African Successes, Volume II: Human Capital 9780226316192

Citation preview

African Successes, Volume II

National Bureau of Economic Research Conference Report

African Successes, Volume II Human Capital

Edited by

Sebastian Edwards, Simon Johnson, and David N. Weil

The University of Chicago Press Chicago and London

Sebastian Edwards is the Henry Ford II Professor of International Economics in the Anderson Graduate School of Management at the University of California, Los Angeles. Simon Johnson is the Ronald A. Kurtz (1954) Professor of Entrepreneurship and professor of global economics and management at the MIT Sloan School of Management. David N. Weil is the James and Merryl Tisch Professor of Economics at Brown University. All three editors are research associates of the NBER.

The University of Chicago Press, Chicago 60637 The University of Chicago Press, Ltd., London © 2016 by the National Bureau of Economic Research Printed in the United States of America 25 24 23 22 21 20 19 18 17 16 1 2 3 4 5 ISBN-13: 978-0-226-31605-5 (cloth) ISBN-13: 978-0-226-31619-2 (e-book) DOI: 10.7208/chicago/9780226316192.001.0001 Library of Congress Cataloging-in-Publication Data Names: Edwards, Sebastian, 1953– editor. | Johnson, Simon, 1963– editor. | Weil, David N., editor. Title: African successes : government and institutions / edited by Sebastian Edwards, Simon Johnson, and David N. Weil. Other titles: National Bureau of Economic Research conference report. Description: Chicago ; London : Chicago and London, 2016– | Series: National Bureau of Economic Research conference report Identifiers: LCCN 2015050080 | ISBN 9780226316222 (cloth : alk. paper) | ISBN 9780226316369 (e-book) | ISBN 9780226316055 (cloth : alk. paper) | ISBN 9780226316192 (e-book) | ISBN 9780226315720 (cloth : alk. paper) | ISBN 9780226315867 (e-book) | ISBN 9780226315553 (cloth : alk. paper) | ISBN 9780226315690 (e-book) Subjects: LCSH: Economic development—Africa. | Africa—Economic conditions—21st century. Classification: LCC HC800 .A56873 2016 | DDC 330.96–dc23 LC record available at http://lccn.loc.gov/2015050080 ♾ This paper meets the requirements of ANSI/NISO Z39.48–1992 (Permanence of Paper).

National Bureau of Economic Research Officers Martin B. Zimmerman, chairman Karen N. Horn, vice chairman James M. Poterba, president and chief executive officer Robert Mednick, treasurer

Kelly Horak, controller and assistant corporate secretary Alterra Milone, corporate secretary Denis Healy, assistant corporate secretary

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

Mohamed El-Erian Linda Ewing Jacob A. Frenkel Judith M. Gueron Robert S. Hamada Peter Blair Henry Karen N. Horn John Lipsky Laurence H. Meyer

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

Directors by University Appointment Timothy Bresnahan, Stanford Pierre-André Chiappori, Columbia Alan V. Deardorff, Michigan Ray C. Fair, Yale Edward Foster, Minnesota John P. Gould, Chicago Mark Grinblatt, California, Los Angeles Bruce Hansen, Wisconsin–Madison

Benjamin Hermalin, California, Berkeley Marjorie B. McElroy, Duke Joel Mokyr, Northwestern Andrew Postlewaite, Pennsylvania Cecilia Rouse, Princeton Richard L. Schmalensee, Massachusetts Institute of Technology David B. Yoffie, Harvard

Directors by Appointment of Other Organizations Jean-Paul Chavas, Agricultural and Applied Economics Association Martin J. Gruber, American Finance Association Arthur Kennickell, American Statistical Association Jack Kleinhenz, National Association for Business Economics William W. Lewis, Committee for Economic Development Robert Mednick, American Institute of Certified Public Accountants

Alan L. Olmstead, Economic History Association Peter L. Rousseau, American Economic Association Gregor W. Smith, Canadian Economics Association William Spriggs, American Federation of Labor and Congress of Industrial Organizations Bart van Ark, The Conference Board

Directors Emeriti George Akerlof Jagdish Bhagwati Carl F. Christ Franklin Fisher

George Hatsopoulos Saul H. Hymans Rudolph A. Oswald Peter G. Peterson

John J. Siegfried Craig Swan

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 co-authors 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

Series Introduction Sebastian Edwards, Simon Johnson, and David N. Weil

1

Volume Introduction Sebastian Edwards, Simon Johnson, and David N. Weil

5

I. Health 1. Evaluating the Effects of Large-Scale Health Interventions in Developing Countries: The Zambian Malaria Initiative Nava Ashraf, Günther Fink, and David N. Weil 2. Prevention of Mother-to-Child Transmission of HIV and Reproductive Behavior in Zambia Nicholas Wilson 3. Stimulating Demand for AIDS Prevention: Lessons from the RESPECT Trial Damien de Walque, William H. Dow, Carol Medlin, and Rose Nathan

13

59

85

4. Alternative Cash Transfer Delivery Mechanisms: Impacts on Routine Preventative Health Clinic Visits in Burkina Faso 113 Richard Akresh, Damien de Walque, and Harounan Kazianga

vii

viii

Contents

II. Gender Issues 5. Girl Power: Cash Transfers and Adolescent Welfare: Evidence from a Cluster-Randomized Experiment in Malawi Sarah Baird, Ephraim Chirwa, Jacobus de Hoop, and Berk Özler 6. Comparing Economic and Social Interventions to Reduce Intimate Partner Violence: Evidence from Central and Southern Africa Radha Iyengar and Giulia Ferrari 7. Family Ties, Inheritance Rights, and Successful Poverty Alleviation: Evidence from Ghana Edward Kutsoati and Randall Morck

139

165

215

III. Education 8. The Surprisingly Dire Situation of Children’s Education in Rural West Africa: Results from the CREO Study in Guinea-Bissau (Comprehensive Review of Education Outcomes) 255 Peter Boone, Ila Fazzio, Kameshwari Jandhyala, Chitra Jayanty, Gangadhar Jayanty, Simon Johnson, Vimala Ramachandran, Filipa Silva, and Zhaoguo Zhan 9. Success in Entrepreneurship: Doing the Math Michael Kremer, Jonathan Robinson, and Olga Rostapshova 10. The Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa: Some Computations Using Data from Ghana Yaw Nyarko Contributors Author Index Subject Index

281

305 347 351 357

Series Introduction Sebastian Edwards, Simon Johnson, and David N. Weil

In the 1950s and early 1960s, there was a great deal of optimism about the prospects for economic development in sub-Saharan Africa. By early in the twenty-first century, the prevailing consensus had become much more negative—and for good reason. Decades of civil war, repeated state failure, corruption, and disappointing private sector performance characterized much of Africa’s postindependence experience. A wave of prominent papers in the economic literature tried to dig more deeply into the causes of these problems, with some scholars putting them in a broader comparative perspective and others focusing on the specifics of the African experience. There was no shortage of deep causes suggested as explanations for repeated African disappointments, including weak rule of law, a lack of democracy, colonial inheritance, the impact of the slave trade, the burden of tropical disease, some form of “resource curse,” and ethno-linguistic divisions among the population. The NBER Africa project, conceived in the middle of the first decade of the twenty-first century, took a different approach. Rather than trying to understand the causes of underperformance in Africa, we decided to focus on finding and understanding more positive aspects of what was happening Sebastian Edwards is the Henry Ford II Professor of International Economics at the University of California, Los Angeles, and a research associate of the National Bureau of Economic Research. Simon Johnson is the Ronald A. Kurtz (1954) Professor of Entrepreneurship and Professor of Global Economics and Management at the MIT Sloan School of Management and a research associate of the National Bureau of Economic Research. David N. Weil is the James and Merryl Tisch Professor of Economics at Brown University and a research associate of the National Bureau of Economic Research. For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see http://www.nber.org/chapters/c13651.ack.

1

2

Sebastian Edwards, Simon Johnson, and David N. Weil

on that continent south of the Sahara, along several different dimensions studied by economists. Our timing proved good for three reasons. First, scholars were turning their attention increasingly to the details of what was working well or at least better with regard to African development. This new wave of research involved working with or, in many cases creating, data sets suited to careful analysis—including sophisticated use of econometrics. We are fortunate to have involved and supported many leading empirical researchers as they broke new ground in multiple directions. Most likely some of this work would have happened in any case, but at the very least we can claim to have provided an appropriate catalyst for many projects. Second, Africa is changing—and the prospects for nearly a billion people are looking up. A decade or more of relatively good performance does not a miracle make, but across a wide range of countries there has been better economic growth, considerable progress with improving public health and other social indicators, and a range of attempts to improve the performance of the state. Many of the chapters in this series of four volumes address some aspect of this profound and important transformation. Third, the study of Africa is becoming much more integrated with the rest of economics. Just as happened earlier with research on Japan and China, a line of work that previously might have been seen as of purely regional interest can now appear in top journals. The broader trend, of course, has been the resurrection of development economics as a vibrant field. But also encouraging—and a central part of the NBER Africa project—has been the increase in interest among nonspecialists regarding what is actually happening on this dynamic and fascinating continent. Volume I in this series focuses on the most basic building blocks of economic development, including the rule of law, civil conflict, and institutions more broadly. Volume II contains chapters on human capital development in Africa, including some important work on public health improvements, but also research into education on the continent. Volume III examines whether Africa can glean any advantage from being relatively late to economic development, including being able to bypass some large investments (for example, in fixed-line telephones or in traditional branch banking) or being able to learn from others (for example, in the development of an export sector). Volume IV looks at whether recent growth can be sustained, in terms of agricultural development and more broadly. The research presented in these volumes covers a wide territory, in geographic and intellectual terms. However, our project was never intended to be comprehensive. Rather we attempted to act as a catalyst for rigorous and innovative thinking applied to recent African development. The work published here far exceeds our earliest expectations—a reflection, primarily, of how many serious scholars are now deeply engaged with these issues.

Series Introduction

3

It has been a great honor for us, as program directors and now as editors of these volumes, to participate in and help to facilitate this surge in serious interest. In terms of making this possible, we must thank, first and foremost, the Bill & Melinda Gates Foundation for their willingness to encourage economists in this endeavor. We have worked with a range of officials at the Foundation over the past decade; their support has been unstinting and their perspective on development is always refreshing. We would specifically like to thank Sara Sievers for many of our early interactions and Oliver Babson for his subsequent engagement. We also greatly appreciate the ideas and energy of Geoffrey Lamb, Kim Hamilton, Negar Akhavi, Adam Gerstenmier, and Mumukshu Patel. We must also recognize the founding insight and energy of Martin Feldstein, who emphasized, among many other relevant and helpful points, that there is an important link between raising the quality of economic research on an issue and improving the odds that policy discussions will become more informed. Since becoming president of the NBER in 2008, James Poterba has also provided us with great leadership and encouragement—including sage advice at every stage of the process. Our four published volumes present results from thirty-nine research teams, with the findings organized along thematic lines. Of the 100 authors whose names appear on one or more of the chapters in our collection, nineteen are from Africa and thirty-two are affiliated with the NBER as either Faculty Research Fellows or research associates. They range from experienced specialists on African development to scholars who had never previously worked on the continent; in seniority they ran the gamut from newly appointed assistant professors to distinguished, established thought leaders. Every project included a significant on-the-ground component. Some research teams combined money we provided with other funds, for example, to support the collection of very large field data sets. In other cases, funding from the NBER made it possible for researchers to interview policymakers, business people (both large and small), farmers, health workers, or others engaged in the process of economic development. Results from the research project were presented at a series of conferences in the United States and in Africa. Given the large size of our project and our explicit goal of including scholars who had not previously been working on Africa, we began with a background conference in February 2008, at the NBER in Cambridge, MA, bringing together researchers and policymakers from the United States, Africa, and Europe. The conference featured presentations on current research and an overview of available data, aimed especially at scholars who were new to the area. Our next meeting, also in Cambridge, MA, in February 2009 was a preconference at which preliminary findings from ten papers were presented. This was followed by a conference in December of the same year, again in

4

Sebastian Edwards, Simon Johnson, and David N. Weil

Cambridge, MA, that took a hybrid form: final versions of five projects were presented, as were preconference versions of another twelve projects. This was followed in July of 2010 by a conference in Accra, Ghana, held in cooperation with the Institute of Statistical, Social, and Economic Research (ISSER). Our final research conference was held in Zanzibar, Tanzania, in August 2011, in cooperation with the Bank of Tanzania. One goal of our project from inception, with full support from the Gates Foundation, was to help connect economic research with the African policy community. We arranged some interactions along these lines throughout our project, but we were fortunate to be able to make a special effort at the end, with a meeting devoted specifically to the Next Macroeconomic Challenges in Africa, held again in Zanzibar, in December 2012, also in cooperation with the Bank of Tanzania. There are many people to thank for the successful completion of this project. Benno Ndulu, governor of the Bank of Tanzania, provided wise guidance throughout the process and particularly helped us to focus on narrowing the gap between research and policy, including our two fruitful meetings in Zanzibar. His team at the Bank of Tanzania was most helpful in many ways, and we have special thanks for Patricia Mlozi, Mechtilda Mugo, Msafiri Nampesya, and Pamella Lowassa-Solomon. Our conference in Accra benefited greatly from the engagement and support provided by Ernest Aryeetey, Kwesi Botchwey, and Robert Osei. And we had great cooperation in many ways across the entire project with African Economic Research Consortium (AERC); thanks to Olu Ajakaiye, William Lyakurwa, and Lemma Senbet for making this possible. We would like to thank everyone who attended our various conferences and who worked hard on all dimensions of these research projects. We are also most grateful for all the inputs received from members of the Advisory Committee, formed especially for this project: Robert Bates, Paul Collier, Martin Feldstein, Benno Ndulu, Franklyn Prendergast, Antoinette Sayeh, Nicholas Stern, and John Taylor. The entire project ran smoothly due to the hard work, precision, and attention to detail of Elisa Pepe at the NBER. Elisa was with us from beginning to end, and words cannot sufficiently express our gratitude for all she has done. Others at the NBER provided outstanding help on a wide variety of tasks. Our meetings in the United States and in Africa were organized with exemplary competence by Carl Beck of the NBER’s conference department. For all their help in the production, management, and dissemination of research, we would also like to thank Alex Aminoff, Laura Bethard, Daniel Feenberg, Helena Fitz-Patrick, Wayne Gray, Steve Harriman, and Alterra Milone. And for their assistance with the data portal, we are grateful to Binh Thanh Le, Dimitry Legagneur, and C. Michelle Tejada.

