close geographic proximity to their institution of training. We do, however, find that Asians, as well as individuals who are underrepresented minorities in science and engineering (nonwhite, nonasian) are less likely to stay in the state or PMSA of training. The latter result may reflect the scarcity and hence wider market for underrepresented minorities receiving Ph.D.s in science and engineering. Being a temporary resident is also a key factor in determining mobility. Compared to citizens, temporary residents are considerably more likely to leave the state as well as to leave the local area. The effect is fairly sizable. Other things being equal, temporary residents are about 6 percent more likely to leave either the state or local area than are citizens. Married Ph.D.s are no more likely to remain in their location of training than are nonmarried Ph.D.s; neither does the presence of children affect mobility, nor is mobility related to being a single parent. However, other things being equal, we find that married women are more likely to stay in state than are unmarried women. There is no indication, holding marital status constant, that women have differential mobility patterns than do men. We also find no support for the hypothesis that mobility decisions are responsive to the present value of moving; in neither instance do we find the coefficients on either age or age-squared to be significant. Preferences as revealed through past mobility patterns play a significant role in determining the location decision. We find that doctorates who earned their Ph.D. in the same state as their college degree are much more likely to remain in the Ph.D.-granting state than are those who changed states between college and graduate school. They are also more likely to stay in the same PMSA. The marginal effects are not inconsequential. Other things being equal, “stayers” are about 11 percent more likely to take an industrial position in state and 5 percent more likely to take a position in the city of training. At the state level we find that individuals who receive their Ph.D. and college degree in the state from which they graduated high school are even more likely to remain in state than are those who moved to

Capturing Knowledge

275

the state to get a college degree and stayed on to receive their Ph.D. At the PMSA level, those who received their degree in the state in which they were born are significantly more likely to remain to take a position in industry. The policy implication is clear: accepting Ph.D. students from in state significantly raises the probably of retention of the highly-skilled work force. At the margin, the cumulative effect of training Ph.D.s who went to both high school and college in the state of doctoral training is 17 percent. For public institutions, this suggests that states capture part of their educational investment. Variables that reflect wider access to networks are generally significant and with the expected sign. Individuals whose primary source of support was a fellowship or dissertation grant are significantly more likely to leave the state of training than the benchmark.14 Individuals trained at top-rated programs15 also are more likely to move, although the effect is fielddependent as well as dependent on the measure of mobility. In five of the ten fields studied (engineering, biology, chemistry, math, and medicine), individuals trained at a top program are significantly more likely to leave their state than are individuals not trained at a top program in their field. And the marginal effects can be quite strong. Turning to equation (2), we find that four of the top program variables are negative and significant as well, suggesting that in smaller geographical areas graduates from top programs leave as well.16 Individuals who worked full or part time during their last year of graduate school are assumed to have more information, other things being equal, concerning jobs in close proximity to their graduate institution. Our results support this hypothesis. We find that those working full or part time are more likely to stay in state and in the primary metropolitan area. The effects are large. For example, those who worked part time their last year in 14. The benchmark is those whose primary source of support during graduate school was other than a fellowship, a dissertation grant, a teaching assistantship, a research assistantship, or employer reimbursement. 15. Top fields are based on the 1995 National Research Council (NRC) rankings for all fields except medicine and agriculture. The rankings for the majority of fields are based on the “scholarly quality” scores in the NRC rankings for each relevant program at the institution. For field definitions that were broader than the program definitions in the NRC rankings (such as biology), we calculated the mean for each rated program applicable to our broader field for each institution. For the fields of medicine and agriculture, we used the 1998 NSF Web CASPAR data to rank institutions, due to the absence of data for these fields in the NRC rankings. Institutions in these fields were ranked by total federal R&D expenditures at each institution. In the case of biology and medicine, which have a very large number of Ph.D. programs, seventy-five institutions were included among the top programs. For smaller fields, such as astronomy, the top category includes the top twenty-five programs. In most other fields, the top category includes the top fifty programs. 16. The engineering, chemistry, and math results persist when we restrict the definition of a top program to one that ranks in the top ten. In addition, using this more restrictive definition of quality, we find that individuals are more likely to leave the state of training if they matriculate from a top computer science or earth science program.

276

Albert J. Sumell, Paula E. Stephan, and James D. Adams

graduate school are 20 percent more likely to remain in state than are those who did not work part time, and 14 percent more likely to remain in the same PMSA. We also know from the SED whether a doctorate with definite plans is “returning to or continuing in pre-doctoral employment.” Not surprisingly, Ph.D.s who indicated they were returning to a previous employer are considerably more likely to remain where they were trained. The marginal effect is particularly strong at the state level (10 percent).17 Student debt level affects mobility, but not in the way hypothesized. Instead, we find that the probability of remaining in one’s location of training depends negatively upon the amount of debt accumulated in graduate school. This counterintuitive result may indicate that students who assumed debt engage in more search activity than do those with no debt, motivated by the need to find a highly remunerative position. Finally, we are interested in knowing the degree to which the attributes of the local area affect the decision to leave the state or metropolitan area. Here we examine two dimensions of this relationship: the presence of innovative activity and the desirability of the state or local area, as proxied by per capita income and educational attainment. At the state level, innovative activity is measured by the count of utility patents granted, as well as by industrial R&D expenditures and academic R&D expenditures.18 In the PMSA equations we use the Milken index and patent counts as measures of innovative activity. In all instances, we control for population and land area. Generally speaking, we find that individuals coming from innovative areas are more likely to accept industrial employment locally. For example, the probability that an individual stays in the city of training is positively related to the number of utility patents granted in the city and the Milken Index.19 At the state level, we find that individuals are more likely to stay if the state has a high level of industrial R&D activity. Somewhat surprisingly, patent counts are not significant at the state level. As a measure of employment opportunities for Ph.D.s in the state (city) of training relative to elsewhere, we construct an index of the relative local absorptive capacity for Ph.D.’s (ABPhDi), measured as the ratio of the flow of new Ph.D.s produced locally to the stock of Ph.D.s working in local in-

17. A doctorate need not remain local, or even in state, to return to or continue in previous employment. In fact, 46 percent of new Ph.D.s who indicate they are returning to or continuing in previous employment leave their state of training after graduation. 18. Data on academic and industrial R&D expenditures come from the National Science Board (2002), and are computed in 1996 constant dollars for the years 1997, 1998, and 1999. 19. The Milken Index, measured by the Milken Institute, is a measure of high-tech concentration in the PMSA. By definition, the Milken Index mean for the United States is equal to 1.0. A metro area with an index higher than 1.0 has a higher high-tech concentration than the United States, a metro area with an index that is lower than 1.0 has a lower high-tech concentration.

Capturing Knowledge

277

dustry relative to the same measure aggregated across the United States. To wit, we define the measure as: ABPhDi  (NPhDIi /PhDIi)/(ΣNPhDIi/ΣPhDIi), where NPhDIi is the number of new Ph.D.s (in all fields) in location i (defined as either the state or PMSA) with plans to work in industry; PhDIi is the total number of all Ph.D.s in location i working in industry. We hypothesize an inverse relationship. We find the variable to be negative and highly significant in predicting the probability that the individual will remain at either the state or local level. Clearly, the ability of the local area to absorb new Ph.D.s is a prime factor in determining whether the individuals stay. Our results also indicate that new Ph.D.s are more likely to stay in their state of training the higher the per capita income in the area. Somewhat surprisingly, we do not find per capita income to be significant in the PMSA equation. In neither instance do we find the educational variables to be significant.20 If higher education were funded at the federal, rather than the state or local level, it would make little difference, from an economic development perspective, whether the newly trained Ph.D.s remained local, or instead left the area of training. However, and as noted earlier, institutions of higher education in the United States are a mixed lot. Public institutions receive funding from the state, and indirectly, local area, in which they are located; private institutions do not. While we do not find a significant difference regarding the decision to stay in state between public and private institutions, we do find a significant difference at the PMSA level. Given the important role that retention plays in leveraging public resources, we reestimate the basic equations, focusing exclusively on public institutions. The results, presented in appendix table 8A.2, are reasonably similar to those presented in table 8.5. The finding that many of the “best” Ph.D.s leave persists when we focus exclusively on public institutions. Specifically, we find that individuals trained at top-rated biology, chemistry, computer science, math, and medical Ph.D. programs are less likely to remain in state than are those coming from non-top-rated programs. Moreover, those who were supported on a fellowship or dissertation grant, an indicator of quality, are more likely to leave. Doctorate recipients from public institutions are more likely to remain in state if they received their undergraduate degree from the same state. Where one went to high school no longer matters when the sample is restricted to individuals who attended public institutions. The public PMSA results are reasonably similar to those for all institutions. 20. These results may reflect our failure to control for the relative values of these variables. Arguably, it is the relative value that affects the decision to stay or leave, not the level of the variable.

278

Albert J. Sumell, Paula E. Stephan, and James D. Adams

8.7 Conclusion and Discussion The movement of the highly educated from universities to firms is one mechanism by which knowledge is transferred. Despite the important role that industrial Ph.D.s can play in economic development, to date we know very little regarding their location decisions. This knowledge gap is especially striking given the focus in recent years on the role that proximity plays in the transmission of knowledge (Feldman 1994; Audretsch and Stephan 1996). To help rectify this deficiency, we measure the degree to which placements are local and what affects the likelihood that a Ph.D. going to work in industry will remain in the same state or metropolitan area. We find that states and local areas capture knowledge embodied in newly-minted Ph.D.s headed to industry, but not at an overwhelming rate. Only about one in three of those going to industry take a job in the state where trained; approximately one in five in the same PMSA. The averages, however, mask wide variations. California retained two out of three of the more than 1,500 Ph.D.s it trained for industry during the period. Indiana retained only one in eight of the 376 it trained. Wide variation exists at the metropolitan level as well: the San Francisco-Oakland-San Jose area retained almost 60 percent of those trained in the metropolitan area who take a position in industry as did the wider New York metropolitan area. By way of contrast, State College, Pennsylvania, retained about 3 percent, as did Champaign-Urbana, Illinois and Lafayette, Indiana. Our research informs the question of whose knowledge is captured. We find that local areas are more likely to retain white students and students having little debt who are returning to a previous position. Being “homegrown” predisposes one to remain as well. Those who receive their Ph.D. in the same state as their undergraduate degree and high school degree are more likely to stay than those who do not. Those who receive their Ph.D. in the same state as their BA degree, as well as in their birth state, are more likely to stay in the PMSA. Graduates from certain fields are especially likely to leave the state: most notably agriculture, chemistry, engineering, computer science, and earth science. Quality matters: top-rated Ph.D. programs are often the ones that are most likely to produce graduates who leave the area. Those supported on fellowships or dissertation grants are more likely to leave the state of training. Graduates from private institutions are also more likely to find industrial employment outside the metropolitan area of training. Not surprisingly, and consistent with a wide body of research on innovation, we find that local areas are more likely to retain new Ph.D.s if the area is high in measures of innovation such as patent counts and R&D expenditures. The relative absorptive capacity of the local community also plays a major role. Champaign-Urbana graduates a large number of new Ph.D.s who want to work in industry; yet relative to the United States, few Ph.D.s work in industry in the city.

Capturing Knowledge

279

8.7.1 Discussion Our results are consistent with the findings of Audretsch and Stephan (1996) concerning the degree to which knowledge is captured locally. To wit, they find only 30 percent of the scientist-firm links they examined to be local; we find that only 25 percent of new Ph.D.s headed to industry stay in the MSA of training. There are at least two distinctions, however, between Audretsch and Stephan’s work and this work. First, university faculty can be on multiple scientific advisory boards; new Ph.D.s can only work for one firm at a time. Second, from the viewpoint of the university, it is entirely different to invest in faculty who establish ties with new firms out of the area while continuing to work at the university than to educate students who leave the area to take a position with a firm. While students who leave may expand and diversify the university’s knowledge and support network, the economic returns to the state from such migration are likely to be relatively low, especially in the short run. Our findings raise the larger question of whether the role of proximity to the university is overemphasized in the transmission of public knowledge from universities to industry. The top source of public knowledge, according to the Carnegie Mellon survey of firms (Cohen, Nelson, and Walsh 2002), is publications and reports. Neither requires proximity to the scientist/engineer. The second source (informal information exchange, public meetings, or conferences and consulting) is facilitated by proximity but proximity is not essential. The next tier includes recently-hired graduate students. Our research shows that, in this respect, proximity does not play a major role. We infer that if firms know what they are looking for, proximity to the university is not that important in the transmission of knowledge. Firms can search for the input. Proximity to the university is most important when the firm does not know what it is seeking or does not want to invest heavily in search, or when the scientists involved in the transmission of tacit knowledge have a strong preference for remaining local, as Zucker, Darby, and Brewer (1998) argue that star scientists had.21 States often invest in higher education with the conviction that it stimulates local economic development. And certainly research supports this conviction. Our work, however, casts doubt on the benefits states realize from one piece of this investment—the education of a doctoral scientific workforce—and suggests that states capture but a portion of the economic 21. This discussion raises the further question of the degree to which spillovers result from nonappropriability. We have argued that tacit knowledge comprises an important component of the knowledge that new Ph.D.s transmit to firms. Yet tacit knowledge, as Zucker, Darby, and Brewer (1998) point out, facilitates excludability. Thus knowledge transmission, to paraphrase the aforementioned authors, can result from the maximizing behavior of scientists who have the ability to appropriate the returns to this tacit knowledge rather than from nonappropriability.

280

Albert J. Sumell, Paula E. Stephan, and James D. Adams

benefits resulting from a trained Ph.D. workforce. What we do not investigate here is why states are able and willing to educate Ph.D.s who leave after graduation. Is the knowledge and technology transfer produced while students are in graduate school sufficient to justify the expenditure? Do graduate students more than compensate for their educational costs, directly through tuition payments and indirectly through their labors in the classroom and the laboratory? Is the halo generated from having a toprated program sufficiently beneficial to the state in terms of general economic development? Do states reap sufficient long-term economic benefits from the networks created by students who migrate? Or, and perhaps what is more likely, do these factors collectively provide sufficient benefits to outweigh the state’s expenditures? Is it, of course, also possible that what we observe is an indication of a disequilibrium that may hasten to adjust as bleak budget prospects lead states to slash budgets for higher education? Can universities such as Illinois and Purdue continue to educate Ph.D.s who overwhelmingly leave the state after graduation? Or are policymakers ignorant of the degree to which it is a leaky system? Groen and White (2001, 24) note that incentives of universities and states with regard to the retention of highly-trained workers differ. They explain: “States have an interest in using universities to attract and retain high-ability individuals because they pay higher taxes and contribute more to economic development. Universities have an interest in their graduates being successful, but little interest in where their students come from or where they go after graduation.” The distinction may be less clear in the post Bayh-Dole world, where public universities promote their science and engineering programs as engines of economic development. One wonders how long these institutions can continue to bake educational cake for other states and countries. The fact that in some instances the institutions are the major supplier of new in-state industrial hires may, of course, mitigate the political pressure to reallocate resources. The implications drawn from this study are somewhat restricted due to the limited scope of the data. For example, the attractiveness of certain regions and cities may have been inflated during the time period of analysis. When we extend the analysis to years following the boom in information technology we may find a somewhat different picture than we do here. Furthermore, the data eliminates Ph.D.s who do not specify a firm as well as Ph.D.s who eventually work in industry after taking a postdoc position. The percent of seasoned Ph.D.s going to industry is much larger than the percent of new Ph.D.s choosing industry, particularly in the life sciences. As a result, if the study were done on location decisions five years following receipt of degree, as opposed to newly-minted PhDs, the conclusions might differ substantially.

samehs_phd

samece_phd

singlepar

married female_married wchild

tempres

permres

nonwhite_asian

age agesq female white* asian

SamePMSA

SameSTATE

Variable

Table 8A.1

Appendix

Age of the individual at the time of Ph.D. Age of the individual squared Dummy variable indicating whether or not an individual is a female Dummy variable indicating whether or not an individual is White Dummy variable indicating whether or not an individual is Asian or Pacific Islander Dummy variable indicating whether or not an individual is a race other than White or Asian Dummy variable indicating whether or not an individual is a permanent resident in the U.S. Dummy variable indicating whether or not an individual is a temporary resident in the U.S. Dummy variable indicating whether or not an individual is married Dummy variable indicating whether or not an individual is a married female Dummy variable indicating whether or not an individual is married with at least one dependent Dummy variable indicating whether or not an individual is not married with at least one dependent Dummy variable indicating whether or not an individual earned their Ph.D. in the same state they went to college Dummy variable indicating whether or not an individual went to high school, college, and earned their Ph.D. in the same state

Independent variables

Dependent variables Dummy variable indicating whether or not an individual has definite plans to remain in the same state in which they earned their Ph.D. Dummy variable indicating whether or not an individual has definite plans to remain in the same PMSA in which they earned their Ph.D.

Definition

Variable definitions and descriptive statistics

0.129 (0.336)

0.182 (0.386)

X

X

X

X

0.245 (0.430) 0.030 (0.170)

X X X

X 0.333 (0.471) 0.613 (0.487) 0.111 (0.315)

0.105 (0.306)

X

X

0.378 (0.485) 0.065 (0.246)

X X X X

—

XX

Same state (Eq. 1)

32.52 (5.043) 1,083.0 (373.94) 0.202 (0.401) 0.555 (0.497)

0.209 (0.4064)

0.367 (0.482)

Mean (Standard deviation)

(continued )

X

X

X

X

X X X

X

X

X

X X X X

XX

—

Same PMSA (Eq. 2)

alleng

agri

astr

supp_other*

supp_employer

supp_RA_trainee

supp_teachasst

supp_fellow

pre_otheremp*

preptemp

preftemp

debtlevel

return

sameb_phd

Variable

Table 8A.1

Definition

Dummy variable indicating whether or not an individual was born, went to high school, college, and earned their Ph.D. in the same state Dummy variable indicating whether or not an individual has definite plans to continue in or return to previous employer Individual’s reported debt level in thousands, measured in $5,000 intervals, at the time of degree Dummy variable indicating whether or not an individual was employed full time one year prior to receipt of Ph.D. Dummy variable indicating whether or not an individual was employed part time one year prior to receipt of Ph.D. Dummy variable indicating whether or not an individual was anything other than full or part time employed one year prior to Ph.D. Dummy variable indicating whether or not individual’s primary source of support during graduate school was fellowship or dissertation grant Dummy variable indicating whether or not individual’s primary source of support during graduate school was teaching assistantship Dummy variable indicating whether or not individual’s primary source of support during graduate school was research assistantship, internship, or traineeship Dummy variable indicating whether or not individual’s primary source of support during graduate school was employer reimbursement or assistance Dummy variable indicating whether or not individual’s primary source of support during graduate school was anything other than employer, research or teaching assistant, trainee, diss. grant, or fellowship Dummy variable indicating whether or not an individual’s field of training was astronomy Dummy variable indicating whether or not an individual’s field of training was in agriculture Dummy variable indicating whether or not an individual’s field of training was engineering

(continued)

0.530 (0.500)

0.030 (0.165)

0.004 (0.063)

0.189 (0.392)

0.050 (0.219)

0.479 (0.500)

0.148 (0.355)

0.133 (0.340)

0.609 (0.508)

0.066 (0.248)

0.324 (0.468)

6.776 (10.76)

0.196 (0.397)

0.085 (0.279)

Mean (Standard deviation)

X

X

X

X

X

X

X

X

X

X

X

X

X

X

Same state (Eq. 1)

X

X

X

X

X

X

X

X

X

X

X

X

X

X

Same PMSA (Eq. 2)

private

topsphys

topsmedi

topsmath

topsearth

topscomp

topschem

topsbiol

topsalleng

topsagri

topsastr

phys

medi

math

earth

comp

chem

biol*

Dummy variable indicating whether or not an individual’s field of training was biology Dummy variable indicating whether or not an individual’s field of training was chemistry Dummy variable indicating whether or not an individual’s field of training was computer science Dummy variable indicating whether or not an individual’s field of training was earth science Dummy variable indicating whether or not an individual’s field of training was mathematics Dummy variable indicating whether or not an individual’s field of training was medicine Dummy variable indicating whether or not an individual’s field of training was physics Dummy variable indicating whether or not an individual’s Ph.D. field was astronomy and their Ph.D. institution was top-ranked in astronomy Dummy variable indicating whether or not an individual’s Ph.D. field was agriculture and their Ph.D. institution was top-ranked in agriculture Dummy variable indicating whether or not an individual’s Ph.D. field was in engineering and their Ph.D. institution was top-ranked in engineering Dummy variable indicating whether or not an individual’s Ph.D. field was biology and their Ph.D. institution was top-ranked in biology Dummy variable indicating whether or not an individual’s Ph.D. field was chemistry and their Ph.D. institution was top-ranked in chemistry Dummy variable indicating whether or not an individual’s Ph.D. field was computer science and their Ph.D. institution was top-ranked in computer science Dummy variable indicating whether or not an individual’s Ph.D. field was earth science and their Ph.D. institution was top-ranked in earth science Dummy variable indicating whether or not an individual’s Ph.D. field was mathematics and their Ph.D. institution was top-ranked in mathematics Dummy variable indicating whether or not an individual’s Ph.D. field was medicine and their Ph.D. institution was top-ranked in medicine Dummy variable indicating whether or not an individual’s Ph.D. field was physics and their Ph.D. institution was top-ranked in physics Dummy variable indicating whether or not an individual received their Ph.D. from a private institution 0.324 (0.468)

0.037 (0.189)

0.021 (0.142)

0.024 (0.154)

0.016 (0.124)

0.046 (0.210)

0.068 (0.251)

0.039 (0.193)

0.354 (0.478)

0.023 (0.149)

0.003 (0.051)

0.065 (0.237)

0.043 (0.195)

0.047 (0.204)

0.025 (0.150)

0.075 (0.255)

0.121 (0.314)

0.060 (0.229)

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X (continued )

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

Definition

Number of patents in thousands granted in the state of the individual’s Ph.D. institution between 1997–1999 Academic R&D expenditures in millions in the state of the individual’s Ph.D. institution between 1997–1999 in thousands of 1996 dollars Industrial R&D expenditures in millions in the state of the individual’s Ph.D. institution between 1997–1999 in thousands of 1996 dollars Geographic size in thousands of square miles of the state of the individual’s Ph.D. institution Population in hundred thousands in 2000 in the state of the individual’s Ph.D. institution Percent of the population age 25 in the state of the individual’s Ph.D. institution with a bachelor’s degree or higher in 1998 Per Capita income in thousands in the state of the individual’s Ph.D. institution in 1994 Ph.D. absorption capacity index in the state of the individual’s Ph.D. institution (see text) Number of patents in hundreds granted in the PMSA of the individual’s Ph.D. institution between 1997–1999 Milken Index in the PMSA of the individual’s Ph.D. institution in 2002 Geographic size in thousands of square miles of the PMSA of the individual’s Ph.D. institution Population in hundred thousands in the PMSA of the individual’s Ph.D. institution in 2000 Percent of the population age 25 in the PMSA of the individual’s Ph.D. institution with a bachelor’s degree or higher in 2000 Per capita income in thousands in the PMSA of the individual’s Ph.D. institution in 1999 Ph.D. absorption capacity index in the PMSA (see text)

(continued)

31.62 (5.863) 3.547 (4.41)

31.572 (6.92)

— —

—

—

—

2.464 (2.116) 25.22 (26.54)

— —

X

X

X

X

X

X

X

X

Same state (Eq. 1)

8.17 (8.68) 1.110 (0.711)

1.129 (0.400)

22.953 (2.570)

25.22 (4.06)

129.696 (99.816)

75.852 (66.31)

28.631 (32.568)

36.539 (28.465)

6.49 (6.66)

Mean (Standard deviation)

X X

X

X

X

X X

—

—

—

—

—

—

—

—

Same PMSA (Eq. 2)

Notes: Asterisk (*) indicates the benchmark or control group. “XX” means the variable is a dependent variable included in the equation. “X” Means the variable is an explanatory variable included in the equation.

ABPhDMSA

pmsapcinc

pmsaperhe

pmsapop

milkenind pmsasize

pmsapats

ABPhDST

STpcinc

STperhe

STpop

STsize

STindRD

STacadRD

STpats

Variable

Table 8A.1

Table 8A.2

Empirical results Sample  placements trained in the continental United States in a public institution Equation (1): Dependent variable  SameSTATE

Equation (2): Dependent variable  SamePMSA

N  6,832 Variable Intercept age agesq female asian nonwhite_asian permres tempres married female_married wchild singlepar samece_phd samehs_phd sameb_phd return debtlevel preftemp preptemp supp_fellow supp_teachasst supp_RA_trainee support_employer astr agri alleng chem math comp earth medi phys topsastr topsagri topsalleng topsbiol topschem topscomp topsearth topsmath topsmedi

N  5,973

Estimate

z-stat

4.0254*** 0.0759 0.0005 0.0152 0.1433* 0.1187 0.0224 0.3443*** 0.0742 0.0971 0.0633 0.3512** 0.5645*** 0.1983 0.0685 0.5790*** 0.0078*** 0.4254*** 0.6526*** 0.4007*** 0.0711 0.1450* 0.1579 1.0445 0.8062** 0.4062* 0.6081** 0.1447 0.4831* 1.1193*** 0.1585 0.0379 0.7449 0.0336 0.1305 0.4799* 0.4939*** 0.4295* 0.1426 0.7983*** 0.6370**

16.16 2.49 0.69 0.01 2.83 0.92 0.04 14.26 0.87 0.39 0.64 3.95 16.37 1.22 0.20 44.55 7.12 33.27 31.95 11.50 0.46 2.95 1.21 1.69 4.17 3.23 5.92 0.28 3.06 9.58 0.28 0.02 0.50 0.01 1.97 3.45 7.67 3.68 0.15 8.90 5.23

a

Estimate

z-stata

3.2767*** 0.0980* 0.0007 0.1222 0.3627*** 0.1546 0.2446* 0.5355*** 0.1158 0.1595 0.0795 0.3785 0.2367 0.1240 0.4018** 0.3927*** 0.0103*** 0.4280*** 0.7237*** 0.1308 0.0886 0.0018 0.0406 0.0444 0.5502 0.1667 0.4540 0.0567 0.2341 1.2857*** 0.0339 0.1738 0.949 0.5447 0.3363*** 0.5912* 0.5684** 0.6812** 0.0452 0.4819 0.8052**

7.77 2.77 0.91 0.55 11.41 1.00 3.00 20.95 1.29 0.62 0.66 2.65 2.00 0.34 5.13 14.19 7.61 20.29 29.67 0.78 0.44 0.00 0.07 0.00 0.61 0.40 2.24 0.03 0.52 7.64 0.01 0.28 0.01 0.59 6.87 3.82 5.45 6.08 0.01 2.23 5.56 (continued )

286

Albert J. Sumell, Paula E. Stephan, and James D. Adams

Table 8A.2

(continued) Equation (1): Dependent variable  SameSTATE

Equation (2): Dependent variable  SamePMSA

N  6,832 Variable topsphys STpats STacadRD STindRD STsize STpop STperhe STpcinc ABPhDST pmsapats milkenind pmsapop pmsasize pmsapcinc pmsaperhe ABPhDMSA –2 Log-likelihood

Estimate 0.0418 0.00025 0.00010** 0.000021** 0.0062*** 0.000014 0.0046 0.0001** 0.2538*** n.a. n.a. n.a. n.a. n.a. n.a. n.a. 8,857.6

N  5,973 z-stat

a

0.03 0.18 4.42 5.90 54.98 0.48 0.11 5.88 7.26 n.a. n.a. n.a. n.a. n.a. n.a. n.a.

Estimate 0.1081 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. 0.0010*** 0.4875*** 0.0000038 0.02050 0.03060** 0.0070 0.0789*** 5,869.2

z-stata 0.15 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. 59.81 33.36 0.02 1.59 6.56 0.66 27.56

Note: n.a.  not applicable. z-stats are based on chi-square distribution. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

a

References Agrawal, A., and R. Henderson. 2002. Putting patents in context: Exploring knowledge transfer from MIT. Management Science 48 (1): 44–60. Almeida, P., and B. Kogut. 1999. Localization of knowledge and the mobility of engineers in regional networks. Management Science 45 (7): 905–17. Anselin, L., A. Varga, and Z. J. Acs. 1997. Local geographic spillovers between university research and high technology innovations. Journal of Urban Economics 42 (3): 422–48. ———. 2000. Geographic spillovers and university research: A spatial econometric perspective. Growth and Change 31 (Fall): 501–15. Audretsch, D., and M. Feldman. 1996a. Innovation clusters and the industry life cycle. Review of Industrial Organization 11 (2): 253–73. ———. 1996b. R&D spillovers and the geography of innovation and production. American Economic Review 86 (3): 630–40. Audretsch, D., and P. Stephan. 1996. Company-scientist locational links: The case of biotechnology. American Economic Review 86 (3): 641–52.

Capturing Knowledge

287

Black, G. 2001. The geography of small firm innovation. Ph.D. diss., Georgia State University Atlanta, Georgia. Black, G., and P. Stephan. 2003. The importance of foreign PhD students to U.S. science. Paper prepared for the conference on “Science and the University” at the Cornell Higher Education Research Institute, Cornell University. 20–21 May, Ithaca, New York. Cohen, W., R. Nelson, and J. Walsh. 2002. Links and impacts: The influence of public research on industrial R&D. Management Science 48 (1): 1–23. Feldman, M., and D. B. Audretsch. 1999. Innovation in cities: Science based diversity, specialization and localized competition. European Economic Review 43 (2): 409–29. Feldman, M. 1994. The geography of innovation. Dordrecht, the Netherlands: Kluwer Academic Publishers. Groen, J. A., and M. White. 2001. In-state versus out-of-state students: The divergence of interest between public universities and state governments. NBER Working Paper no. 9603. Cambridge, MA: National Bureau of Economic Research, April. Jaffe, A. 1989. Real effects of academic research. American Economic Review 79 (5): 957–70. Link, A. 1995. A generosity of spirit: The early history of research triangle park. Chapel Hill: The research triangle foundation of North Carolina. Chapel Hill, NC: Research Triangle Foundation. Mansfield, E. 1995. Academic research underlying industrial innovations: Sources, characteristics, and financing. Review of Economics and Statistics 77 (1): 55–65. National Association for Law Placement. 1998. Class of 1997 employment report and salary survey. Washington, D.C.: National Association for Law Placement. National Science Foundation. 2005. Interstate migration patterns of recent recipients of Bachelor’s and Master’s degrees in science and engineering. Available at http://www.nsf.gov/statistics/nsf05318/sect3.htm. Powell, W., K. Koput, L. Smith-Doerr, and J. Owen-Smith. 1998. Network position and firm performance: Organizational returns to collaboration in the biotechnology industry. In Research in the Sociology of Organizations, vol. 16, ed. S. B. Andrews and D. Knocke, 129–59. Greenwich CT: JAI Press. Sjaastad, L. A. 1962. The costs and returns of human migration. Journal of Political Economy 70 (5): 80–93. Stephan, P., A. Sumell, G. Black, and J. Adams. 2004. Doctoral education and economic development: The flow of new PhDs to industry. Economic Development Quarterly 18 (2): 151–67. Stephan, P., and G. Black. 1999. Bioinformatics: Does the U.S. system lead to missed opportunities in emerging fields? A case study. Science and Public Policy 26 (6): 382–89. Stern, S. 1999. Do scientists pay to be scientists? NBER Working Paper no. 7410. Cambridge, MA: National Bureau of Economic Research, October. Thompson, P. 2006. Patent citations and the geography of knowledge spillovers: Evidence from inventor- and examiner-added citations. The Review of Economics and Statistics 88:383–88. Zucker, L., and M. Darby. 1997. The economists’ case for biomedical research: Academic scientist-entrepreneurs and commercial success in biotechnology. In The future of biomedical research, ed. C. Barfield and B. Smith, 42–66. Washington, D.C.: American Enterprise Institute for Public Policy Research and the Brookings Institute. Zucker, L., M. Darby, and M. Brewer. 1998. Intellectual capital and the birth of the U.S. biotechnology enterprise. American Economic Review 88 (1): 290–396.