Volume Introduction Sebastian Edwards, Simon Johnson, and David N. Weil

Volume I in this series deals with some fundamental issues for African development: Will there be social peace? Can government create (or at least not destroy) legitimate political, economic, and social institutions? Is international aid effective? How will the private sector develop? The issues in volume II are just as important for economic growth, and arguably even more important for human development. Will there be major improvements in public health, reaching at least some minimally acceptable level? Will girls and women be protected and find opportunities as societies change? Can educational attainment be raised to a level commensurate with hoped-for economic growth? The selection of chapters in this volume was shaped by several different considerations. Part of our mandate, as reflected in the titles of the four volumes in this series, was to look for cases where things were working well in Africa. But we did not take this as our only criterion, and we certainly did not ask our researchers to spin their results to come out a certain way. Some of the chapters described below look at interventions where the results are mixed, or at areas where it is too soon to tell if progress is likely. In one case (chapter 8), the chapter reports nothing but bad news. Our overriding goal was to support good research—and good researchers—addressing Sebastian Edwards is the Henry Ford II Professor of International Economics at the University of California, Los Angeles, and a research associate of the National Bureau of Economic Research. Simon Johnson is the Ronald A. Kurtz (1954) Professor of Entrepreneurship and Professor of Global Economics and Management at the MIT Sloan School of Management and a research associate of the National Bureau of Economic Research. David N. Weil is the James and Merryl Tisch Professor of Economics at Brown University and a research associate of the National Bureau of Economic Research. For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see http://www.nber.org/chapters/c13385.ack.

5

6

Sebastian Edwards, Simon Johnson, and David N. Weil

important topics. The studies in this book fit nicely within the great NBER tradition of uncovering new facts and helping to shift our understanding of how economies actually function. Health According to United Nations (UN) estimates, the population of subSaharan Africa was around 900 million in 2013. This will rise to around 2.4 billion by 2050 and, in the UN projections, to close to 4 billion by 2100.1 Long-run projections should always be viewed with some skepticism, but it is striking that in the years after 2050, the UN’s demographers think that African population growth will be greater (in percentage terms and in terms of numbers) than anywhere else in the world.2 Whatever you think of the specific forecasts, as population growth in Asia and elsewhere slows in coming decades, the number of people living in Africa south of the Sahara seems likely to still be increasing at a rapid pace. Will these Africans be healthy, experiencing long life expectancies and relatively low morbidity rates? Or will they struggle to gain access to decent health care? Tropical diseases, most notably malaria, have long burdened African societies; in the last several decades Africa has also become the focus of the global HIV epidemic. Chapter 1, “Evaluating the Effects of Large-Scale Health Interventions in Developing Countries: The Zambian Malaria Initiative,” by Nava Ashraf, Günther Fink, and David N. Weil, suggests some grounds for optimism. A broad approach to prevention and treatment of malaria appears to have made real headway over the past decade, with total deaths from the disease falling by half over an eight-year period. The Zambian experience may offer lessons in terms of what works best, for example, by comparing the provision of insecticide-treated mosquito bed nets and indoor residual spraying. At least along this dimension of health, the improvements are real and likely to be lasting. Unfortunately, HIV-AIDS could prove harder to bring under control, in part because it involves difficult taboo topics and, arguably, requires changing attitudes and behavior as well as government policy. At the same time, chapter 2 presents evidence—also from Zambia—that it is possible to reduce the transmission of the HIV virus from mother to child. Nicholas Wilson’s “Prevention of Mother-to-Child Transmission of HIV and Reproductive Behavior in Zambia” explains the importance of this problem. He also lays 1. These numbers are from the UN’s 2012 demographic projections, for example, as presented in their Wall Chart: http://www.un.org/en/development/desa/population/publications /pdf/trends/WPP2012_Wallchart.pdf. 2. See, for example, the UN’s World Population Prospects, The 2012 Revision: Volume I: Comprehensive Tables, http://esa.un.org/wpp/Documentation/pdf/WPP2012_Volume-I_Com prehensive-Tables.pdf.

Volume Introduction

7

out some ways in which progress has been made, while not underestimating the serious difficulties that remain to be overcome. In keeping with our project’s attempt to encourage scholars to push hard at the research frontier that is relevant for practical problems in Africa, Wilson also emphasizes what can and cannot yet be established beyond a reasonable doubt. There is a great deal here to attract the attention of researchers and, hopefully, to inspire policymakers—as well as foundations and other sources of potential funding. But we may also need to consider different and somewhat more controversial approaches that are complementary to traditional public health measures. In chapter 3, “Stimulating Demand for AIDS Prevention: Lessons from the RESPECT Trial,” Damien de Walque, William H. Dow, Carol Medlin, and Rose Nathan report on their efforts in Tanzania to develop effective ways to change sexual behavior. Conditional cash transfers have proved effective in antipoverty programs elsewhere, and the authors apply a version of this philosophy to reduce the kind of risk taking that contributes to the spread of disease. The authors make a strong case for further study of incentive-based approaches to public health. This is exactly the approach reported in chapter 4, “Alternative Cash Transfer Delivery Mechanisms: Impacts on Routine Preventative Health Clinic Visits in Burkina Faso,” by Richard Akresh, Damien de Walque, and Harounan Kazianga. The authors explore the effects of changing the degree of conditionality regarding transfers, as well as the household member (mother vs. father) who receives transfer. They find that the former has a large effect on outcomes, but not the latter. The work contributes to a better understanding of household decision making as well as improving the design of health interventions. Gender Issues The empowerment of women is in itself an aspect of human development, but it is also an important input into several dimensions of human capital accumulation, including child health, education, and the spread of HIV. For this reason, we have grouped together several chapters that specifically address the role of women in Africa. Continuing in the vein of program design explored in chapters 3 and 4 is “Girl Power: Cash Transfers and Adolescent Welfare. Evidence from a Cluster-Randomized Experiment in Malawi” (chapter 5). Here Sarah Baird, Ephraim Chirwa, Jacobus de Hoop, and Berk Özler examine how best to empower and otherwise help adolescent girls. The authors examine a range of important outcome variables, including access to financial resources, improved schooling outcomes, decreased teen pregnancies and early marriages, and better health. They pay particular attention to the status of these young women, how they are treated, and the extent to which they have access

8

Sebastian Edwards, Simon Johnson, and David N. Weil

to resources within their households. The bottom line is that well-designed programs can make a significant difference at the individual level. In chapter 6, Radha Iyengar and Giulia Ferrari examine whether microfinance and associated counseling can also help women. In “Comparing Economic and Social Interventions to Reduce Intimate Partner Violence: Evidence from Central and Southern Africa,” the authors flag the importance of finding ways to reduce domestic violence. This is a difficult issue on which to make progress, but the authors appear to be on their way to some important findings. We are very pleased to have been able to support the development of this potentially high-impact line of research. The definite gains available through well-designed programs, particularly those focused around health and microfinance, stand in contrast to frustrations that often arise from attempts to simply change the law in ways that would help women and children. Edward Kutsoati and Randall Morck report (in chapter 7) that despite seemingly plausible changes in inheritance law, conventional practices with regard to widows and children can still predominate. “Family Ties, Inheritance Rights, and Successful Poverty Alleviation: Evidence from Ghana” carefully assesses whether formal law matters much in this important area. Traditional norms, which are often complex, can be hard to change. Education Education is the most traditionally measured form of human capital, and three of the chapters in this volume focus on it directly. Perhaps the most sobering chapter in this volume and arising from our broader research project is chapter 8, “The Surprisingly Dire Situation of Children’s Education in Rural West Africa,” by Peter Boone, Ila Fazzio, Kameshwari Jandhyala, Chitra Jayanty, Gangadhar Jayanty, Simon Johnson, Vimala Ramachandrin, Filipa Silva, and Zhaoguo Zhan. The authors implemented a baseline study of education in Guinea-Bissau, one of the poorest countries in the world. The goal was to find pockets of success and to use careful observation of those experiences to design an intervention that would aim to make formal education more effective. Shockingly, the authors found a pattern of literacy and numeracy that is uniformly bad. There simply is no success story for this issue in this country. Undaunted, this research team is nevertheless pressing ahead to assess some specific changes and forms of outside support that could make a difference. The baseline is well established and quite dismal. The question now is: What will make a difference? Education matters for development—in Africa, as elsewhere—and this is evident both from very specific, focused data, as well as from broader assessments. In chapter 9, “Success in Entrepreneurship: Doing the Math,” Michael Kremer, Jonathan Robinson, and Olga Rostapshova assess the

Volume Introduction

9

math skills of small-scale retail entrepreneurs in western Kenya. Even a little bit of math—elementary or perhaps middle-school level—makes a big difference, in part because it leads to better management of the business. It is hard to grow without profits, and it is hard to make money without at least a reasonable amount of elementary formal education. Chapter 10, “The Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa: Some Computations Using Data from Ghana,” should be read alongside the two previous chapters. Chapters 8 and 9 stress the importance of primary education; chapter 10 is about the social benefits that can accrue from building up a university system. Yaw Nyarko takes into account the flows of highly educated people out of the country—that is, the “brain drain”—but also the remittances they send back to the country. There are some fascinating questions here, including Professor Nyarko’s central issue: Is the brain drain all bad, or could it be considered part of the solution—at least along the current plausible path of development for many African countries? Conclusion Africa’s human capital deficit, measured in terms of either health or schooling, is daunting. The chapters in this volume explore the importance of human capital for growth more generally, as well as shining light on what leads to successful human capital accumulation. Africa’s unique health environment, characterized by both historical tropical diseases and the new scourge of HIV, poses particular challenges. And because state action is often required for human capital development, dysfunctional governments can cause great damage in this arena. At the same time, as several of the chapters in this volume show, relatively new techniques of targeted transfers and other forms of individual-level incentives demonstrate great promise in terms of encouraging individuals to make greater investments in their own—and their family’s—health and education. Subsequent volumes in this series offer complementary insights into African development. Volume III focuses on specific aspects of modernization, including finance and its interaction with mobile telephony. Volume IV examines whether African growth will prove sustainable.

1

Evaluating the Effects of Large-Scale Health Interventions in Developing Countries The Zambian Malaria Initiative Nava Ashraf, Günther Fink, and David N. Weil

1.1

Introduction

The National Malaria Control Program (NMCP) of Zambia is one of Africa’s largest malaria prevention and treatment initiatives. In 2005, the NMCP set the goal of achieving a 75 percent reduction in malaria incidence and a 20 percent decrease in under- five mortality within five years through a combination of insecticide- treated nets, indoor residual spraying, and deployment of rapid diagnostic tests and front- line combination therapy drugs (Government of Zambia, Ministry of Health 2005, 2006). The total 2008 malaria prevention and control budget, including in-kind contributions, was estimated at $59.8 million,1 including significant aid from the Global Fund to Fight HIV/AIDS, Tuberculosis and Malaria, the United States Agency for International Development (USAID), the World Bank, the World Health Organization (WHO), and the Bill and Melinda Gates Nava Ashraf is associate professor in the Negotiations, Organizations, and Markets Unit at Harvard Business School and a faculty research fellow of the National Bureau of Economic Research. Günther Fink is associate professor of international health economics at the Harvard School of Public Health. David N. Weil is the James and Merryl Tisch Professor of Economics at Brown University and a research associate of the National Bureau of Economic Research. We are extremely grateful to David Fiocco and Richard Sedlmayr, who did all the field work required to assemble the data. We are also grateful to Dr. Elizabeth Chizema, Dr. Pascalina Chanda, Henry Kansembe, and Dr. Victor Mukonka of the Ministry of Health, to Busiku Hamainza and the National Malaria Control Programme as a whole, and to John Miller of Malaria Control and Evaluation Partnership Africa (MACEPA). Primary funding for this project came from the NBER Project on African Development Successes. Additional funding for data workshops in Zambia was provided by MACEPA, Malaria Consortium, World Bank, and the Zambia Ministry of Health. For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see http://www.nber .org/chapters/c13371.ack. 1. Government of Zambia, Ministry of Health (2008). All money figures are in US dollars.

13

14

Nava Ashraf, Günther Fink, and David N. Weil

Foundation (through the Malaria Control and Evaluation Partnership in Africa [MACEPA]). Figure 1.1 tells the story of the program’s success according to data in Zambia’s national health statistics system, the Health Management Information System (HMIS), which we discuss in detail in the next section. The annual number of malaria deaths in the country decreased by at least half over the period 2000– 2008, during which population rose by 30 percent, implying a reduction in the death rate of over 60 percent. As will be seen below, the number of inpatient visits for malaria declined by a comparable magnitude, implying a reduction in morbidity as well as mortality. Evidence from the 2001 and 2007 waves of the Demographic and Health Survey confirms the picture painted by the HMIS. The percentage of children under five reported with a fever over the two weeks preceding the interview dropped from 44.6 percent in 2001 to 17.9 percent in 2007, a reduction of close to 60 percent. (As a comparison, the fraction of children suffering from diarrhea fell by only a quarter, from 41.9 percent to 31.8 percent, over the same time period.) The progress made in all- cause child mortality between the two surveys is remarkable: The under- five mortality rate fell from 168 per thousand live births in 2001 to 119 in 2007. As discussed below, the latter figure may not even reflect the full mortality reduction to date. This decline was not solely the result of the malaria initiative, however, since other health campaigns were taking place at the same time. There are different ways to quantify the magnitude of Zambia’s recent success in health improvement. The reduction in under- five mortality represents approximately 25,000 children’s lives saved per year. To compare the mortality improvement with more “economic” outcomes, we can do a back- of-the- envelope calculation using the Human Development Index (HDI), which weighs economic and noneconomic characteristics into an overall measure of quality of life. Using the formula for the HDI, one can calculate the amount of income growth that would be equivalent to a particular rise in life expectancy at birth. A conservative estimate, using just the data on under- five mortality, is that life expectancy at birth in Zambia rose by 2.25 years over the period 2000– 2007.2 Plugging this into the HDI formula shows that an equivalent rise in HDI would be achieved if income per capita grew by 25 percent.3 In the research program of which this chapter is a part, we study the Zambian malaria initiative with three broad goals. First, we want to systematize and improve the quality of available data on both inputs to malaria control 2. This is based on multiplying a one in twenty reduction in under- five mortality by 2000 life expectancy at birth, which was approximately forty- five years. 3. The HDI is the sum of three terms, two of which are (e0 – 25)/60 and ln(y)/[ln(40,000) – ln(100)], where e0 is life expectancy at birth and y is real GDP per capita. The change in income that has the same impact on HDI as a particular change in life expectancy is given by the equation ∆ln(y) = [(ln(40,000) – ln(100)]/60 × ∆e0.

Effects of Large-Scale Health Interventions in Developing Countries

Fig. 1.1

15

Malaria deaths in HMIS

in Zambia and the outcomes of the malaria control program. As will be seen below, much of the available data were not easily interpretable before we began our efforts, so that the review and consolidation of existing data sources became a substantial part of this research endeavor. The second goal of the project is to use the available data from Zambia to examine the causal relationship between inputs into malaria control and health outcomes. Much of the monitoring of the campaign’s progress has focused solely on the input and implementation side, with outcome measures such as the number of bed nets distributed or houses sprayed. Jointly analyzing data on inputs and health outcomes allows for some measurement of how well the program is doing in achieving its stated overall health goal, and possibly also for inference about cost effectiveness. Finally, our third goal (which we do not advance in the current chapter) is to use the experience of the current malaria campaign in Zambia to understand the economic effects of malaria and of its control. The rest of this chapter is structured as follows. In section 1.2, we discuss our data sources regarding both health outcomes and inputs into malaria control. A good deal of our effort in this project has gone into improving the quality of the data in Zambia’s HMIS, an administrative record system that has the potential to yield richly detailed data, but is also subject to a number

16

Nava Ashraf, Günther Fink, and David N. Weil

of problems. We discuss the HMIS data, how we have tried to help improve it, and the picture of changing malaria impact painted in this improved data. Section 1.3 describes the background to and scope of the current malaria initiative. Section 1.4 presents data on how the different components of the initiative have been rolled out, focusing in variation among different regions. In section 1.5 we attempt to assess the link between data on the rollout—that is, inputs to reduced malaria—and data on improvements in malaria mortality and morbidity, using both the HMIS and the Zambian Demographic and Health Surveys (DHS). Section 1.6 concludes by discussing avenues for future research and also the challenge of sustaining the progress made in Zambia over time. 1.2

Data on Malaria and Other Health Outcomes

In this chapter we rely on two data sources: first, the 2001 and 2007 waves of the Demographic and Health Survey (DHS), and second, the Zambian Health Management Information System (HMIS).4 1.2.1

DHS Data

Our first sources of data are the 2001 and 2007 waves of the Zambian Demographic and Health Survey (ZDHS). For our analysis we use the children recode files, which contain detailed information on all children under age five at the date of the interview as well as a complete list of household and respondent characteristics. We have 13,219 child observations, 6,877 from 2001, and 6,342 from 2007. To link the DHS households to the National Malaria Control Centre (NMCC)’s rollout data, we used ArcGIS to map DHS cluster locations into the respective districts. All seventy- two districts were covered in the 2001 survey; seventy out of seventy- two were covered in 2007. 1.2.2

HMIS Data

The national Health Management Information System (HMIS) captures an impressive amount of routine health data. The database was first introduced in 1995 to collect disease data, service delivery information, and clinic operations reports. It provides data on health outcomes in the vast majority of Zambia’s health facilities. These range in size from hospitals (located in sixty of the seventy- two districts in Zambia) to small health posts staffed by a single nurse or community health worker.5 4. Government of Zambia, Central Statistical Office et al. (2003, 2009). A third potential source of data is the Malaria Indicators Survey (MIS) conducted in 2006 and 2008 (Government of Zambia 2006, 2008). The MIS collects data on household uptake of antimalarial measures (such as use of bed nets and IRS) and health outcomes such as child sickness. However, this data was not available for our use at the time of this writing. 5. Lusaka’s referral hospitals are not included in the HMIS in order to prevent redundancy; theoretically, every person seen in such a hospital should have already been recorded at another hospital or health center.

Effects of Large-Scale Health Interventions in Developing Countries

17

The database has recently been subject to major technical revisions, resulting in a system with a monthly reporting structure and new management software that captures additional indicators, including separate measures for confirmed and unconfirmed malaria cases. Officially, the new HMIS became the primary reporting system beginning January 2009. Most districts transitioned into the new system by reporting in both formats for some part of 2008. Using historical data however, the analysis presented in this chapter is based on files from the “old” HMIS. The following description relates to this original system. Traditionally, health data were passed from each of the reporting health facilities to the respective district office (seventy- two), and then passed on from the district to the regional offices (nine in Zambia). The facilities kept both hard copy patient logs and tally sheets that track clinic functions. At the end of each quarter, facilities reported their summary data to district offices. District health information officers were charged with collecting the reports from each health facility and compiling district reports. They were responsible for ensuring all health facilities had reported, as well as cleaning each facility’s data. Provincial data management specialists collected data from the districts and compiled a provincial data set disaggregated at the facility level. Before forwarding the data on to the national level, the provincial officers once again cleaned the data and verified it for completeness. The provincial data sets were compiled into a unified national data set at the Ministry of Health’s headquarters. This data set remained disaggregated at the facility level. Given the multilevel reporting system, the potential for error was relatively large in the original system, and the quality of health facility data was dependent on staff and their commitment to record keeping. Some health facilities had been meticulous about their record keeping, plotting their health outcomes manually and discussing them in meetings. Others had not been able to keep records in conjunction with patient visits, or had delegated reporting requirements to semiliterate staff. In some cases, tally sheets were only updated on a monthly basis and figures estimated when regular reporting was limited. The transfer from health facility paper records to electronic district summaries was also error prone. The original HMIS database had no built-in consistency checks for disease data, and data were entered only once, without systematic data verification.6 Quarterly updates from districts to provinces were then processed centrally, with only the most recent quarter received from the provinces appended to previous data. However, even after initial submission deadlines, data staff 6. The data verification and cleaning exercises by officers at the district and provincial levels did not follow systematic guidelines, and the quality of these activities depended heavily on their training. In addition, as the task division followed geographical boundaries, workloads were split very unevenly: some district officers were responsible for less than ten facilities, others for more than forty.

18

Nava Ashraf, Günther Fink, and David N. Weil

at subnational levels continued to work on assembling missing data and auditing reports. As a result, changes subsequently made at the district or province level were lost in the national data set. When data were cleaned at the district and provincial levels, revised data sets were not returned to the original sources of the data. Consequently, major inconsistencies existed between the data sets at different levels of the health system. Improvement of the HMIS Ultimately, the HMIS is the only source of consistent national health data. Selected surveys (especially the Demographic and Health Surveys and the Malaria Indicator Surveys) provide more accurate snapshots of malaria levels in select districts at given times, but only the HMIS allows for detailed, localized analysis of malaria levels over time. With this in mind, a major goal of this project was to improve and validate the HMIS data as much as possible. To facilitate this data improvement goal, the project team conceptualized a series of malaria data verification workshops; in collaboration with the National Malaria Control Centre, one HMIS workshop was organized in each of Zambia’s nine provinces. In preparation for these workshops, we systematically scanned HMIS data for outliers and suspicious data points (duplicate figures, significant variance between quarters or years, reporting inconsistencies). District health officials were asked to find missing reports and justify all irregular data ahead of the workshops. This data audit served as one of the final updates of health data from 2000 to 2008, ahead of its transition into the new HMIS system.7 The workshops also provided an opportunity to reconcile incomplete or mismatching data on malaria prevention and treatment campaigns at the district level. Districts were asked to visit each health facility in an attempt to capture all malaria interventions within the facility’s catchment area by government and nongovernment partners. District pharmacists provided distribution data on treatment courses and diagnostic tests at local health facilities, while environmental health technicians provided data on IRS activities. During the workshops, these data were pooled with health outcomes presented and discussed by the district staff. Subsequently, it could be used by the project team to plausibility- check the national data sets. In total, the nine provincial workshops were conducted at a total cost of approximately $200,000. Funding was provided by the Malaria Consortium and MACEPA, as well as the National Malaria Control Centre with support 7. These resulting changes sometimes had major implications for national- level trends. For instance, in the uncorrected HMIS, under- five malaria deaths rose from 3,342 in 2006 to 3,783 in 2007. These figures were reported in the 2008 World Malaria Report, among other places. The workshops showed that the reported increase was due to three districts with erroneous figures. In the fully corrected HMIS, under- five malaria deaths fell from 3,235 in 2006 to 2,680 in 2007.