9 Instruments of Commerce and Knowledge Probe Microscopy, 1980–2000 Cyrus C. M. Mody

9.1 Introduction Universities have long struggled to define their relation to the business world (Geiger 2004). How much should professors be involved in commercial activities? How much influence should firms have on university policies? How much should universities themselves be run like for-profit businesses? Today, much of the U.S. science and engineering workforce is trained at institutions where professional start-up companies are the norm and where much corporate research has been outsourced to academic groups. This situation differs markedly from that of the period between 1945 and about 1970, when federal funding for basic research was so large that academics did not need to rely on corporate support, and companies like IBM and AT&T could run big laboratories doing high-quality basic research (Mirowski and Sent 2007). As with any trend in higher education, emotions run high in discussions of university-industry relations. Critics see “academic capitalism” (Slaughter and Leslie 1997) as leading to the exploitation of students, the neglect of teaching, and the distortion of scientific knowledge to meet commercial patrons’ needs (Bok 2003; Kirp 2003). Proponents see it as enticing professors out of their ivory tower and making universities more responsive to changing public (and market) demands (Etzkowitz 2002). Proponents call for more direct incentives for professors to commercialize their discoveries; critics call for a return to a (probably nonexistent) golden Cyrus C. M. Mody is an assistant professor of history at Rice University. This work was made possible by funding from NBER, the NSF, the American Institute of Physics, the IEEE History Center, the Lemelson Center, and the Chemical Heritage Foundation. It would have been impossible without the generous cooperation of my interviewees.

291

292

Cyrus C. M. Mody

age when academic scientists did not need to justify every discovery in terms of its commercial applications.1 Most of these arguments are couched in very abstract terms. Where we do have detailed, empirical studies of academic capitalism, they have tended to be broad surveys or examinations of a small selection of regions and universities, especially Stanford, MIT, Silicon Valley, and Route 128 (Vettel 2006; Lécuyer 2006; Kenney 2000). These studies reveal much, but they neglect key aspects of the way science is actually conducted. Most researchers participate in networks that are geographically dispersed and that include colleagues in both academia and industry and from a variety of disciplines. To understand the commercialization of academic knowledge, we need a multiinstitutional, multidisciplinary, multiregional unit of analysis—what I will call an “instrumental community.” By this I mean the porous group of people commonly oriented to building, developing, using, selling, and popularizing a particular technology of measurement.2 Such communities are instrumental primarily in focusing on new research tools—microscopes (Rasmussen 1997), fruit flies (Kohler 1994), tobacco mosaic virus (Creager 2002), lab rats (Rader 2004), ultracentrifuges (Elzen 1986), and so forth. Because such communities usually include academic and commercial participants, though, they will often seek ways to morph those tools into industrially-relevant devices. Thus, such communities are also instrumental in focusing on new ways of doing or making things. There are a number of excellent case studies of various instrumental communities, spanning from the seventeenth century to the 1960s (Shapin and Schaffer 1985; Jackson 2000; Pantalony 2004; Bromberg 1991; Lenoir and Lécuyer 1995). Yet there have been virtually no studies of instrumental communities that have arisen since the late 1970s. We know that there have been significant changes in legislation, federal funding, corporate research, and the demographics of science in the past three decades. We do not know how those changes have affected the operation of instrumental communities, nor how they have affected relationships between corporate and academic members of those communities. This chapter aims to bring these issues to the fore through a case study of the development and commercialization of the scanning tunneling microscope (STM) and its nearrelatives, the atomic force microscope (AFM) and magnetic force microscope (MFM)—known collectively as probe microscopes.3 In 1981, there 1. Shapin (2003) casts doubt on the assumptions underlying both these positions. 2. An “instrumental community” bears a close resemblance to the “innovation communities” analyzed by Shah (2003). “Instrumental community” is—so far as I know—my own formulation, but others have covered very similar ground, especially Blume (1992) and Shinn (1997). 3. The technical details of the microscopes are important to this story, but can be glossed for the purposes of this chapter. Basically, all scanning probe microscopes bring a small, solid probe very close (usually to within a nanometer—one billionth of a meter) to a sample and measure the strength of different kinds of interactions between probe and sample to deter-

Instruments of Commerce and Knowledge

293

was only one, homemade, unreliable STM at the IBM research lab in Zurich. Today, through the joint efforts of corporate and academic researchers, there are thousands of AFMs, MFMs, and STMs at universities, national labs, and industrial research and quality control facilities. High school students make STMs from Legos, while chip manufacturers use million-dollar AFMs on the factory floor. One AFM has even made it to the surface of Mars. Using the instrumental community as a unit of analysis allows us to approach the same issues that motivate the other chapters in this volume, but from a different perspective and with a different methodology. Looking at the dynamics of an instrumental community can help us understand several things: some reasons why professors commercialize their research; some ways the training of graduate students and postdocs is linked to the needs of companies their supervisors are associated with (Davis, chapter’s, this volume); ways to interpret regions’ rates of retention of their science and engineering graduates (Sumell, Stephan, and Adams, chapter 8, this volume); and ways gender and ethnic diversity can make for more robust (and commercial) knowledge (Whittington, chapter 6, this volume). Unfortunately, an instrumental community is a nebulous, intangible, unstable grouping that would be very difficult to study via the methodologies of the other chapters in this volume. Surveys of scientists and engineers, for instance, work well when there are institutions that map closely to the group being surveyed: universities offer an infrastructure for surveying recent Ph.D.s.; professional societies (and their journals) offer an infrastructure for surveying members of specific disciplines; funding agencies offer mine the height (and other characteristics) of the sample. The probe is then rastered much like the pixels on a TV screen and a matrix of values for the strength of the tip-sample interaction is converted into a visual “picture” of the surface. Different probe microscopes use different kinds of tip-sample interactions to generate their images. The earliest probe microscope, the STM, works by putting a voltage difference between the tip and a metal or semiconductor sample; when the tip is brought close to the sample, some electrons will quantum mechanically “tunnel” between them. The number of electrons that do so (the “tunnel current”) is exponentially dependent on the distance between tip and sample; also, the stream of tunneling electrons is very narrow. Thus, an STM has ultrahigh resolution both vertically and laterally—most STMs can actually see individual atoms on many samples. Today, the STM’s younger cousin, the atomic force microscope, is more commonly used. An AFM uses a very small but flexible cantilever as a probe. As the tip of the cantilever (usually weighted with a small pyramid of extra atoms) is brought close to the surface, the cantilever bends due to the attraction or repulsion of interatomic forces between tip and sample. The degree of bending is then a proxy for the height of the surface. Originally this bending was measured by putting an STM on the back of the cantilever; today the deflection is detected by bouncing a laser off the cantilever and measuring the movement of the reflected spot. Another common and industrially-relevant tool, the magnetic force microscope, works in a similar way, but uses a magnetic tip to map the strength of magnetic domains on a surface, rather than surface height. Both the AFM and MFM have slightly less resolution than the STM (i.e., they cannot usually see single atoms); yet because they (unlike the STM) can be used on insulators as well as conductors, and in air and fluids as well as vacuum, they have become much more popular.

294

Cyrus C. M. Mody

an infrastructure for surveying their grantees. Often, no institution maps well to an instrumental community. Part (but only part) of the probe microscopy community was affiliated with a professional society—the American Vacuum Society—and probe microscopists tended to publish in many different journals and get funding from many different sources. The closest the community came to owning an institution were the annual (later biennial) STM Conferences. Yet even these only drew part of the community and, over the years, evolved into general nanotechnology conferences rather than meetings devoted to a specific class of instrumentation. I have opted to study (and partially define) the instrumental community through extensive oral history interviewing. Interviews can be a problematic methodology: people misremember or mislead, interviewers ask leading questions, important people can often spare little or no time for an interview, one cannot interview everyone, or even know exactly who to talk to. Still, interviews allow scientists and engineers to map out their relevant communities for themselves, rather than having an institution do it for them. That is, they can tell the interviewer who they were working with, who they thought of as peers and competitors, with whom they were sharing ideas, and so forth. By then interviewing those people, the historian can delineate the network of relationships that make up an instrumental community. In the end, interview and survey data should be complementary. My interviews, for instance, revealed some rather counterintuitive motivations for people to found start-up companies. Unfortunately, interviews cannot show how common such motivations are across the science and engineering community. Such motivations can, however, be folded into future survey questionnaires. Similarly, many of my interviews were informed by quantitative data from other studies. For instance, many interviewees in this study were people who were postdoctoral fellows when they first built or used a probe microscope. Those people could easily have been invisible in my study had I not seen quantitative data on the evolving nature of the postdoc as an institution of American science. Letting the participants map out the boundaries of their instrumental community allows us to see the wide variety of relationships linking nodes in this network: student-teacher, buyer-seller, funder-grantee, supervisorpostdoc, inventor-early adopter, editor-author, and so forth. Indeed, it is the diversity of such linkages in the network that makes an instrumental community robust. Instrumental communities are also made more robust by the variety of the linkages connecting universities and corporations. Both proponents and critics of academic capitalism in science tend to focus on a small subset of corporate-academic relationships: professorial start-ups, patenting of academic research, and corporate sponsorship of academic research. This chapter will explore a much wider range of relationships, however.

Instruments of Commerce and Knowledge

295

These include: researcher sabbaticals from firms to universities and vice versa, technology transfer through firms’ hiring of former graduate students and universities’ hiring of former corporate researchers, corporate sponsorship of community-building activities such as conferences, corporate influence over researchers’ choices of materials to characterize with their microscopes, corporate supply of parts for building microscopes (and academic feedback to the design of those parts), and corporate sponsorship of intramural research to stimulate formation of an extramural academic market. Most of these kinds of relationships are invisible in the debate about academic capitalism. Yet we shall see that other forms of commercialization of academic research (e.g., professorial start-ups) actually have their roots in these less-noticed dynamics of an instrumental community. 9.2 Inventing and Community-Building Invention, though often praised in the abstract, can be problematic for corporate scientists and engineers. Inventions can emerge from digressions from assigned tasks, and may not meet any commercial objective. Corporate inventors often need to defy their managers to promote their innovations. The STM was this kind of institutional orphan. Its inventors, Gerd Binnig and Heini Rohrer, had been tasked in 1978 with finding new ways to characterize thin films for an advanced supercomputer project on which IBM had staked much of its reputation. Yet by the time they came up with the STM, the supercomputer project had been canceled (Binnig and Rohrer 1985, 1987). Binnig and Rohrer’s response was threefold. First, they temporarily hid the STM from managerial oversight. Second, they began querying IBM colleagues about new applications for their microscope, eventually attracting interest from the company’s semiconductor surface scientists. Their third critical strategy was to cultivate an extramural, academic community. By convincing colleagues at universities to replicate the instrument, they could point to extramural interest as a reason why their managers should let them continue developing the STM. Indeed, the interest in STM both inside and outside Big Blue convinced IBM’s senior research managers that the STM—despite the absence of commercial relevance—should become a major corporate project. Multiple groups of scientists at the IBM laboratories in Zurich; Yorktown Heights, New York; and San Jose, California were recruited to build STMs and make discoveries that would bring credit to the instrument and to the company. In turn, IBM’s research archrival, Bell Labs, saw a need to steal Big Blue’s thunder and began recruiting its own cadre of STMers. The dynamics of building the STM community show how the corporate and academic worlds are interpermeated much more than is noticed in debates about academic capitalism. Binnig and Rohrer could quickly find

296

Cyrus C. M. Mody

and convince academics to build their own STMs because of networks of personnel exchange between IBM and various universities. Some replicators were professors taking sabbaticals at IBM, some were academics Rohrer had known from his own sabbaticals at universities, and some were people who had been postdocs at IBM or currently had students serving postdoctoral appointments there.4 Similarly, interest in the STM grew within IBM and Bell Labs not because it could solve commercially-relevant problems, but because it could generate credible knowledge within academic disciplines such as physics and surface science.5 Accolades from an academic audience—evidenced by standing-room-only crowds at American Physical Society meetings, the awarding of the Nobel Prize to Binnig and Rohrer in 1986, and the growth of academic STM—were largely the aim of IBM’s STM program. Among other things, prestige within a hot new instrumental community like tunneling microscopy allowed IBM to recruit the best graduate students as postdocs and junior researchers. Some of those persons in turn built the second and third generations of IBM’s tunneling microscopes. 9.3 Dynamics of Community By 1986, the STM was no longer in any danger. An instrumental community was beginning to take shape, and was even beginning to organize itself through an annual conference series. In the first few years, those conferences (and the community as a whole) was dominated by corporate groups, especially from IBM and Bell Labs. Early academic STMers, such as Paul Hansma at the University of California at Santa Barbara (UCSB), Calvin Quate at Stanford, and John Baldeschwieler at Caltech, were important contributors to the community. Yet these academics struggled to compete with better-resourced corporate groups. The IBM and Bell Labs STMers also had the advantage of proximity to other STM groups housed in their same buildings. While there was often intense competition between groups that were working for the same organization, their copresence did allow the tacit knowledge (Polanyi 1962; Collins 1975) needed to build an STM to flow more quickly at IBM and Bell Labs than at more isolated locales such as Stanford and UCSB. Bell Labs and IBM were both traditionally strong in computing and microelectronics research. Their early dominance of the STM community 4. The source material for this study is a collection of interviews with over 150 probe microscopists conducted between 2000 and 2004. I will reference specific oral histories using an alphanumeric code listed in the appendix to this chapter. Information about the corporateacademic network of sabbaticals and hires came from, among others, , , and . 5. There is rich historical material on the large, corporate labs of the twentieth century: Wise (1985); Riordan and Hoddeson (1997); Bassett (2002); Knowles and Leslie (2001).

Instruments of Commerce and Knowledge

297

meant, therefore, that early STMers largely used their microscopes to study materials used in microelectronics manufacturing. In particular, Binnig and Rohrer were most successful in enrolling colleagues interested in the surface structure of metals and semiconductors. A few semiconductor surfaces (especially of silicon) became yardsticks for measuring whether a group had a working STM or not—until a group’s STM had resolved single atoms of silicon, its builders could not enter the top tier of STM builders.6 Both corporate and academic STMers were evaluated in this way. However, there was considerably more local knowledge about preparing silicon specimens at Bell Labs and IBM than in academic groups such as Quate’s or Baldeschwieler’s. Thus, the accreditation standards of the early STM community favored corporate groups. Binnig and Rohrer, however, were keen to undo their own company’s lead in STM research. Thus, they began looking for new applications for the STM that would not interest IBM management, but where academic researchers could move forward quickly. In Europe, for instance, Binnig and Rohrer collaborated with academics to explore applications for the STM in electrochemistry and biophysics. In the United States, Binnig visited Stanford for more than a year to help Quate’s group think of new uses for the STM, while traveling around to help other groups get their microscopes running. And Rohrer dispatched Binnig to Santa Barbara (where Rohrer had taken a sabbatical several years earlier) to convince Paul Hansma to adapt the STM to do vibrational spectroscopy of molecules. Soon, the American STM community began to segregate into two moieties—surface science STMers, dominated by (but not exclusive to) corporate and national laboratories on the East Coast; and nonsurface scientists, dominated by (but not exclusive to) universities on the West Coast.7 These two moieties continued to share a great deal. Members of each occasionally collaborated, and a few people moved from one to the other. More importantly, the basic design of the STM was—in the late 1980s—common to both, so design innovations in one moiety could be transported to the other. This meant that opportunities for copresence—conferences and visits and sabbaticals between labs—continued to be useful for both moieties until the early 1990s. Yet the two moieties did differ markedly on some points of STM design and use. In particular, surface science STMers built their microscopes for compatibility with ultrahigh vacuum (UHV) chambers, so as to keep their metal and semiconductor samples pristine. These chambers were, however, large, finicky, expensive, and time-consuming. Academics 6. , . 7. Crucial corporate members of the latter, predominantly academic, moiety were Quate’s allies within IBM: Dan Rugar, John Foster, and Tom Albrecht (former students who worked at IBM Almaden); Kumar Wickramasinghe (a former postdoc, later at IBM Yorktown); and Gerd Binnig (who took a sabbatical at Stanford from 1985 to 1986).

298

Cyrus C. M. Mody

like Quate and Hansma, who were less interested in studying pristine semiconductor samples, therefore developed easier, cheaper variants of STM that did not require a vacuum chamber, such as doing tunneling microscopy of samples exposed to open air, or immersed in water, oil, or a variety of different gases.8 Surface science STMers had a well-defined set of questions to ask and materials to study. By branching into new uses of STM, Quate, Hansma, and Baldeschwieler freed themselves from the constraints of surface science, but they also forfeited the structure that a discipline like surface science can supply. They could better afford to temporarily put aside discipline than the younger STMers in the corporate labs because Hansma, and especially Quate and Baldeschwieler, were all tenured faculty with long track records of inventing instruments, getting grants, and winning acclaim from their colleagues. The STM was, for them, a chance to start over, rather than (as it was for the young corporate STMers) a chance to start off. Still, by going down this road, they now had little idea what materials to look at, what questions to ask, how to interpret their data, or what audiences might be interested in their work. To answer those questions, they encouraged their students to quickly build a wide variety of microscopes and to playfully use them to characterize haphazard materials—leaves of houseplants, polaroids, bone from ribeye steaks, ice, the electrochemistry of Coke versus Pepsi, and so forth.9 This undisciplined, shoestring bricolage extended even into microscope-building: the Baldeschwieler group made STM probes from pencil leads, for instance, while the Hansma group made AFM tips from hand-crushed pawn shop diamonds, glued to tin foil cantilevers with brushes made from their own eyebrow hairs. Yet such indiscipline could damage STM’s acceptance by new disciplinary audiences, since the STMers’ ways of preparing samples and interpreting images might not be credible to biologists, electrochemists, materials scientists, geologists, and so forth. Thus, Quate, Hansma, and other academic STMers began bringing representatives (postdocs or young professors) from potential new disciplinary audiences in to work with their students, learn how to use the microscope, show the group how to prepare samples, and then proselytize for the technique within their home community. Quate tended to recruit postdocs himself and share them with other Stanford faculty. Hansma actively sought collaborations with young faculty both at UCSB and elsewhere, but he was more haphazard about postdocs, taking in people who brought their own money and expertise but not seeking them out. Nevertheless, he had a long string of such visitors, since by the late 1980s graduating Ph.D.s in molecular biology or electrochemistry 8. , . 9. , .

Instruments of Commerce and Knowledge

299

could see that if they learned to use STM or AFM they would be able to understand their discipline’s canonical samples in a way that none of their disciplinary colleagues would. At both Stanford and UCSB, some of these people took a microscope with them when they left, some founded their own microscope-building groups at other universities, and some used their knowledge of probe microscopy as a tool for gaining acceptance among disciplinary colleagues and securing tenure from their universities.10 Thus, the differences between the two moieties were as much about pedagogy and career arc as they were about samples, designs, and audience. In groups such as Quate’s and Hansma’s, graduate students were trained to build instruments quickly and collaboratively, to think primarily about novel design rather than use. Postdocs in those groups, meanwhile, were trained to develop new uses for the microscopes, and to integrate them into various established disciplines—STM for biology, materials science, electrochemistry, and so forth. In the corporate labs, postdocs and young staff scientists also underwent a kind of training. At the time, Bell Labs was considered one of the preeminent research institutions in the world in a variety of fields, especially solid state physical sciences. In a few areas, especially those relevant to the STM such as surface science and semiconductor physics, IBM ran neckand-neck with Bell Labs. Thus, postdocs in these organizations had the opportunity to do exciting, cutting-edge work. But they also had to compete hard to remain in that rarefied world. To do so, they needed to convince the large numbers of senior managers and surface scientists/semiconductor physicists within these organizations that they could contribute rigorous knowledge to those disciplines. The young corporate STMers therefore learned to build and use microscopes geared specifically to the questions and materials of surface science. Indeed, the most helpful managers were those who directed postdocs to see the disciplinary apparatus of surface science as a way to define problem areas—a kind of functional equivalent of the research/career plans that Davis (chapter 3, this volume) discusses. After they had established themselves they could branch out somewhat, but early on the young corporate STMers all built relatively similar microscopes to look at the same handful of samples—though with enough variation to demonstrate their builders’ individual initiative, creativity, and experimental ingenuity.11 In other words, the instrumental community growing around the STM included elements of pedagogy at all participating sites, rather than just in the academic groups—the STM was a technology for turning young researchers into full-fledged scientists as much as a new technique for char10. , , . The propagation of a technique through the cascade of postdocs and collaborators away from one of the centers of an instrumental community is described in Kaiser (2005). 11. .

300

Cyrus C. M. Mody

acterizing materials. Analysts of academic capitalism should keep this in mind—universities have no monopoly on scientific training. Moreover, in this particular case the pedagogical uses of the STM encouraged a wider division of labor in the instrumental community. Because young corporate STMers had such a monopoly on metal and semiconductor samples, graduate students building STMs were instead encouraged to expand the instrument’s capabilities into new areas. Critics of academic capitalism often complain that corporate influence can restrict academic researchers’ focus too narrowly—that only those lines of research that might be profitable are pursued. In some cases, such influence clearly can be detrimental to the conduct of science. In other cases, such as the early STM community, corporate influence actually prompted academic research to adopt a diversity of approaches and a more expansive outlook. 9.4 Building and Buying Until 1986, all probe microscopes (whether corporate or academic) were home-built, in that they were put together by the groups that were using them. Yet home-built instruments were not made entirely from scratch— some components were made by hand, but most were bought from commercial suppliers. The STM designs were strongly shaped by the commercial availability of components such as operational amplifiers, and high-grade materials such as platinum-iridium alloy. In some cases, STM builders simply ordered these items from catalogs. In other cases, they were active consumers, lobbying companies to modify products (vacuum chambers, piezoelectric crystals, video output devices, etc.) to suit their needs.12 Thus, STMers were both consumers and producers of equipment. This is important to note because most proponents of academic capitalism focus solely on academics as producers of marketable knowledge and goods. Their recommendations for achieving a more commercial university therefore center on stimulating professorial start-up companies. Yet universities may find that the best way to gain influence over an instrumental community is by encouraging professors to be savvy, active consumers who can trade their expertise for favorable deals from manufacturers. In a number of indirect and often counterintuitive ways, commerce supplied the infrastructure needed to make the STM community grow. Information about commercial sources of reliable components and materials was 12. . Much recent history of technology has focused on the active role of users. For consumers’ adaptations of artifacts for uses that manufacturers were unaware of, or even opposed, see Kline and Pinch (1996). For users’ pressure on companies (often—as in instrumental communities—through threats to form their own cooperatives or firms), see Fischer (1992). For an overview of different kinds of user activity, see the essays in Oudshoorn and Pinch (2003).

Instruments of Commerce and Knowledge

301

a major topic of gossip among early STM builders. Those who had built working instruments offered blueprints and recommended particular commercial suppliers of components to new members of the community. Those newcomers, anxious to make up for lost time, rarely questioned their predecessors’ advice, so that STM-building came to resemble doing a project from Popular Mechanics. STM-building became standardized through STMers’ nearly ritualistic allegiance to recommended suppliers, even when the technical rationale for their components disappeared. For instance, IBM’s STMers used a trademarked rubber called Viton (from Dupont) to dampen vibration, because Viton could survive ultrahigh vacuum.13 Later, as IBM’s blueprints disseminated, Viton continued to be widely used even in academic STMs that operated in air or fluid, not vacuum.14 A commercial infrastructure also helped STMers standardize the materials they looked at with their microscopes. As Daniel Lee Kleinman (2003) has noted, corporate influence over the choice of research materials is pervasive but indirect. Tapping into the right commercial infrastructure can be crucial to growing an instrumental community. Reliable, cheap commercial sources of materials give newcomers easy access to research, and give the rest of the community a yardstick by which to measure newcomers’ progress. Among corporate surface science STMers, this yardstick was provided by a few key semiconductor samples. Newcomers had to prove that their microscopes could resolve individual atoms of those canonical semiconductor samples to gain entry to the instrumental community. Atoms can (ordinarily) only be seen on semiconductor samples when they are kept in ultrahigh vacuum. Thus, when academic STMers designed microscopes for use in air and water, they could no longer use the canonical semiconductor surfaces as a standard of microscope-building ability. Quate, Hansma, and others looked desperately for new yardstick materials. Gold, paraffin, and graphite vied for the job, but graphite won out partly because ultrapure samples could be obtained cheaply from commercial sources.15 Union Carbide used graphite to make monochromators for neutrons, an application requiring extraordinarily pure samples; hence, they rejected large amounts of slightly imperfect graphite still pure enough for STMers. The Quate group heard about this and alerted other academic groups who then called Union Carbide’s graphite man, Arthur Moore, to get cheap, more or less standardized samples. By 1989, the STM community was awash in graphite, such that talks about that material outnumbered talks on semiconductors at the annual STM conferences. Sometimes, the industrial relevance of materials was a direct influence acting on academic microscopists, feeding back into the designs of their in13. , , . 14. For similar instances of practices spreading through an experimental community through transmission of knowledge about particular brands, see Jordan and Lynch (1998). 15. .

302

Cyrus C. M. Mody

struments. Cal Quate, for instance, framed his STM work within Stanford’s long tradition of industrial ties and his own involvement in developing acoustic microscopy in the 1970s as a nondestructive characterization tool for manufacturing.16 Nondestructive testing held tremendous promise for microelectronics, where chips must be inspected throughout the manufacturing process, yet where traditional testing tools (especially electron microscopy) require breaking and discarding expensive silicon wafers. Quate moved into STM believing it could be the next generation nondestructive evaluation tool for the microelectronics industry. He was then able to inspire his former students and postdocs at microelectronics giants such as IBM to follow his lead and join the STM community.17 The STM, though, requires a conducting (metal or semiconductor) sample, whereas most microelectronic materials have an insulating oxide layer. This was unproblematic for corporate surface scientists tasked with generating basic knowledge about materials like silicon and gallium arsenide. Yet STM’s restriction to conducting materials blocked its use in nondestructive testing or in industrial quality control more generally. So when IBM allowed Gerd Binnig to take a sabbatical at Stanford from 1985 to 1986, he and Quate adapted the STM so it could use interatomic forces to map insulating materials, calling their new invention the atomic force microscope (AFM). Thus, Quate positioned his research much further downstream in IBM’s R&D cycle than most of IBM’s own STMers and, together, IBM and Stanford dramatically shifted the world of academic and corporate probe microscopy. 9.5 Commercialization and Gray Markets What we have seen so far, then, are the more intricate, unglamorous ways corporate and academic actors are linked within an instrumental community: through pedagogy, through institutional politics, through commercial infrastructures, and through tacit knowledge. These are not the relationships that exercise most analysts of academic capitalism. Instead, both proponents and critics tend to focus on large corporate buy-ins to academic departments, professors keeping research secret so they can patent it, and corporations and universities colluding to suppress unfavorable results. One topic central to the academic capitalism debate will occupy the rest of this chapter—the commercialization of academic research and the founding of professorial start-up companies. The commercialization process was not sudden, dramatic, and profit-driven, but built slowly and qui16. , , . “Nondestructive testing” means that the process of quality control testing does not damage the item being tested. Products can be taken off the assembly line while half-finished, inspected, then returned to the assembly line. See Quate (1985) for a brief description of scanning acoustic microscopy at Stanford. 17. Quate’s optimism for STM derived from its ultrahigh resolution and the fact that (ideally) the STM tip does not touch (and thereby mar) the sample surface.

Instruments of Commerce and Knowledge

303

etly from the kinds of practices I have described thus far. Commercialization of probe microscopy was driven less by profit-seeking and more by the desire of elite STMers to grow a larger instrumental community in which their groups would be centers of expertise. This desire was strong among both corporate and academic groups, though the motivations differed in the two moieties. At IBM and (to a lesser extent) Bell Labs, research managers wanted to increase the number of in-house STM groups, so as to keep the center of the STM community within the corporation. Academics like Quate and Hansma wanted to grow the STM (and AFM) community because they were looking for new applications and audiences, and because they wanted to build a critical mass of researchers committed to nonsurface science probe microscopy. Both Bell Labs and IBM built something like an internal free market for tunneling microscopy, with multiple groups in different parts of the organization given similar tasks and competing for the attention of senior managers. Both companies also developed an infrastructure for STM research that allowed new lab groups to get up to speed very quickly. For instance, Bell Labs housed several (varying between two and four) STMs in an old tractor shed on the edge of its property. There, microscope builders could very quickly trade ideas, materials, blueprints, and software—very much in the same way that Quate’s students worked on multiple microscopes at once and cannibalized parts and design ideas from one project to another.18 IBM took the internal STM market/infrastructure to even greater lengths. IBM had been first into STM, yet it took other IBM groups just as long—almost two years in some cases—as everyone else to replicate Binnig and Rohrer’s microscope. Thus, senior management cast about for ways to package the tacit knowledge of instrument-building and reduce replication time. The preferred strategy was to make semistandardized, batch-produced (Scranton 1997) STM packages available to its researchers. The first was the “Blue Box” designed by Othmar Marti, a Swiss graduate student doing doctoral work at IBM Zurich.19 The Blue Box was primarily an electronics package—researchers constructed the hardware themselves, often using Binnig and Rohrer’s designs. The STM electronics presented a significant challenge; complicated feedback circuitry brings the probe to the surface, reads out and controls the tunnel current, and rasters the tip without crashing. The success of the Blue Box in allowing newcomers to work around these difficulties inspired a more ambitious effort at IBM Yorktown. There, Joe Demuth, manager of an STM group, assigned his postdocs to work with Yorktown’s Central Scientific Services (CSS) shop to develop and batch-produce complete STMs to “sell” to other Yorktowners.20 18. , . 19. , . 20. , , .

304

Cyrus C. M. Mody

By 1990, about a dozen of these CSS STM’s were in use at Yorktown and the nearby Hawthorne facility; some also accompanied former IBM postdocs when they left to become professors.21 Yorktown management encouraged use of the CSS STM by making its purchase a zero-cost budget item for each research group. Still, groups had to invest labor—usually a postdoc—to make the microscope productive. This confronted its postdoc users with a dilemma. They needed to creatively solve technical problems and display initiative to managers to advance to staff positions. This meant they needed to radically reconstruct the CSS STM to show off their skills. Postdocs also found that the CSS STM pulled them into intense institutional politics. Postdocs using the CSS STM found that competing Yorktown groups viewed them as partisans of Demuth’s style of microscopy. Thus, there was great institutional pressure on these postdocs to disavow the CSS STM by rebuilding and transforming it.22 The culture of research at Yorktown made it impossible to view the CSS microscope as a ready-touse black box (Latour and Woolgar 1986), and therefore precluded any possibility of its commercialization outside IBM. In contrast, commercialization was more successful from academic STM and AFM groups largely because of the outward-looking, multidisciplinary style they had cultivated in order to avoid competing with the surface scientists at IBM and Bell Labs. People like Binnig, Rohrer, Quate, and Hansma were extraordinarily open with newcomers, freely offering blueprints and advice in order to build a critical mass of nonsurface science probe microscopists. Thus, the circulation of materials and ideas—a kind of “gray market”—became the norm in academic STM and AFM. Software (to control probe and display images) was particularly easy to distribute, and the groups that gave it away both won goodwill within the instrumental community, and ensured access to modifications to the software that their collaborators came up with.23 Sometimes code was given for free, sometimes at nominal cost. Profit was not the motive for dissemination. A well-traveled hardware innovation was the microfabricated AFM cantilever. One perceived defect of early AFMs was that probes were laboriously handmade from small strips of aluminum foil with a tiny sliver of diamond glued on one side and a tiny shard of glass on the other.24 Although these cantilevers could yield exquisite AFM images, each required considerable time and training to make, and results were so particular to one cantilever and its maker that images taken with different cantilevers were diffi21. . 22. , . 23. . 24. Diamonds were used as tips because their sharp points were less likely to wear down from repeated use than other materials. The glass on the back of the cantilever acted as a small mirror, bouncing laser light into a photodiode; the position of the reflected beam in the photodiode indicated how much the cantilever was bending (i.e., a proxy for how much the surface was pulling or pushing on the diamond tip).

Instruments of Commerce and Knowledge

305

cult to compare. Handmade cantilevers sufficed early on, when every image was new and spectacular. But as the technique matured, AFMers sought standardization. The Quate group delivered this by integrating itself with microlithography expertise at Stanford and around Silicon Valley. Over several years, Quate sent his students to other electrical engineering professors at Stanford to learn the microelectronics industry’s techniques for patterning and etching silicon. The students adapted those techniques to make batches of small, standardized silicon cantilevers. By 1990, Quate was sending surplus probes to friends and collaborators, sometimes so he and his students could share authorship on those collaborators’ papers. Quickly, Quate-type probes became essential to AFM research.25 Quate’s and Hansma’s multidisciplinary collaborations prepared the ground for commercialization in other ways. These collaborators would usually found their own STM or AFM groups at other universities and effectively advertise for the technique, building interest in probe microscopy among biologists, electrochemists, mineralogists, and so forth. Eventually interest from those disciplines would turn them into markets for commercial STMs and AFMs. At the same time, Quate’s and Hansma’s graduate students learned to deal with potential “customers” from other disciplines and to design microscopes with their needs in mind. The leap from these practices to outright commercialization was very small. The first to make this leap was a Quate student, Doug Smith, who founded the Tunneling Microscope Company in 1986. Smith had only one employee, a fellow student who helped put together scanners, and he recruited customers by word of mouth. He viewed the company less as an ongoing enterprise than as a way to sweeten the hardships of graduate school—a well-circulated story is that he sold just enough microscopes to buy a BMW before taking a postdoctoral fellowship. Quate himself pushed Smith to separate scholarship and business more cleanly: “Dr. Quate said ‘graduate students work, eat, and sleep, and most of the time they go hungry.’ You can’t have a company and be a graduate student at the same time, so Doug had to finish up and move out.”26 On the demand side, Smith’s customers were in much the same position as the postdocs at IBM who were presented with the CSS STM. They saw a commercial STM as a way to quickly catch up and join a hot new instrumental community. Yet they knew that if they were to join the elite of that community, they would have to demonstrate instrument-building virtuosity on their own. Thus, like the batch-produced IBM instruments, Smith’s commercial STMs were more starter kits than black-boxed devices. To use the instrument, customers needed to construct much of it themselves.27 All 25. , , . 26. . 27. , , .

306

Cyrus C. M. Mody

Smith sold was the microscope “head”—the piezoelectric scanner, tip, base, and vibration-isolating stacks of Viton. Customers built the electronics themselves, customizing the microscope for their own applications. Today, when customers buy an off-the-shelf AFM, they generally are buying all of the expertise of microscope-building so that they will not have to develop it themselves. Early on, that expertise was exactly what customers of commercial STMs did not want to buy. Instead, they were purchasing time, membership in an instrumental community, and a platform on which to demonstrate their own instrument-building expertise. Consumers wanted to be on an equal footing with producers, whether in a commercial or academic setting. 9.6 Digital Instruments Commercialization of the STM was accomplished through a series of minute, unremarkable steps. Very little separated the home-built STM (made with parts ordered from catalogs, often using blueprints given by colleagues) from the STMs sold by Smith and (internally) IBM. The next step was only slightly more dramatic—the founding of organizations dedicated wholly to manufacturing and selling probe microscopes. Critics and proponents of academic capitalism both lay heavy emphasis on the founding of start-ups. It is seen both as the best means to extract profit from academic work, and as the ultimate distraction from the university’s pedagogical mission. Yet both these views neglect the realities of how start-ups operate within an instrumental community. Profit is often the least visible (and least successful) motivation for founding a start-up. Start-ups often extend and enhance the pedagogical culture of an instrumental community rather than despoiling it. Digital Instruments (DI), the first true start-up in the probe microscopy community, was the brainchild of Virgil Elings, one of Paul Hansma’s colleagues in the physics department at UCSB. Elings’ first contact with the STM community, though, was Niko Garcia, a Spanish academic with close ties to IBM, who came to give a lecture at UCSB. After talking with Garcia and Hansma and attending the 1986 STM Conference in Spain, Elings saw a market for an off-the-shelf STM and offered to cofound a company with Hansma. Hansma was even more wary than Quate of commerce encroaching on his lab’s activities, so he declined. However, he gave Elings the same advice and schematics he made available to other STMers.28 With this, Elings and his son built a prototype in their garage and entered it in a junior high science fair (where it took last place, since, as the judges pointed out, “everybody knows you can’t see atoms”).29 28. , . 29. , .

Instruments of Commerce and Knowledge

307

For Elings, building the prototype was a chance to make sure the Hansma design was commercializable, but also to test—and discard— many axioms of STM-building that had accrued since 1981. Elings saw STM-builders’ trade secrets as geared to instruments that were finicky and difficult to operate; and he saw possession of these trade secrets as limiting the STM community to those deemed serious enough to build their own microscopes. Elings wanted, eventually, to make STMs for nonbuilders who demanded a simple to operate black box. Thus, he delighted in debunking the STM-builders’ recipes by creating a more streamlined, easy to use, more durable tool. Elings wanted DI to be the first to market a commercial microscope in time for the annual STM Conference in 1987. His initial plan was to sell a computer-controlled microscope (hence Digital Instruments). However, by the time he brought in a former student, Gus Gurley, as cofounder, it was too late for Gurley to write the necessary software in time. Instead, Elings marketed the analog Nanoscope I as DI’s first product. Probe microscopists from this era—both builders and buyers—remember their first acquaintance with the Nanoscope as a turning point. Now, for the first time, researchers could join the STM community without having to build any part of their microscope. Moreover, unlike Smith’s clients, DI’s customers did not need to have personal ties to the community. People could (and did) simply call up Digital Instruments and order a microscope. Yet though it marked an important shift, the Nanoscope I still illustrates the gradual, emergent character of commercialization. Like the CSS STM and Doug Smith’s instrument, the Nanoscope I was more a kit than a fullfledged, black-boxed research tool. Indeed, Elings now calls this era at DI the “toy business”—both for the Nanoscope I’s immature design, and for its lack of serious applications.30 In following Hansma’s lead, Elings designed an air STM, rather than the expensive, narrowly-focused ultrahigh vacuum instruments used at IBM and Bell Labs. This made sense in opening up a broad market, since few disciplines were willing to deal with or pay for an ultrahigh vacuum chamber (which, in any case, ruined samples relevant to almost everyone except surface scientists). Yet it was unclear in the 1980s what air STM could be used for, or what the images it produced meant. Only in 1991 to 1992 did a consensus develop that air STM was often not relying on tunneling for its contrast mechanism, and that many well-publicized air STM images (particularly of DNA) were erroneous. As a result, most air STMers abandoned the technique and followed Quate and Hansma to AFM, usually by buying one of DI’s newly-available Multimodes (capable of running an STM or AFM). At this point, Elings had no sales force. He simply advertised in Physics 30. , .

308

Cyrus C. M. Mody

Today (“$25,000 for atomic resolution”) and orders came in. Instruments were FedExed to buyers, who put them together and got the microscope running on their own. Despite this minimal marketing and customer service (and limited product utility), the toy business was successful. An advertisement from 1990 estimates that in the first three years, DI sold more than 300 Nanoscopes at $25,000 to $35,000 each.31 The probe microscopy community expanded quickly, and the center of gravity shifted as well. As more people bought instruments, AFM and air STM began to outweigh ultrahigh vacuum STM, and the corporate labs became less dominant. High demand created a waiting list for DI’s instruments, prompting a policy that researchers who wanted a microscope quickly could promise to name DI’s founders or employees as coauthors on papers generated with Digital’s products. 9.7 The Start-Up Era The end of the toy business roughly corresponded to the end of DI’s monopoly on commercial probe microscopy. By 1990, there were several new STM and AFM manufacturers whose products and strategies differed considerably. Some competed with Digital Instruments for the generalpurpose microscope market, some targeted specific disciplinary niches. Some made easy to operate black boxes, some built “open architectures” for researchers who wanted to tinker with and modify the device. Some survived, others floundered. All were small companies, mostly founded out of universities specifically to make probe microscopes, though a few moved into microscopy from other product lines. No big firms made more than desultory attempts to sell STMs or AFMs—though a few (Hitachi, IBM, Perkin Elmer) started down that road. These start-ups were founded for a very diverse set of reasons. Most debates about academic capitalism simply assume that, under the right set of incentives, professorial start-ups are inevitable; for good or ill, the inducements for professors to commercialize their work will be irresistible. Maybe so, but this assumption looks less reliable when the contingent and often counterintuitive reasons why people found start-ups are examined. Digital Instruments provides a striking example. Digital Instruments and Elings thrived at UCSB’s disreputable margins. When he arrived in the late 1960s as a brash, confrontational professor, it was hoped Elings would build UCSB’s reputation in high energy physics. His swagger, though, led to conflict with his department, which sidelined him into running its lucrative but unloved Master’s of Scientific Instrumentation program.32 Un31. From FASEB Journal, v. 4, n. 13 (1990), p. 1. 32. .

Instruments of Commerce and Knowledge

309

cowed, Elings transformed the master’s program into his personal empire and a source for patents and start-up companies. In the master’s program, students from many educational backgrounds (biologists, engineers, even psychology majors) learned to build all kinds of measurement technologies—not just research instruments but also meters and tools for industry. Elings developed a pedagogical method that prized tacit over formal knowledge, participation over instruction. Instead of using textbooks and lectures, he simply connected students with professors on campus who needed instruments built and let them learn by doing. Because student projects were based on finding solutions to real problems faced by local researchers, they often yielded technologies Elings could market to those researchers’ subdisciplines. Students learned to understand customers’ needs and design technologies to answer them. This made former master’s students the most important source of early employees for Digital Instruments. So UCSB did, in a way, encourage creation of DI, though no school would replicate their path. By sidelining a brilliant but difficult professor to the poorly-regarded master’s program, they encouraged him to reject campus culture, denigrate academically-instilled formal knowledge, and be receptive to the commercial possibilities of the tacit knowledge his students accrued. Moreover, in making clear that Elings’ commercial ventures hindered his academic career, the UCSB physicists made it more likely that Digital Instruments would be his bridge to leaving academia. Tension between Elings and UCSB even smoothed technology transfer from Hansma to DI, since Elings’ hostility toward academic researchers meant he rejected Hansma’s designs until they had been engineered to look more like commercial products than most home-built instruments. Disgruntlement of a different kind motivated the engineers who worked for DI and its competitors. Several of the early STM and AFM manufacturers were founded in the heart of the West Coast military-industrial complex. Graduate students who had grown accustomed to the picturesque surroundings and lifestyle of southern California often sought employment nearby, usually with defense firms like Lockheed and Hughes. Yet defense work galled many of these engineers, driving them to probe microscopy. Much of Elings’ early workforce came to DI for this reason; and in Los Angeles Paul West, one of John Baldeschwieler’s former postdocs at Caltech, grew so frustrated with defense work that he started his own probe microscopy company, Quanscan. As the West Coast start-ups matured, they took on large numbers of students from their affiliated academic groups as collaborators or summer employees. As Sumell, Stephan, and Adams (chapter 8, this volume) point out, that kind of corporate experience during graduate school can be a strong inducement to stay in-region. And, in fact, by the early 1990s Quate, Hansma, and Baldeschwieler grad-

310

Cyrus C. M. Mody

uates were routinely staying on the West Coast either to work at (respectively) PSI, DI, or Topometrix, or to join new STM and AFM start-ups founded by engineers who had left those companies.33 The STM and AFM start-ups positioned themselves relative to each other in ways that mirrored the relationships between the academic groups with which they were associated. Digital Instruments may not have been officially affiliated with the Hansma group at first (indeed, even at the best of times, there was always some suspicion between the two groups)—and Elings was certainly proud of the ways the Nanoscope differed from Hansma’s microscopes—but DI’s products bore an obvious genealogical kinship with Hansma’s STMs and AFMs, and DI drew on Hansma’s reputation in the community. In return, Hansma’s design innovations spread much farther and faster than those of professors not affiliated with a start-up. Hansma’s peers in the probe microscopy elite noticed how DI helped spread Hansma’s ideas and resolved to affiliate with their own microscope manufacturers. Usually this meant helping former students and postdocs found start-ups. For instance, two Stanford postdocs, Sung-Il Park and Sang-Il Park (no relations) started Park Scientific Instruments (PSI) in 1989 with Quate’s assistance and quickly became the major employer of Quate group veterans. Park Scientific Instruments’ designs traveled much more directly from Stanford than DI’s designs did from UCSB. Moreover, the research on new AFM applications conducted at Park Scientific often picked up just where Quate’s own research left off.34 Commercialization was the continuation of academic science by other means. Similarly, just as John Baldeschwieler’s group at Caltech always lagged behind Quate’s and Hansma’s in popularizing its discoveries and innovations, the company he helped Paul West found, Quanscan, lagged behind DI and Park Scientific in marketing commercialized versions of the Caltech designs. But not all probe microscope companies were founded as proxies for the competition between academic groups in an instrumental community. Some were started to extend academic collaborations. For instance, Stuart Lindsay, a physics professor at Arizona State University, had been an early collaborator of Hansma’s, helping to adapt the STM for electrochemistry and biophysics. Once Hansma’s designs were commercialized by DI, Lindsay pressed Elings to adapt the Nanoscope for Lindsay’s colleagues in electrochemistry and biophysics—to no avail, since Elings was usually hostile to adapting the Nanoscope for anyone (DI had a strict no-custominstruments policy), especially when the suggestion came from outside the 33. That is, by the late 1990s one could see a probe microscopy cluster forming, primarily around Santa Barbara and Los Angeles, with start-ups like Pacific Nanotechnology, Quesant, Asylum Research, and Nanodevices founded by veterans from DI and Topometrix. Engineers who left Park Scientific tended to drift into established Silicon Valley firms such as KLA-Tencor. 34. , , .

Instruments of Commerce and Knowledge

311

company. So Lindsay founded his own company, Molecular Imaging, to make attachments to the Nanoscope that would make it more compatible with electrochemistry and biophysics—attachments that DI grudgingly distributed for a few years until it developed its own competing line.35 Lindsay’s other motivations for founding a start-up say a great deal about how commercialization and pedagogy fit together in instrumental communities. Long before Molecular Imaging, his group—like Hansma’s at UCSB—had become a center for distributing blueprints and (especially) software to new STM builders. One of Lindsay’s technicians, Uwe Knipping, developed one of the first and most sophisticated computercontrolled microscopes. Knipping’s software formed the basis for Lindsay’s academic network-building, but it also caught the eye of two local entrepreneurs, Larry and Darryl McCormick, who founded a company, Angstrom Technology, to commercialize it.36 As it turned out, Knipping’s architecture was far too sophisticated for a commercial instrument, and the enterprise failed. But Lindsay had gotten a taste for how network-building in the academic domain might be enhanced by commercialization. So a few years later when he encountered a former postdoc of his who was having trouble finding work, Lindsay decided to put the former postdoc in charge of starting a new company— Molecular Imaging—as an extension of Lindsay’s research at Arizona State University (ASU).37 This is probably the classic—if woefully understudied—story of commercialization: a professor’s technicians and graduate students make a widget, then the professor’s colleagues call up asking for their own widget (or blueprints thereof), a student starts making batches of widgets in their garage, and eventually, whether to help position the professor within that instrumental community or to give lab personnel needed work, the widget-making is spun off into its own organization. In only a few cases were probe microscope start-ups inspired primarily by profit. Paul West’s Quanscan, for instance, always had the most venture capital, the most MBAs, and the slickest advertising. Yet West’s own motivation was more personal than commercial: he saw entrepreneurship as an intellectual challenge and a path to personal growth.38 Even so, Quanscan was actually at a disadvantage relative to competitors because it was run more like a for-profit business. The venture capitalists continually interfered in operations, the MBAs had trouble understanding the values of an instrumental community that they had never participated in, and the advertising alienated many potential customers. In contrast, Park Scientific, like DI, was a more rough-hewn affair. Sung-Il Park’s barber was hired as the office manager, for instance, and the 35. . 36. , . 37. , . 38. , , .

312

Cyrus C. M. Mody

senior executives were Quate students with no business training.39 Indeed, the Parks carved a market niche by letting it be known they were more interested in technically sweet innovations than in mundane moneymaking. As “gentleman scientists,” they could speak as peers with other researchers and capture customers’ trust. Park Scientific gained a reputation for making builders’ instruments—well-crafted, reliable, with enough idiosyncracies and innards showing to be reminiscent of a microscope made by a graduate student. Park was even willing to work with individual customers to build a microscope for a specific application (something DI never did)— if the engineering required a certain finesse. Through most of the 1990s this strategy kept PSI running near DI, but Digital ultimately won out because of its skeptical (sometimes hostile) attitude toward the expertise of its customers. Where Park Scientific was willing to relive its Quate group origins by respecting the knowledge of foreign disciplines, DI made one type of microscope for everyone, and hid the workings of that instrument completely from customers’ view. This attitude allowed DI to break into the industrial market, where it could sell many more, and more expensive, microscopes to companies that usually wanted low-level technicians to learn how to use the instrument in a day or two—market conditions for which PSI was wholly unprepared. As much as DI eventually prospered by distancing itself from Hansma’s academic model, though, the company’s success hinged equally on continual sharing of culture, people, and inventions between start-up and academic lab. Both Elings and Hansma saw tacit, rather than formal, knowledge as primary in instrument-building—Elings because of his work in the instrumentation master’s program, Hansma because the contours of the STM community had pushed him to encourage multidisciplinary collaborations and undisciplined instrument-building. This shared emphasis on the tacit meant both men took in people with diverse and unusual educational backgrounds: junior high students, river guides, undergraduates, yoga instructors, retirees, psychology majors, and historians.40 This diversity was almost unthinkable at other centers of probe microscopy such as IBM or Bell Labs. Age and gender diversity followed along with diversity in educational background. There were, to be sure, a few women and young (i.e., collegeeducated but no Ph.D.) people in the corporate labs, but they tended to exit to academia somewhat more quickly than their male colleagues (women usually to run their own academic groups, college graduates to go get Ph.D.s). In the Hansma group and DI, very young people did much of the daily work, while women and older people were often the source of crucial innovations. A retired teacher named Sam Alexander, for instance, to whom 39. , . 40. , , , , , .

Instruments of Commerce and Knowledge

313

Hansma had donated part of his laboratory, came up with the optical detection scheme that is now the basis of DI’s AFMs. And Helen Hansma, Paul’s first wife, used her Ph.D. in biochemistry to transition Paul’s group into biophysics at the same time that she transitioned back into lab work after many years of rearing children and teaching in the Santa Barbara school system.41 As members of DI and the Hansma group became aware of parallels between their organizational styles, they appropriated these similarities to accelerate the two-way flow of people, materials, designs, and knowledge. After the initial phase (when most DI employees were Elings’ former master’s students), several Hansma graduates, postdocs, and collaborators took high-ranking jobs at DI. Individuals on both sides collaborated to transform Hansma’s research into commercial products; for instance, the Hansma AFM (on which DI’s fortunes eventually rested) was turned into a product through negotiations between Barney Drake (Hansma’s technician) and James Massie (a former Elings student) over which elements of the Hansma design were indispensable and which were too finicky for anyone but the graduate students who built them.42 Hansma also, for the first time in his career and at Elings’ behest, began patenting his research.43 As DI’s sales increased, the Hansma group kept its place at the forefront of the AFM community through its steady supply of DI instruments and the ability of Hansma’s students and postdocs to go up the road to DI to scavenge parts and advice.44 That is, whatever his initial reservations about commercialization, Hansma came to see the partnership with DI as a way to position himself—intellectually and socially— within his instrumental community. In turn, once the toy business ended in the early 1990s, Elings began to imitate Hansma’s tactic of bringing in collaborators from various disciplines. Digital Instruments built its own group of researchers from biophysics, magnetics, and polymer chemistry, who (like Hansma’s postdocs) worked with instrument-builders, developed and published on new STM and AFM applications, and traveled to give talks and attend conferences to spread word about the technique.45 Though DI was a profit-making ven41. That is, Paul Hansma’s lab, and DI, upended the whole notion of an educational mismatch (Bender and Heywood, chapter 7, this volume). For Hansma and Elings, AFM-building was not a static process for which any educational background could be mismatched. Rather, diverse educational backgrounds offered an engine for pushing AFM-building in new, unexpected directions. 42. , . 43. Quate, Hansma, Lindsay, Binnig, and the other more outward-looking probe microscopists all patented their work, especially after 1986. Many of those patents were probably overlapping and difficult to defend. The point of the patents, though, was initially to tie the academic group to the company with which it was most closely affiliated. Later on, patents were used more to raise the stakes for other academic groups and start-ups to join this elite club. 44. . 45. , .

314

Cyrus C. M. Mody

ture, its success arose partly from academic activities: doing basic research, publishing articles, training and “graduating” employees, giving talks. These practices were then widely emulated by the other start-ups. 9.8 Conclusion So what does probe microscopy tell us about commercialization of academic knowledge and the value of corporate-academic linkages? First, the development of probe microscopy shows how thoroughly—yet intricately and indirectly—the corporate and academic worlds are connected. The locus of academic research is much wider than the university campus, just as the locus of commerce is wider than the for-profit business. Instrumental communities are distributed across academic and corporate institutions. Commercialization—the transformation of academic research into commerce—is not a simple pipeline from university to firm. Commercialization can play many roles within an instrumental community, and academic research can be traded for many things other than money. Attempts, therefore, to directly stimulate and accelerate the transformation of academic research into cash may well backfire. As we have seen, it was the looser, indirect ties between corporate and academic groups that fostered the growth of STM and AFM and encouraged start-ups to emerge from universities, rather than direct pressure from corporations or overt incentives from governments and universities. Thus, proponents of academic entrepreneurialism should be wary of focusing too narrowly on increased profit as the fruit of a commercialized university. As we have seen, trading goes on all the time in instrumental communities. The token of exchange is usually a mix of knowledge, prestige, personnel, time, materials, money, opportunity, and so forth. The popularity of various forms of barter changes as the instrumental community changes. Commercialization can restrict some exchanges and make money-based trades more prevalent. Few instrumental communities, though, reach the point where their products can be sold for money. Even within the probe microscopy community, only the atomic force microscope and the magnetic force microscope have been commercial successes. The STM, which provided the first product for microscope manufacturers, was effective in training engineers to build microscopes, but never found industrial application. University administrators who hope that stimulating professors to turn a gray market into a profit-making start-up will bring real patent revenue to their school will almost always be disappointed. Moreover, development of an entrepreneurial instrumental community may require that its members be drafted from less profitable fields where commercialization did not occur. The STM and AFM community, for instance, initially drew on its members’ expertise in low-energy electron

Instruments of Commerce and Knowledge

315

diffraction, sandwich tunnel junction spectroscopy, and field ion microscopy—instrumental communities with poor records of commercialization. Later, STM and AFM pulled in participants from many fields (surface science, biophysics, mineralogy, electrochemistry, polymer science—some more commercialized than others) who aided groups like Quate’s and Hansma’s in their gray market activities. Instrumental communities and disciplines that are not conducive to profit-making nevertheless provide the infrastructure and knowledge/labor pool for communities in which profit may be enormous. Policymakers should not think they can predict which will be which; nor are they likely to succeed if they encourage only the one at the expense of the other. Policymakers may be best advised to encourage professors to foster gray markets within their instrumental communities—whether as consumers, producers, or both. Gray market activities of trading research materials, people, and components of technologies enlarge the outlook of academic research and allow academic scientists to be influential even when they are not profitable. Finally, both opponents and supporters of corporate involvement in university life have seized on grains of truth. Supporters have it right that corporate-academic linkages are desirable, even necessary, for research and innovation. There was no golden age when faculty operated independent of firms, pursuing disinterested research. Knowledge production in physics, engineering, and chemistry was always aided by faculty consulting and trading of personnel and ideas. The oft-criticized commercialism of the “biotech revolution” merely extended long-standing entrepreneurial practices into molecular biology. The STM and AFM case does, however, give reason for opposing the notion that universities should be run as businesses, squeezing profit where they can and operating along the rational lines of modern management. The probe microscopy community developed rapidly because participants could point to different institutional poles—corporations, universities, national labs. At times, innovation occurred because these poles were opposed—as when Hansma and Quate shifted from surface science and UHV STM to new designs and applications. At other times, innovation occurred because participants strung out hybrid forms between these poles—the gray market of software trading, the CSS STM, and the toy business. Instrumental communities rely on a variety of actors contained in different kinds of institutions. If all these institutions are run on the same highly-managed, profit-driven model, then the movement of people and ideas—and the production of new technologies—will likely be hindered.

316

Cyrus C. M. Mody

Appendix Interviewees listed by alphanumeric, name, positions held over the period covered by the interview, and date of the interview. All interviews conducted by the author. AG1: Andy Gewirth: Hansma collaborator; University of Illinois; 6/26/01 BD2: Barney Drake: Hansma group technician; UCSB; 10/18/01 BH1: Bob Hamers: Yorktown researcher; University of Wisconsin; 5/9/01 BP1: Becky Pinto: Stanford; Park Scientific; KLA-Tencor; 2/3/04 BS1: Brian Swartzentruber: Bell Labs technician; University of Wisconsin; Sandia National Laboratory; 1/10/03 BW2: Bob Wolkow: IBM Yorktown; Bell Labs; NRC Canada; 5/22/01 CG1: Christoph Gerber: IBM Zurich technician; 11/12/01 CP1: Craig Prater: Hansma graduate student; Digital Instruments engineer; 3/19/01 DB1: Dawn Bonnell: Yorktown postdoc; University of Pennsylvania; 2/26/01 DB2: Dan Bocek: UCSB undergraduate; DI engineer; Asylum Research; 3/23/01 DB3: David Braunstein: Stanford; Park Scientific; IBM San Jose; 4/3/01 DC1: Don Chernoff: Sohio Research; Advanced Surface Microscopy; 9/5/01 DF1: Dave Farrell: Burleigh Instruments; 5/29/01 DR1: Dan Rugar: Quate student; Almaden researcher; 3/14/01 FG1: Franz Giessibl: IBM Munich; Park Scientific; Uni Augsburg; 11/16/01 GA1: Gary Aden: Topometrix executive; 3/12/01 HG1: Hermann Gaub: Ludwig-Maximilians Universität; 11/14/01 HH1: Helen Hansma: UCSB professor; 3/19/01 JA1: John Alexander: Angstrom Technology; Park Scientific; KLATencor; 10/15/01 JB1: John Baldeschwieler: Caltech; 3/28/01 JD2: Joe Demuth: Yorktown manager; 2/22/01 JF1: John Foster: Quate student; Almaden researcher; 10/19/01 JG1: Jim Gimzewski: IBM Zurich researcher; UCLA; 10/22/01 JG3: Joe Griffith: Bell Labs; 2/28/01 JH1: Jan Hoh: Hansma postdoc; Johns Hopkins; 6/10/02 JM1: John Mamin: UC Berkeley; IBM Almaden; 3/15/01 JM3: James Massie: Elings master’s student; DI engineer; 10/18/01 JN1: Jun Nogami: Quate postdoc; Michigan State; 6/28/01 JV1: John Villarrubia: Yorktown postdoc; National Institute of Standards and Technology; 6/28/00 JW1: Jerome Wiedmann: Elings master’s student; DI employee; 10/18/01

Instruments of Commerce and Knowledge

317

MA1: Mike Allen: UC Davis; Digital Instruments; Biometrology; 10/12/01 MK1: Mike Kirk: Quate student; Park Scientific Instruments; KLATencor; 10/12/01 MS1: Miquel Salmeron: Lawrence Berkeley National Laboratory; 3/9/01 MT1: Matt Thompson: Digital Instruments; 2/26/01 NB1: Nancy Burnham: Naval Research Lab postdoc; Worcester Polytechnic; 2/20/01 OM1: Othmar Marti: IBM Zurich student; Hansma postdoc; University of Ulm; 11/16/01 PH1: Paul Hansma: UC San Barbara; 3/19/01 PW2: Paul West: Caltech; Quanscan; Topometrix; Thermomicroscopes; 3/30/01 RC1: Rich Colton: Naval Research Lab; Baldeschwieler collaborator; 6/27/02 RT1: Ruud Tromp: Yorktown researcher; 2/23/01 SG1: Scot Gould: Hansma student; DI employee; Claremont McKenna; 3/27/01 SL1: Stuart Lindsay: Hansma collaborator; Arizona State; Molecular Imaging; 1/6/03 SM2: Sergei Magonov: Digital Instruments; 3/21/01 TA1: Tom Albrecht: Quate student; Almaden researcher; 3/14/01 TB1: Thomas Berghaus: Uni Bochum; Omicron; 11/19/01 TJ1: Tianwei Jing: Arizona State; Molecular Imaging; 1/7/03 VE1: Virgil Elings: UC Santa Barbara; Digital Instruments; 3/20/01

References Bassett, R. K. 2002. To the digital age: Research labs, start-up companies, and the rise of MOS technology. Baltimore, MD: Johns Hopkins University Press. Binnig, G., and H. Rohrer. 1985. The scanning tunneling microscope. Scientific American 253 (2): 50–56. ———. 1987. Scanning tunneling microscopy: From birth to adolescence. Reviews of Modern Physics 59 (3): 615–25. Blume, S. 1992. Insight and industry: On the dynamics of technological change in medicine. Cambridge, MA.: MIT Press. Bok, D. 2003. Universities in the marketplace: The commercialization of higher education. Princeton: Princeton University Press. Bromberg, J. L. 1991. The laser in America, 1950–1970. Cambridge, MA: MIT Press. Collins, H. M. 1975. The seven sexes: A study in the sociology of a phenomenon, or the replication of experiments in physics. Sociology 9 (2): 205–24. Creager, A. N. H. 2002. The life of a virus: Tobacco mosaic virus as an experimental model, 1930–1965. Chicago: University of Chicago Press. Elzen, B. 1986. Two ultracentrifuges: A comparative study of the social construction of artefacts. Social Studies of Science 16 (4): 621–62.