Effects of Large-Scale Health Interventions in Developing Countries Table 1.1

19

National-level HMIS data sets, before and after malaria data review workshops

Health facilities Quarterly reports Outpatient observations Inpatient observations Death observations

Preworkshop data seta

Additionsb

1,473 43,455 2,066,964 743,650 365,589

28 1,901 76,036 23,041 9,132

Editsc

Postworkshop data setd

32,510 8,638 2,142

1,501 45,356 2,143,000 766,691 374,721

a As of April 16, 2009; data from previous workshops had not yet been received at national level. b Previously missing data entered. c Previously nonmissing data modified. d As of December 31, 2009.

from a World Bank grant. The workshops brought together a diverse array of district health personnel, including district information officers, malaria focal point persons, and district directors of health. Provincial and national health officers also attended each workshop. In total, over 250 attendees were trained in various aspects of data collection, cleaning, and analysis. The workshops resulted in a more complete, correct, and consistent HMIS database for the years 2000 through 2008, as outlined in table 1.1. In terms of completeness, the original national- level HMIS files that we received ahead of the workshops had 3,318 records marked as missing (out of a total of approximately 45,000 facility quarterly reports that should have been present). Over the course of the malaria data verification workshops, 1,901 missing quarterly reports could be retrieved. Also, data were added on twenty- eight facilities that had no HMIS records altogether, bringing the total number of facilities in the system to 1,501. Reviewing apparent data entry errors ahead of the workshops allowed district staff to conduct major revisions. Finally, the most recent versions of the district’s HMIS databases were disseminated to the national level, thereby ensuring consistent insights for both local and central decision making. While the workshops resulted in a higher quality HMIS data set, they also built local capacity through several days of training on how to compile and analyze integrated health databases. Participants were encouraged to challenge each other’s presentations, and lively debates were common. In many cases, these workshops provided the first opportunity for district medical officers to use data to consider the effects of interventions they had implemented and create strategic plans for future interventions. Data analysis capabilities empower local staff and generate decentralized interest in accurate and complete data; this was considered a promising approach by the Ministry of Health and its partners, and follow-up workshops have since been initiated in several provinces.

20

1.2.3

Nava Ashraf, Günther Fink, and David N. Weil

Remaining Issues in the HMIS Data

Diagnosis and Access to Health Facilities The number of malaria cases that are reported in HMIS is potentially biased by several factors that cannot be readily uncovered in the data. The first issue is inconsistencies in diagnostic practices over time. The HMIS is supposed to report all outpatient and inpatient visits, broken down by diagnosis. The problem is that the way in which malaria is diagnosed has changed over the period we examine. Because of both lack of diagnostic technology and differential treatment guidelines, a majority of fevers in Zambia were traditionally classified (and treated) as malaria. As discussed in further detail below, the rollout of the national antimalaria initiative has included the massive deployment of rapid diagnostic tests (RDTs) to health centers, in order to economize on malaria treatment and properly treat nonmalaria fevers. The deployment of RDTs would have led to a reduction in diagnosed cases of malaria even if there was no true change in disease prevalence, as well as an increased diagnosis of other conditions, such as respiratory tract infections. Since the method of diagnosis is not tracked in the original HMIS, quantifying the magnitude of this “diagnostic effect” is not straightforward. Another bias from diagnosis concerns HIV/AIDS. With HIV/AIDS highly stigmatized, anecdotal reports from the clinics suggest that a considerable fraction of HIV deaths were officially classified as malaria mortality. Given the large inflow of foreign funding for HIV treatment over the last five years, HIV mortality has fallen substantially, which may have contributed to the officially recorded declines in malaria mortality. Another issue is that the accessibility of the health system changed over time. This is most dramatically shown in outpatient data for adults in rural facilities. A large increase in this variable is recorded in 2006, which can likely be attributed to the abolition of user fees in rural health clinics in the same year. Prior to 2006, all public- sector patients were expected to pay nominal fees for consultations, diagnostic tests, and some drugs supplied in government clinics. In 2006, all consultation and diagnostic fees were removed for patients living in designated rural areas. It is likely that prior to the elimination of user fees, many would-be outpatients had resorted to selftreatment. To minimize the bias from such contextual factors, we focus on severe cases—specifically, inpatient malaria cases, reported malaria deaths, and reported total deaths. In the HMIS, diagnostic codes for inpatients are entered only at the time of discharge or death. We think that even prior to the abolition of user fees or the advent of RDTs, severe cases would have received inpatient treatment and would have been correctly diagnosed most of the time (prior to the advent of RDTs, malaria diagnoses could be confirmed using microscopy at many clinics and hospitals treating severe malaria).

Effects of Large-Scale Health Interventions in Developing Countries

21

Extent of HMIS Coverage A potential worry about the HMIS is that it only measures cases that enter the government’s health system. To the extent that this is a small fraction of total cases, one would worry both that it is nonrepresentative and, more seriously, that the fraction of cases captured by the HMIS varies over time. In the latter case, the use of the HMIS to measure trends in disease prevalence would be seriously compromised. One way to measure the fraction of cases that the HMIS captures is to focus on deaths, because in this case there is the possibility to use other data sources as a benchmark. Figure 1.2 shows the under- five death rate (annual deaths of children under five in the HMIS per 1,000 children; population data are estimates from Ministry of Health). The data for the full sample show a relatively steady decline from 8.46 in 2000 to 5.05 in 2007, followed by a precipitous decline to 3.26 in 2008. This last figure may represent incomplete reporting for 2008. In the high- quality sample (where we use only consistently reporting facilities as discussed below), the trend from 2000 to 2007 is roughly similar, but the decline in 2008 is smaller, suggesting that increasing underreporting might indeed be an issue for the aggregate 2008 data. We can compare the count of deaths in the HMIS to both the DHS and to other mortality estimates. Most estimates of child mortality are expressed

Fig. 1.2

Deaths per 1,000 children under five in HMIS

22

Nava Ashraf, Günther Fink, and David N. Weil

in terms of deaths before age five per 1,000 live births. To convert the HMIS data to this metric, we simply multiply the number of under- five deaths per 1,000 by five (this is a slight overestimate because the number of children declines with age due to population growth and mortality). Thus in the HMIS data, the child mortality rate was approximately 42.3 per 1,000 children under five in 2000, and 25.3 in 2007; deaths declined by 40.2 percent. In the DHS, deaths per thousand live births were calculated at 168 in 2001– 2002 and 119 in 2007, implying a decline of 29 percent.8 Other published estimates of child mortality show numbers similar to the DHS for 2000, but do not show the same decline as observed in the HMIS and DHS. For example, the World Health Organization (WHO)’s World Health Statistics 2009 lists the under- five mortality rate for Zambia as 178 in 2000 and 170 in 2007. It is not clear what the source for these numbers is, although other WHO publications rely on the DHS estimate for the year 2000. We suspect that the lack of decline in these published figures reflects nonavailability of data, rather than information from an alternative source. Using either the DHS or WHO number for 2000 as a benchmark implies that the HMIS in that year is capturing between a fifth and a quarter of deaths under age five. In terms of the change between 2000 and 2007, the HMIS shows a larger decline than the DHS, although in both cases the magnitude is impressive. Further, it is important to note that the DHS measures under- five mortality by looking at the experience of all children born in the last five years. Thus the number from the 2007 DHS includes children born in 2003, whose early life (when mortality is highest) was not impacted by the malaria initiative. Thus it is expected that the decline in the HMIS would be larger. If instead of using the HMIS death rate in 2007, we use the average death rate from 2003 to 2007, then the decline in under- five mortality in the HMIS is 29 percent, exactly matching the DHS. The low representation of total deaths in the HMIS suggests several possible biases. Presumably the fraction of disease episodes and deaths that reach a clinic or hospital is higher in urban regions than in remote, rural areas.9 Since malaria is higher in rural regions, this suggests that malaria deaths are understated in the HMIS data. Regarding the change in malaria prevalence as a result of the antimalaria initiative, sorting out the bias is more difficult. On the one hand, the intensity of the program was probably highest in the same areas (places with health facilities nearby) that are overrepresented in the HMIS. This might suggest that the HMIS data would overstate the impact of the program. On the other hand, areas that were near clinics, especially cities, were likely relatively better served and had lower 8. Zambia DHS Final Report, March 2009, http://www.measuredhs.com/pubs/pdf/FR211 /FR211[revised- 05– 12– 2009].pdf. 9. In 2005, 36.5 percent of the population lived in urban areas (though presumably a smaller fraction of the under- five population).

Effects of Large-Scale Health Interventions in Developing Countries

Fig. 1.3

23

Deaths by province in DHS versus HMIS

malaria impact prior to the initiative. Thus the HMIS data may be underweighting places where the largest scope for impacting malaria incidence. To further investigate the bias from undercounting in the HMIS, figure 1.3 compares under- five mortality by province in the DHS and HMIS. The data from the DHS are the under- five mortality rates for 2007. From the HMIS, we take under- five deaths in 2007, divided by an estimate of underfive population, then multiplied by five to make a figure consistent with the DHS measure. The extent to which the ratio of HMIS to DHS deaths differs among provinces is quite surprising. At the high end, in the Copperbelt province, HMIS deaths are 32 percent of those implied by the DHS. At the low end, in Lusaka, deaths in the HMIS are only 6 percent of those implied by the DHS. The explanation for the low fraction of deaths recorded in the HMIS in Lusaka is that the city’s two major hospitals are not included in the HMIS since they are considered referral hospitals and not primary health facilities. However, most critical patients in the Lusaka area would end up in one of the referral hospitals. Theoretically, all patients are supposed to be seen at another health center prior to admission to these hospitals, and the hospital is supposed to report back to the referring center with the patient’s outcome for entry into the HMIS. The available data suggests that the final treatment outcomes at the referral centers do not make it back into the referring center’s HMIS records.

24

Nava Ashraf, Günther Fink, and David N. Weil

Fig. 1.4

Mortality changes: HMIS versus DHS

Note: Horizontal axis is change in the under-five mortality rate divided by the 2001 level. Vertical axis is change in HMIS deaths per 1,000 children divided by the 2001 level.

To check whether undercounting in the HMIS might bias conclusions regarding changes in malaria over time, figure 1.4 compares the provincelevel change in under- five mortality in the DHS between 2001 and 2007 with the change in deaths in the HMIS per 1,000 children under five. The fit is relatively good (correlation 0.76). This gives us confidence that changes in the HMIS reflect actual changes in malaria mortality and morbidity. Nonreporting Facilities A final concern with the corrected HMIS is the potential confusion between health facilities that have no cases to report and cases where data are actually missing. Upon conclusion of the malaria data verification workshops, the HMIS had data on 1,501 health facilities reporting at some point between 2000 and 2008. Each quarterly record has number of in- and outpatients under and over age five, as well as the number of deaths for up to fifty- nine diseases—the total number of diseases recorded in the HMIS is seventy- one. About 20 percent of reporting health facilities deal only with

Effects of Large-Scale Health Interventions in Developing Countries

25

outpatients. Many of the remaining facilities have incomplete reporting. To make sure our results are not affected by differences in reporting, we construct a high- quality sample consisting of 253 large hospitals with at least one patient reported in each quarter between 2000 and 2008—a sample of 9,108 quarter- year observations at the hospital level. To the extent that missing entries represent true zeros (no report because there was no malaria), this sample will lead to an underestimate of the true effects, as hospitals with few malaria patients might discontinue reporting malaria due to the observed declines. Table 1.2 shows data on inpatient visits and deaths broken into age groups (under five and over five), separately for malaria and for all nonmalaria conditions. We show the data both for all facilities and for the set of “alwaysreporting” facilities. The table shows that, as discussed below, the decline in malaria prevalence has come at the same time as a dramatic improvement in health more generally, due primarily to a massive scale-up of HIV treatment and child health programs. Another notable feature is that among the group of all facilities, there is an apparently discontinuous drop in both inpatient visits and mortality, both for malaria and other conditions, in the year 2008. Among the alwaysreporting facilities, there is a smaller drop in malaria cases and deaths as well as in nonmalaria conditions. It is possible that this is due to reporting problems in non- always- reporting facilities. There is a particular drop off in the number of facilities providing non- zero reports in the last two quarters of 2008. We believe this is due to both the switch over to the new HMIS and to delays in facility reports reaching Lusaka. For these reasons, we assign little credence to the all- facilities drop in mortality in 2008. In figure 1.5 we look at an alternative to the always- reporting facilities. We construct a chain index by looking for every pair of adjacent years at the full set of facilities that report in both years. The overall story told in this data is not very different from the always- reporting facilities. From 2000 to 2008, under- five malaria deaths fall by 66 percent and deaths of people age five and above fall by 59 percent. The figure also shows a spike in over- five outpatients diagnosed with malaria in 2006, a phenomenon that may be associated with the abolition of user fees, as discussed above. As a final, very conservative way to look at the decline in malaria, we examine the ratio of malaria deaths to nonmalaria deaths in figure 1.6. We do this for different age groups and for both all facilities and the alwaysreporting facilities. If there were no actual improvement in nonmalaria health outcomes, and the decline in nonmalaria mortality in the HMIS reflected only reporting problems, then this measure would nonetheless correctly measure the decline in malaria mortality. As the figure shows, the ratio of malaria to nonmalaria mortality fell by between one- third and one- half over this period.

2000 2001 2002 2003 2004 2005 2006 2007 2008

2000 2001 2002 2003 2004 2005 2006 2007 2008

148,116 183,368 169,333 172,021 137,414 156,309 152,304 157,118 124,813

11,684 12,103 12,090 11,224 9,810 9,641 9,315 9,816 6,423

Nonmalaria inpatient deaths under age five

Nonmalaria inpatients under age five

96,569 117,214 114,759 124,665 99,938 96,014 96,977 83,599 56,693

Malaria inpatients age five and over

225,018 248,121 287,986 286,737 280,107 298,237 291,504 297,832 231,925

Nonmalaria inpatients age five and over

All facilities

5,039 5,598 4,937 4,808 4,056 3,489 3,235 2,684 1,680

134,516 167,814 162,760 167,919 136,623 139,808 141,312 119,618 69,637

Malaria inpatient deaths under age five

All facilities

15,713 16,699 18,913 19,758 20,967 20,709 20,052 17,652 12,887

Nonmalaria inpatient deaths age five and over

3,882 4,352 4,530 4,727 4,352 3,979 3,096 2,356 1,566

Malaria inpatient deaths age five and over

Inpatient and mortality data from the HMIS

Malaria inpatients under age five

Table 1.2

70,757 75,893 67,356 70,220 57,332 68,430 65,545 69,290 68,627

Nonmalaria inpatients under age five

65,457 75,308 71,440 72,718 56,524 56,557 56,251 50,092 32,004

Malaria inpatients under age five 47,818 51,158 48,857 51,079 40,159 38,952 40,056 34,703 26,972

Malaria inpatients age five and over

5,644 5,380 4,872 4,618 3,906 3,842 3,568 4,477 3,269

Nonmalaria inpatient deaths under age five

105,645 109,316 116,814 117,027 111,896 120,392 117,190 119,387 112,440

Nonmalaria inpatients age five and over

Always-reporting facilities

2,223 2,619 2,117 1,919 1,626 1,357 1,295 1,160 861

Malaria inpatient deaths under age five

Always reporting facilities

7,507 7,426 7,538 8,401 8,033 8,064 7,708 6,939 6,193

Nonmalaria inpatient deaths age five and over

1,745 1,885 1,740 1,895 1,572 1,629 1,383 1,070 807

Malaria inpatient deaths age five and over

Fig. 1.5

Malaria cases and deaths, chained index

Source: HMIS.

Fig. 1.6

Ratio of malaria to nonmalaria mortality

Source: HMIS.

28

1.2.4

Nava Ashraf, Günther Fink, and David N. Weil

Seasonality in the HMIS

As a check on the quality of the data in the HMIS, and also to exploit some of its richness, we look at seasonal variation in the malaria incidence. Table 1.3 shows the results of a basic regression with quarterly under- five mortality as dependent variable with quarter and year fixed effects.10 Figure 1.7 plots the coefficients on the quarter dummies for both total mortality and malaria mortality. The seasonal fluctuations are strong, and are consistent with Zambia’s climate, as discussed below. With an average under- five death rate of 1.75 per 1,000, a negative coefficient of – 0.7 in quarter three implies that the death burden in the best quarter is only about 60 percent of the death burden in the worst quarter (quarter 1). These seasonal fluctuations are even more pronounced for malaria. Overall, about half of the seasonal fluctuation in under- five mortality is driven by malaria. In columns (2) and (4) of table 1.3, we divide the sample in the middle and estimate quarter dummies separately for the two halves. Figure 1.8 shows the interacted quarter dummies for malaria mortality while figure 1.9 shows the quarter dummies for all- cause mortality. What we find is that the seasonality of both has gone down. Indeed, the coefficients show that the decline in seasonality in malaria deaths (a change in .227 deaths per thousand in the worst quarter relative to the best) is more than two- thirds of the decline in the seasonality of total deaths (.317 deaths per thousand in the worst quarter relative to the best). Malaria accounted for approximately 29 percent of under- five deaths over the entire sample period, so the large share of malaria in the decline in seasonality is not simply a reflection of the overall decline in the death rate. 1.3

Origins and Scope of the Zambia Malaria Control Program

Zambia is a landlocked country in southern Africa with three distinct seasons: a hot, dry season from late August to October; a warm, rainy season from approximately November to April; and a cool, dry season stretching from May to early August. In the cool season temperatures can be as low as 10 degrees Celsius,11 with both the lack of rain and the cool temperatures hindering mosquito reproduction; as such, reported cases of malaria are the lowest during the third quarter of the year. Traditionally malaria transmission is highest in the first and fourth quarters, peaking in March toward the end of the warm, rainy season. The swampy Luapula Province in Zambia’s north remains the region’s hotspot, though malaria is traditionally endemic throughout the country. 10. Note that this is deaths in a quarter divided by the under- five population, so its mean is one- quarter of the under- five death rate reported in table 1.2 and figure 1.1. 11. Data provided by the Zambian Meteorological Department.

Table 1.3

Seasonality and changes over time Under-five deaths per 1,000 (all causes)

Second quarter Third quarter Fourth quarter

(1)

(2)

(3)

(4)

–0.326*** (0.0486) –0.726*** (0.0488) –0.431*** (0.0488)

–0.324*** (0.0652) –0.584*** (0.0656) –0.318*** (0.0657) 1.186*** (0.0869) –0.00536 (0.0976) –0.317*** (0.0978) –0.251** (0.0978) 0.0697 (0.0726) –0.149** (0.0726) –0.346*** (0.0726) 0.608*** (0.0744) 0.534*** (0.0744) 0.411*** (0.0745) 0.334*** (0.0746)

–0.136*** (0.0202) –0.360*** (0.0203) –0.235*** (0.0203)

–0.111*** (0.0270) –0.258*** (0.0271) –0.152*** (0.0272) 0.522*** (0.0359) –0.0563 (0.0404) –0.227*** (0.0405) –0.185*** (0.0405) 0.0438 (0.0300) –0.0832*** (0.0300) –0.121*** (0.0300) 0.249*** (0.0308) 0.209*** (0.0308) 0.174*** (0.0308) 0.120*** (0.0308)

Pre-2004 Second Q. * pre Third Q. * pre Fourth Q. * pre Year 2001 Year 2002 Year 2003 Year 2004 Year 2005 Year 2006 Year 2007 Year 2008 Constant Observations R-squared

Under-five deaths per 1,000 (malaria)

0.0697 (0.0728) –0.148** (0.0728) –0.346*** (0.0728) –0.733*** (0.0728) –0.808*** (0.0728) –0.931*** (0.0728) –1.008*** (0.0730) –1.344*** (0.0746) 2.584*** (0.0595) 2,562 0.550

1.178*** (0.0668) 2,562 0.554

0.0438 (0.0302) –0.0830*** (0.0303) –0.121*** (0.0302) –0.255*** (0.0302) –0.294*** (0.0302) –0.330*** (0.0303) –0.384*** (0.0303) –0.505*** (0.0310) 0.872*** (0.0247) 2,562 0.417

Source: HMIS. Note: Standard errors in parentheses. All estimates include district fixed effects. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

0.316*** (0.0276) 2,562 0.426

Fig. 1.7

Seasonality of mortality in the HMIS

Source: See table 1.3.

Fig. 1.8

Change in seasonality of malaria mortality

Source: See table 1.3.

Effects of Large-Scale Health Interventions in Developing Countries

Fig. 1.9

31

Change in the seasonality of all-cause mortality

Source: See table 1.3.

As discussed in Packard (2007), the malaria situation in Zambia prior to the current campaign was rather critical from a recent history perspective. Annual malaria incidence rose from 121 per 1,000 in 1976 to 376 per 1,000 in 2000. Among the factors that contributed to this deterioration were Zambia’s role as a front- line state in the struggle against apartheid, international constraints on the use of DDT, and the country’s broader economic decline (gross domestic product [GDP] per capita as purchasing power parity [PPP] rose from $954 at independence in 1964 to $1,235 in 1970 and peaked at $1,474 in 1976. It then declined, reaching a nadir of $829 in 1995 before rising to $946 in 2003).12 Zambia’s malaria control program had relied extensively on indoor residual spraying, but by 1990 spraying had ceased altogether. In addition, resistance to chloroquine started to emerge rapidly across the country. The beginning of the current antimalaria initiative was a result of a confluence of factors both in Zambia and elsewhere in the world. In particular, 12. Penn World Tables, version 6.2. Variable RGPCH in year 2000 constant international dollars. The GDP per capita rose a further 16 percent in total from 2003 to 2007 (World Development Indicators database), on the back of soaring copper prices. The price of copper fell by 60 percent in the year to February 2009, suggesting that Zambia will be particularly hard hit in the current world slowdown.

32

Nava Ashraf, Günther Fink, and David N. Weil

the development of new technologies and a new alignment of priorities within the development community led to a desire to undertake a demonstration case showing the possibility of rapidly scaled-up malaria control. Resources would be concentrated on a single country with the goal of producing a tangible success, which would then serve as a model for neighboring countries. Zambia was chosen as the test case because it was perceived as having the institutional capacity and political will to successfully undertake such a program, and also because its climate made it all the more likely that success could be accomplished. 1.3.1

Elements of the Program

Zambia’s integrated malaria control program is one of the world’s largest national treatment and prevention plans. The program is led by the National Malaria Control Centre (NMCC), a subdivision of the Department of Public Health and Research within Zambia’s Ministry of Health. The national secretariat is responsible for overall program administration throughout the country and disburses funds to districts on a programmatic basis. The NMCC works in partnership with Medical Stores (a quasi-private national distribution program for all drugs and medical supplies in the public sector) to ensure that each district and referral hospital receives adequate supplies of drugs and diagnostic tools. Seventy- two district medical offices (previously “district health offices”) directly implement most national malaria prevention and treatment programs. The districts are grouped into nine provinces, each with a Provincial Health Office responsible for supervising district health programs. The program is funded by a collaboration between the government of Zambia and national as well as international partners. From its own revenues and various support lines, the government of Zambia budgeted $25.4 million for the Department of Public Health’s Malaria Control & Management activities in 2008.13 This amounted to 61 percent of the department’s budget. The total 2008 malaria prevention and control budget, including in-kind contributions from nongovernmental institutions, was estimated at $59.8 million,14 though actual contributions may not have always matched pledges or projections. The largest component of the 2008 budget was the provision of insecticidetreated mosquito nets (ITNs). The $32.0 million allocated to this activity accounted for 54 percent of the NMCP’s overall 2008 budget, although budget allocations did not always match actual expenditure, sometimes due to donors failing to deliver on their commitments. For example, in 2008, the government budgeted for 3.5 million nets but ended up distributing only about one million; only one- sixth of the Ministry of Health funds bud13. Government of Zambia (2008). 14. Government of Zambia, Ministry of Health (2008).

Effects of Large-Scale Health Interventions in Developing Countries

33

geted for nets in Lusaka, Copperbelt, and Central provinces were actually expended. Of all the malaria prevention modalities of the program, bed nets suffer from the greatest problem in terms of a mismatch between distribution and effectiveness. The standard guideline is that every person living in a household not sprayed with indoor residual spraying should sleep under a bed net. However, utilization remains well below the 85 percent target. Some report sleeping under a bed net to be uncomfortably hot or claustrophobic, while others report irritation to the chemical treatments. There are also frequent reports of people not using nets at all, sometimes keeping them packaged as a sign of wealth or using them for other purposes (such as wedding veils and fishing nets), but there has been no systematic study to measure full utilization levels. The other primary preventive intervention, indoor residual spraying (IRS), was carried out in thirty- six of Zambia’s seventy- two districts in 2008,15 targeting primarily urban and periurban areas with relatively high population densities. Its projected cost was $8.6 million in 2008. Treatment is another major component of the national malaria control program. The NMCC’s strategic plan targets achieving prompt and effective case management (PECM), with a goal of ensuring that at least 80 percent of malaria patients receive effective treatment within twenty- four hours of the onset of symptoms. After noting decreasing efficacy of sulphadoxine/ pyrimenthamine (SP) and chloroquine, Zambia became one of the first countries to introduce artemisinin- based combination therapy (specifically artemether plus lumefantrine, with the brand name Coartem®) (Chanda et al. 2007; Zurovac 2007). The ACTs, free in the public sector, became the first- line treatment for all malaria cases during the 2002– 2003 malaria transmission season, but it was not until the 2005– 2006 season that the drug reached all districts in the country. Until 2007, the country faced continuing challenges to retain national stocks. The national supply stabilized in 2007; since that time there have not been complete national stock- out periods, though logistical challenges in distribution to the provinces, districts, and health facilities still remain (Sipilanyambe et al. 2008). The total allocation for PECM in 2008 was $11.1 million, approximately 19 percent of the total program budget. Of this money, $2.6 million was designated for malaria diagnostics, primarily for the purchase of two million rapid diagnostic tests (RDTs). The procurement of 3.8 million courses of Coartem® was the largest component of the case management budget, costing $5.3 million (note that Coartem® is heavily subsidized by its manufacturer, Novartis). In the public sector, national treatment guidelines dictate Coartem® as 15. As described in further detail below, IRS spraying was originally only targeted to a handful of urban areas across the country and only slowly scaled up over time.