318

Cyrus C. M. Mody

Etzkowitz, H. 2002. MIT and the rise of entrepreneurial science. London: Routledge. Fischer, C. S. 1992. America calling: A social history of the telephone to 1940. Berkeley: University of California Press. Geiger, R. L. 2004. Knowledge and money: Research universities and the paradox of the marketplace. Stanford: Stanford University Press. Jackson, M. 2000. Spectrum of belief: Joseph von Fraunhofer and the craft of precision optics. Cambridge, MA.: MIT Press. Jordan, K., and M. Lynch. 1998. The dissemination, standardization, and routinization of a molecular biological technique. Social Studies of Science 28 (5/6): 773–800. Kaiser, D. 2005. Drawing things apart: The dispersion of feynman diagrams in postwar physics. Chicago: University of Chicago Press. Kenney, M., ed. 2000. Understanding silicon valley: The anatomy of an entrepreneurial region. Stanford: Stanford University Press. Kirp, D. L. 2003. Shakespeare, Einstein, and the bottom line: The marketing of higher education. Cambridge, MA: Harvard University Press. Kleinman, D. Lee. 2003. Impure cultures: University biology and the world of commerce. Madison, WI: University of Wisconsin Press. Kline, R., and T. Pinch. 1996. Users as agents of technological change: The social construction of the automobile in the rural United States. Technology and Culture 37 (4): 763–95. Knowles, S., and S. W. Leslie. 2001. “Industrial Versailles”: Eero Saarinen’s corporate campuses for GM, IBM, and AT&T. Isis 92 (1): 1–33. Kohler, R. 1994. Lords of the fly. Chicago: University of Chicago Press. Latour, B., and S. Woolgar. 1986. Laboratory life: The construction of scientific facts. Princeton, NJ: Princeton University Press. Lécuyer, C. 2006. Making Silicon Valley: Innovation and the growth of high tech, 1930–1970. Cambridge, MA.: MIT Press. Lenoir, T., and C. Lécuyer. 1995. Instrument makers and discipline builders: The case of nuclear magnetic resonance. Perspectives on Science 3 (2): 276–345. Mirowski, P., and E.-M. Sent. 2007. The commercialization of science and the response of STS. In Handbook of science and technology studies, ed. E. Hackett, O. Amsterdamska, M. Lynch, and J. Wacjman, 635–89. Cambridge, MA: MIT Press. Oudshoorn, N., and T. Pinch, eds. 2003. How users matter: The co-construction of users and technologies. Cambridge, MA.: MIT Press. Pantalony, D. 2004. Seeing a voice: Rudolph Koenig’s instruments for studying vowel sounds. American Journal of Psychology 117 (3): 425–42. Polanyi, M. 1962. Personal knowledge: Towards a post-critical philosophy. New York: Harper Torchbooks. Quate, C. F. 1985. Acoustic microscopy: Recollections. IEEE Transactions On Sonics and Ultrasonics 32 (2): 132–35. Rader, K. 2004. Making mice: Standardizing animals for American biomedical research, 1900–1955. Princeton, NJ: Princeton University Press. Rasmussen, N. 1997. Picture control: The electron microscope and the transformation of biology in America, 1940–1960. Stanford: Stanford University Press. Riordan, M., and L. Hoddeson. 1997. Crystal fire: The birth of the information age. New York: Norton. Scranton, P. 1997. Endless novelty: Specialty production and American industrialization, 1865–1925. Princeton, NJ: Princeton University Press. Shah, S. 2003. Community-based innovation and product development: Findings

Instruments of Commerce and Knowledge

319

from open source software and consumer sporting goods. Ph.D. diss. MIT, Sloan School of Management, Cambridge, MA. Shapin, S. 2003. Ivory trade. London Review of Books 25 (17): 15–19. Shapin, S., and S. Schaffer. 1985. Leviathan and the air-pump: Hobbes, Boyle, and the experimental life. Princeton, NJ: Princeton University Press. Shinn, T. 1997. Crossing boundaries: The emergence of research-technology communities. In Universities and the global knowledge economy: A triple helix of university-industry-government relations, ed. H. Etzkowitz and L. Leydesdorff, 85–96. London: Pinter. Slaughter, S., and L. L. Leslie. 1997. Academic capitalism: Politics, policies, and the entrepreneurial university. Baltimore, MD: Johns Hopkins University Press. Vettel, E. J. 2006. Biotech: The countercultural origins of an industry. Philadelphia: University of Pennsylvania Press. Wise, G. 1985. Willis R. Whitney, General Electric, and the origins of U.S. industrial research. New York: Columbia University Press.

10 International Knowledge Flows Evidence from an Inventor-Firm Matched Data Set Jinyoung Kim, Sangjoon John Lee, and Gerald Marschke

10.1 Introduction This chapter uses U.S. patent records to examine the nature and extent of knowledge spillovers from outside of the United States to U.S. industry. Because of their implications for economic development and science and technology policy, knowledge spillovers within a country or across borders have received considerable attention in the literature (e.g., Griliches 1992). Knowledge spillovers between innovating firms on opposite sides of a national boundary can occur via arm’s-length communication (scholarly publications, the material published in patent applications, and the like) or through person-to-person contacts in informal settings.1 Knowledge spillovers across countries may accompany the migration of workers who relocate across international borders or collaborations between workers across borders, which is a focus of this chapter. We examine whether international migration of researchers or the international location of subsidiaries by U.S. firms facilitates knowledge transfers across borders. Understanding how knowledge spillovers across countries work is of interest because of the role spillovers may play in economic growth and because of their implications for science and technology policy Jinyoung Kim is a professor of economics at Korea University. Sangjoon John Lee is an assistant professor of economics at Alfred University. Gerald Marschke is an associate professor of economics at SUNY Albany and a research fellow of the NBER and IZA. Kim was supported by the Korea Research Foundation Grant funded by the Korean Government (KRF-2007-332-B00062). 1. See Cohen, Nelson, and Walsh (2002) on various means by which innovating firms access know-how developed externally. See Agrawal, Cockburn, and McHale (2003) for evidence of the importance of social networks in promoting diffusion. Von Hippel (1988) documents how direct informal contacts between researchers affect knowledge spillovers.

321

322

Jinyoung Kim, Sangjoon John Lee, and Gerald Marschke

(Freeman 2005; Regets 2007). Knowledge spillovers from the U.S. and Europe may be an important factor for the impressive growth rates enjoyed in countries such as South Korea and Taiwan (Hu and Jaffe 2003). Understanding the consequences of the immigration of scientists and researchers to the United States for U.S. R&D productivity, for wages and job prospects of native workers, and for national security has important implications for policy-making in the immigration, labor market, and education arenas. Studies in both the economics and sociology of innovation literatures argue that new technologies are frequently tacit and difficult to transmit to the uninitiated via spoken or written communication (Polyani 1958, 1966). Often the most efficient means of transmission across organizational boundaries for tacit knowledge is via person-to-person contact involving a transfer or exchange of personnel. Recent findings that technological diffusion appears to be geographically limited (e.g., Jaffe 1989; Jaffe, Trajtenberg, and Henderson 1993; Audretsch and Feldman 1996; Zucker, Darby, and Brewer 1998; Mowery and Ziedonis 2001; Branstetter 2001; Keller 2002; Thompson and Fox-Kean 2005) are often interpreted as evidence of the tacitness of knowledge (e.g., Feldman 1994). More direct evidence exists that person-to-person interaction is important for the diffusion of technology. Cohen, Nelson, and Walsh (2002) surveyed R&D managers on the means by which they gather and assimilate new technologies. They find that firms access externally-located technology partly through the hiring of and collaboration with researchers from the outside. Moreover, they find that hiring/collaboration with outside researchers is complementary to other means of accessing externally produced knowledge, such as through informal communications with outsiders and more formal (such as consulting) relationships with outsiders. Almeida and Kogut (1999) find that scientific references that firms cite in their patent applications reflect the employment histories of their inventors, suggesting that ideas in the semiconductor industry are spread by the movement of key engineers among firms, especially within a geographical area.2 Zucker, Darby, and Armstrong (2001) find evidence of a payoff to firms that seek interactions with outside researchers. They find a positive impact on patent productivity for biotech firms that collaborate with university researchers on research and scholarly publications. The previously mentioned literature is at the least suggestive of the im2. See also the (indirect) evidence of a link between scientific mobility and technological diffusion in Kim and Marschke (2005) and Moen (2005). Kim and Marschke find that firms are more likely to patent in environments where scientists are likely to switch employers, suggesting that workers do transmit technological know-how when they move from one employer to another. Technical knowledge acquired by the scientist that can be transmitted to future employers is a form of general human capital. Thus, scientists would be willing to pay by accepting lower wages to acquire technological knowledge that they can exploit with multiple employers. Moen finds some evidence of this: he shows that technical workers in R&D intensive firms in Norway accept lower wages early in their career in exchange for higher wages later.

International Knowledge Flows

323

portance of the movement of technically trained personnel and extramural collaboration in facilitating knowledge transmission. We argue then that if one can measure the movement of researchers one can get a sense for the direction and magnitude of knowledge transmission. This chapter details the construction of a researcher-based data set and then describes its use in an analysis of the influence of foreign R&D on U.S. innovation. This chapter is part of a larger project that empirically examines issues related to the labor market for scientifically and technically trained personnel. The first half of the chapter describes the construction of these data. The inventors behind a patented invention, as well as their home addresses, are listed on each U.S. patent, as is the firm to which the patent is assigned and the assignee’s nationality of incorporation. The firm to which the patent is assigned is in most cases the employer of the persons named in the inventor field. We match names in the inventor fields of patents to construct a panel data set of inventors that contains the patents in each year of the inventors’ careers. The resulting data set allows us to track researchers geographically over the course of their career. These data afford us a window on the migration of technological human capital across national borders, one possible mechanism by which technology diffuses internationally. Patent applications disclose any previous relevant inventions. Through its citations to previous patents each patent documents the “prior art” upon which the new innovation builds, and because we know each cited patent’s assignee type, we know in which sector and country the prior art originated. These citations provide an additional window on the pathways of knowledge (for evidence that citations proxy for knowledge flows, see Jaffe, Fogarty, and Banks [1998], and Duguet and MacGarvie [2005]). In the final stage of constructing our patent-inventor data set we merge in citations made by the patent for each patent to which the inventor is named. One use to which we wish to put our data is in understanding the factors that influence the innovating firm’s accessing of recent innovations developed externally. A focus of this part of the analysis is the pharmaceutical and semiconductor industries, two industries that are especially prolific generators of innovations and patents and produce relatively homogenous outputs based on globally standardized technologies. Thus, the last stage of data construction involves carefully matching the inventor data to data on publicly traded firms in these two industries. After detailing our data construction efforts, we put our data to use investigating the international transmission of technology through scientific labor markets. For each patent assigned to a U.S. firm, we can determine the country of the inventor’s residence at the time of patent application, and whether they had ever been named as an inventor on a patent while residing abroad. Inventing in a foreign country can be regarded as evidence of an inventor’s exposure to research abroad. We also investigate which U.S. firms in our two industries cite foreign-assigned patents as prior art and thus build upon innovations originating abroad.

324

Jinyoung Kim, Sangjoon John Lee, and Gerald Marschke

Our main findings are the following: we find that there has been an increase in recent years of U.S. innovating firms employing or collaborating with researchers with foreign experience. This increase appears to work primarily through an increase in U.S. firms’ employment of foreignresiding researchers; the fraction of research-active U.S. residents with foreign research experience is low and appears to be falling, suggesting that U.S. pharmaceutical and semiconductor firms are going to foreign countries to employ such researchers as opposed to such researchers immigrating to the United States to work for U.S. firms. We find, however, that inventors migrating to the United States with past foreign experience show the fastest growth in patent productivity and their patents receive more citations, possibly because of either accelerating knowledge spillovers or more selective migration of high-productivity inventors. In addition, we investigate the firm-level determinants of accessing non-U.S. technological know-how. We find, for example, that employing or collaborating with researchers with research experience abroad seems to facilitate this access. Also, in the semiconductor industry, smaller and older firms (and in the pharmaceutical industry, younger firms) are more likely to make use of the output of non-U.S. R&D. The chapter is organized as follows. Sections 10.2 and 10.3 describe the sources for the data construction and the construction itself. Section 10.4 details some descriptive statistics of the data set. Section 10.5 describes our analysis on the influence of foreign R&D on U.S. innovation. Section 10.6 concludes. 10.2 Data Sources The data set we have created contains measures—patents and patent citations—of the R&D productivity of individual researchers between 1975 and 1998. For patents assigned to publicly traded firms in the U.S. pharmaceutical (Primary Standard Industrial Classification [SIC] code 2834) or semiconductor industry (Primary SIC code 3674), these data also contain information on the patents’ assignees (e.g., firm size and R&D expenditures). Budgetary and time constraints limited the number of industries that we could include in our analysis. The pharmaceutical and semiconductor industries were selected because they are especially prolific generators of innovations3 and their products are relatively homogeneous4 compared to those of other industries. 3. Based on NBER-Case Western University data of U.S. patents from 1963 to 1999, 15.3 percent of industry patents were granted to the firms in the pharmaceutical industry and 14.8 percent were granted to those in the semiconductor industry. 4. In a cross-sectional analysis involving multiple industries, differing technologies and patent propensities make interpretation of results difficult. By limiting analysis to the patents and their inventors in a specific industry, we resolve heterogeneity in the propensity to patent across industries, thus making comparisons of patents and citations more meaningful.

International Knowledge Flows Table 10.1

325

Variables from each data source

Data source

Variables

Patents BIB

Patent ID number, application year, inventors’ names, address, city, state, country, assignee ID, and assignee name Firm name, primary and other corresponding SIC codes, R&D expenditures, sales, number of employees, capital, and subsidiaries of the firms Firm name and ownership changes due from merger and acquisition, and obsolete securities due to bankruptcy or dissolution Founding year of firm Citing patent number and cited patent number

Compact D/SEC

S&P Thomas Register Citation

The data for this study come from five sources: (a) Patent Bibliographic data (Patents BIB) released by the U.S. Patent and Trademark Office (USPTO), which contains bibliographic information on all U.S. utility patents issued from 1969 to 2002; (b) the Compact Disclosure/Securities and Exchange Commission (D/SEC) database from 1989 to 1997, which contains firm information taken primarily from 10-K reports filed with the Securities and Exchange Commission; (c) the Standard & Poor’s Annual Guide to Stocks-Directory of Obsolete Securities, which includes a history of firm name changes, and of mergers and acquisitions; (d) the Thomas Register, Mergent, and Corptech data, which report a firm’s founding year, and finally (e) the National Bureau of Economic Research (NBER) PatentCitations data collected by Hall, Jaffe, and Trajtenberg (2001), which contain all citations made by patents granted from 1975 to 1999. These data sources are described in detail following and the variables used in our study from each data source are in table 10.1. 10.2.1 Patent Bibliographic Data (Patents BIB) Patents BIB is one of the Cassis Series of optical disc products released by the USPTO. Patents BIB contains bibliographic information for U.S. utility patents issued since January 1969. The information includes the patent ID number, dates of the patent’s application and granting, patent assignee, and geographic information on all inventors involved. The original optical disc we use covers patents issued between 1969 and 2002, and contains over 3 million U.S. patents granted. We use only the patents granted after January 1975 because detailed geographic information for all inventors is available in Patents BIB only for patents granted after that date. Most foreign innovating firms (especially those in Western Europe and in Japan) apply for patents in the United States in addition to their home countries so that U.S. patent data reflect nearly the universe of patented innovations. Over this period the USPTO granted 2,493,610 patents (U.S. Patent No. 3,858,241 through 6,351,850), which together list 5,105,754 inventors (an average of 2.05 inventors listed per patent).

326

Jinyoung Kim, Sangjoon John Lee, and Gerald Marschke

10.2.2 Compact D/SEC The Compact D/SEC contains about 12,000 firms that have at least $5 million in assets and at least 500 shareholders of one class of stock of U.S. companies traded on the American Stock Exchange, the NASDAQ, the New York Stock Exchange, or the Over-the-Counter equities market. The data set provides financial and other information obtained from annual reports, 10-K and 20-F filings, and proxy statements for those companies. Most of the companies included are American. Company records include directory information, primary and secondary SIC codes, brief business descriptions, names of subsidiaries, names of top executives, ownership data, financial data, and excerpts from annual reports and other SEC reports. 10.2.3 Standard & Poor’s Annual Guide to Stocks (S&P) The firm-level information from the Compact D/SEC data cannot be directly matched to assignees in the Patents BIB data because parent firms patent sometimes under their own names and other times under the names of their subsidiaries. Mergers and acquisitions at both the parent firm and subsidiary levels and name changes further complicate linking the patent to firm-level data. To track the ownership of firms over the entire period of our study, we use the information in the Standard & Poor’s Annual Guide to Stocks. The S&P data provide histories of firm ownership changes due to mergers and acquisitions, bankruptcy, dissolution, and name changes, updated through December 2002. 10.2.4 NBER Patent-Citations Patent applicants are legally obligated to disclose any knowledge they have of previous relevant inventions. Citations are of two kinds: to science (or prior science publications) and to technology (or previous patents). The patent examiner may add to the application relevant citations omitted by the applicant. Thus, through the patent citations each patent documents the prior art upon which the new innovation builds. Through the citations we can trace knowledge flow, measure the closeness of technological innovations, and measure an innovation’s impact. The data collected by Hall, Jaffe, and Trajtenberg (2001a, 2001b) contain all citations made and received by patents granted between 1975 and 1999. Their data contain a total of 16,522,438 citation records; the mean number of citations received by a patent is 5.07, ranging from a minimum of 1 and a maximum of 779, respectively. The number of patents granted to the firms identified in the pharmaceutical and semiconductor industries between 1975 and 1999 is 244,158. The mean citations received by a patent in these two industries is 8.13, ranging from a minimum of 1 and a maximum of 631.

International Knowledge Flows

327

10.3 Data Set Construction Process This section discusses key issues that arise in assembling our data set from these five sources. The assembly requires three steps. First, we create an inventor identifier in Patents BIB because of the nonuniqueness of inventors’ names. The primary challenge in this step is identifying who is who among inventors with same or similar names. Second, we identify each firm’s ownership structure of subsidiaries and their name changes over the data period to construct firm-level data, using the Compact D/ SEC and the S&P data. In the final step, we combine the inventor data and the firm data and then add the patent citation data where each citing patent that was granted between 1975 and 1999 is matched to all patents cited by the patent. 10.3.1 Identifying the Same Inventor among “Same/Similar” Names Over 5.1 million inventor names are contained in the U.S. patent data from January 1975 through February 2002. Each inventor name record includes the last name, first name, middle name, and suffix (Jr., Sr., etc.) of the inventor, as well as his or her city, state, and country of residence at the time of the granting of the patent. Identifying the same inventor in different records with same or similar names (for example, John Maynard Keynes, John M. Keynes, John Keynes, and John Keyens) is not an easy task. Our matching method uses as much information in the patent data as possible to increase the number of names matches without losing matching accuracy. Our name-matching methodology is similar to that in Trajtenberg, Shiff, and Melamed (2006). To start, we treat each entry that appears in the inventor name field of every patent in the Patents BIB data as a unique inventor. Given N number of names in this name pool, we pair each name with all other names, which generates N(N–1)/2 number of unique pairs. The 5.1 million names in the Patents BIB data (2.05 inventors per patent) thus produce 13 trillion unique pairs. For each pair, we consider the two names as belonging to the same inventor if the Soundex codes of their last names and their full first names are the same, and at least one of the following three conditions is met: (a) the full addresses for the pair of names are the same; (b) one name from the pair is an inventor of a patent that is cited by another patent whose inventors include the other name from the pair; or (c) the two names from the pair share the same coinventor. In implementing the second and third conditions, we make comparisons based on whether the first and last names are spelled identically. After our name matching procedure is completed, we go back and check that these conditions are still valid based on the inventor identifier constructed by the matching procedure. If not, we repeat the name-matching process to create a new inventor identifier. Soundex is a coded index for last names based on the way a last name

328

Jinyoung Kim, Sangjoon John Lee, and Gerald Marschke

sounds in English rather than the way it is spelled. Last names that sound the same, but are spelled differently, like Smith and Smyth, have the same Soundex code. We use the Soundex coding method to expand the list of similar last names to overcome the potential for misspellings and inconsistent foreign name translations to English; misspellings are common in the USPTO data as are names of non-Western European origin (see the appendix for the detailed Soundex coding method). We also consider a pair of names as a match if two have the same full last and first names as spelled in the Patent BIB data, and at least one of the following conditions is met: (a) the two have the same zip code; (b) they have the same full middle name; or (c) they reside in the same metropolitan statistical area (MSA). As an additional step beyond the aforementioned pairwise comparisons, we treat a pair of inventors as mismatched if the middle name initials of the pair are different. Table 10.2 illustrates our name-matching procedure. Inventors 001 and 002 in table 10.2 have the same last and first names, and share the same coinventor. Thus, the two records in this pair are treated as the same inventor. Inventors 002 and 003 do not have the same full middle name but share the same zip code, and thus the two inventors are treated as the same inventor. Although inventors 002 and 005 share the same zip code, the middle name initials are different. Therefore, the pair is not considered a match (they would not be considered a match by our algorithm even if their street addresses were identical, possibly a case of a parent and a child). Imposing Transitivity Transitivity is imposed in the following sense: if name A is matched to name B and name B is matched to name C, name A is then matched to name C. We iterate this process until all possible transitivity matches are completed. After the transitivity procedure, we assign the same inventor ID number for all the names matched. For instance, inventors 001 and 003 are not linked in the initial round of name matching, but they are matched through transitivity because inventors 001 and 002 are matched and inventors 002 and 003 are matched. Table 10.2

Examples of name matching

Initial ID

Inventor name

Coinventor

Middle name

ZIP

Final ID

001 002 003 004 005 006

Adam Smith Adam Smith Adam Smith Adam Smith Adam Smith Adam Smyth

John Keynes John Keynes — — John Keynes John Keynes

— Emmanuel E Emmanuel J —

20012 14228 14228 14214 14228 14228

001 001 001 001 005 001

International Knowledge Flows

329

Imposing transitivity, however, poses a possibility of name mismatch. Suppose, for example, Adam E. Smith and Adam Smith are matched in one pair, and Adam J. Smith and the same Adam Smith are matched in another pair. According to our transitivity procedure, Adam E. Smith and Adam J. Smith are identified as a match although their middle name initials are different. The number of matches through transitivity suffering from this problem appears to be trivial, however: we find 126 cases where two inventors are matched (although their middle names were different) out of 2.3 million uniquely identified inventors. Upon further investigation of these cases, we found the mismatches are of three kinds. In the first kind, some middle names in the Patents BIB data are incorrectly coded. For instance, our transitivity procedure matched the names “Laszlo Andra Szporny” and “Laszlo Eszter Szporny,” which appear to belong to the same inventor according to other information. We found that the middle names attributed to him are the first names of the next coinventors listed on his patents, suggesting that “Andra” and “Eszter” are not his middle names. In the second kind of mismatch, an inventor with two middle names is coded in the Patents BIB data with one middle name in some cases and with the other middle name in other cases. In the third kind, a mismatch occurs when two inventors with the same last and first name but different middle names appear in the same patent. We corrected by hand instances of the first two kinds of mismatch, but dropped from our data the observations displaying the third kind of mismatch. Trajtenberg, Shiff, and Melamed (2006) assign scores for each matching criterion and consider a pair matched only if its total score from all matching criteria exceeds a threshold. Because the choice of weights and the score threshold for a match is largely arbitrary, we do not use this scoring method in our data construction. Our method also differs in that we do not use as a matching criterion whether two inventors share the same assignee because name matching based on this criterion might bias our measure of mobility among inventors. Instead, we apply the rule that two inventors are not treated as a match if their middle name initials differ. From our experience with the patent data, imposing this rule is effective because the Soundex coding system sometimes so loosely specifies names that apparently different last names are considered a match. In the end, because of these differences, the number of distinct inventors identified with our procedure is a little higher than the number of distinct inventors reported in Trajtenberg, Shiff, and Melamed (2006). We identified 1.72 million unique inventors (34 percent) out of 5.1 million names in the entire patent data, while Trajtenberg, Shiff, and Melamed found 1.6 million distinctive inventors (37 percent) out of 4.3 million names. Note that our patent database is larger because it includes additional years, 2000 to 2002.

330

Jinyoung Kim, Sangjoon John Lee, and Gerald Marschke

Matching Accuracy We evaluated the accuracy of the algorithm using the curricula vitae (CV) of a sample of 100 inventors (which we will call our “benchmark sample”) obtained from the internet. While the benchmark inventors may not be representative of the inventors in our data in many ways—the benchmark inventors are more likely to have academic ties and are more prolific than the inventors in our data set5—they still tell us something useful about the errors produced by our algorithm. The algorithm may over or undermatch patents and inventors. An inventor is subject to undermatching error if the matching algorithm fails to group all of the inventor’s patents under one inventor identification number. The algorithm correctly groups 685 (89 percent) of the 769 patents that belong to the 100 inventors in the CV sample. In other words, a reassignment of as few as eighty-four patents would eliminate the undermatching error in the benchmark sample. Yet partly because the benchmark inventors are so prolific, we find that the algorithm has failed to match at least one patent for each of thirty-eight of the 100 benchmark inventors. The fewer patents an inventor has, the smaller the opportunity for the algorithm to misassign one of his patents,6 suggesting that the undermatching error problem in the actual data is smaller than the undermatching error rate in the benchmark sample, at least in terms of the fraction of inventors affected.7 Overmatching occurs when a patent is assigned to the wrong inventor. Overmatching error is found in eight of the 100 benchmark inventors. To these eight inventors, the algorithm should have assigned 158 patents, but instead assigned 308 patents. Two inventors whose last names are Johnson and Smith account for 138 of 150 overmatched patents, suggesting that overmatching error arises in our data for researchers with common names. How might the under- and overmatching errors affect our results? Error rates affect the accuracy of an estimate at a point in time, as when we estimate the fraction of patents in a given year that name an inventor with foreign research experience. On estimates of trends, however, matching errors may have qualitatively little effect if error rates are not changing over time. 5. The inventors in our CV sample have, on average, over seven patents each, compared to a little over one patent each among our matched data. 6. The average number of patents per inventor affected by undermatching error was 9.6 compared to 6.6 for those not affected by undermatching error. The average number of patents per inventor in our data set is less than two. 7. Undermatching error appears to be due primarily to the importance that the algorithm places on the inventor’s middle name for matching and the inconsistency with which an inventor’s middle name is represented from one patent to the next patent. We are currently testing an improved version of the matching algorithm that allows for more variation in the way the middle name appears. Preliminary testing suggests that the new algorithm significantly reduces undermatching error at little cost in increased overmatching error.

International Knowledge Flows

331

One concern that has been expressed to us is that inventor names on U.S. patents have become less varied. East Asian surnames, which are increasing in frequency among inventors on U.S. patents, tend to vary less than Western European names. Therefore, the overmatching error rate may be increasing. If names are getting more similar, we may be lumping under a single ID an inventor with foreign experience and an inventor without foreign experience. This will increase the frequency with which, for example, we find an inventor team with foreign experience and cause us to find a positive trend in foreign influence even if there is not. Our algorithm, however, did not overassign patents to any of the eleven benchmark inventors with surnames of Asian ancestry, possibly because there is sufficient variation in their first names, which are also used in the matching. 10.3.2 Identifying the Ownership Structure and Combining Patent-Inventor Data with Firm Data Because parent firms patent sometimes under their own names and at other times under the names of their subsidiaries, combining the Patents BIB data with firm-level data in the Compact D/SEC data is not straightforward. Mergers and acquisitions at both the parent firm and subsidiary levels—common in these two industries during the 1990s—and name changes complicate linking the patent to firm-level data. (The USPTO does not maintain a unique identifier for each patenting assignee at the parent firm level nor does it track assignee name changes.) Thus, to use the firm-level information available in the Compact D/SEC data, the names of parent firms and their subsidiaries and the ownership of firms must be tracked over the entire period of the study.8 To start, we identify mergers and acquisitions, and name changes of firms in the two industries, pharmaceutical preparation (Primary SIC code 2834) and semiconductor and related devices (3674), over the period between 1989 and 1997, using the Standard & Poor’s data. We also identify the ownership structure of subsidiaries of firms using subsidiaries information available from the Compact D/SEC from 1989 to 1997.9 We can then relate each assignee in the patent data to a firm in the Compact D/SEC data, which enables us to match each patent to a firm in the Compact D/SEC data. We then combine firms’ founding years, obtained from Thomas Register, Mergent, and Corptech, with the other firm-level information. 8. NBER-CWRU researchers created a database of parent firms and their subsidiaries for all the names among USPTO patent assignees. However, they only linked subsidiaries based on the corporate ownership structure as it existed in 1989. 9. The subsidiary list reported in the Compact D/SEC is not always complete. For example, some subsidiaries appear intermittently and some firms report subsidiaries every other year. Hence, if a firm is reported as a subsidiary of another firm once between the years 1989 to 1997, we consider it a subsidiary of that firm for the entire period.