34

Nava Ashraf, Günther Fink, and David N. Weil

the first line of treatment, with quinine (available in tablets and intravenously) reserved for those who fail to respond to Coartem®. Chloroquine and Fansidar® (a combination of pyrimethamine and sulphadoxine) are no longer to be used for malaria treatment in the public sector, though Fansidar® remains the indicated preventive treatment given to pregnant mothers. In practice, however, Fansidar® is often used as a first- or second- line treatment for patients, especially for those who have negative malaria test results. When ACTs were introduced in 2003, the high initial cost of the drug prompted an effort to improve diagnostics to control the drug prescriptions. In addition to promoting the increased availability of microscopes and trained laboratory technicians, rapid diagnostic tests (RDTs) were first purchased in 2004 for facilities where microscopes were unavailable. The RDTs are simple testing devices allowing malaria diagnosis by a health worker with limited training in just fifteen minutes. Global production of RDT kits rose from roughly 3 million in the year 2000 to 28 million units in 2005 (Frost and Reich 2008). There is an important interaction between availability of RDTs and use of ACTs. When RDTs are not available, there is a strong tendency for health workers to treat all fevers as malaria (as it was previously suggested by the WHO), and dispense ACTs accordingly (Lubell et al. 2008; Rolland et al. 2006). When RDTs are present, a significant fraction of fevers are determined not to be malaria. Until 2007, the national Integrated Management of Childhood Illnesses (IMCI) guidelines dictated that all children under age five with fevers were to be presumptively treated for malaria. National malaria policy guidelines revised in 2008 state that it is against national policy to dispense Coartem® to any patient weighing more than 5 kg (about three months old) without a confirmed malaria diagnosis through RDT or microscope. However, reaching compliance among clinicians with this remains one of the biggest challenges in the national program. A 2007 study found that of 58.4 percent of patients with a negative blood smear and 35.5 percent of those with a negative RDT result were prescribed an antimalarial drug (Hamer et al. 2007). The 2008 performance assessments at health facilities reveal that some centers are still not using RDTs at all. Unlike Coartem®, which is on a “push” system and comes to the clinic whether they request it or not, RDTs are on a “pull” system and need to be requested. Data compiled at our malaria workshops shows that the availability of RDTs is associated with greatly lower- reported cases of malaria and lower use of ACTs. The wholesale price of RDT is about $.70, and for ACTs is roughly in the same range (Frost and Reich 2008). Even though these prices would seem to say that there is no point in using RDTs before dispensing ACTs in order to save on costs, the public health benefits of not overusing ACTs are enormous, since restricting use will prevent the development of resis-

Effects of Large-Scale Health Interventions in Developing Countries

35

tance. In addition, ruling out malaria allows for better management of negative cases. Smaller components of the 2008 budget included: Information, Education, Communication/Behavior Change Communication (IEC/BCC), Advocacy ($2.2 million), Monitoring and Evaluation (M&E) ($2.2 million), Operations Research ($986,178), Emergency and Epidemic Malaria Preparedness Plan ($982,000), and Entomological Activities ($800,400). Also, $877,507 was allocated for program management at the national secretariat. A 2008 health facility census recorded 1,554 health facilities, including public, private, and church- run health centers. Two- thirds of these facilities are located in Zambia’s rural areas. There are ninety- eight referral hospitals in the country.16 All districts have an active cadre of lay community health workers (CHWs) who supplement permanent health centers. However, the level of care CHWs can provide varies widely by district. A current Home Management of Malaria initiative (HMM) seeks to train CHWs to administer rapid diagnostic tests and provide artemisinin- based combination treatment to patients at their homes; the program is currently being scaled up, but continues to face logistical difficulties in providing drugs and supplies to all trained volunteers. 1.3.2

Nonmalaria Changes in the Zambian Health Environment

As mentioned above, the antimalaria initiative was not the only change in the Zambian health environment over the period we examine. During the period 2003– 2006 there were a series of other health programs occurring, mostly in HIV, tuberculosis, and child health. In addition, the reduction in malaria may have led to decreases in other diseases either through reduced comorbidity or through the freeing up of resources within the health care system. In 2004, funds from the United States President’s Emergency Plan for AIDS Relief (PEPFAR) began arriving in Zambia. According the US Embassy, PEPFAR funds in Zambia were $149 million in 2006, which was spent on prevention, treatment, and care. One targeted area for spending was prevention of mother- to-child transmission, which has the potential to sharply reduce child mortality. While the efforts to combat child mortality by reducing tuberculosis and the transmission of HIV to children have undoubtedly contributed to the declines in child mortality observed, the interactions between these two diseases and malaria are hard to pin down. One of the main effects of childhood exposure to malaria is anemia, which makes children more vulnerable to other diseases such as tuberculosis and diarrhea. The same could clearly be said the other way around: progress made in terms of diarrheal diseases or tuberculosis means healthier children with better immune systems. 16. Government of Zambia, Ministry of Health (2008).

36

Nava Ashraf, Günther Fink, and David N. Weil

Table 1.4

Rollout data at the national level Number of bed nets distributed

Population covered by spraying

RDTs distributed

112,020 557,071 176,082 516,999 1,163,113 2,446,102 964,748

— 324,137 679,582 1,163,802 2,836,778 3,286,514 5,558,822

0 0 0 172,257 25,700 243,600 2,015,500

2002 2003 2004 2005 2006 2007 2008 Source: NMCC.

1.4 1.4.1

Rollout of the Program National Data

Table 1.4 shows the rollout of the three key pieces of the malaria prevention program: ITNs, IRS, and RDTs.17 The table shows the degree to which the program accelerated in 2006 and 2007. Half a million nets were distributed in 2003, but only 176,082 in 2004 (the low number has been attributed to foreign donors failing to provide them that year). The number returned to half a million in 2005, then 1.2 million in 2006 and 2.4 million in 2007.18 As discussed above, the decline in bed net distribution in 2008 was not planned as of the beginning of that year. There were to have been 3.5 million nets distributed. Starting in early 2007, all bed nets distributed were long- lasting insecticide- treated nets (LLITNs), which according to the WHO and NMCC can last for multiple washes and several years (beyond their expected lifetime of three years) without insecticide retreatment (Government of Zambia, Ministry of Health 2008). In the initial design of the malaria initiative, IRS was to be restricted to only areas of very high population density. The scope of IRS was expanded in discrete jumps through 2008 (see figure 1.10). The number of households increased dramatically (almost tenfold) between 2004 and 2008. By the latter year, roughly 43 percent of the population was covered by spraying. Indoor residual spraying is generally conducted at the end of the calendar year, before the onset of the rainy season when malaria is high, and must be reapplied each year. The shift in strategy toward increased reliance on IRS was driven by two 17. Steketee et al. (2008). 18. The NMCC and WHO say any long- lasting insecticide- treated nets are effective if distributed in the last three years. Earlier nets given that were not treated lasted for shorter periods of time, but some retreatment kits were distributed for those nets.

Effects of Large-Scale Health Interventions in Developing Countries

37

factors: first, it is seen as more cost effective. Second, reports of low utilization of bed nets were perceived as a major problem. Indoor residual spraying, once applied, does not require active uptake by the household. The last column of table 1.4 shows the rapid growth of RDT use, which only happened at the very end of our sample period. According to the 2007 DHS, 91 percent of women in urban areas and 84 percent rural took some form of antimalarial treatment during their last pregnancy and 68 percent (61 percent rural) received IPT during an antenatal visit. On average, 38 percent of women used IPT in the 2001 DHS. 1.4.2

Regional Variation in the Rollout

The fundamental strategy of the antimalaria initiative was to push distribution and use of bed nets in high- malaria, rural areas. Initially, IRS was targeted only at urban areas; later the scope of IRS was expanded to include half the districts in the country. ITNs The NMCC goal is to ensure that 100 percent of households in non-IRStargeted areas have at least one mosquito net for every two people, with utilization rates of at least 85 percent. Insecticide- treated mosquito nets are distributed by the NMCC directly, and through a number of partners to specific populations. Programs target mothers and infants, vulnerable populations (orphans, economically deprived populations, HIV+/AIDS patients), and the general public through commercial and subsidized sales, targeted distributions, and free mass community distributions. Table 1.5 compares the DHS and NMCC database at the provincial level. The first two columns are based on NMCC data on the number of nets distributed (total and per capita) by province. The 2001 DHS finished collecting data in May 2002, while the 2007 DHS began collecting data in April of that year. The number of nets in the table is the total from quarter 3 of 2002 through quarter 1 of 2007. The next three columns use data from the 2001 and 2007 waves of the DHS on the fraction of children under five living in a house with at least one bed net. It is not clear whether one would expect a bigger correlation between bed net distribution and the level of bed net ownership in 2007 or the change in bed net ownership between 2001 and 2007. In any case, both correlations are high: .73 and .62, respectively. The last three columns show data on whether the child slept under a bed net. Overall, bed net use more than doubled over this period, reaching 43 percent. Across provinces, the correlation between the change in bed net ownership and the change in bed net use is .72. The correlation between bed net distribution 2002– 2007 and the level of bed net use in 2007 is .63, while the correlation between bed net distribution and the change in bed net use is .57.

188,405 224,425 206,439 417,351 264,591 244,078 290,202 342,484 598,199 2,776,174

Total nets distributed between 2001 and 2007 DHSa

0.15 0.12 0.12 0.43 0.16 0.15 0.39 0.22 0.64 0.26

Nets distributed per person between 2001 and 2007 DHSa,c

b

a

0.23 0.31 0.25 0.36 0.31 0.21 0.36 0.18 0.32 0.28

Percentage of children in household owning at least one net 2001b

ITN distribution, ownership, and use

Source: NMCC. Source: Zambia DHS (2001, 2007). c Source: Central Statistical Office for population.

Central Copperbelt Eastern Luapula Lusaka Northern Northwestern Southern Western Total

Table 1.5

0.68 0.74 0.71 0.86 0.68 0.57 0.73 0.60 0.87 0.72

Percentage of children in households owning at least one net 2007b 0.45 0.43 0.46 0.50 0.37 0.36 0.38 0.42 0.55 0.43

Change in ownership DHS 2001– DHS 2007b 0.13 0.20 0.20 0.31 0.20 0.17 0.27 0.10 0.22 0.20

Percentage of children sleeping under net 2001b

0.37 0.43 0.37 0.74 0.30 0.41 0.43 0.25 0.55 0.43

Percentage of children sleeping under net 2007b

0.25 0.23 0.17 0.42 0.10 0.24 0.16 0.16 0.33 0.23

Change in child net use DHS 2001 to DHS 2007b

Effects of Large-Scale Health Interventions in Developing Countries Table 1.6

Province Central Copperbelt Eastern Luapula Lusaka Northern Northwestern Southern Western

39

Indoor residual spraying coverage (2007) and self-reported coverage Fraction of population officially covered by spraying in 2006

Percentage of children in 2007 DHS living in sprayed households

Urbanization (2000)

0.12 0.63 0.00 0.00 0.73 0.00 0.09 0.16 0.00

0.12 0.41 0.02 0.01 0.29 0.04 0.14 0.13 0.02

0.24 0.78 0.09 0.13 0.82 0.14 0.12 0.21 0.12

Sources: Central Statistical Office, NMCC, DHS (2007).

Indoor Residual Spraying Table 1.6 compares data from the DHS and the NMCC on IRS by district. We use data spraying in 2006 (before the rainy season), which was the last spraying before the 2007 DHS. The table also shows urbanization rates from the 2000 census (Government of Zambia, Central Statistical Office 2003). The nine provinces fall into three groups. Four provinces, all with very low rates of urbanization, had no official spraying, and fewer than 5 percent of households report having received spraying.19 In three provinces, official data show 9– 16 percent of households sprayed, and DHS data report roughly commensurate coverage. Finally, the two most highly urbanized provinces were targeted for intensive spraying: Copperbelt (63 percent of households) and Lusaka (73 percent). In both provinces there is a significant shortfall between official estimates and the DHS. This is particularly severe in Lusaka, where only 29 percent of children in the 2007 DHS reportedly lived in sprayed structures. Figure 1.10 shows the rollout of spraying at the district level. The initial five districts targeted in 2003 were Kabwe, Kitwe, Livingstone, Lusaka, and Ndola. These are urban areas, where little net distribution was happening at the time, and where spraying was considered a relatively economical option due to the relatively high population densities. The first scale-up came in 2005, with spraying extended to Chililabombwe, Chingola, Kalulushi, Luanshya, Mufulira, Chongwe, Kafue, Solwezi, Kazungula, and Mazabuka. In general, spraying was targeted to urban areas where health facilities reported high levels of malaria incidence, 19. There was private- sector spraying being done in some districts that did not overlap with government spraying until potentially 2008. The main private- sector spraying without government involvement was in areas where there were small mines of various minerals (for example, Mumbwa district).