332

Jinyoung Kim, Sangjoon John Lee, and Gerald Marschke

As the final step, we add information on all citations from the NBER Patent-Citations data collected by Hall, Jaffe, and Trajtenberg (2001), where each citing patent that was granted between 1975 and 1999 is matched to all patents cited by the patent. 10.4 Descriptive Statistics Figure 10.1 describes the distribution of U.S. patents granted by year of application. The figure shows the surge in patenting that began in the mid1980s. The applicant flow for the previous twenty-five years had been remarkably stable. The possible causes of this patent surge have been discussed in Hall and Ziedonis (2001), Kortum and Lerner (1999, 2003), and Kim and Marschke (2004). Because it covers the surge, the mid-1980s through the late 1990s is an interesting period to examine and is the period we study here.10 Figure 10.1 shows that the annual number of patents granted dips sharply after 1997. This dip reflects a lag between the application and granting dates. About 70 to 80 percent of all patent applications ultimately granted are granted within the first three years of the application and 97 percent of all patent applications are granted within the first four years of the application date (Hall, Griliches, and Hausman 1986). For this reason, the last year covered by our analysis is 1997. Between January 1975 and February 2002, 45.5 percent were granted to U.S. assignees and 37.4 percent were granted to foreign assignees (see table 10.3) with the rest unassigned. In figure 10.2, we report the number of patents granted to firms in each of our two industries. Note that in both industries the number of patents granted annually rose over the period we study: the annual number of patents granted between 1989 and 1998 rose from about 1,000 patents annually, but by a factor of two in the pharmaceutical industry and nearly seven in the semiconductor industry. Table 10.4 shows that the number of inventors named as an inventor to at least one patent assigned to a firm in one of our two industries is 59,292 out of the 2,299,579 unique inventors in our data (25,609 inventors in the pharmaceutical and 33,683 in the semiconductor industry). Inventors working in the pharmaceutical and semiconductor industries are named as inventors on more patents on average than inventors in other industries (see table 10.4). An inventor in a pharmaceutical firm is named as an inventor on average on 2.80 patents over our sample period, whereas an inventor in the semiconductor industry appears on average on 2.60 patents. We identified pharmaceutical and semiconductor firms in the Compact 10. The application rate has since appeared to level off, somewhat. The number of patents granted in 2004, 2005, and 2006 were approximately 165,000, 144,000, and 174,000, respectively.

International Knowledge Flows

333

D/SEC data by their primary SIC. We identified 447 parent firms and 5,331 subsidiary firms in the pharmaceutical industry and 332 parent firms and 4,211 subsidiary firms in the semiconductor industry. Firm information starts in 1989 because we had access to the Compact D/SEC data only beginning in 1989. We dropped all patent applications filed after 1997 because, as discussed previously, starting with application year 1998 the patent time series tailed off due to the review lag at the USPTO. Some sample statistics from the firms in the two industries in our data— the number of selected firms and the number of employees, sales, and R&D

Fig. 10.1

Table 10.3 Assignee U.S. U.S. U.S. Foreign Foreign Foreign Others Others Total

Number of patents granted by year of application (1975–2001)

Number of patents by assignee type (January 1975–February 2002) Description Assigned to U.S. organization and state/local governments Assigned to a U.S. resident (individual) Assigned to a U.S. Federal Government organization

# Observations

Percentage

1,090,194

43.7

15,849 30,431

0.6 1.2

Assigned to a non-U.S., nongovernment organization Assigned to a non-U.S. resident (individual) Assigned to a non-U.S. government organization (all levels)

914,826

36.7

7,873 8,613

0.3 0.4

Unassigned Missing observations

412,621 13,203

16.6 0.5

2,493,610

45.5

37.4

17.1

334

Jinyoung Kim, Sangjoon John Lee, and Gerald Marschke

Fig. 10.2 Table 10.4

Number of patents granted by year of application in two industries Patent statistics for all inventors (January 1975–February 2002)

Inventors No. of patents per inventor

Total

Pharmaceutical

Semiconductor

2,299,579 2.22

25,609 2.80

33,683 2.60

expenditures—are reported in table 10.5. For the year 1997, for example, the data show 221 firms in the pharmaceutical and 151 firms in the semiconductor industry, with 177 firms and 135 firms, respectively, reporting positive R&D expenditures. Pharmaceutical firms are larger in terms of number of employees, sales volume, and R&D expenditures. 10.5 International Knowledge Flows A small literature uses patents to examine international knowledge flows. Jaffe and Trajtenberg (1999) find, for example, that patents are more likely to cite other patents from the same country and that citations to other countries’ patents occur after a lag, suggesting that international borders impede or slow knowledge diffusion. Singh (2007) uses patent citations to examine knowledge flows between foreign-based subsidiaries of multinational corporations (MNCs) and their host countries. He finds that host country patents cite the patents of local foreign MNC subsidiaries at high rates. He also finds that MNC patents cite host country patents at an even higher rate, especially when the host country is technologically

No. of firms

88 88 146 151 161 179 184 193 221 209

71 67 87 93 107 114 131 136 151 154

Year

1989 1990 1991 1992 1993 1994 1995 1996 1997 1998

1989 1990 1991 1992 1993 1994 1995 1996 1997 1998

Table 10.5

70 65 86 92 107 112 124 123 141 125

85 85 137 145 155 170 171 170 196 180

Employment

71 67 84 91 103 108 127 131 147 153

78 81 124 123 132 149 150 158 193 186

Sales

No. of firms reporting

55 52 70 79 95 100 115 122 135 139

69 64 98 109 126 136 142 152 177 170

R&D

25,530 30,203 28,720 28,546 28,395 33,344 43,822 62,304 87,658 102,042

Semiconductor industry 3,108 275,944 3,273 309,869 3,492 410,065 3,244 423,890 2,919 477,073 2,057 590,914 2,050 720,290 3,201 746,105 3,277 1,081,964 3,328 1,095,652

R&D

85,151 78,612 95,712 101,187 105,501 104,193 124,575 126,510 194,237 210,768

Sales

Pharmaceutical industry 5,903 895,924 5,722 794,134 4,741 884,836 4,694 987,210 4,297 1,609,557 4,668 1,670,350 4,460 1,924,112 4,114 1,996,128 4,078 2,161,856 4,391 2,341,263

Employment

Mean

Summary statistics from the pharmaceutical and semiconductor industry samples (Units in sales and R&D: $1,000)

10,043 10,102 13,365 13,226 13,307 6,922 7,156 14,466 14,750 14,013

13,058 13,726 12,690 12,374 11,764 13,395 13,263 12,652 13,980 14,307

Employment

885,337 959,953 1,521,497 1,698,277 2,044,170 2,588,848 3,103,003 3,243,984 4,956,544 5,255,021

1,940,353 2,036,458 2,187,517 2,404,976 7,440,985 7,693,383 7,897,844 8,085,530 9,156,677 10,268,340

Sales

Standard deviation

64,583 82,405 81,478 94,122 105,515 119,556 151,508 215,290 336,703 392,838

180,615 188,652 217,398 237,929 250,355 255,514 323,828 355,602 797,319 830,591

R&D

336

Jinyoung Kim, Sangjoon John Lee, and Gerald Marschke

advanced. He interprets this to mean that significant amounts of knowledge flow from MNCs to the host country, and even greater quantities flow from the host country to the MNC. He also finds that inventor flows between MNCs and host countries correlate with the direction of the flow of research personnel.11 We focus on knowledge flows to the United States from abroad and are interested in how they have changed in recent years, especially over the period of the patent surge that began in the mid-1980s. We use the geographic mobility of researchers and their location at the time of invention to track the transmission of foreign knowledge from other countries to the United States. In addition, we test if the international migration of researchers facilitates knowledge transfers across borders. Table 10.6 shows the annual number of unique inventors named on U.S. domestic patents for the years 1985 through 1997. It also shows the percentages of inventors who at the time of the patent application (a) resided in a foreign country, (b) resided in the United States and had been previously listed as a foreign-residing inventor on a successful patent application, and (c) resided in the United States but had never been previously listed as a foreign residing inventor on a successful patent application. Because our data included patents granted in 1975 and later, we imposed a cutoff for the patents used to define whether an inventor has foreignexperience at the time of the patent’s application. We consider as “foreignexperienced” only those inventors who are currently foreign residents or had been foreign residents sometime in the ten-year period prior to the date of the patent’s application, because ten years still leaves us a long period over which to conduct our analysis and because knowledge acquired in a foreign country far in the past may not be very valuable. Table 10.6 shows a dramatic increase in the number of unique inventors on U.S. domestic patents between 1985 and 1997, from 42,368 to 119,556, which translates to an average annual growth rate of 9 percent. This increase in this period is expected given the timing of the patent surge. Among those inventors with foreign experience, the percentage of inventors with current foreign addresses increased steadily during the period from 8.15 percent to 9.11 percent while the percentage of U.S.-residing inventors with foreign experience increased from 0.99 percent in 1985 to 1.30 percent in 1992, then dropped to 1.01 percent in 1997. Overall, the percentage of inventors with foreign experience increased (from 9.14 percent in 1985 to 10.13 percent in 1997). Table 10.6 shows that the growth in the number of inventors in the pharmaceutical (13 percent annually) and semiconductor (31 percent annually) 11. Singh also uses the USPTO patent data to produce an inventor panel. His namematching strategy differs in several ways from ours (see Singh). Singh does not report on the accuracy of his match.

42,368 44,828 48,810 54,947 59,164 63,812 67,657 73,640 80,428 90,910 104,775 104,829 119,556

1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997

— — — — 2,143 2,259 3,332 3,876 4,505 5,320 6,629 4,894 6,093

Pharma

— — — — 1,139 1,362 2,791 3,370 4,190 5,739 7,450 7,916 9,993

Semi 8.15 8.30 8.21 8.49 8.60 8.02 7.76 7.86 8.06 8.44 8.78 9.19 9.11

All (%) — — — — 14.47 17.35 19.09 20.38 25.88 26.86 28.87 31.55 29.71

Pharma — — — — 9.04 7.78 6.02 7.15 7.06 14.76 15.18 13.26 15.31

Semi 0.99 1.07 1.13 1.13 1.17 1.22 1.26 1.30 1.21 1.20 1.13 1.07 1.01

All (%) — — — — 2.01 1.51 1.23 1.21 1.31 0.98 0.87 0.90 0.75

Pharma — — — — 1.14 1.25 1.22 1.13 1.03 0.94 0.86 0.78 0.80

Semi

Current U.S. residents w/ foreign experience

90.86 90.63 90.66 90.37 90.23 90.76 90.98 90.85 90.73 90.36 90.08 89.75 89.87

All (%)

— — — — 83.53 81.14 79.68 78.41 72.81 72.16 70.25 67.55 69.54

Pharma

— — — — 89.82 90.97 92.76 91.72 91.91 84.30 83.96 85.95 83.89

Semi

Current U.S. residents w/o foreign experience

Notes: Columns (2)–(4) show the number of unique inventors in all U.S. domestic patents, in pharmaceutical patents, and in semiconductor patents, respectively. In columns (8)–(10), we report the percent of inventors with current addresses in the U.S. who have at least one patent in the past ten years while residing at a foreign address.

All

Current foreign residents

Percentage of inventors by foreign-experience type (%)

Inventors on U.S. domestic patents with foreign experience

Number of inventors

Year

Table 10.6

338

Jinyoung Kim, Sangjoon John Lee, and Gerald Marschke

industries has been significantly faster than for all industries combined. In the pharmaceutical industry, the share of inventors with foreign experience grew rapidly although the increase is mostly in the share of inventors with current foreign addresses and there is a decrease in the fraction of U.S.residing inventors with past foreign experience. This finding is not surprising given the increasing rate at which U.S. pharmaceutical firms have been citing new laboratories abroad (Chacar and Lieberman 2003) and findings that collaborations among academic scientists have become more dispersed, possibly due to improvements in telecommunications (Adams et al. 2004). The semiconductor industry shows a similar pattern, but the changes are less pronounced than in pharmaceutical industry.12 Figure 10.3, panel A shows the average annual patent productivity of inventors in U.S. domestic patents by foreign-experience type for all patents. Panels B and C of figure 10.3 repeat the analysis of panel A, but for the pharmaceutical and semiconductor industries alone. The calculations in these figures are based on inventors named to at least one patent in each year and to at least one additional patent during the previous ten-year period. The latter restriction is imposed because inventors with foreign experience have at least one patent earlier by design. We first note in these figures that patent productivity for all three types has been increasing. However, the growth rate of patent productivity is the highest for the U.S.residing inventors with past foreign experience and those inventors in later years have significantly higher patent-inventor ratio than other types of inventors. On the other hand, the growth in patent productivity among current foreign residents has been the slowest. There are a number of possible explanations for this. First, inventors with higher productivity are more likely to migrate to the United States, especially in recent years, because of better compensation for skilled labor in the U.S. labor market or because of U.S. immigration policies. Second, as shown in table 10.6, the share of current foreign residents has been rising while that of U.S. residents with foreign experience has been falling, especially in our two industries. These changes may be associated with a more selective migration of researchers with higher productivity. Third, foreign experience somehow improves the productivity of researchers (proportionally more in recent years), or inventors with foreign experience happen to be working in technological areas with higher patent propensities. Figure 10.3, panels B and C, show qualitatively similar changes in productivity among the different types of inventors. Note that because we are looking within an industry with rather homogenous technology, these gaps are less likely due to heterogeneity in technology class. Panel C of figure 12. Phene and Almeida (2003) note a dramatic rise in overseas patenting by the five leading U.S. semiconductor companies between 1986 and 1995.

International Knowledge Flows

Fig. 10.3

339

Patent-inventor ratio by foreign-experience type

10.3 more clearly demonstrates the changes in patent productivity across the types of inventors in the semiconductor industry. Where figure 10.3, panels A through C track the productivity of inventors by their patent output, figure 10.4, panels A through C track how the quality of inventors’ output changes by inventor type. There is evidence that citations received reflect the economic value of the patent (Trajtenberg 1990). Figure 10.4, panel A shows the citations received in the five-year period following application per patent by inventor type for all industries over time. This figure covers only years of application through 1992 be-

340

Jinyoung Kim, Sangjoon John Lee, and Gerald Marschke

Fig. 10.4

Citations per patent by foreign-experience type

cause the NBER citation data contain citations made by patents granted in years up to 1999 and we take into account the five-year period of citation and a two-year gap between application and granting dates. Between 1985 and 1992, the citations per patent rose for all three classes of inventors. Throughout the 1985 to 1992 period, the average citations per patent produced was the highest and grew fastest for U.S. residents with foreign experience and was the lowest and grew slowest for foreign residing inventors. In 1992, the number of citations attracted by the average patent of a U.S.-residing inventor with foreign-patenting experience, U.S.-residing inventor without foreign patenting experience, and foreign-residing inven-

International Knowledge Flows

341

tor, was about 6.5, 5, and 3.5, respectively. Thus, taken together, Figures 10.3, panel A, and 10.4, panel A, show that U.S.-residing inventors with foreign experience produce more patents on average and patents of higher quality than the other two classes of inventors between 1985 and 1992. Figure 10.4, panels B and C, conduct the analysis separately for the pharmaceutical and semiconductor industries. The semiconductor industry shows the same ordering of inventor types, though the levels are higher for each type. The pharmaceutical industry, however, shows no clear and consistent distinction between the two classes of U.S.-residing inventors. Figure 10.4, panel C, does show that foreign-residing inventors produce the lowest quality patents, as measured by citations, and of U.S.-residing inventors, those with foreign patenting experience produce more valuable patents. Thus it appears that in both industries in the late 1980s and into the 1990s, U.S.-residing inventors with foreign experience have higher patent rates. In the semiconductor industry, their patents also attract more citations. Is a patent from a domestic firm more likely to cite foreign-assignee patents when its inventors have foreign experience? We are interested in learning if knowledge spillovers from foreign countries are facilitated by direct exposure to inventors with foreign experience. Table 10.7 presents the results of our estimation of the determinants of accessing foreign Table 10.7

Determinants of citations to foreign-assigned patents Dependent variable  CITE_FRGN Pharmaceutical

FRGN_EXP Log INVENTOR Log EMPLOYEE Log R&D/INV Log NSIC Log MEXP Log FIRMAGE Observations R2

Semiconductor

(1)

(2)

(3)

(1)

(2)

(3)

0.2295 8.54 — — — — — — — — — — — —

0.2225 7.48 0.0366 0.97 0.0200 0.92 0.0019 0.56 0.0061 0.19 0.3289 5.67 0.0505 2.54

0.2299 5.77 0.0415 1.08 0.0172 0.76 0.0015 0.43 0.0017 0.05 0.3260 5.57 0.0516 2.54

0.2615 4.91 — — — — — — — — — — — —

0.2514 4.43 0.0566 2.75 0.0167 0.87 0.0028 0.95 0.0490 1.46 0.3212 2.73 0.1151 2.95

0.2899 4.70 0.0546 2.66 0.0139 0.74 0.0029 0.98 0.0487 1.44 0.3168 2.61 0.1160 2.91

1,430 0.0325

1,247 0.1794

1,215 0.1772

4,316 0.0237

4,186 0.1157

4,112 0.1165

Notes: Rows show the estimated coefficient and the t statistic for each regressor. The result for a constant term is suppressed. Column (3) shows the results from a regression that omits patents for which an inventor is listed as an inventor on a cited patent. The t statistic is based on the Huber-White sandwich estimator of variance.

342

Jinyoung Kim, Sangjoon John Lee, and Gerald Marschke

knowledge in the pharmaceutical and semiconductor industry. The unit of observation in the regression is a patent applied for in year 1997. The dependent variable is the fraction of citations to the patent that are assigned to foreign assignees (CITE_FRGN). The means and standard deviations of the independent and dependent variables, along with their definitions, are described in table 10A.1.13 The key regressor in these regressions is a binary variable that takes 1 if at least one inventor on the patent is currently residing or formerly resided in one of the foreign countries where foreign assignees of cited patents are located (FRGN_EXP). Note that this regressor reflects not just whether an inventor has foreign experience but which country the inventor has experience from. We speculate that knowledge spillover is country-specific. The regressions in table 10.7 also include as right-hand side variables firm-level characteristics in year 1997. A measure of the size of the research operation, proxied by the number of unique inventors named to patents awarded to the firm in 1997 (INVENTOR), is included to examine whether large-scale R&D enterprises are more likely to rely on foreign knowledge. We use the number of employees (EMPLOYEE) as an alternative measure of organizational size at the firm level. Included are the R&D-inventor ratio (R&D/INV) and the number of business lines in the firm (NSIC), measured by the number of secondary SIC’s identified with the firm. We include the R&D-inventor ratio (R&D/INV) as a regressor because a highly capitalized firm may rely on more advanced technology and thus may be more open to foreign technology. We include NSIC as a regressor to estimate the impact of economies of scope in the firm’s use of foreign knowledge. Our regressions also include the median experience of all inventors in the firm (MEXP) and years elapsed since the founding year of the firm (FIRMAGE). Column (1) in table 10.7 for each industry panel shows the estimated relationship between the fraction of a patent’s citations to foreign-assigned patents and the existence of foreign-experienced inventors using ordinary least squares. Column (2) for each industry panel reports the estimates of the determinants of the citation to foreign patents. One concern for our regression is that inventors are more likely to cite their own past patents than other inventors’ patents, which may drive the estimated relationship between our dependent variable and the key regressor, FRGN_EXP. In column (3) for each industry panel we thus exclude patents that have the same inventors as those in their cited patents. The results in table 10.7 show that a patent by inventors with foreign experience in both industries is more likely to cite patents assigned to foreign firms from the same country where the inventors are residing or resided in 13. Note that the means of the variables reported in table 10.7 are not the averages across firms because our regressions are at the patent level, not at the firm level. For instance, the mean value of INVENTOR and EMPLOYEE is greater than the firm mean in 1997 because larger firms tend to have more patents.

International Knowledge Flows

343

the past: FRGN_EXP has a significantly positive effect in all models. This effect is still significant with the data without self-citing patents. The coefficient estimates suggest that having an inventor on the patent with patenting experience in a particular (non-U.S.) country increases the fraction of citations to that country’s patents by between .22 and .29. The results show a negative effect of the size of the R&D enterprise on the fraction of citations to foreign patents in the semiconductor industry. There is no significant effect of the size of the R&D enterprise in the pharmaceutical industry. On the other hand, the coefficient estimate on the firm size variable (EMPLOYEE) is insignificant in all models. The coefficient estimate on log R&D/INV is generally positive but insignificant in all regressions. The coefficient estimate on log NSIC is never significant by conventional criteria of significance. The coefficient estimate on log MEXP is negative and significant for both industries. This may partly reflect that it is more costly for older inventors to learn new technologies from abroad, or it may be due to a vintage or a composition effect (e.g., areas of technology that experienced innovators innovate in are somehow more domestic). The coefficient estimate on log FIRMAGE is significant for both industries but has different signs for the two industries. The effect is negative in pharmaceutical industry while it is positive in semiconductor industry. That is, we find that in the semiconductor industry older firms, and in the pharmaceutical industry, younger firms, are more likely to make use of the output of non-U.S. R&D. 10.6 Conclusion We describe the construction of a panel data set that links inventors to the U.S. pharmaceutical and semiconductor firms for whom they work. These data contain measures of inventors’ R&D productivity—patents and patent citations—as well as information on the firms to which their patents are assigned. In this chapter we use these data to examine the role of research personnel as a pathway for the diffusion of ideas from foreign to U.S. innovators. We envision that local knowledge abroad that is tacit can be accessed or imported in two ways. The first is for U.S. firms to move closer to the sources of local foreign knowledge, possibly by setting up subsidiaries abroad (see Phene and Almeida 2003) and by hiring local scientists, or by sending firms’ U.S. scientists abroad to the subsidiaries. In our analysis this is captured by the number of inventors who are foreignresiding at the time of invention. The second way is for firms to hire scientists who had previously worked in laboratories abroad; that is, for scientists with a foreign background to move to U.S. firms on U.S. soil. This is captured by the number of inventors who are residing in the United States but have foreign experience. Table 10.6 suggests that U.S. domestic firms are relying more on the first way than the second way of accessing foreign sources of knowledge. In any

344

Jinyoung Kim, Sangjoon John Lee, and Gerald Marschke

year, the fraction of inventors who are residing in the United States and have foreign patenting experience is always less than .02. The fraction of inventors on U.S. patents assigned to U.S. firms who are abroad at the time of invention rises from about .08 in 1987 to about .09 in 1997. In the pharmaceutical and semiconductor industries, however, foreign-residing inventors are used more extensively and the growth has been more dramatic. Between 1989 and 1997, the fraction of inventors who were foreign-residing at the time of invention rose from .15 to .30 and from .09 to .15 for the pharmaceutical and semiconductor industries, respectively. This may be consistent with the argument by Phene and Almeida (2003) that the ability of local subsidiaries of multinational enterprises to tap the stock of knowledge in their host countries increases as part of a maturation process. While U.S. innovating firms’ employment of migrant workers with foreign research experience has fallen in relative terms, compared to foreignresiding and U.S. domestic researchers without foreign experience, these migrant workers are highly productive. Moreover, their productivity has increased over the period that we study, possibly because of either accelerating knowledge spillovers or more selective migration of high-productivity inventors. In table 10.7, we present evidence that either employing researchers abroad or foreign-experienced researchers in the United States contributes to the import of foreign knowledge. The citation of the patent of a foreign assignee by a U.S. firm’s patent represents either an assimilation of foreign knowledge by the parent firm in the U.S. or the assimilation of the foreign knowledge by foreign subsidiaries of the U.S. firm. Evidence from the international business literature suggests that the multinational corporation’s home base learns from its foreign-based subsidiaries (Singh 2007; Kogut and Zander 1993; Dunning 1992). Thus, we interpret the evidence presented in table 10.7 as evidence that when U.S. innovating firms employ either foreign-experienced researchers in the United States or those at a foreign-based subsidiary, transmission of foreign knowledge from their foreign origins to the United States results. Table 10.6 and 10.7 together, we believe, suggest that this transmission has been increasing in the U.S. pharmaceutical and semiconductor industries from the late 1980s through the late 1990s. Our findings are consistent with reports that during this period the U.S. pharmaceutical industry has been increasing the pace at which it is establishing laboratories on foreign soil. Our findings are also consistent with arguments that foreign subsidiaries of U.S. semiconductor firms are becoming more adept at extracting foreign-based knowledge. We anticipate this data set will be useful in addressing other important questions. These data will allow us to investigate the consequences of the mobility of R&D personnel on firm R&D. What is the impact, for example, of the arrival of a researcher with a particular set of R&D experiences on the character and quantity R&D done by a firm? We will be able to address this question because we know each researcher’s patenting history, both in

International Knowledge Flows

345

terms of quantity, and we also know the kinds of technologies underlying the innovations. This data set will allow us to directly observe the importance of interfirm mobility for technological diffusion. From the perspective of the researcher, this data set will allow us to examine the determinants of interfirm mobility. The panel nature of these data will allow us to investigate the productivity profiles of researchers working in industry over their careers. Because we observe all the inventors responsible for a patent, we will be able to use this data set to investigate how firms organize the R&D enterprise, the extent of collaboration among researchers who are geographically dispersed, and the extent of interaction among researchers with different backgrounds.

Appendix The Soundex Coding System The Soundex is a coded index for last names based on the way a last name sounds rather than the way it is spelled. Last names that sound the same, but are spelled differently, such as Smith and Smyth, have the same Soundex code. We use the Soundex coding method to expand the list of similar last names to overcome the potential for misspellings and inconsistent foreign name translations into English; misspellings are common in the USPTO data, as are names of non-Western European origin. A Soundex code for a last name takes an upper case initial followed by 6-digit numeric codes. For example, the Soundex code for Keynes is K520000. The rules for generating a Soundex code are14: 1. Take the first letter of the last name and capitalize it. 2. Go through each of the following letters, giving them numerical values from 1 to 6 if they are found in the Scoring Letter table (1 for B, F, P, V; 2 for C, G, J, K, Q, S, X, Z; 3 for D, T; 4 for L; 5 for M, N; 6 for R; 0 for Vowels, punctuation, H, W, Y). 3. Ignore any letter if it is not a scoring character. This means that all vowels as well as the letters h, y, and w are ignored. 4. If the value of a scoring character is the same as the previous letter, ignore it. Thus, if two “t”s come together in the middle of a name they are treated as a single “t” or a single “d”. If they are separated by another nonscoring character then the same score can follow in the final code. The name Pettit is coded as P330000. The second “t” is ignored but the third one is not, since a nonscoring “i” intervenes. 14. The strings of “–, ., , /, (,), %, ?, #, &, “, _” in all name fields have been translated to blank space in advance and then last names are Soundex coded.

346

Jinyoung Kim, Sangjoon John Lee, and Gerald Marschke

Table 10A.1

Variable definitions and sample statistics Mean (standard deviation) Definition

CITE_FRGN FRGN_EXP

INVENTOR EMPLOYEE R&D/INV

NSIC MEXP

FIRMAGE

Fraction of citations to patents that are assigned to foreign assignees  1 if at least one inventor is residing or has resided in the past in one of the foreign countries where foreign assignees of cited patents are located Number of all inventors in the patenting firm Number of employees in the patenting firm Real R&D expenditures in 1996 constant dollars divided by the number of inventors in the patenting firm (thousands of dollars per inventor) Number of secondary SICs assigned to the patenting firm Median experience of all inventors in the patenting firm where experience is measured as the number of years elapsed after the application year of an inventor’s first patent Years elapsed since the founding year of the patenting firm

Pharmaceutical

Semiconductor

0.5505 (0.3319) 0.0734 (0.2609)

0.4760 (0.2850) 0.0290 (0.1677)

326.0 (195.7) 35,979 (21,833) 31.67 (24.51)

923.5 (728.6) 41,538 (52,501) 12.04 (27.34)

3.791 (1.991) 5.292 (1.582)

3.154 (1.944) 3.832 (1.067)

77.40 (51.51)

36.17 (23.40)

5. Add the number onto the end of the Soundex code if it is not to be ignored. 6. Keep working through the name until you have created a code of 6 characters maximum. 7. If you come to the end of the name before you reach 6 characters, pad out the end of the code with zeros. 8. You may choose to ignore a possessive prefix such as “Von” or “Des.” See National Archives and Records Administration (1995) for the detailed method.

References Adams, J. D., G. C. Black, J. R. Clemmons, and P. E. Stephan. 2004. Scientific teams and institution collaborations: Evidence from U.S. universities, 1981– 1999. NBER Working Paper no. 10640. Cambridge, MA: National Bureau of Economic Research, July.