40

Nava Ashraf, Günther Fink, and David N. Weil

Fig. 1.10

Indoor residual spraying distribution by region

so that spraying is often concentrated around the health facilities reporting to the NMCC via the HMIS. 1.5

Assessing the Link from Rollout to Incidence

Ideally, we would like to be able to use data on inputs to better health (that is, preventive measures or disease treatments) and health outcomes

Effects of Large-Scale Health Interventions in Developing Countries

41

in order to learn the efficacy of different inputs. Such an endeavor faces obvious problems with identification. Clearly, the timing and spatial distribution of health interventions are not random. Different modalities are used in different locations because health planners make optimizing choices of what will work best in a given area. Similarly, the provision of resources may respond to perceived needs. For example, extra resources may be pushed to areas where health conditions have deteriorated or are forecast to deteriorate in the future. Finally, the efficiency with which health resources are provided may be correlated with other factors that directly affect health. For example, a district with an especially competent public health staff may be able to obtain additional resources, but may also have had a lower rate of disease incidence even in the absence of these additional resources ( in the field we noted that drug supplies, record keeping and quality, take-up of new treatment guidelines and drugs, and involvement of community health volunteers and neighborhood health committees were all highly dependent on the performance of health staff at all levels of the system, and that the level of performance varied significantly). Given these program allocation mechanisms, we are only able to identify the causal effect of health interventions to the extent that there is some (measureable) randomness to the pattern by which such interventions are applied. If there is statistical power to identify the effects of inputs to better health on health outcomes, it will generally only be in cases where these inputs deviate from the optimal plan or when inputs respond to conditions in some noncontinuous fashion. Since we do not have a formal model of optimizing choice of treatments, it is not possible to formally specify deviation from that optimal plan and use these. In this chapter, we focus on the presentation and discussion of observed correlations between the rollout of different malaria control modalities and available health outcomes. 1.5.1

Bed Nets

In addition to the other statistical problems discussed above, assessment of the link from ITN distribution to health outcomes is complicated by the facts that ITNs have limited effective lifetimes, and that the length of time over which they remain effective has been changing. ITNs in the DHS The first measure of malaria we use in the DHS is a binary indicator that equals one if the child had a fever over the two weeks preceding the interview. The data are pooled from the 2001 and 2007 waves of the DHS; the unit of observation is a child under age five. We include dummies for individual years of age and a dummy for being in the 2007 wave. In addition, all specifications control for district fixed effects, sex, mother’s age, mother’s age squared, mother’s education, mother’s marital status, mother’s employ-

42

Nava Ashraf, Günther Fink, and David N. Weil

ment status, urban, female household head, number of household members, and household assets (electricity, radio, television, refrigerator, and bicycle). In the first two columns of table 1.7, we use indicators for whether a household owns a bed net and for whether the child slept under a bed net the previous night as measures of input to malaria control, respectively. In both cases, we do not think that the estimated coefficient can be interpreted structurally, because both ownership and use of the bed net are affected by disease conditions and other household characteristics related to health outcomes. In the third column, we use bed net distribution per capita as recorded by the NMCC in the district in the period between the two DHS surveys as the independent variable. The variable is zero for all 2001 observations. The coefficient is quite significant, and implies that a distribution of one net per person in the district (100 percent coverage) lowers fever prevalence by about 20 percentage points. In column (4) we instrument for Table 1.7

Bed nets, child fever, and child diarrhea (DHS) Child had fever over last two weeks

Dependent variable Household owns net Child slept under net Bed net distribution pc Child age 1 Child age 2 Child age 3 Child age 4 2nd wave dummy Constant Observations R-squared

(1)

(2)

(3)

–.0213* (0.0111)

Diarrhea (4)

(5)

–0.921*** (0.267)

(6) –0.124 (0.255)

–0.0106 (0.0110)

0.0639*** (0.0122) 0.0142 (0.0132) –0.0728*** (0.0125) –0.121*** (0.0127) –0.243*** (0.0115) 0.546*** (0.0752)

0.0641*** (0.0122) 0.0150 (0.0134) –0.0730*** (0.0127) –0.120*** (0.0129) –0.250*** (0.0109) 0.548*** (0.0759)

–0.209*** (0.0487) 0.0648*** (0.0122) 0.0150 (0.0132) –0.0713*** (0.0125) –0.121*** (0.0127) –0.197*** (0.0169) 0.548*** (0.0749)

11,193 0.129

11,027 0.128

11,193 0.131

0.0672*** (0.0154) 0.00604 (0.0150) –0.0798*** (0.0154) –0.122*** (0.0159) 0.101 (0.107) 0.406*** (0.117) 11,193 –0.513

–0.0286 (0.0703) 0.181*** (0.0252) –0.0307 (0.0252) –0.195*** (0.0247) –0.275*** (0.0216) –0.0736*** (0.0257) 0.735*** (0.138)

0.181*** (0.0315) –0.0319 (0.0297) –0.196*** (0.0262) –0.275*** (0.0236) –0.0335 (0.0967) 0.773*** (0.196)

11,187 0.065

11,187 0.063

Notes: All specifications control for district fixed effects, sex, mother’s age, mother’s age squared, mother’s education, mother’s marital status, mother’s employment status, urban, female household head, number of household members, and household assets (electricity, radio, television, refrigerator, and bicycle). Robust standard errors in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

Effects of Large-Scale Health Interventions in Developing Countries Table 1.8

43

First-stage results for columns (4) and (6) in table 1.7 Dependent variable ITN per capita Child age 1 Child age 2 Child age 3 Child age 4 2nd wave dummy Constant Observations R-squared Kleibergen-Paap F-stat Kleibergen-Paap weak identification p-value

HH owns bed net (1) 0.230*** (0.0662) 0.00210 (0.0104) –0.0106 (0.00925) –0.0107 (0.0104) –0.00518 (0.00934) 0.322*** (0.0231) –0.146* (0.0804) 11,193 0.317 11.65 0.0014

Notes: Includes control for district fixed effects, sex, mother’s age, mother’s age squared, mother’s education, mother’s marital status, mother’s employment status, urban, female household head, number of household members, and household assets (electricity, radio, television, fridge, and bike). Robust standard errors in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

household ownership with per capita distribution in the household’s district (the first stage is reported in table 1.8). In columns (5) and (6) we do a placebo test, using diarrhea in the two weeks preceding the interview as a dependent variable; we use ordinary least squares (OLS) in column (5) and apply the same IV strategy used in column (4) in column (6). Diarrhea is an important health outcome, but reductions in prevalence should not be related to bed net distribution. The bed nets have no significant effect. One concern with the results in table 1.7 is that, as mentioned above, the distribution of nets is not random. Since all the regressions include district fixed effects, the fact that more nets are distributed in districts with permanently higher malaria is not a concern. However if nets are distributed in response to temporary changes in malaria prevalence, our results could be biased. For example, suppose that nets are targeted to districts experiencing temporarily high malaria prevalence. In this case, our estimate would be biased to show nets being more effective than they really are. To attempt to remove this bias, in table 1.9 we control for baseline fever prevalence, that is,

44

Nava Ashraf, Günther Fink, and David N. Weil

regress the change in fever between the two DHS surveys on bed net coverage as well as the fraction of children with fevers in the two weeks prior to the 2001 DHS survey. The coefficient on baseline prevalence is slightly less than one, indicating a moderate degree of convergence over time. The coefficient on bed net distribution per capita falls by roughly half but remains significant while the coefficient on household bed nets becomes insignificant. Finally, in table 1.10 we aggregate to the level of districts and do a regression in first differences. The dependent variable is the change in fever prevalence between the two waves of the DHS. The measure of malaria control is the change in bed net ownership, in column (1), and ITN distribution in the five years prior to the survey in the other columns. The results look very similar to the individual- level regressions. In particular, the effect of ITN rollout falls by about half once we control for baseline fever prevalence, but remains significant. In table 1.11 we look at child mortality data in the DHS for evidence of the effects of ITNs. As in tables 1.7 and 1.9, we report in columns (1) Table 1.9

Control for baseline level in microlevel regression (DHS) Child had fever over last two weeks

Dependent variable Household owns net

(1)

(2)

(3)

–0.0141 (0.0105)

Child slept under net

Diarrhea (4)

–0.695 (0.496)

(6) –0.0429 (0.597)

–0.00428 (0.00895)

Bed net distribution

–0.104*** (0.0364) Child age 1 0.0640*** 0.0644*** 0.0645*** 0.0665*** (0.0121) (0.0121) (0.0121) (0.0148) Child age 2 0.0119 0.0129 0.0124 0.00701 (0.0130) (0.0130) (0.0130) (0.0131) Child age 3 –0.0706*** –0.0704*** –0.0700*** –0.0770*** (0.0109) (0.0109) (0.0110) (0.0121) Child age 4 –0.122*** –0.120*** –0.122*** –0.122*** (0.0115) (0.0118) (0.0116) (0.0126) Baseline fever pre-eval. 0.867*** 0.888*** 0.806*** 0.393 (0.0944) (0.0933) (0.0973) (0.400) Observations R-squared

(5)

11,193 0.136

11,027 0.135

11,193 0.136

11,193 –0.229

–0.00656 (0.0922) 0.181*** 0.181*** (0.0328) (0.0326) –0.0312 –0.0315 (0.0323) (0.0312) –0.194*** –0.195*** (0.0281) (0.0265) –0.276*** –0.276*** (0.0263) (0.0260) 0.168 0.143 (0.227) (0.495) 11,187 0.065

11,187 0.065

Notes: All specifications control for district fixed effects, sex, mother’s age, mother’s age squared, mother’s education, mother’s marital status, mother’s employment status, urban, female household head, number of household members, and household assets (electricity, radio, television, fridge, and bike). Robust standard errors in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

Effects of Large-Scale Health Interventions in Developing Countries Table 1.10

45

District-level differences, DHS Change in fever prevalence

Dependent variable Change in ownership

(1)

(2)

(3)

(4)

(5)

(6)

–0.188** (0.0739)

–0.0778** (0.0357) –0.977*** (0.0983)

–0.0778** (0.0367) –0.976*** (0.0981) –0.0452 (0.0430)

–0.0711* (0.0399) –1.045*** (0.124) –0.0427 (0.0416) –0.0341 (0.0340)

–0.222*** (0.0253)

0.191*** (0.0428)

0.193*** (0.0424)

0.220*** (0.0510)

–0.0761* (0.0402) –1.048*** (0.125) –0.0346 (0.0407) –0.0392 (0.0328) –0.0545 (0.0558) 0.222*** (0.0515)

–0.136 (0.0935)

ITN rollout per capita Baseline fever Change in mother’s education Change in mother working Change in average age Constant Observations R-squared

–0.214*** (0.0446) 70 0.044

70 0.107

70 0.617

70 0.622

70 0.627

70 0.637

Note: Robust standard errors in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

and (2) results in which the independent variable of interest is whether the household owns a bed net and whether children slept under a bed net the previous night. (Note that this latter variable applies to the household in which the child was born, not the child him/herself. We cannot use the childspecific use variable here since it is missing for all the deceased children). We do not interpret these estimates structurally, since there are biases that go in both directions: households that care more (or know more) about their children’s health are more likely to have their children sleep under a bed net, which would yield an overestimate of the true effect. On the other hand, one might also expect some learning from the parents’ side: parents who have lost one child to malaria in the last five years might be more likely to make their remaining children sleep under a net than parents who have not lost a child, which would yield a selection bias going in the opposite direction. In column (3), we use district ITN coverage as the explanatory variable. The coefficient is borderline significant, but of a large magnitude (–.044). It implies that a full coverage with bed nets (one net per capita) in the years prior to the surveys lowers child mortality by 4.4 percentage points from an average baseline level of 12.7 in the 2001 survey. One concern with the mortality regressions is that mortality covers the whole five years prior to the survey, while all the household- level infor-

46

Nava Ashraf, Günther Fink, and David N. Weil

Table 1.11

Bed nets and child mortality Death of child

Dependent variable HH owns bed net

(1)

(2)

Sample restrictions Constant Observations R-squared

(5)

–0.0361* (0.0216) –0.0185*** (0.00489) –0.0265*** (0.00805) last 3 years

–0.0486*** (0.00608)

ITN district coverage

2nd wave

(4)

–0.00968 (0.00690)

Kids in HH slept with bed net

Female

(3)

–0.0199*** (0.00538) –0.0364*** (0.00629) none

–0.0199*** (0.00535) –0.0307*** (0.00587) none

–0.0443* (0.0255) –0.0199*** (0.00539) –0.0289*** (0.00926) none

0.270*** (0.0478)

0.254*** (0.0474)

0.271*** (0.0479)

0.232*** (0.0426)

0.0116 (0.00808) –0.00540** (0.00214) –0.0107*** (0.00361) last 3 years older 2 0.0171 (0.0176)

13,201 0.032

13,201 0.036

13,201 0.032

12,835 0.034

11,941 0.022

Notes: All specifications control for age and district fixed effects, mother’s age, mother’s age squared, mother’s education, mother’s marital status, mother’s employment status, urban, female household head, number of household members, and household assets (electricity, radio, television, fridge, and bike). Robust standard errors in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

mation we have (ownership and use of bed nets, assets, etc.) relates to the time of the survey. When restricting the analysis to the three years prior to the survey, similar results emerge; with a one- year restriction, the sample becomes too small for identification. One last interesting finding is the negative estimate on females. This matches raw DHS data: under- five mortality in 2007 was estimated at 151 for males and 124 for females. ITNs in the HMIS Using the HMIS, we can take advantage of much higher frequency data on disease impact and more carefully explore time variations in the program rollout relative to the DHS data. Table 1.12 presents descriptive statistics for our panel of HMIS data. We match malaria inpatients, malaria deaths, and nonmalaria deaths, all for children under five, with data on net distribution. Since program rollout data is available only at the district level, we aggregate the facility- based HMIS data at the same level. Tables 1.13A and 1.13B show regressions of health outcomes in a district on net distribution in the same calendar year as well as the two preceding

Effects of Large-Scale Health Interventions in Developing Countries Table 1.12

47

Descriptive statistics of district-level panel HMIS Obs.