International Knowledge Flows

347

Agrawal, A. K., I. M. Cockburn, and J. McHale. 2003. Gone but not forgotten: Labor flows, knowledge spillovers, and enduring social capital. NBER Working Paper no. 9950. Cambridge, MA: National Bureau of Economic Research, September. Almeida, P., and B. Kogut. 1999. Localization of knowledge and the mobility of engineers in regional networks. Management Science 45 (7): 905–17. Audretsch, D. B., and M. P. Feldman. 1996. R&D spillovers and the geography of innovation and production. The American Economic Review 86 (3): 630–40. Branstetter, L. G. 2001. Are knowledge spillovers international or intranational in scope? Journal of International Economics 53 (1): 53–79. Chacar, A. S., and M. B. Lieberman. 2003. Organizing for technological innovation in the U.S. pharmaceutical industry. Geography and Strategy: Advances in Strategic Management 20 299–322. Cohen, W. M., R. R. Nelson, and J. P. Walsh. 2002. Links and impacts: The influence of public research on industrial R&D. Management Science 48 (1): 1–23. Duguet, E., and M. MacGarvie. 2005. How well do patent citations measure flows of technology? Evidence from French innovation surveys. Economics of Innovation and New Technology 14 (5): 375–93. Dunning, J. H. 1992. Multinational enterprises and the global economy. Wokingham, England: Addison-Wesley. Feldman, M. P. 1994. The geography of innovation. Boston: Kluwer Academic Publishers. Freeman, R. B. 2005. Does globalization of the scientific/engineering workforce threaten U.S. economic leadership? NBER Working Paper no. 11457. Cambridge, MA: National Bureau of Economic Research, June. Griliches, Z. 1992. The search for R&D spillovers. Scandinavian Journal of Economics 94 (Supplement): 29–47. Hall, B., Z. Griliches, and J. Hausman. 1986. Patents and R&D: Is there a lag? International Economic Review 27 (2): 265–83. Hall, B., A. Jaffe, and M. Trajtenberg. 2001a. Market value and patent citations: A first look. NBER Working Paper no. 7741. Cambridge, MA: National Bureau of Economic Research, June. ———. 2001b. The NBER patent citation data file: Lessons, insights and methodological tools. NBER Working Paper no. 8498. Cambridge, MA: National Bureau of Economic Research, October. Hall, B., and R. Ziedonis. 2001. The determinants of patenting in the U.S. semiconductor industry, 1980–94. RAND Journal of Economics 32 (1): 101–28. Hu, A. G. Z., and A. B. Jaffe. 2003. Patent citations and international knowledge flow: The cases of Korea and Taiwan. International Journal of Industrial Organization 21 (6): 849–80. Jaffe, A. B. 1989. Real effects of academic research. The American Economic Review 79 (5): 957–70. Jaffe, A. B., M. S. Fogarty, and B. A. Banks. 1998. Evidence from patents and patent citations on the impact of NASA and other federal labs on commercial innovation. Journal of Industrial Economics 46 (2): 183–205. Jaffe, A. B., and M. Tratjenberg. 1999. International knowledge flows: Evidence from patent citations. Economics of Innovation and New Technology 8:105–36. Jaffe, A., M. Trajtenberg, and R. Henderson. 1993. Geographic localization of knowledge spillovers as evidenced by patent citations. Quarterly Journal of Economics 108 (3): 577–98. Keller, W. 2002. Geographical localization of international technology diffusion. American Economic Review 92 (1): 120–42. Kim, J., and G. Marschke. 2004. Accounting for the recent surge in U.S. patenting:

348

Jinyoung Kim, Sangjoon John Lee, and Gerald Marschke

Changes in R&D expenditures, patent yields, and the high tech sector. Economics of Innovation and New Technology 13 (6): 543–58. ———. 2005. Labor mobility of scientists, technological diffusion, and the firm’s patenting decision. The RAND Journal of Economics 36 (2): 298–317. Kogut, B., and U. Zander. 1993. Knowledge of the firm and the evolutionary theory of the multinational corporation. Journal of International Business Studies 24 (4): 625–45. Kortum, S., and J. Lerner. 1999. What is behind the recent surge in patenting? Research Policy 28 (1): 1–22. ———. 2003. Unraveling the patent paradox. Paper presented at the American Economic Association Annual Meeting. Washington, D.C. Moen, J. 2005. Is mobility of technical personnel a source of R&D spillovers? Journal of Labor Economics 23 (1): 81–114. Mowery, D.C., and A. A. Ziedonis. 2001. The geographic reach of market and nonmarket channels of technology transfer: Comparing citations and licenses of university patents. NBER Working Paper no. 8568. Cambridge, MA: National Bureau of Economic Research, October. National Archives and Records Administration. 1995. Using the Census Soundex. General information leaflet 55. College Park, MD: National Archives and Records Administration. Phene, A., and P. Almeida. 2003. How do firms evolve? The patterns of technological evolution of semiconductor subsidiaries. International Business Review 12 (3): 349–67. Polanyi, M. 1958. Personal knowledge: Towards a post-critical philosophy. Chicago: University of Chicago Press. Polanyi, M. 1966. The tacit dimension. Garden City, NY: Doubleday. Regets, M. C. 2007. Research issues in the international migration of highly skilled workers: A perspective with data from the United States. Division of Science Resources Statistics, National Science Foundation Working Paper, SRS 07-203. Singh, J. 2007. Asymmetry of knowledge spillovers between MNCs and host country firms. Journal of International Business Studies 38 (5): 764–86. Thompson, P., and M. Fox-Kean. 2005. Patent citations and the geography of knowledge spillovers: A reassessment. American Economic Review 95 (1): 450–60. Trajtenberg, M. 1990. A penny for your quotes: Patent citations and the value of innovations. RAND Journal of Economics 21 (1): 172–87. Trajtenberg, M., G. Shiff, and R. Melamed. 2006. The names game: Harnessing inventors’ patent data for economic research. NBER Working Paper no. 12479. Cambridge, MA: National Bureau of Economic Research, August. Von Hippel, E. 1988. The sources of innovation. New York: Oxford University Press. Zucker, L. G., M. R. Darby, and J. S. Armstrong. 2001. Commercializing knowledge: University science, knowledge capture, and firm performance in biotechnology. NBER Working Paper no. 8499. Cambridge, MA: National Bureau of Economic Research, October. Zucker, L. G., M. R. Darby, and M. B. Brewer. 1998. Intellectual capital and the birth of U.S. biotechnology enterprises. American Economic Review 88 (1): 290–306.

11 The Growing Allocative Inefficiency of the U.S. Higher Education Sector James D. Adams and J. Roger Clemmons

11.1 Introduction This chapter presents new evidence on the productivity of U.S. universities. Our interest in this subject originates with recent developments in U.S. higher education that strike us as noteworthy and perhaps troubling. First, despite their high state, growth of employment and output in top U.S. research universities has slowed down in recent years.1 And second, growth of university research has not kept pace with that of industrial research. This appearance of strain is linked to changes in funding, in which the fedJames D. Adams is a professor of economics at the Rensselaer Polytechnic Institute and a research associate of the National Bureau of Economic Research. J. Roger Clemmons is a coordinator of statistical research at the Institute for Child Health Policy, College of Medicine of the University of Florida. The Andrew W. Mellon and Alfred P. Sloan Foundations have generously supported this research. John Marsh provided research assistance and Jason Todd Abaluck provided data on student characteristics. We thank Ronald G. Ehrenberg, Irwin Feller, Amanda Goodall, Richard Jensen, and Donald Vitaliano for comments on earlier drafts, as well as two reviewers. Special appreciation goes to Richard Freeman for support and guidance. This chapter has benefited from presentations at Rensselaer Polytechnic Institute, Cornell University, meetings of the NBER-Science and Engineering Workforce Project, and the NBER Higher Education Meetings. Any remaining errors are our responsibility. 1. Data on the top 200 universities worldwide in The Times Higher Education Supplement (2004 and 2005) suggest first the preeminence of U.S. universities, and second the erosion of this preeminence. Fifty U.S. schools are in the top 200. Where a lower rank is better, the mean for twenty-seven U.S. privates is 67.7 in 2004 and 60.7 in 2005; for twenty-three publics the rank is 72.5 in 2004 and 94.8 in 2005; the mean U.S. rank falls from 69.9 to 76.4. Shanghai Jiao Tong University (2003 and 2005) ranks 100 schools worldwide in 2003 to 2005. In 2003 fiftyeight U.S. universities are in the top 100, while fifty-three appear in 2005. The rankings of U.S. universities improve, but since several publics drop out, it is not clear what to make of this. Both rankings are controversial. The Times uses employer evaluations while the Shanghai ranking uses a weighted average of objective data on prizes, papers, citations, and the like. I thank Amanda Goodall for these references.

349

350

James D. Adams and J. Roger Clemmons

eral share of university R&D has declined over time. Given the trends and the reliance that firms place on universities, an analysis seems warranted, to see whether the slowdown reflects a fundamental decline in university prospects. We find that research productivity grows at a healthy rate but the allocation of R&D has grown less efficient over time. While this has interfered with aggregate productivity growth, increasing budget stringency, especially in public universities, may be the root cause of the problem. The empirical analysis is based on a panel of 102 top U.S. universities, sixty-eight of which are public and thirty-four private, whose outputs and inputs we observe during 1981 to 1999. A key feature of our analysis is its separation of productivity into research and teaching, with most of our emphasis placed on research owing to data availability. The approach assumes that research and teaching activities are on the whole separable. In one sense, though, our approach makes a virtue out of necessity. Price index numbers for research and teaching that could combine the two into a single index are missing for higher education.2 The definition of productivity is output per faculty-equivalent engaged in research and teaching. Research output is papers and citations, teaching output consists of undergraduate and graduate degrees, and numbers of faculty are divided into researchers and teachers. Equipped with these measures, we begin the empirical work with a description of research and teaching productivity. Next we decompose productivity growth into sources within and between universities, and also groups of public and private universities. Finally, using regression analysis, we examine the determinants of productivity in individual universities. Beginning with trends, we find that faculty in top 102 schools grow at 0.6 percent per year, while research faculty, a close approximation to researchers in science and engineering, grows at 1.4 percent a year. Both are low compared with growth of scientists and engineers in U.S. industry. In all universities during 1981 to 1999, full-time faculty grow at 1.5 percent a year, while all faculty grow at two percent (National Science Board 2004, vol. 2, table 5-17). By comparison, growth in the industrial science and engineering workforce is 4.9 percent a year during 1980 to 2000 (National Science Board 2004, vol. 1, chapter 3). The university sector is a less important employer of U.S. scientists and engineers by 2000 than it was in 1981. Also, we find that researchers increase more rapidly than teachers. By our reckoning, researchers grow at 1.4 percent a year while teachers grow at 0.3 percent. At the same time, papers per researcher grow at 1.4 percent a year and citations to these papers grow at 6.7 percent. Research productivity is clearly rising. A cautionary note is that growth in citations and real 2. In the future, the R&D satellite accounts at the U.S. Bureau of Economic Analysis could fill this gap.

Growing Allocative Inefficiency of the U.S. Higher Education Sector

351

research growth are not necessarily the same, given the falling cost of citations and worldwide growth in the number of citing researchers.3 Research productivity in private universities is roughly twice that of public universities. The growth rate of research productivity is also greater in private universities (where papers and citations grow at 2.2 and 8.6 percent per year) than in public universities (where growth is respectively 1.2 and 6.2 percent). The growth rate of research productivity is therefore twothirds to one-third higher in private universities. Findings on teaching productivity are as follows. The 102 universities produce 4.5 undergraduate degrees per teacher and 2.6 graduate degrees. Undergraduate degrees are 50 percent lower per teacher in private universities, but then graduate degrees per teacher are 50 percent higher in these universities. So, productivity in public and private institutions is roughly equal. Over time, however, teaching productivity drops slightly in private universities, while it increases at one percent a year in public schools.4 These quantity indexes do not capture changes in the value of higher education, nor do they capture changes in quality, but they represent a start on the problem of measuring teaching productivity. Besides the study of trends, we examine sources of growth in aggregate productivity. By this we mean a shift-share analysis that decomposes aggregate growth into growth within universities, growth between universities, and the covariance of growth in shares and productivity growth. Findings from the decomposition are these. Across all universities the withinuniversity component of growth accounts for more than 100 percent of growth in research output. The between-university contribution is smaller but remains positive. But the covariance of growth in research shares with growth in research productivity is negative. This implies that research shares grow faster in universities where productivity growth is slower. The decomposition yields similar results within groups of private and public universities. The covariance term is always negative and research grows faster in universities where research productivity grows more slowly. This result suggests growing allocative inefficiency in research in higher education. Analysis of sources of growth in teaching productivity tells a similar story. More than 100 percent of growth is accounted for by the within component, the between component is small but positive, and the covariance term is strictly negative. Regression analysis of research and teaching productivity concludes the empirical work. We find that R&D, endowment, and postdoctoral students increase research productivity but that research is subject to de3. See the remarks of Hall, Jaffe, and Trajtenberg in ch. 13 of Jaffe and Trajtenberg (2002). 4. The comparison between top-ten research universities and non-top-ten schools is similar.

352

James D. Adams and J. Roger Clemmons

creasing returns. In public universities (but not private) there is evidence that graduate students contribute to research productivity. The nonfederal R&D stock in a university is linked to a decline in research productivity. This result disappears when fixed effects are included so that we are unable to identify a within-university effect of nonfederal R&D. One interpretation is that nonfederal funds are subject to earmarking and are awarded under less stringent competitive conditions. Another is that the goal of nonfederal funds is less to produce research than to produce information. Regardless of the interpretation, the share of nonfederal funds in university R&D stocks grows by 19 percent over the sample period. Overall, it comprises 40 percent of funding in the publics, versus 20 percent in the privates. It could be a factor in productivity differences among public and private universities. Regression analysis finds that undergraduate teaching productivity increases with enrollment, and (in public universities) with graduate assistants. In public universities state appropriations are linked to a decline in undergraduate degrees per teacher. Production is not subject to decreasing returns to the same degree as research, suggesting that variation in university size is primarily a matter of teaching and not research. Graduate teaching productivity increases with graduate students and R&D. However, the output of graduate degrees decreases with the nonfederal share of R&D, suggesting that unlike federal R&D, nonfederal funds are not for the support of graduate students. Reassuringly, graduate students are at least as important in their own education as they are in faculty research. The rest of the chapter consists of five sections. Section 11.2 describes productivity measurement and presents identities that decompose productivity growth into within, between, and covariance components. In addition, the section specifies productivity regressions. Section 11.3 discusses the database and presents descriptive statistics. Section 11.4 carries out the decomposition analysis of productivity growth. Regression findings are presented in section 11.5. Section 11.6 is a discussion and conclusion, with emphasis on the challenges facing public universities in the United States. 11.2 Analytical Framework 11.2.1 Productivity Definitions The productivity index that we use in this chapter is output per faculty member.5 But university faculties produce both research and teaching. Can labor productivity be measured separately for both? Our best but also 5. We rely on labor productivity for the usual reason in productivity studies, that we lack data on physical capital stocks that would give us indices of total factor productivity.

Growing Allocative Inefficiency of the U.S. Higher Education Sector

353

very imperfect answer is yes. We can exploit expenditure shares on research and teaching to construct estimates of research and teaching faculty-equivalents and labor productivity in research and teaching. This of course assumes that these outputs are separable in production. While the assumption seems reasonable for research and undergraduate teaching, it is less promising for research and graduate education. To an unknown extent these are jointly produced, but for practical reasons we set this complication to one side. First, undergraduate teaching dominates most universities and this conforms to the assumption of separability. Second, statistics of teaching expenditures by universities do not distinguish undergraduate and graduate students. Estimated teaching faculty exceeds the number of undergraduate teachers. The result is a downward bias in undergraduate teaching productivity. Third, the proportion of graduate teaching in all teaching is higher in universities of the first rank. Omitting graduate teaching would bias teaching productivity comparisons between schools. A related reason for including graduate students is that top U.S. research universities have increasingly emphasized graduate teaching. Omitting graduate education would underestimate the growth of teaching productivity. So while research and graduate education have joint production aspects, there are reasons for provisionally treating the two as separable. We therefore use the following indexes of labor productivity in research and teaching: (1)

Xjit LPjit   . Ljit

j  R, I

Output and faculty form the numerator and denominator of (1). Subscript j  R, I stands for research (R) and instruction (I ), subscript i indexes universities, and t stands for time. 11.2.2 Decomposition of Productivity Growth Section 11.4 uses a shift-share analysis to decompose research and teaching productivity growth into within, between, and covariance components.6 We apply this decomposition to the explanation of productivity growth in universities and groups of public and private universities. To simplify notation we drop subscript j  R, I and let LP stand for either research or teaching. Also, let LPt represent the weighted average of productivity across universities and let LPit stand for productivity of university i. Finally, let sit  Qit / ΣNi1Qit be the share of university i in total output ΣNi1Qit. The share variable serves as a weight in the decomposition. After some algebra, which is shown in the first part of the appendix, we reach 6. See, for example, Foster, Krizan, and Haltiwanger (2001).

354

(2)

James D. Adams and J. Roger Clemmons

LPt  LPt  LPt1  Σi sit1LPit  Σi sit LPit  Σi sit(LPit1  LPt1).

⎫ ⎬ ⎭

⎫ ⎬ ⎭

⎫ ⎬ ⎭

Within-University

Covariance

Between-University

The change in aggregate productivity consists of three terms. The first is the sum of changes in productivity within universities weighted by their share in output. This is the within-university component. The second is the covariance of changes in shares with changes in productivity. It answers the question, is growth in share positively or negatively associated with productivity growth? The third term is the between-university component. It is the sum of changes in shares times the difference between individual and average productivity. This captures whether more efficient universities on average gain or lose share. Equation (2) applies to individual universities, but we are also interested in groups of private and public universities. The second part of the appendix shows that (3)

LPt  [t1·LPAt  (1  t1)·LPBt ]  t·(LPAt  LPBt )

⎫ ⎬ ⎭

⎫ ⎬ ⎭ Within-Group

 t·(LP

A t1

 LP

Covariance

B t1

).

⎫ ⎬ ⎭ Between-Group

The first term is the within-group component. It is the average across the two groups of growth in productivity within each group using withingroup average productivity growth. The second is the covariance component: growth in group A’s share times the gap between growth in its productivity and group B’s. The third term is the between-group component: the increase in group A’s share times the difference in its initial productivity and that of group B. We use (2) and (3) to decompose productivity growth in higher education in section 11.4. 11.2.3 Productivity Regressions Section 11.5 undertakes regression analysis of labor productivity. For this purpose, as noted, productivity is derived from separable production functions for research and teaching. We assume that labor productivity in research takes an almost Cobb-Douglas form: F tu (ARitLRit) (RKNF it  Kit ) e LPRit  QRit /LRit   , LRit R

R

(4)

 F tu  ARit LRit1(RK NF . it  K it ) e R

R

R

R

Rit

R

Rit

Growing Allocative Inefficiency of the U.S. Higher Education Sector

355

The term ARit captures productivity-augmenting features of universities.7 NF We decompose R&D of a university (KRit) into the nonfederal stock (KRit ), on which we allow a discount or premium R and the federal stock (KFRit). F 8 The effective R&D stock is KRit  R KNF Rit  KRit. Also included in (4) are time trend t and uRit, the error term in research productivity. Besides R, the parameters include R, the output elasticity of labor; R, the output elasticity of the R&D stock; and R, the coefficient of time trend. The error term uRit consists of a sum of variance components, (5)

uRit  vRi  vRt  eRit.

In (5) the error consists of components for university vRi and time vRt as well as the innovation eRit. We sometimes include fixed effects in (4) to absorb university and time. As long as the innovation is unanticipated, in these equations it will be orthogonal to predetermined variables on the right of (4). Returning to productivity ARit, one determinant of it is an indicator of public or private control Ci. This affects productivity through governance and selectivity. Endowment Eit is used to hire star faculty and buy back time, so we expect it to increase productivity. And both postdoctoral and graduate students Mit and Git could augment faculty time. Research laboraugmentation follows the constant-elasticity function, ARit  BR e

E it M it G it .

RCCi

RE

RM

RG

Inserting this into (4), substituting (5) for the error, rearranging, and taking logarithms we reach the nonlinear regression (6)

ln (QRit / LRit)  R ln(BR)  Rt  RRCCi  RRE ln(Eit / LRit1)  RRM ln(Mit / LRit1)  RRG ln(Git / LRit1) F  R ln[(RKNF Rit  KRit) / LRit1]

 [R(1  RE  RM  RG)  R  1]ln(LRit1)  vRi  vRt  eRit. Section 11.5 reports estimates of (6). When constant returns to scale hold, the coefficient on the logarithm of LRit vanishes. Otherwise its sign captures 7. The (almost) Cobb-Douglas assumption means that Hicks-neutral shifts cannot be distinguished from factor augmentation. For convenience we treat all shifts as labor augmenting. 8. This functional form allows a direct comparison between the effects of a dollar of nonfederal and federal R&D stock. As far as we are aware, use of this device appears first in Griliches (1986), who used it to distinguish the effects of basic and applied research on firm productivity.

356

James D. Adams and J. Roger Clemmons

the direction of divergence from constant returns.9 Notice that we lag LRit on the right by one year to limit division error bias. Teaching productivity can be similarly modeled. Assuming a CobbDouglas production function for baccalaureate and graduate degrees, we obtain the following specification for teaching productivity (7)

(AIitLIit)I S itIe I tuIit LPIit  QIit / LIit    AIitI LIitI1S itIe ItuIit. LIit

As with (4), we consider the error term to consist of a sum of variance components: (8)

uIit  vIi  vIt  eIit.

The error again consists of components for university vit and time vit and the innovation eRit. Thus, we sometimes include fixed effects to absorb university and time. As long as the innovation is unanticipated, in these equations it will be orthogonal to predetermined variables on the right of (7). Labor augmentation AIit depends on teaching skill and other aspects of teaching. Included are enrollments or stocks of students in residence Sit; time trend t; and uIit, the error term in teaching productivity. Parameters are I, the output elasticity of labor; I, the output elasticity of enrollment; and I, the coefficient of time trend. Determinants of instructional labor-augmentation AIit again include public or private control Ci. A second determinant, in public universities, is state teaching appropriations per teacher Tit. This could be destined for the reduction of class size. If so, we expect it to reduce degrees per teacher. Alternatively, state appropriations could alter the composition of education in favor of graduate education. But in addition, Tit could increase the quality of education. And third, graduate students Git per teacher could substitute for faculty in undergraduate teaching. Thus, instructional laboraugmentation is represented by the constant-elasticity function, AIit  BI eICCiT itITG itIG. Next insert AIit and the equation error (8) into (7) and take logarithms: (9)

ln(QIit / LIit)  I ln(BI)  It  IICCi  IIT ln(Tit /LIit1)  IIG ln(Git /LIit1)  I ln(Sit /LIit1)  [I (1  IT  IG)  I  1] ln(LIit1)  vIi  vIt  eIit.

9. Adams and Griliches (1998) regress the logarithm of research output on the logarithm of R&D stock. They find that the specification exhibited diminishing returns at the universityfield level and constant returns at the field level. They also consider the role of graduate students in R&D. But they did not examine labor productivity, because such data were not available.

Growing Allocative Inefficiency of the U.S. Higher Education Sector

357

We also include the logarithm of R&D stock in some of the graduate student equations, using the same functional form K ln(KRit)  I ln (I KNF Rit  KFRit ), as in (6). Section 11.4 reports estimates of (9). If constant returns holds, then the coefficient on LIit disappears; otherwise its sign captures the divergence from constant returns. As before, we lag LIit to limit division error bias. 11.3 Description of the Data 11.3.1 Database of Universities This study is based on 110 universities that account for most academic research in the United States. The primary data sources that we use are the Institute for Scientific Information (ISI) for research outputs, the Integrated Postsecondary Education Data System (IPEDS) data from the National Center for Education Statistics (NCES) for finances, faculty, salaries, and degrees; and the National Science Foundation (NSF) CASPAR database for academic R&D and graduate students. Since data are missing for eight universities, this study examines 102 schools. Allowing for lags we observe universities during 1982 to 1999. Thus, before missing values are removed, the data form a panel of 1,836 observations (eighteen years times 102 universities).10 Included in the panel are faculty counts, research and teaching expenditures, research outputs consisting of papers and citations, and teaching outputs consisting of baccalaureate and graduate degrees. We use the expenditure data to allocate faculty between research and teaching. These data yield labor productivity statistics in research and teaching. In addition, we construct R&D stocks, endowment, stocks of graduate students, undergraduate enrollments, and indicators of public-private control.11 The rest of this section describes the variables and calculations that we have performed using them. 11.3.2 Faculty Statistics The data include estimates of faculty counts by university. We use tenure-track and non-tenure-track faculty counts from the National Center for Education Statistics’ (NCES) Faculty Salary Survey, available through the Integrated Postsecondary Education System (IPEDS). Figure 11.1 shows tenure-track and non-tenure-track faculty over time. Non10. Because research and teaching faculty are lagged one year on the right of equations (5) and (7), the 1981 data are excluded from the regressions. 11. The R&D is overcounted because of transfers between universities. Such transfers should be deducted from the R&D of sending universities and added to the R&D of receiving universities, but this is not the current practice.

358

James D. Adams and J. Roger Clemmons

Fig. 11.1

Average number of faculty, 102 universities, 1981–1999

tenure-track faculty grow at a slightly faster rate than tenure-track faculty, but not by enough to change the non-tenure-track share, which remains at nine percent throughout the period.12 Because faculty engages in research and teaching and these tend to be competing uses of time, we would like to obtain faculty-equivalents in these activities. If these were mutually exclusive, then production functions for research and teaching would be separable. This assumption is not as reasonable for graduate education, where teaching and research are to an extent jointly produced.13 But as noted in section 11.2, it is necessary to tolerate some inaccuracy in the allocation of faculty to research and teaching. 12. National Science Foundation data show that the share of part-time faculty during 1981 to 1999 rises from 19 percent to 28 percent in research universities (National Science Board 2004, vol. 2, table 5-17). We studied the use of part-time faculty using the biennial NCES Fall Staff Surveys from 1987 to 1997. Leaving aside graduate assistants, we find that the thirtyfour privates use a higher proportion of part-time faculty than the sixty-eight publics. However, the part-time proportion grows faster, by 24 percent versus 10 percent. This suggests that the Salary Survey may understate relative faculty growth in public universities. But the Fall Staff Survey data are rather noisy; and they fail to classify graduate assistants by teaching and research function. The evidence presented in table 11.9 suggests that graduate students are an important substitute for faculty in public universities. 13. Modern graduate education is often credited to the nineteenth century chemist Justus von Liebig, who learned how to combine graduate teaching with laboratory research. See the entry on von Liebig in the Encyclopedia Britannica and Mokyr (2002).

Growing Allocative Inefficiency of the U.S. Higher Education Sector

359

The data on teaching expenditures do not distinguish undergraduates from graduates, and removing graduate education as an output biases the contributions of different universities. Thus, we employ research and teaching expenditure to separate faculty into research and teaching components. Note that these categories exclude administration, sports, and auxiliary enterprises such as food and dormitory services, hospitals, and student organizations. This seems correct since the primary activities of faculty are teaching and research. Notice also that research expenditures REXPit include separately budgeted expenditures that are internal and external to the university. However, research is almost entirely in science and engineering, so that research faculty are a close approximation to researchers in science and engineering. Instructional expenditures IEXPit include expenditures for credit and noncredit instruction. This includes all instruction: academic, occupational, vocational, special session, community, and remedial and tutorial instruction. Also included are research and public service that are not separately budgeted. One problem is that both research and teaching expenditure include spending on capital and auxiliary personnel. Thus, use of the research expenditure share could yield a biased estimate of research faculty. To guard against this we include R&D stock (which includes capital expenditures) as well as graduate students and postdocs in the regressions. By this account the separation of research and teaching is imperfect. But as an assumption, it is clearly an improvement on perfect multitasking. That assumption argues that faculty members simultaneously teach and perform research. We replace it with a better—even if imperfect— approximation, that the proportion of research faculty equals the proportion of research expenditures in both research and teaching expenditures REXPit /(REXPit  IEXPit). Research and teaching faculty LRit and LIit in university i at time t are to a first approximation: (10)

REXPit LRit   Lit. REXPit  IEXPit LIit  Lit  LRit

In (10) Lit is total faculty in university i at time t. Research and teaching faculty are LRit and LIit, the denominators of labor productivity in research and teaching in equations (1), (4), (6), (7), and (9). From what has gone before, research faculty are a close approximation to research scientists and engineers, consistent with the definition of research output. There is however, a possible bias in this, which suggests that researchers are overestimated and teachers underestimated. Because the research skill price exceeds that of teaching, research expenditures buy fewer researchers and teaching expenditures buy more teachers than (10) would suggest. But

360

James D. Adams and J. Roger Clemmons

Fig. 11.2

Ratio of research to total expenditures, 102 universities, 1981–1999

since we know rather little about the research premium we cannot correct this bias.14 Figure 11.2 charts the course of the expenditure proportion REXPit / (REXPit  IEXPit). For all universities, the curve’s fish-hook shape reflects the decline in research funding from 1981 to 1983 and its subsequent recovery and expansion. But the overall curve conceals differences between public and private universities. In both cases the expenditure share declines through 1983, but afterwards the pattern differs. The research share in private schools recovers to 0.38 in 1988 but then declines. This is consistent with reductions in overhead rates for private schools in the late 1980s (Ehrenberg 2003). The overall pattern in private universities is one of decline, from 0.41 in 1981 to 0.36 in 1999. In contrast, the research share in public universities rises from 0.33 in 1983 to 0.40 in 1999 and the overall pattern is one of increase. Table 11.1 reports means and growth rates of faculty, the research expenditure proportion, and researchers and teachers. It does so for all universities, as well as public and private. Universities employ an average of 14. Let f  REXP/(REXP  IEXP) as in equation (10), and let  (wR – wI) / wI  0, where wR  research wage and wI  teaching wage and let n  measured total faculty. Then it can be shown that the true number of researchers is n∗R  f /[ f  (1 – f )(1  )]n and the true number of teachers is n∗I  [(1 – f )(1  )] / [ f  (1 – f )(1  )]n. But unfortunately the value of is unknown, including its variation by university.

Growing Allocative Inefficiency of the U.S. Higher Education Sector Table 11.1

361

Faculty by research and teaching function, public and private universities, 1981–1999 University classification

Faculty indicator

All

Public

Private

Means Tenure-track  Non-tenure-track faculty Research expenditure proportion Research faculty-equivalents Instructional faculty-equivalents

1,048 0.379 381 667

1,218 0.381 444 774

703 0.376 252 451

Annual percentage growth rates Tenure-track  Non-tenure-track faculty Research expenditure proportion Research faculty-equivalents Instructional faculty-equivalents

0.6 0.5 1.4 0.3

0.4 1.0 1.6 0.1

1.4 0.5 0.8 1.8

Notes: The universities are 110 top U.S. research universities, less eight schools with incomplete data. Means and growth rates of the expenditure proportion are weighted by expenditure.

1,048 faculty. The research expenditure proportion is 38 percent and an estimated 381 faculty are engaged to do research while 667 teach. Public universities employ 1,218 faculty, of which 444 are researchers and 774 teachers. Employment in private schools is 703, of which 252 are researchers and 451 teachers. Table 11.1 also presents growth rates. Researchers grow faster than teachers by 1.4 percent a year versus 0.3 percent. Thus, research-intensity of faculty is growing. Growth of researchers is faster in public universities, while growth in teachers is faster in private universities.15 Figures 11.3 and 11.4 are graphs of research and teaching faculty. To concentrate on cumulative growth and facilitate comparison we normalize each time series by its 1981 value. Figure 11.3 shows that research faculty rise by almost 30 percent in the publics but by less than 15 percent in the privates. Figure 11.4 reveals that teachers grow by more than 30 percent in the privates but decline slightly in the publics. For all universities, cumulative growth in researchers is 25 percent by 1999 (fig. 11.3) but only five percent for teachers (fig. 11.4). This suggests that the mix of faculty in top U.S. universities is becoming more research-oriented. 11.3.3 Research and Teaching Outputs To calculate labor productivity in research and teaching we require output measures. We treat papers and citations as research outputs, comparable 15. Because research expenditures that are not separately budgeted are recorded as instructional expense, the figures for instruction may include cross-subsidization of research by teaching.

Fig. 11.3

Research faculty equivalents, 102 universities, 1981–1999 (1981  1.0)

Fig. 11.4 Instructional faculty equivalents, 102 universities, 1981–1999 (1981  1.0)

Growing Allocative Inefficiency of the U.S. Higher Education Sector

363

with patent statistics in industry. The articles derive from agriculture, astronomy, biology, chemistry, computer science, earth sciences, economics and business, engineering, mathematics and statistics, medicine, physics, and psychology. These fields account for nearly all research carried on in universities and are closely linked to total research expenditures. The universities publish 2.4 million papers during 1981 to 1999 and the papers receive 18.8 million citations. For each paper we calculate the fraction that a given university contributes. If two schools are listed each is assigned half of the paper, if three are listed each is assigned one-third, and so on. Citations received are similarly assigned and in this way we limit the problem of multiple counting of research output. The fractions are summed across fields by year to arrive at fractional paper-equivalents of a university per year. Fractional citations are similarly summed, and the citations are accumulated over the first five years since publication, yielding a five-year window on citations received. This right-truncates the citations. The five-year window also cuts off citations in 1995, the last year for which a complete record exists. Despite this, the five-year window standardizes citations received and provides a quality dimension for research output. Baccalaureate and graduate degrees are currently our indicators of teaching output. At the present time we lack a quality indicator such as cost or forward value of a degree.16 The data are taken from NCES-IPEDS degree surveys. The upper half of table 11.2 reports mean research output consisting of papers and five-year citations, and teaching output consisting of baccalaureate and graduate degrees. As before, we report data for all universities, as well as public and private. Universities publish 1,183 papers per year: the papers account for 4,948 citations over their first five years. Private universities publish slightly more total papers and public universities slightly less, but private schools have a decided advantage in citations (Adams and Griliches 1998), which probably signals differences in faculty quality as reflected in salary (Ehrenberg 2003). Universities produce 3,010 baccalaureate degrees and 1,747 graduate degrees per year. Reflecting their size and specialization in undergraduate education, public universities produce 3,795 baccalaureate degrees and 1,721 graduate degrees. Private universities produce 1,417 baccalaureate degrees and 1,758 graduate degrees; they specialize in graduate education. 11.3.4 Labor Productivity in Research and Teaching The lower half of table 11.2 reports means of productivity by type. The data show an 85 percent advantage of private universities in papers (7.4 16. One idea is to use National Association of Colleges and Employers (NACE) data on starting salaries by major, but these are not available for use by academic researchers.

364

James D. Adams and J. Roger Clemmons

Table 11.2

Research and teaching outputs and productivity, public and private universities, 1981–1999 University classification

Faculty indicator

All

Public

Private

Mean research output Papers Five-year citations

1,183 4,948

1,173 4,170

1,204 6,526

Mean teaching output Baccalaureate degrees Graduate degrees

3,010 1,747

3,795 1,741

1,417 1,758

Weighted mean research productivity Papers/research faculty Five-year citations/research faculty

3.1 10.3

2.6 7.4

4.8 20.4

4.5 2.6

4.9 2.2

3.1 3.9

Weighted mean teaching productivity Baccalaureate degrees/teaching faculty Graduate degrees/teaching faculty

Notes: Means of research and teaching productivity are weighted by faculty size.

papers versus 2.6 papers per faculty), and an almost three-to-one advantage in citations (20.4 citations versus 7.4 citations per researcher).17 In table 11.2 total degrees per teacher are similar across university type. Any differences show up in undergraduate and graduate productivity.18 Indeed, the total degree gap is small considering the concentration of private schools on costly graduate education. The smaller output of undergraduate degrees per faculty in these institutions again indicates their specialization in graduate education. Figures 11.5 and 11.6 are graphs of research productivity over time. Again, the series are normalized by 1981 values. All the series on papers per researcher in figure 11.5 grow through 1995 and flatten afterwards. Private universities grow faster, with the divergence taking place during 1981 to 1995. By 1999, papers per research faculty grow by 20 percent in public universities but by 40 percent in private universities. Figure 11.6 reports citations received per faculty. The data series end in 1995, given the five-year window on citations. Again, a gap opens up between privates and 17. Means weighted by size of research faculty. Equally-weighted means for public and private institutions are 3.8 and 4.9 papers per researcher, and 17.4 and 25.3 five-year citations per researcher. We prefer weighted means, which give larger universities more weight and offer a clearer picture of overall research productivity. 18. Since the data do not allow us to distinguish undergraduate teachers from graduate teachers, we are double-counting teachers in computing teaching productivity. Thus, it is not all clear that fewer undergraduate degrees are produced per undergraduate teacher in private schools, or that fewer graduate degrees are produced per graduate teacher in public schools.

Fig. 11.5

Papers per research faculty, 102 universities, 1981–1999 (1981  1.0)

Fig. 11.6 Five-year citations per research faculty, 102 universities, 1981–1995 (1981  1.0)

366

James D. Adams and J. Roger Clemmons

Table 11.3

Annual percentage growth rates in research and teaching productivity, public and private universities, 1981–1999 University classification

Productivity statistic

All

Public

Private

Percentage growth in research productivity Papers/research faculty Five-year citations/research faculty

1.4 6.7

1.2 6.2

2.2 8.6

Percentage growth in teaching productivity Baccalaureate degrees/teaching faculty Graduate degrees/teaching faculty

0.8 1.1

1.2 1.4

0.6 0.1

Notes: The table covers 1981 to 1999 for papers and 1981 to 1995 for citations. Productivity growth rates are weighted by faculty size. All growth rates are in percents per year.

publics during 1981 to 1995. By 1995 citations per researcher in public universities grow by 80 percent, but by 220 percent in private universities. Table 11.3 provides more evidence on the increasing productivity gap between public and private universities. Annual growth in papers is 1.4 percent in all institutions and growth in citations is 6.7 percent. Comparable figures in public universities are 1.2 percent (papers) and 6.2 percent (citations). Productivity growth in private universities equals 2.2 percent (papers) and 8.6 percent (citations). The bottom half of the table shows growth in teaching productivity in all universities of about one percent a year. The data show a decline in teaching productivity in private universities of –0.6 to –0.1 percent, compared with a rise of 1.2 to 1.4 percent in public universities. But again these measurements lack a quality dimension. Trends in baccalaureate and graduate degrees per teacher are shown in figures 11.7 and 11.8. The figures show that all the growth in teaching productivity occurs in public universities. Comparing these with figures 11.6 and 11.7 we see that as measured, productivity growth is faster in research than teaching. 11.3.5 Other Data We collected several other variables, including faculty salary, academic R&D stocks, endowment, and state teaching appropriations, all expressed in thousands of 1992 dollars. In addition, we collected lagged stocks of graduate students from the NSF-CASPAR database. Table 11.4 reports means of faculty compensation, consisting of wages plus fringe benefits, by faculty rank and university type. Mean compensation averages 65,000 in 1992 dollars. Compensation is higher in private universities, especially at the full professor level, so that the wage trajectory is much steeper in these universities. Figure 11.9 shows that compensation also rises at a faster rate in private universities. Both patterns are familiar,

Fig. 11.7 Baccalaureate degrees per instructional faculty, 102 universities, 1981– 1999 (1981  1.0)

Fig. 11.8 Graduate degrees per instructional faculty, 102 universities, 1981–1999 (1981  1.0)

368

James D. Adams and J. Roger Clemmons

Table 11.4

Faculty compensation by rank, public and private universities, 1981–1999 University classification Faculty indicator

All

Public

Private

Means Assistant professor Associate professor Full professor All ranks

49.1 59.3 81.9 64.7

48.7 58.5 79.4 62.6

50.0 61.2 87.4 69.5

Notes: Faculty compensation is expressed in thousands of 1992 dollars and includes fringe benefits in addition to wages.

Fig. 11.9

Faculty compensation, 102 universities, 1981–1999 (1981  1.0)

but what is not as well known is how closely the public-private wage differential tracks the differential in public-private research productivity (but not teaching productivity). This advantage of private universities is of course related to their financial resources. Past R&D funding contributes to current research output and it also indicates research excellence. For both reasons it is correlated with research productivity. The R&D stock is the lagged stock of research funding received over the previous eight years, depreciated at 15 percent per year, and expressed in thousands of 1992 dollars. The R&D pertains to the same

Growing Allocative Inefficiency of the U.S. Higher Education Sector

Fig. 11.10

369

Nonfederal share in R&D stock, 102 universities, 1981–1999

fields of science and schools that yield the research output statistics.19 The source of the R&D data is the NSF-CASPAR database. We divide the R&D stock into federal and nonfederal components. This is a likely factor in research productivity because nonfederal money could be less subject to competitive pressures than federal grants and because it may consist of contracts that provide information and advice rather than publications.20 Figure 11.10 show that nonfederal R&D contributes 20 percent of the private university stock but 40 percent of the public university stock. The share of nonfederal R&D grows relative to the federal stock and is 19 percent higher by 1999. Endowment is used to attract highly skilled faculty and to support research. For both reasons, endowment per faculty should increase research productivity. Endowment could also reduce size of classes or support students, although we fail to find evidence for this. State appropriations could reduce class size and degrees per faculty member but they could also expand graduate programs. These data derive from NCES-IPEDS surveys. 19. The twelve fields are agriculture, astronomy, biology, chemistry, computer science, earth sciences, economics and business, engineering, mathematics and statistics, medicine, physics, and psychology. 20. It is for this reason that we think that recent findings (De Figueiredo and Silverman 2006) that 5 to 6 percent of federal R&D dollars are earmarked and a source of inefficiency represent an understatement of the problem. We agree that the federal question is interesting, but we also believe that replacement of federal funds by nonfederal funds may be the larger issue.

370

James D. Adams and J. Roger Clemmons

The lagged stock of graduate students helps to produce research and undergraduate teaching. It should increase research and teaching productivity, but besides this it is an output (Adams and Griliches 1998). The graduate student data are drawn from the NSF-CASPAR database for the twelve sciences in this study. Also taken from this source is the stock of postdoctoral students, another input into research. 11.4 Decomposition of Aggregate Productivity Growth Following equation (2), table 11.5 reports decompositions of aggregate productivity growth in research and teaching. The table contains three panels corresponding to all universities, as well as public and private. The top line of each panel reports aggregate productivity growth. This is arithmetic rather than percentage growth. It is the sum of the change in productivity over all universities in a given set. By (2), the within-university, covariance, and between-university components sum to the total except for rounding error. The shares of each component in aggregate productivity growth are shown in parentheses. The within-university component dominates. It is usually positive: the exception is a small decline in teaching productivity within private universities. The covariance term is always negative: this implies that output share grows more rapidly in universities where productivity grows more slowly. The between-university component is usually positive: output shares grow in universities whose productivity is above average. One exception to this is a slight decline in the between-university component of citations. We would like to compare table 11.5 with decompositions for the private sector. Foster, Haltiwanger, and Krizan (2001) offers the closest comparison. In their findings for industry the within-establishment component is a much smaller share of productivity growth.21 This is partly because net entry contributes to industry growth. Entry is identically zero for top universities but besides this, the covariance term is positive in industry and negative in higher education. In summary, while entry and betweenestablishment reallocation increase private sector growth, they are either not a factor (entry) or they decrease growth in universities (covariance). Table 11.6 studies growth in groups of public and private universities. The decomposition follows equation (3). Within-group productivity growth is positive but the covariance and between-group terms are negative in seven out of eight cases. The results imply that the share in research and teaching rises faster for the group whose productivity grows more slowly (covariance component), and that the share grows faster for the group whose productivity is less (between-group component). In research it is the less efficient 21. See Foster, Haltiwanger, and Krizan (2001, 322, table 8.4, line 2).

0.509 (1.00) 0.520 (1.02) 0.297 (0.58) 0.287 (0.56)

1.534 (1.00) 1.880 (1.23) 0.514 (0.34) 0.168 (0.11)

Public universities (N  68) Total productivity growth Within university Covariance Between university

Private universities (N  34) Total productivity growth Within university Covariance Between university 20.019 (1.00) 22.878 (1.14) 2.582 (0.13) 0.278 (0.01)

5.969 (1.00) 5.933 (0.99) 0.589 (0.10) 0.625 (0.10)

8.625 (1.00) 9.998 (1.16) 1.518 (0.18) 0.145 (0.02)

Five-year citations/ res. faculty

0.377 (1.00) 0.176 (0.46) 0.190 (0.50) 0.010 (0.03)

0.976 (1.00) 1.041 (1.07) 0.189 (0.19) 0.123 (0.13)

0.585 (1.00) 0.801 (1.37) 0.251 (0.43) 0.035 (0.06)

Bacc. degrees/teach. faculty

0.064 (1.00) 0.051 (0.80) 0.334 (5.22) 0.219 (3.42)

0.518 (1.00) 0.626 (1.21) 0.145 (0.28) 0.037 (0.07)

0.470 (1.00) 0.512 (1.09) 0.221 (0.47) 0.178 (0.38)

Grad. degrees/teach. faculty

Notes: Productivity growth is the difference over 1981 to 1999. It is the arithmetic difference XT  X1 and not (XT  X1) / X1. The decomposition follows equation (2). The sum of the components may differ slightly from the total because of rounding error. Shares in total productivity growth in parentheses.

0.701 (1.00) 0.846 (1.21) 0.374 (0.53) 0.229 (0.33)

All universities (N  102) Total productivity growth Within university Covariance Between university

Papers/res. faculty

Aggregate productivity growth in university research and teaching

University classification

Table 11.5

372

James D. Adams and J. Roger Clemmons

Table 11.6

Aggregate productivity growth within and between groups of public and private universities

Productivity statistic All universities (N  102) Total productivity growth Within group Covariance Between group

Papers/ res. faculty

Citations/ res. faculty

Bacc. degrees/ teach. faculty

Grad. degrees/ teach. faculty

0.701 (1.00) 0.755 (1.08) 0.022 (0.03) 0.032 (0.05)

8.625 (1.00) 9.339 (1.08) 0.424 (0.05) 0.291 (0.03)

0.585 (1.00) 0.709 (1.21) 0.069 (0.12) 0.055 (0.09)

0.470 (1.00) 0.403 (0.86) 0.030 (0.06) 0.097 (0.21)

Notes: Productivity growth is the difference over 1981 to 1999. It is the arithmetic difference XT  X1 and not (XT  X1) / X1. The decomposition follows equation (3). The sum of the components may differ slightly from the total because of rounding error. Shares in total productivity growth in parentheses.

group of public universities whose share increases, while in teaching it is the apparently less efficient group of private universities. 11.5 Regression Findings The empirical work concludes with regression analysis of research and teaching productivity. Tables 11.7 and 11.8 contain findings on research productivity in public and private universities. The dependent variable in 7.1 through 7.3 is the logarithm of papers per research faculty. The dependent variable in 7.4 through 7.6 is the logarithm of five-year citations to the papers per research faculty. Equations 7.3 and 7.6 include university fixed effects while the rest exclude these effects. Consider papers per researcher in public universities. The coefficient of time trend is negative and significant in 7.1 and 7.2 but is positive and significant in 7.3. This is consistent with the shift of research toward less productive universities. Table 11.5 has shown that as a result, within-university growth accounts for more than 100 percent of growth. This negative between effect is included in 7.1 and 7.2 but is omitted from the within regression 7.3. Besides trend, the table includes the logarithm of R&D stock per researcher, and it also includes the logarithm of lagged researchers, as a check on returns to scale. The nonfederal coefficient is significantly less than that of federal R&D and it approximates zero in the citation regressions.22 The R&D elasticity is always positive. The coefficient of lagged researchers is negative, suggesting decreasing returns to scale throughout. Equation 7.2 adds endowment, graduate students, and postdoctoral students to 7.1.23 The effect of R&D stock declines but remains positive and 22. The negative sign on nonfederal R&D does not hit a boundary because nonfederal funds are small. 23. To be more precise, graduate and postdoctoral students are averages of stocks over the previous three years.

1982–1999 No 0.015*** (7.6) 0.478*** (5.0) 0.455*** (21.7) 0.021*** (3.3) 0.431*** (15.7) 0.138*** (14.2) 0.224*** (15.1) 68 1,054 0.325 0.760

7.2 1982–1999 Yes 0.009*** (8.6) 0.566*** (4.1) 0.297*** (16.3) 0.019*** (4.4) 0.277*** (12.8) 0.004 (0.8) 0.400*** (32.4) 68 1,054 0.073 0.988

7.3 1982–1995 No 0.016*** (3.6) 0.043 (0.9) 0.831*** (20.6) — — — — — — 0.301*** (9.0) 68 831 0.573 0.573

7.4 1982–1995 No 0.002 (0.4) 0.113* (2.0) 0.544*** (13.8) 0.050*** (4.5) 0.312*** (6.3) 0.218*** (12.4) 0.325*** (11.2) 68 831 0.534 0.679

7.5

7.6 1982–1995 Yes 0.056*** (24.5) 0.738* (2.0) 0.272*** (6.7) 0.005 (0.6) 0.178*** (3.5) 0.042*** (3.3) 0.807*** (30.1) 68 831 0.122 0.983

Citations per research faculty

Notes: Dependent variables are logarithms of papers and citations per research faculty-equivalent. So as to avoid division error bias, research facultyequivalents used in the right-hand side variables are lagged one year relative to research faculty equivalents on the left. t-statistics in parentheses. ***Significant at the one-tenth of one percent level. **Significant at the one percent level. *Significant at the five percent level.

Number of universities Number of observations Root mean squared error Adjusted R2

Log (research faculty1)

Log (postdoctoral students per research faculty1)

Log (graduate students per research faculty1)

Log (endowment per research faculty1)

Log (stock of R&D per research faculty1) (βR)

Nonfederal stock of R&D per research faculty1 (δR)

1982–1999 No 0.026*** (10.9) 0.477*** (6.3) 0.707*** (31.4) — — — — — — 0.245*** (13.3) 68 1,054 0.406 0.625

7.1

Papers per research faculty

Public universities: Nonlinear least squares (NLLS) research productivity equations, papers and citations per research faculty

Time period University fixed effects Time trend

Variable or statistic

Table 11.7

1982–1999 No 0.017*** (5.4) 1.352*** (3.9) 0.443*** (13.4) 0.094*** (5.4) 0.077** (2.7) 0.263*** (11.7) 0.214*** (10.6) 34 475 0.260 0.705

8.2 1982–1999 Yes 0.003 (1.8) 0.315 (1.6) 0.304*** (11.4) 0.144*** (6.9) 0.068* (2.1) 0.031 (1.5) 0.381*** (15.7) 34 475 0.072 0.977

8.3 1982–1995 No 0.015*** (3.9) 0.467*** (8.4) 0.793*** (16.9) — — — — — — 0.193*** (5.1) 34 475 0.499 0.584

8.4 1982–1995 No 0.008* (2.1) 0.627*** (56.8) 0.325*** (10.6) 0.121*** (5.1) 0.036 (1.0) 0.539*** (17.2) 0.100*** (3.5) 34 475 0.364 0.779

8.5

8.6 1982–1995 Yes 0.042*** (14.1) 0.891 (1.7) 0.295*** (6.8) 0.104*** (3.1) 0.072 (1.4) 0.033 (1.0) 0.553*** (14.1) 34 475 0.117 0.977

Citations per research faculty

Notes: Dependent variables are logarithms of papers and citations per research faculty-equivalent. So as to avoid division error bias, research facultyequivalents used in the right-hand side variables are lagged one year relative to research faculty equivalents on the left. t-statistics in parentheses. a By coincidence data on endowments of private universities end in 1995 so that numbers of observations on papers and five-year citations are the same. ***Significant at the one-tenth of one percent level. **Significant at the one percent level. *Significant at the five percent level.

Number of universities Number of observationsa Root mean squared error Adjusted R2

Log (research faculty1)

Log (postdoctoral students per research faculty1)

Log (graduate students per research faculty1)

Log (endowment per research faculty1)a

Log (stock of R&D per research faculty1) (βR)

Nonfederal stock of R&D per research faculty1 (δR)

1982–1999 No 0.024*** (8.3) 0.617*** (3.8) 0.699*** (21.0) — — — — — — 0.264*** (10.9) 34 475 0.318 0.558

8.1

Papers per research faculty

Private universities: NLLS research productivity equations, papers and citations per research faculty

Time period University fixed effects Time trend

Variable or statistic

Table 11.8

Growing Allocative Inefficiency of the U.S. Higher Education Sector

375

significant. Since R&D stock supports graduate and postdoctoral students, part of its effect is mediated by these variables, which are accordingly positive and significant. In 7.2, endowment has a small positive effect. The sign and significance of lagged researchers again suggests diminishing returns. We include fixed effects in the within-university equation 7.3, which is otherwise the same as 7.2. The elasticities of the R&D stock, graduate students, and postdoctoral students decline in the within-university dimension but remain positive and significant. Endowment is now negative and significant, which is puzzling. Diminishing returns are stronger than before. Equations 7.4 through 7.6 report citation regressions whose setup follows 7.1 through 7.3. Compared to the earlier results trend growth is higher, but consistent with table 11.5 it is still higher in the within regression 7.6. The discount of nonfederal R&D is even greater than for papers, but this effect disappears in 7.6. The elasticity of R&D stock is higher than in the papers regressions, which suggests that part of R&D’s effect occurs through research quality. Diminishing returns to R&D continues to prevail. The contribution of postdoctoral students (but not graduate students) to research productivity remains positive and significant once fixed effects are included. Table 11.8 reports similar results for private universities. Equations 8.1 through 8.2 and 8.4 through 8.5 are the total specifications for papers and citations. As in table 11.7, the coefficient of time trend reverses sign when fixed effects are included in 8.3 and 8.6. When fixed effects are included, as in 8.3 and 8.6, the elasticity of the R&D stock declines but this coefficient remains significant. The estimate of the nonfederal coefficient is imprecise: in the papers equations 8.1 and 8.3 it is significantly less than 1.0, but in 8.2 this difference is not significant. The nonfederal effect is significantly less than zero in 8.4 and 8.5 but does not differ from 1.0 in the within equation 8.6. Overall, as in table 11.7, the nonfederal R&D coefficient is less than or equal to that of federal R&D. Endowment is consistently stronger in table 11.8, implying that private universities are adept at harnessing endowment to raise their research productivity. The coefficient of postdoctoral students increases but the graduate student coefficient decreases compared with table 11.7. Thus, private universities rely more on postdoctoral students to produce their research. Finally we turn to tables 11.9 and 11.10, which contain regression findings for teaching productivity. The dependent variable in 9.1 through 9.3 and 10.1 through 10.3 is the logarithm of baccalaureate degrees per teacher. In 9.4 through 9.6 and 10.4 through 10.6 it is the logarithm of graduate degrees per teacher. We begin with undergraduate productivity in public universities. Equation 9.1 includes time trend, the logarithm of undergraduate enrollments

OLS 1982–1999 No 0.001 (0.6) 0.604*** (23.9) 0.007 (1.5) 0.354*** (16.1) 0.243 (11.8) — — — — 0.069*** (4.9) 68 886 0.240 280.7 0.655

9.2 OLS 1982–1999 Yes 0.007*** (6.7) 0.487*** (12.2) 0.003 (0.5) 0.065* (2.2) 0.017 (0.9) — — — — 0.354*** (7.8) 68 886 0.069 409.2 0.971

9.3 OLS 1982–1999 No 0.011*** (4.0) — — — — 0.409*** (16.7) — — — — — — 0.066*** (3.8) 68 886 0.346 144.3 0.327

9.4 NLLS 1982–1999 No 0.008*** (3.8) — — 0.031*** (4.8) 0.248*** (8.7) 0.068** (2.6) 0.308*** (4.5) 0.130*** (5.0) 0.007 (0.4) 68 886 0.320 — 0.426

9.5

NLLS 1982–1999 Yes 0.001 (0.8) — — 0.020** (3.2) 0.408*** (10.6) 0.061** (3.2) 0.090 (1.2) 0.268*** (9.3) 0.135*** (7.5) 68 886 0.088 — 0.957

9.6

Grad. degrees per teaching faculty

Notes: Dependent variables are logarithms of undergraduate and graduate degrees per teaching faculty-equivalent. To avoid division error bias, teaching faculty used in the right-hand side variables are lagged one year relative to teaching faculty on the left. t-statistics in parentheses. ***Significant at the one-tenth of one percent level. **Significant at the one percent level. *Significant at the five percent level.  F-statistic is significant at the one-tenth of one percent level.

Number of universities Number of observations Root mean squared error F Adjusted R2

Log (teaching faculty1)

Log (stock of R&D per teaching faculty1) (βI)

Nonfederal stock of R&D per teaching faculty1 (δI)

Log (state appropriations per teaching faculty1)

Log (graduate students per teaching faculty1)

Log (endowment per teaching faculty1)

Log (undergrad. enrollment per teaching faculty1)

OLS 1982–1999 No 0.003 (1.5) 0.790*** (31.9) — — — — — — — — — — 0.106*** (8.1) 68 886 0.276 353.5 0.544

9.1

Bacc. degrees per teaching faculty

Public universities: Ordinary least squares (OLS) and NLLS teaching productivity equations, baccalaureate and graduate degrees per teaching faculty

Estimation method Time period University fixed effects Time trend

Variable or statistic

Table 11.9

OLS 1982–1999 No 0.007* (2.4) 0.731*** (18.2) 0.062*** (4.2) 0.072** (3.7) 0.022*** (8.9) — — — — 0.207*** (10.2) 34 475 0.236 84.2 0.513

10.2 OLS 1982–1999 Yes 0.000 (0.1) 0.658*** (7.1) 0.073** (3.3) 0.021 (0.7) 0.003 (1.4) — — — — 0.359*** (4.0) 34 475 0.067 294.6 0.960

10.3 OLS 1982–1999 No 0.003 (0.5) — — — — 0.388*** (11.7) — — — — — — 0.228*** (5.9) 34 475 0.462 50.6 0.239

10.4 NLLS 1982–1999 No 0.013** (3.2) — — 0.118*** (3.9) 0.146** (3.0) 0.014** (3.0) 0.492 (0.9) 0.285*** (6.2) 0.217*** (6.4) 34 475 0.428 — 0.347

10.5

NLLS 1982–1999 Yes 0.006*** (3.5) — — 0.025 (1.1) 0.308*** (8.6) 0.002 (0.9) 0.130 (0.5) 0.154*** (6.2) 0.282*** (12.6) 34 475 0.079 — 0.978

10.6

Grad. degrees per teaching faculty

Notes: Dependent variables are logarithms of undergraduate and graduate degrees per teaching faculty-equivalent. To avoid division error bias, teaching faculty used in the right-hand side variables are lagged one year relative to teaching faculty on the left. t-statistics in parentheses. ***Significant at the one-tenth of one percent level. **Significant at the one percent level. *Significant at the five percent level.  F-statistic is significant at the one-tenth of one percent level.

Number of universities Number of observations Root mean squared error F Adjusted R2

Log (teaching faculty1)

Log (stock of R&D per teaching faculty1) (βI)

Nonfederal stock of R&D per teaching faculty1 (δI)

Log (state appropriations per teaching faculty1)

Log (graduate students per teaching faculty1)

Log (endowment per teaching faculty1)

Log (undergrad. enrollment per teaching faculty1)

OLS 1982–1999 No 0.002 (0.8) 0.631*** (16.4) — — — — — — — — — — 0.175** (8.1) 34 475 0.262 105.3 0.398

10.1

Bacc. degrees per teaching faculty

Private universities: OLS and NLLS teaching productivity equations, baccalaureate and graduate degrees per teaching faculty

Estimation method Time period University fixed effects Time trend

Variable or statistic

Table 11.10

378

James D. Adams and J. Roger Clemmons

per teacher, and following equation (7) the logarithm of teachers, to test for the returns to scale to teaching.24 Time trend is insignificant. The logarithm of enrollment is positive and significant, and its coefficient is robust in 9.3 to the inclusion of fixed effects. We would expect it to be robust given that students are inputs into their own education (Rothschild and White 1995; Winston 1999). The coefficient of teaching faculty is positive and significant in 9.1, suggesting increasing returns. However, when fixed effects are included in 9.3 this sign reverses. Thus, unlike research, where returns are decreasing, the evidence on returns to scale is mixed in undergraduate teaching. Equation 9.2 includes the logarithms of graduate students, endowment, and state appropriations per teacher. Graduate students play a significant role in public undergraduate education but it is perhaps not surprising that endowment has little effect. State appropriations reduce degrees per faculty, but the interpretation of this is unclear. Equation 9.3 adds fixed effects to 9.2. Enrollment and graduate students remain important determinants of baccalaureate degrees within universities, but state appropriations drop out. The graduate teaching equations conclude table 11.9. Equation 9.4 includes trend, graduate students, and lagged teachers. Trend is positive and significant, graduate students are a key input into their own education, and the sign of lagged teachers provides some evidence of diminishing returns. Equation 9.5 adds state appropriations per teacher. These increase output of graduate degrees, the opposite of 9.2. Together this suggests that state support substitutes graduate students for undergraduates. Because R&D hones the research skills of graduate students, equation 9.5 also includes the logarithm of the stock of R&D. The coefficient of nonfederal stock has a negative effect on graduate degrees; this is insignificant in 9.6. Federal R&D supports graduate education while nonfederal R&D does not. Equation 9.6 adds fixed effects to 9.5. Coefficients of graduate students and R&D stock remain significant, but the signs of endowment, state appropriations, and lagged teachers change. In particular, the evidence on decreasing returns in this table is fragile and conflicting. Along with the evidence on decreasing returns to research, it suggests that variation in university size is primarily due to teaching. Table 11.10 reports findings for private universities. Main differences from table 11.9 are as follows: first, there is evidence for decreasing returns to undergraduate teaching in private universities. Second, unlike their role in public universities, graduate students are not a significant input for undergraduate education. As before, graduate degrees do not increase with nonfederal R&D. 24. To be precise, undergraduate enrollment is the average undergraduate enrollment over the previous three years.

Growing Allocative Inefficiency of the U.S. Higher Education Sector

379

11.6 Discussion and Conclusion This chapter finds evidence of growing allocative inefficiency in U.S. higher education. Our most compelling evidence for this claim derives from research output, which is better measured than teaching output at this time. We find that universities whose productivity grows less rapidly experience more rapid growth in research share. The allocation of research between public and private universities has also grown less efficient over time. While the share of public universities grows more rapidly, their research productivity grows more slowly. On top of this the betweenuniversity component is negative: the public university share grows though their research productivity is less. One suspect that might explain this growing inefficiency is nonfederal R&D. Its more rapid growth and its much larger role in public universities fit the patterns that we observe. In support of this view, tables 11.7 and 11.8 show that nonfederal R&D stock decreases research productivity. Whether this result is due to less competitive conditions attending nonfederal grants or whether nonfederal awards produce less research by intention, we cannot say. According to tables 11.7 and 11.8, private university endowments also contribute to the gap in public-private research productivity. Our findings for teaching productivity are similar, but we are less convinced by them. For starters, the quality dimension of instruction is missing. Falling class size could reflect a rising demand for quality due to growth in wealth at the top of the distribution. This indicates that families partly control the allocation of students to schools. Surely this moderates allocative inefficiency in teaching. A deeper interpretation of the observations might instead point to the financial fortunes of public and private universities over the past quarter century. The public-private comparisons in this chapter are consistent with rising teaching pressures on public universities that could well discourage more productive researchers from applying for positions. This decline in competitiveness might explain the increasing reliance, especially by state universities, on nonfederal R&D that appears to detract from researchproductivity. On that interpretation, the rising allocative inefficiency of research that we uncover results from funding pressures that render state universities less competitive, and drive them to less productive funding sources. This view of the matter implies a stunning reversal of fortune for public universities. Starting from the Morrill Act of 1862 and the Hatch Act of 1887, state universities offered practical education in the agricultural and mechanical arts to support local industry. For more than a century this formula has achieved great successes (Huffman and Evenson 1993; Adams 2002). But in our own time it appears to have been less successful. This can perhaps be traced to aging of the population and to the rising mobility of

380

James D. Adams and J. Roger Clemmons

students, both of which weaken the appeal of state finance of universities. If this interpretation is correct, then it suggests a different and more privatized approach to funding universities that would place greater reliance on parental finance of teaching, and federal and private foundation finance of research. In any event, some solution seems urgent if the United States is to retain its preeminence in higher education, and subsequently in academic and industrial science, technology, and innovation.

Appendix Productivity Decomposition Section 11.4 uses the shift-share analysis described in Foster, Haltiwanger, and Krizan (2001) to decompose productivity growth into within, between, and covariance components for universities. This section explains the algebra underlying equations (2) and (3) of the text. Decomposition among Individual Universities Let LPt represent mean labor productivity across universities, LPit stand for productivity of a university, and sit  Qit /ΣNi1Qit represent the share of a university in total output. Then (A1)

LPt  LPt  LPt1  Σi sitLPit  Σi sit1LPit1  Σi sit LPit  Σi sit1LPit1  Σi sit1LPit  Σi sit1LPit  Σi sit1LPit  Σi sitLPit  Σi sit1LPit  Σi sit1LPit  Σi sitLPit  Σi sit1LPit  Σi sitLPit1  Σi sitLPit1  Σi sit1LPit1  Σi sit1LPit1  Σi sit1LPit  Σi sitLPit  Σi sitLPit1.

To (A1) we add the term: Σi sit LPit1.

(A2)

Equation (A2) equals zero because LPt–1 can be factored out and the sum of the changes in shares is zero. Combining terms in the result yields equation (2) of the text: (A3)

LPt  LPt  LPt1

⎫ ⎬ ⎭

⎫ ⎬ ⎭

⎫ ⎬ ⎭

 Σi sit1LPit  Σi sitLPit  Σi sit(LPit1  LPt1). Within-University

Covariance

Between-University

Growing Allocative Inefficiency of the U.S. Higher Education Sector

381

Decomposition among Groups of Universities We are also interested in contributions of groups of universities A and B to productivity growth. Let A and B exhaust the set of universities. Then aggregate labor productivity growth is (A4)

LPt  LPt  LPt1  Σi sit LPit  Σi sit1LPit1  (ΣAsit LPit  ΣAsit1LPit1)  (ΣB sit LPit  ΣB sit1LPit1).

Notice that the sit weights do not add to 1.0 within groups. The following equation rewrites the weighted averages of labor productivities in (A4) in terms of within-group averages: (A5)

LPt  (t ΣAsAit LPit  t1ΣAs Ait1LPit1)  [(1  t )ΣB s Bit LPit  (1  t1)ΣB s Bit1LPit1].

The three new terms in (A5) are: (A6)

t  ΣAQit  (ΣAQit  ΣBQit)

(A7)

sAjt  Qjt  (ΣAQit), j ∈ A sBjt  Qjt  (ΣBQit), j ∈ B.

Factor total output from the denominator of (A4). Then multiply and divide by the sum of output in each group using the within-group weights (A7), yielding (A5). As a result we can rewrite (A5) as (A8)

LPt  (tLPAt  t1LPAt1)  [(1  t)LPBt  (1  t1)LPBt1].

The top line of (A8) is (A9)

tLPAt  t1LPAt1  t1 · LPAt  t · LPAt  t · LPAt1.

The bottom line of (A8) equals (A10) (1  t)LPBt  (1  t1)LPBt1  (1  t1) · LPBt  (1  t) · LPBt  (1  t) · LPBt1. Substitute (A9) and (A10) into (A8) and combine terms using (1 – t)  –t. We reach (A11) LPt  [t1 · LPAt  (1  t1) · LPBt ]  t · (LPAt  LPBt )  t · (LPAt1  LPBt1).

⎫ ⎬ ⎭ Between-Group

(A11) is equation (3) of the text.

⎫ ⎬ ⎭

⎫ ⎬ ⎭ Within-Group

Covariance

382

James D. Adams and J. Roger Clemmons

References Adams, J. D. 2002. Comparative localization of academic and industrial spillovers. Journal of Economic Geography 2 (3): 253–78. Adams, J. D., and Z. Griliches. 1998. Research productivity in a system of universities. Les Annales D’Economie et de Statistique 49/50:127–62. De Figueiredo, J. M., and B. S. Silverman. 2006. Academic earmarks and the returns to lobbying. Journal of Law and Economics 49 (October): 597–625. Ehrenberg, R. G. 2003. Studying ourselves: The academic labor market. Journal of Labor Economics 21 (2): 267–88. Foster, L., J. Haltiwanger, and C. J. Krizan. 2001. Aggregate productivity growth: Lessons from microeconomic evidence. In New developments in productivity analysis, studies in income and wealth, vol. 63, ed. C. R. Hulten, E. R. Dean, and M. J. Harper, 303–72. Chicago: University of Chicago Press. Griliches, Z. 1986. Productivity, R&D, and basic research at the firm level in the 1970s. American Economic Review 76 (1): 141–54. Jaffe, A. B., and M. Trajtenberg. 2002. Patents, citations and innovations: A window on the knowledge economy. Cambridge, MA: MIT Press. Huffman, W. E., and R. E. Evenson. 1993. Science for agriculture: A long-term perspective. Ames, IA: Iowa State University Press. Mokyr, J. 2002. The gifts of Athena: Historical origins of the knowledge economy. Princeton, NJ: Princeton University Press. National Science Board. 2004. Science and engineering indicators 2004. (vol. 1, NSB 04-1; vol. 2, NSB 04-1A). Arlington, VA: National Science Foundation. Rothschild, M., and L. J. White. 1995. The analytics of the pricing of higher education and other services in which the customers are inputs. Journal of Political Economy 103 (3): 573–86. Shanghai Jiao Tong University, Institute of Higher Education. (2003 and 2005). Academic ranking of world universities. Available at http://ed.sjtu.edu.cn/ ranking.htm. Times Higher Education Supplement. (2004 and 2005). Available at http:// www.thes.co.uk/international comparisons. Winston, G. C. 1999. Subsidies, hierarchy and peers: The awkward economics of higher education. Journal of Economic Perspectives 13 (1): 12–36.

Contributors

James D. Adams Department of Economics Rensselaer Polytechnic Institute 3504 Russell Sage Laboratory Troy, NY 12180-3590

Hanley Chiang Mathematica Policy Research, Inc. P.O.Box 2393 Princeton, NJ 08543-2393

Keith A. Bender Department of Economics University of Wisconsin-Milwaukee P.O. Box 413 Milwaukee, WI 53201

J. Roger Clemmons Institute for Child Health Policy University of Florida, School of Medicine 1329 SW 16th St., Room 5130 Gainesville, FL 32608

George J. Borjas Kennedy School of Government Harvard University 79 JFK Street Cambridge, MA 02138

Geoff Davis Google 1600 Amphitheatre Parkway Mountain View, CA 94034

John Bound Department of Economics University of Michigan Ann Arbor, MI 48109-1220

Richard B. Freeman National Bureau of Economic Research 1050 Massachusetts Avenue Cambridge, MA 02138

Tanwin Chang Adaptive Optics Associates 10 Wilson Road Cambridge, MA 02138

Donna K. Ginther Economics Department 333 Snow Hall University of Kansas 1460 Jayhawk Boulevard Lawrence, KS 66045-7585

383

384

Contributors

Daniel L. Goroff Department of Mathematics Harvey Mudd College 301 Platt Boulevard Claremont, CA 91711

Cyrus C. M. Mody Department of History-MS 42 Rice University P.O. Box 1892 Houston, TX 77251-1892

John S. Heywood Department of Economics P.O. Box 413 University of Wisconsin-Milwaukee Milwaukee, WI 53201

Paula E. Stephan Andrew Young School of Policy Studies Georgia State University University Plaza Atlanta, GA 30303-3083

Shulamit Kahn Department of Finance and Economics Boston University School of Management 595 Commonwealth Avenue Boston, MA 02215 Jinyoung Kim Department of Economics Korea University 5-1 Anam-Dong, Sungbuk-Ku Seoul, Korea 136-701 Sangjoon John Lee College of Business Alfred University 1 Saxon Drive Alfred, NY 14802 Gerald Marschke Department of Economics University at Albany State University of New York Albany, NY 12222

Albert J. Sumell Department of Economics Youngstown State University 1 University Plaza Youngstown, OH 44555 Sarah Turner Department of Economics University of Virginia P.O. Box 400182 Charlottesville, VA 22904-4182 Patrick Walsh Department of Economics St. Michael’s College 1 Winooski Park Colchester, VT 05439 Kjersten Bunker Whittington Department of Sociology Reed College 3203 SE Woodstock Boulevard Portland, OR 97202-8199

Author Index

Acs, Z. J., 260 Adams, J. D., 356n9, 363, 370, 379 Adams, S. J., 241n4 Agrawal, A., 260 Aisenburg, N., 199 Allen, J., 231 Allison, P. D., 166, 199 Almeida, P., 261, 322, 338n12, 344 Altonji, J.G., 132n5 Anderson, M. S., 104 Anselin, L., 260 Ansell, C. K., 217 Armstrong, J. S., 322 Audretsch, D., 260, 261, 279, 322 Banks, B. A., 323 Bartel, A. P., 116 Bassett, R. K., 296n5 Battu, H., 229, 231 Belfield, C. R., 229, 231 Belman, C. R., 229 Bender, K. A., 116, 235, 240n3 Bercovitz, J., 198 Berelson, B., 66n9 Berkley, J., 223n23 Binning, G., 295 Black, G., 258n2, 260, 270, 270n12 Blank, D., 93 Blume, S., 292n2 Bok, D., 291 Bonaccorsi, A., 197 Borghans, L., 231

Borjas, G., 87n23, 132n4, 133, 133n6, 134, 136n7 Bound, J., 87 Bowen, W., 60n2, 66n9, 66n11, 92n28 Branstetter, L. G., 322 Brass, D. J., 223 Brewer, M., 260, 279, 279n21, 322 Bromberg, J. L., 292 Bruinshoofd, A., 231 Buchel, F., 231 Caplow, T., 198 Card, D., 132n5, 133, 136n6 Carre, F., 212 Carrington, W., 73 Cavender, J., 116 Chacar, A. S., 338 Chang, T., 62n5 Chevalier, A., 231 Chiang, H., 62n5 Cialdini, R., 115 Clark, A. E., 231, 240 Cohen, M. B., 117, 279 Cohen, W. M., 322 Cohen, W. R., 195, 259 Cole, J., 197 Cooley, E., 21 Cox, B. G., 165 Creager, A. N. H., 292 Daraio, C., 197 Darby, M., 259n4, 260, 270, 279n21, 322

385

386

Author Index

Davis, G., 103 de Grip, A., 231 Detragiache, E., 73 De Vries, R., 104 De Young, B. R., 117 Diener, E., 115 Ding, W. W., 197, 198 Dolton, P., 231 Drucker, P. F., 115 Duguet, E., 323 Dunning, J. H., 344 Edmonston, B., 132n4 Ehrenberg, R., 133n8, 360, 363 Elzen, B., 292 Etzkowtiz, H., 196, 197, 199, 291 Evenson, R. E., 379 Everhart, S. S., 230 Feldman, M, 198, 260, 322 Finn, M., 68n12, 131n2, 138n12 Fischer, C. S., 300n12 Florida, R., 195 Fogarty, M. S., 323 Foster, L., 353n6, 370, 371n21, 380 Fox, M. F., 197, 198, 199 Fox-Kean, M., 322 Freeman, R. B., 41, 62, 90, 92, 93, 101, 116, 132n3, 133n6, 136n7, 151n26, 183n15, 230, 231, 321 Friedberg, R. M., 132, 132n5 Geiger, R. L., 291 Gereffi, G., 64n4 Ginther, D. K., 166, 184 Goe, R., 195 Goldberg, C., 164 Goldsmith, S., 21 Goolsbee, A., 94n30 Graham, L., 197, 224 Griliches, Z., 321, 332, 355, 356n9, 363, 370 Groen, J. A., 280 Groot, W., 229, 231 Grossman, J. B., 132n5 Hall, B. H., 201, 326, 332 Haltiwanger, J., 353n6, 370, 370n21, 380 Hamermesh, D., 132, 240 Harary, F., 216 Harrington, M., 199 Harris, R. D., 231 Hausman, J., 332 Hayes, K. J., 166, 184

Hechinger, F., 84n21 Henderson, R., 260 Heywood, J. S., 116, 229, 235, 240n3 Hoddeson, L., 296n5 Hoffman, C. M., 131n1 Hu, A. G. Z., 322 Huffman, W. E., 379 Hunt, J., 132 Iaffaldano, M. T., 116 Ibarra, H., 223 Ionescu-Pioggia, M., 103 Jackson, M., 292 Jacobsen, J. P., 115 Jaffe, A. B., 201, 260, 322, 323, 326, 334, 351n3 Jordan, K., 301n14 Kahn, S., 166, 184 Kahneman, D., 115 Kaiser, D., 299n10 Kannankutty, N., 197 Katz, L. F., 133n6, 133n7 Keller, W., 322 Kemelgor, C., 197, 199 Kenney, M., 292 Kim, B., 230 Kim, J., 322n2, 332 Kirp, D. L., 291 Kleinman, D. L., 195, 301 Kline, R., 300n12 Knowles, S., 296n5 Kogut, B., 261, 322, 344 Kohler, R., 292 Koput, K. W., 22, 211 Kortum, S., 332 Krizan, C. J., 353n6, 370, 370n21, 380 Kruytbosch, C., 197 Kurzweil, M., 6n11, 60n2 LaLonde, R. J., 132n5 Latour, B., 304 Lazear, E. P., 116, 241n4 Lécuyer, C., 292 Lemieux, T., 133 Lenoir, T., 292 Lerner, J., 332 Leslie, L. L., 291 Leslie, S. W., 296n5 Levin, S. G., 133, 229, 230, 231 Lieberman, M. B., 338 Link, A., 258

Author Index Long, J. S., 164, 166, 184, 196, 197, 199 Lowell, L., 68 Lynch, M., 301n14 Maasen van den Brink, H., 231 MacGarvie, M., 323 Maeroff, G., 61n3 Mansfield, E., 258n3 Marschke, G., 322n2, 332 Martinson, B. C., 104 McGee, G., 116, 198 McGinnis, R., 166, 199 McGoldrick, K., 229 McPherson, M., 223 Melamed, R., 327, 329 Merton, R. K., 220 Mirowski, P., 291 Mitchell, S. B., 165 Moguerou, P., 116, 118 Mokyr, J., 358n13 Moody, J., 216 Moonesinge, R., 165 Morgan, R., 197 Moshavi, D., 231 Mowery, D. C., 322 Muchinsky, P. M., 116 Murray, F., 197, 224 Nelson, R., 195, 259, 279, 322 Nohria, N., 223 Oh, H., 230 Oswald, A. J., 231 Oudshoorn, N., 300n12 Owen-Smith, J., 195, 196, 197, 201, 213, 213n14 Pace, E., 84n21 Packer, K., 196 Padgett, J. F., 217 Page, K. L., 196, 211 Pantalony, D., 292 Petty, M., 116 Phene, A., 338n12, 344 Pinch, T., 300n12 Pion, G., 103 Podolny, J. M., 196, 211 Polyani, M., 322 Porter, K. A., 217 Powell, W. W., 196, 197, 201, 211, 212, 213, 213n14, 217, 259n4 Presley, J., 21 Preston, A. E., 183, 230, 234, 245

387

Quate, C. F., 302n16 Rader, K., 292 Rasmussen, N., 292 Rayman, P., 212 Raymond, C., 84 Regets, M., 85n22, 87n23, 103, 321 Riordan, M., 296n5 Robst, J., 229 Rohrer, H., 295 Rosen, S., 116 Rosenberg, N., 195 Rosser, S. V., 182 Rothschild, M., 378 Roy, A. D., 42 Rubb, S., 247 Rudensine, N., 66n9 Samuelson, P. A., 132 Sattinger, M., 229 Schaffer, S., 292 Schoeni, R. F., 132n5 Schwarz, N., 115 Scranton, P., 393 Seaman, P., 229 Sent, E.-M., 291 Shapin, S., 292, 292n1 Shauman, K. A., 118, 164, 197, 198 Shiff, G., 327, 329 Shinn, T., 292n2 Shkolnikov, V., 84 Simar, L., 197 Singh, J., 334 Sjaastad, L. A., 261 Skillman, G. L., 115 Slaughter, S., 291 Sloane, P. J., 229, 231, 240 Smith, J. P., 132n4 Smith-Doerr, L., 196, 197, 200, 211, 212, 224, 225 Smith-Lovin, L., 223 Snyder, T. D., 131n1 Solomon, L. C., 231 Stephan, P. E., 197, 198, 229, 230, 257, 258n2, 261, 270, 279 Stern, S., 258n1 Stevens, R., 66n11 Stigler, G., 93 Stuart, T. E., 197, 198 Tarnow, E., 117 Teitelbaum, M. S., 230 Terborg, J. R., 231

388

Author Index

Thompson, P., 260, 322 Thurow, L., 230 Tobin, E., 60n2, 66n11 Topel, R. H., 132n5 Trajtenberg, M., 201, 322, 326, 327, 329, 334, 351n3 Tsang, M. C., 229, 230, 231 Turner, S., 87, 92 Vallas, S. P., 195 Van der Velden, R., 231 Varga, A., 260 Vettel, E. J., 292 Vignoles, A., 231 Vishwanath, T., 73 Vogel, G., 100 Walsh, J., 259, 279, 322 Ward, M., 240 Weakliem, D. L., 208n12

Webster, A., 196 Welch, F., 133, 144, 145 White, D. R., 216 White, L. J., 378 White, M., 280 Whittington, K. B., 196, 197, 200, 212, 213, 214n16, 217, 222, 223, 224, 225 Winston, G. C., 378 Wise, G., 296n5 Witte, M., 92n28 Wolbers. M., 231 Woolgar, S., 304 Xie, Y., 118, 164, 197, 198 Zander, U., 344 Ziedonis, A. A., 322, 332 Zucker, L., 259n4, 260, 279, 322, 279n21 Zuckerman, H., 197, 198

Subject Index

Page references followed by f or t refer to figures and tables, respectively. Academia: industry and, 198–99; patenting collaboration in, 216–18; patents and, 197 Academic capitalism, 291 Academic collaborations, 310 Academic patenting collaboration networks, 218. See also Patenting collaboration networks Academic science, structure of, 218–20 Alexander, Sam, 312–13 American Competitiveness Initiative, 1 American Vacuum Society, 294 Assignment theory, 229 Asylum Research, 310n33 Atomic forced microscopes (AFMs), 292; start-ups for, 309–10 Baldeschwieler, John, 296, 309 Bell Labs, 295–96, 303 Binning, Gerd, 295, 301 China, 81–83; expansion of university system in, 12 Collaboration networks. See Patenting collaboration networks Collaborations, academic, 310 Committee of Visitors (COV), 22–23 Community-building, invention and, 295– 96. See also Instrumental communities Compact D/SEC data set, 326

Crowd-out effects, 87–89 Cumulative Index (NSF), 20 Digital Instruments (DI), 306–8, 310, 310n33, 312–14 Dissemination. See Publishing Doctorate students: earnings of, 8; job satisfaction and, 8. See also Ph. D. graduates Drake, Barney, 313 Earnings, doctorate scientists and, 8 Eastern Europe, 84–85 Elings, Virgil, 306–7, 309, 310, 312, 313 Employment, 11–12 Engineering: employment in, 11–12; U.S. loss of dominance in, 12–14. See also Sciences Engineering programs, foreign students in, 5 European Union, expansion of university system in, 12 Fellowships, 19–20 Foreign-born students: Ph. D. graduates and, 59–60; in science and engineering programs, 5; in sciences vs. humanities, 64; supply of, 5; in United States, 7; wage impact of, 156–58 F-1 visas, 62

389

390

Subject Index

Garcia, Niko, 306 Gender, network structures and, 222–24. See also Women, in sciences Geographic mobility, Ph. D. graduates and, 8–9. See also Location decisions Geographic proximity, role of, in knowledge transfers, 260–61 Graduate Research Fellowships (GRFs), 2, 20; alternative policy scenarios for, 45– 52; awards per science and engineering baccalaureates, 23, 23f; demographic determinants of winning, 33–37; determinants of number of applicants for, 37–41, 40t; determinants of winning, 20, 27–37; disciplinary distribution over time of, 24–25, 24f; historical review of, 21–28; measured skills of awardees of, 41–43, 42f; minorities and, 26–28; numbered value of awards over time of, 22–23, 22f; percentage of women fellows, 25–26, 26f; proportion of applicants gaining awards, 24–25, 25f Graduate students: fellowships for, 19–20; stipends for, 19 Gray markets, for probe microscopes, 302– 6; startup era of, 308–14

Instrumentation innovation, 9 Invention: community-building and, 295–96 Inventions: instrumental communities and, 295–96; patent statistics for, 334t Inventors, 332–34, 333t; overseas, 9–10 Japan, expansion of university system in, 12 Job satisfaction, 8; expectations and, 231; overeducation and, 231; underutilized skills and, 231. See also Mismatches KLA-Tencor, 310 Knipping, Uwe, 311 Knowledge flows, international, 334–43, 337t Knowledge spillovers: construction process for data set for, 327–32; data sources for, 324–26; descriptive statistics for, 332–34, 333f, 333t; introduction to, 321–24 Knowledge transfers: Ph. D. graduates and, 259–60; role of geographic proximity in, 260–61 Korea, expansion of university system in, 12

Hansma, Helen, 313 Hansma, Paul, 296, 298–99, 301, 305, 309, 310, 311, 312, 313 High-skill labor markets: data for, 134–42; impact of foreign students on salaries in: 131–34; postdoctoral appointments, 151–56; regression analysis of, 142–51; simulation of wage impact of foreign student influx in, 156–58. See also Ph. D. graduates

Labor markets. See High-skill labor markets Labor productivity. See Productivity Langer, Robert, 216–17, 218 Lindsay, Stuart, 310–11 Location decisions, 257–59; descriptive statistics for, 281–84t; determinants of, 261–62; results of study of, 271–77, 285–86t; of science and engineering Ph. D. graduates in industry, 262–71, 265–67t; top twenty-five metropolitan areas, 269t; variable definitions for, 281–84t

IBM, 295–96, 303 India, expansion of university system in, 12 Industrial patenting collaboration networks, 220–22. See also Patenting collaboration networks Industrial science, structure of, 220–22 Industry: patenting collaboration in, 216–18 Industry, impact of organizational structure of research in, 7–8 Industry, women in, 199 Instrumental communities, 292–95, 292n2; dynamics of, 296–300; invention and, 295–96; moieties, 297–98. See also Probe microscopes

Magnetic force microscopes (MFMs), 292 Massie, James, 313 McCormick, Darryl, 311 McCormick, Larry, 311 Microfabricated AFM cantilevers, 304–5 Migration. See Location decisions Migration patterns, of Ph. D. graduates in industry, 262–71, 265–67t. See also Location decisions Mismatches: consequences of, 230; data for study of, 232; determinants of, 250–52; explanations for, 230; influence of experiences on penalty for, 248–50; introduction to, 229–31; methodology for

Subject Index study of, 232–37; other indicators for, 241–44; panel data for, 246–48; results of study of, 237–41; role of reasons for, 244–46. See also Job satisfaction Molecular Imaging, 311 Nanodevices, 310n33 Nanoscope I, 307, 310 National Postdoctoral Association, 6 National Science Foundation (NSF): reports on women and minorities in science by, 163–65 National Science foundation (NSF), 2–5, 20; Cumulative Index of, 20. See also Graduate Research Fellowships (GRFs) NBER Patent-Citations, 326 Networks. See Patenting collaboration networks Nonsurface STMers, 297–98 Overseas inventors, 9–10 Pacific Nanotechnology, 310n33 Park, Sang-Il, 310 Park, Scientific Instruments (PSI), 310, 311–12 Park, Sung-Il, 310, 311–12 Patenting, 8; academia and, 197; data for, 213–15; institutional environments and, 198; overview of, 195–96; patterns of, across sectors and disciplines, 201– 13; publishing and, 197–98 Patenting collaboration networks, 333f, 333t; academic and industry, 216–81; constructing sample of, 215–16; data for, 213–15; descriptive statistics for, 332–34; gender and, 222–24 Patents: number of, for pharmaceutical and semiconductor industries, 334; structures of academic, 218–20 Patents BIB, 325 Ph. D. degrees: China and, 81–85; cross sectional analysis of, 68–71, 69t; distribution of, by country, 67–68; distribution of, by country, field, and program quality, 71–73, 72t; Eastern Europe and, 84–85; former Soviet Union and, 84– 85; Iran and, 83–84; opportunity costs, for U.S. students, 89–94 Ph. D. graduates, 5; crowd-out effects of, U.S. students by foreign students and,

391

87–89; foreign-born, 59–60; geographic mobility and, 8–9; impact of foreignborn students on earnings of, 131–34; increase in, 62, 62f; in industry, migration patterns of, 262–72; knowledge transfers and, 259–60; location decisions of, 257–60; role of, in knowledge transfer, 259–60; top twenty-five destination metropolitan areas for, 269t; U.S.-born graduates, 5–6; U.S.-born vs. foreign born, 64–65, 65f. See also High-skill labor markets Postdoctoral researchers: causal relationships of success measures and, 118–20, 119t; components of success measures for, 109–11, 110t; concerns of, 100; contracts for, 115; correlates of success measures for, 111–18, 112–14t, 121–25; demographics and, 118; distribution of measures of recommended practices for, 106–9, 107f, 107t, 108t; impact of foreign students on earnings and, 151– 56; impact of recommended practices for, 109–11; impact of structured plans and professional opportunities on, 6; improving training for, 102–3; introduction to, 99–100; measures of recommended practices for, 106; measuring quality of experiences of, 103–6; ownership of intellectual property and, 117; professional development for, 116; regression results for, 115–19; research/career plans and, 115–16; salary and benefits for, 116; time and decline in productivity of, 117–18; working conditions of, 100–101 Postdoctoral work, 6 Private universities: baccalaureate degrees per instructional faculty of, 367f; decomposition of aggregate productivity growth at, 370–72, 371t, 372t; faculty compensation at, 368f, 368t; graduate degrees per instruction faculty of, 367f; nonfederal share in R&D funding at, 368–69, 369f; OLS and NLLS teaching productivity at, 377t; regression findings for research and development productivity at, 372–79, 374t; research and etching productivity of, 10; research outputs of, 361–66, 362f; teaching outputs of, 361–66, 362f. See also Public universities; Universities

392

Subject Index

Probe microscopes, 292; building and buying of, 300–302; commercialization of, 302–6; Digital Instruments (DI) and, 306–8; gray markets for, 302–6; startup era of, 308–14 Productivity: analytical framework for, 352– 57; decomposition equations for, 380– 81; decomposition of growth of, 353– 54; definitions for, 352–53; regressions, 354–57; of U.S. universities, introduction to, 349–52; of women in sciences, 354–57. See also Universities, U.S. Productivity puzzle, 198 Public universities: baccalaureate degrees per instructional faculty of, 367f; decomposition of aggregate productivity growth at, 370–72, 371t, 372t; faculty compensation at, 368f, 368t; graduate degrees per instructional faculty of, 367f; nonfederal share in R&D funding at, 368–69, 396f; OLS and NLLS teaching productivity at, 376t; regression findings for research and development productivity at, 372–79, 373t; research and teaching productivity of, 10, 362f; research outputs of, 361–66; teaching outputs of, 361–66, 362f. See also Private universities; Universities Publishing: factors influencing, 199–200; overview of, 195–96; patents and, 197– 98; patterns of, across sectors and disciplines, 201–13 Quanscan, 309, 310, 311 Quate, Calvin, 296, 301, 305, 309, 310 Quesant, 310n33 Research: federal funds to universities for, 62, 62f; organizational structure of, women and, 7–8 Researchers. See Postdoctoral researchers Rising Above The Gathering Storm: Energizing and Employing America for a Brighter Economic Future (National Academy of Science), 11 Rohrer, Heini, 295 Scanning tunneling microscopes (STMs), 9, 292–93, 292n3; building and buying of, 300–302; commercialization of, 302–6; Digital Instruments (DI) and, 306–8; gray markets for, 302–6

Science programs, foreign students in, 5 Sciences: employment in, 11–12; U.S. loss of dominance in, 12–14. See also Engineering; Women, in sciences Scientific papers, 8 Smith, Doug, 305–6, 307 Soundex coding system, 345–46 Standard&Poor’s Annual Guide, 326 Star scientists: in academic science, 218; in industrial science, 220–22 Stipends, 19; model of optimal policy for, 43–56 Success measures, 121–25 Surface STMers, 297–98 Survey of Doctoral Recipients (SDI), 134– 35, 165 Survey of Earned Doctorates (SED), 134 Taiwan, expansion of university system in, 12 Tenure-track jobs, women and, 7 Topometrix, 310, 310n33 Tunneling Microscope Company, 305 Undergraduate degrees: attainment of, for U.S. residents, 85–87; cross sectional analysis of, 68–70, 69t United States: education market of, foreign students and, 61–65; loss of dominance in science and engineering, 12–14 Universities: degree receipt among U.S. students in, 85–94 Universities, U.S.: analytical framework for productivity of, 352–57; baccalaureate degrees per instructional faculty of, 367f; changes in enrollment of Chinese students in, 81–83; changes in enrollment of Eastern European students in, 84–85; changes in enrollment of former Soviet Union students in, 84–85; changes in enrollment of Iranian students in, 83–84; cross-sectional distribution by country in doctoral attainment at, 67–73; database of, 357; decomposition of aggregate productivity growth at, 370–72, 371t, 372t; expansion of, 5–6; faculty compensation at, 368t; faculty compensation of, 368f; faculty statistics for, 357–61; graduate degrees per instructional faculty of, 367f; growth in representation of foreign students enrolled in, 73–81; insti-

Subject Index tutional context of, 65–67; internationalization of, 59–61; productivity of, 349–52; regression findings for research and development productivity at, 372–79; research outputs of, 361– 63; teaching outputs of, 361–63. See also Private universities; Public universities West, Paul, 309, 310, 311 Women, in sciences, 7, 163–94; academic career ladders for, 169–75; analysis of promotion of academic scientists, 175– 81; children and, 174, 182–83; data for,

393

165–69; industry and, 199; likelihood of publishing or patenting for, 8; literature review of, 197; marriage and, 182; marriage and career outcomes, 7, 172– 74; methodology for study of, 165–69; productivity of, 197, 198; research organizational structure and, 7–8; scientific dissemination and, 198–201; tenure-track jobs and, 7, 172–74, 173t, 185–87t, 188–90t, 190–92t Women in Engineering and Computer Science (WECS), 25–26 Women in Engineering Program (WENG), 25