Mean

Std. dev.

Min.

Max.

Malaria inpatients under five Malaria deaths under five Other deaths under five Nets (’000) District population (’000)

645 645 645 648 648

1922 55 142 9 157

1531 55 181 20 153

0 0 0 0 19

8169 325 1770 182 1341

Malaria inpatients per 1,000 children under five Malaria deaths per 1,000 children under five Other deaths per 1,000 children under five Nets per capita

645 645 645 648

72.01 1.91 4.67 0.06

46.54 1.37 3.59 0.12

0 0 0 0

343.87 8.81 21.57 0.71

years. Table 1.13A looks at absolute numbers of cases and nets distributed; in table 1.13B, cases are normalized by the under- five population, and nets are normalized by district population. Because malaria incidence peaks in the first quarter of the year, while nets are distributed throughout the year, our expectation is that the greatest impact on disease in a year should be net distribution in the previous year. Tables 1.13A and 1.13B show that this expectation holds true. The coefficient on net distribution in the previous year is always negative and significant as a predictor of malaria cases; nets in the current year are not. Interestingly, nets distributed two years earlier are also often insignificant as a predictor of malaria cases. The interpretation of the coefficient on once- lagged nets in columns (2) and (3) of table 1.13A is as follows: 100,000 nets distributed lead to a reduction of about 900 under- five malaria inpatients and to a reduction of twenty- five child deaths reported at health facilities in the average district. Recall that the HMIS records only about one- fourth of deaths. If we assume that the reduction in HMIS mortality in the regressions is only 25 percent of the true benefit, 100,000 nets mean approximately 100 child lives saved. Table1.13B, where cases and nets are scaled by population, yields estimates similar in magnitude. Full coverage of bed nets (one per person) lowers malaria inpatients by twenty- six cases per 1,000, which is a reduction of 35 percent relative to the mean. Similarly, full bed net coverage lowers malaria deaths per 1,000 by a factor of 0.8, which corresponds to a reduction of 42 percent relative to the mean. 1.5.2

Indoor Residual Spraying

The manner in which IRS was rolled out suggests that it might be a good candidate for econometrically identifying the effects of the program on disease outcomes. In the initial design of the malaria initiative, IRS was to be restricted to only areas of very high population density. The scope of IRS was expanded in discrete jumps in 2008. Further, unlike bed nets, IRS must be reapplied each year to be effective. The IRS is generally conducted at the

Table 1.13A

ITN distribution and malaria Malaria inpatients under five (1)

Malaria deaths under five (2)

Other deaths under five (3)

Nets current year Nets previous year

–8.888*** (2.307)

Observations R-squared

2263*** (53.05) 573 0.872

Malaria deaths under five (5)

Other deaths under five (6)

–0.0913 (0.0767) –0.342*** (0.0813) –0.315*** (0.0972) 68.70*** (3.014)

–0.101 (0.169) –0.210 (0.165) –0.488* (0.260) 167.0*** (5.257)

–0.255*** (0.0651)

–0.143 (0.156)

68.53*** (2.744)

166.9*** (4.921)

–0.680 (1.875) –9.422*** (2.643) –4.856 (3.881) 2264*** (60.00)

573 0.741

573 0.904

501 0.881

Nets two years ago Constant

Malaria inpatients under five (4)

501 0.774

501 0.904

Note: Robust standard errors in parentheses are clustered at the district level. All specifications include year and district fixed effects. Nets are in thousands. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

Table 1.13B

ITN Distribution and malaria relative to population Malaria inpatients per 1,000 children under five (1)

Malaria deaths per 1,000 children under five (2)

Other deaths per 1,000 children under five (3)

Nets per capita L1 nets per capita

–26.25*** (9.279)

–0.778*** (0.271)

–0.709 (0.769)

59.75*** (2.751)

1.316*** (0.0930)

3.673*** (0.191)

L2 nets per capita Constant Observations R-squared

573 0.811

573 0.634

573 0.744

Malaria inpatients per 1,000 children under five (4)

Malaria deaths per 1,000 children under five (5)

Other deaths per 1,000 children under five (6)

6.088 (9.872) –30.14** (12.74) –33.50 (36.40) 89.62*** (2.836)

–0.121 (0.309) –0.852** (0.382) –0.0370 (0.817) 2.295*** (0.0940)

–1.543 (1.102) –1.797* (1.077) –3.839** (1.557) 5.334*** (0.185)

501 0.824

501 0.637

501 0.771

Note: Robust standard errors in parentheses are clustered at the district level. All specifications include year and district fixed effects. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

49

Effects of Large-Scale Health Interventions in Developing Countries

end of the calendar year, before the onset of the rainy season when malaria is high. IRS in the DHS Table 1.14 shows regressions of child fever on spraying, similar to the regressions for bed nets reported in table 1.7. When spraying is included alone on the right- hand side, the estimated coefficient is positive, implying that districts that were sprayed in 2007 had a worse time trend for malaria (i.e., slower decline in malaria over time) than those that were not sprayed in that year. This result is driven by a strong negative correlation between the initial fever burden and the rollout of the spraying. By the time of the 2007 DHS, spraying was done in fifteen districts, which on average had a fever prevalence of 37 percent in 2001, significantly below the national average of 45 percent in that year. In 2007, the average fever prevalence in the spraying target areas was actually slightly above the nonspraying areas. This result looks the same when we run the regressions at the individual level (households reporting whether or not they have been sprayed over the twelve months preceding the interview) as shown in column (2). Table 1.14

IRS results, DHS Child had fever over last two weeks

Dependent variable Percentage of district population sprayed

(1)

(2)

(3)

0.102*** (0.0192)

Household sprayed (self-report)

0.0482** (0.0195)

–0.0162 (0.0199)

Fraction of households sprayed in cluster Age child 2nd wave dummy Constant Observations R-squared

(4)

–0.0367*** (0.00277) –0.283*** (0.0122) 0.614*** (0.0733)

–0.0367*** (0.00277) –0.257*** (0.0108) 0.622*** (0.0733)

11,524 0.123

11,523 0.121

–0.0206*** (0.00330)

–0.00778 (0.0394) –0.0207*** (0.00331)

0.361*** (0.0919)

0.357*** (0.0921)

5,671 0.047

5,672 0.046

Note: All specifications control for district fixed effects, sex, mother’s age, mother’s age squared, mother’s education, mother’s marital status, mother’s employment status, urban, female household head, number of household members, and household assets (electricity, radio, television, fridge, and bike). Robust standard errors in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

50

Nava Ashraf, Günther Fink, and David N. Weil

To minimize the potential convergence bias, we focus on the 2007 survey only in columns (3) and (4). In column (3), we use spraying at the household level only; in column (4), we take the fraction of households sprayed within the cluster (typically fifteen households with children per cluster) as explanatory variable. The coefficient becomes negative, but is not significant. Within a given district, targeted households and clusters appear to have similar fever prevalence rates to nontargeted ones. We are not quite sure what to make out of this result; if the numbers are correct, it could either be that spraying misses its target (people get infected outside, or spraying is done before the rain and then washed away), or, alternatively, that households that do not get spraying have lower risk or more actively engage in other unobserved preventive measures. IRS in the HMIS As with the ITN analysis, the use of the HMIS has the main advantage of offering higher frequency data when it comes to evaluating the disease impact of the IRS spraying. Given that spraying loses its protective effect within about a year, close to all spraying in Zambia arranged through the NMCP is done in the last quarter of the each year when the rains start and mosquito populations rapidly reemerge after the dry season. Under ideal conditions, IRS spraying is supposed to protect household members throughout the rainy seasons, and to be repeated at the end of each year with the new rainfalls. In table 1.15A, we try to identify the effects of IRS spraying on the levels of malaria inpatients, malaria deaths, and deaths due to other causes with a simple IRS spraying target dummy. The IRS target dummy variable equals 1 if the district was in the spraying program in a given year, and is zero otherwise. As figure 1.10 shows, the rollout of the IRS implemented by the NMCP spraying was incremental; any district enrolled in the program since 2003 has been receiving spraying in all subsequent years. In columns (1)–(3), we regress health outcomes on spraying without controlling for the contemporaneous net distribution. The effects of IRS spraying on malaria inpatients and malaria deaths are negative, but only marginally significant. The estimated coefficient in column (2) implies that being a target district is associated with 22.5 fewer malaria deaths per year. Once we add controls for bed net rollout in columns (4)–(6), the estimated effects on the number of under- five malaria inpatients and deaths become larger and more significant. A simple comparison of the estimated coefficients in column (5) implies that the effect of becoming a target for spraying is comparable to the distribution of 100,000 nets in the preceding year in the average district. The results become much weaker, however, when we express patients and mortality numbers in population terms, and regress disease burden per capita on per capita measures of bed net and spraying rollout in table 1.15B. While the coefficient on bed net distribution remains highly significant (and

Table 1.15A

Spraying only versus spraying and ITN

Spraying target dummy

Malaria inpatients under five (1)

Malaria deaths under five (2)

–241.5 (189.1)

–22.57* (12.15)

Other deaths under five (3)

Malaria inpatients under five (4)

Malaria deaths under five (5)

Other deaths under five (6)

0.539 (17.62)

–24.72** (12.12) –0.298*** (0.0702) 68.51*** (2.513)

–0.278 (17.28) –0.113 (0.147) 166.9*** (4.905)

573 0.766

573 0.905

42.18*** (2.690)

129.7*** (11.44)

–308.9* (176.4) –9.351*** (2.324) 2262*** (51.98)

573 0.760

573 0.905

573 0.873

Lag 1 bed nets in ’000 Constant

1713*** (70.43)

Observations R-squared

573 0.866

Note: Robust standard errors in parentheses are clustered at the district level. All specifications include year and district fixed effects. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

Table 1.15B

Fraction sprayed

Spraying only versus spraying and ITN Malaria inpatients per 1,000 children under five (1)

Malaria deaths per 1,000 children under five (2)

Other deaths per 1,000 children under five (3)

Malaria inpatients per 1,000 children under five (4)

Malaria deaths per 1,000 children under five (5)

Other deaths per 1,000 children under five (6)

6.199 (9.660)

–0.416 (0.370)

0.792 (0.559)

55.53*** (3.029)

1.273*** (0.0892)

3.421*** (0.208)

2.526 (9.760) –25.38*** (9.548) 59.29*** (3.555)

–0.558 (0.372) –0.984*** (0.257) 1.419*** (0.0968)

0.722 (0.543) –0.484 (0.704) 3.492*** (0.190)

Nets per capita Constant Observations R-squared

573 0.809

573 0.656

573 0.787

573 0.811

573 0.661

573 0.787

Note: Robust standard errors in parentheses are clustered at the district level. All specifications include year and district fixed effects. Net distribution and spraying numbers reflect program activities in the preceding year. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

52

Nava Ashraf, Günther Fink, and David N. Weil

Fig 1.11

Health facilities and spraying in the Chingola district 2008

Source: NMCC. Crosses represent health facilities, and black dots sprayed structures. Black lines are district boundaries.

similar in magnitude to the ITN regressions where we do not control for spraying), the spraying coverage does not appear to have any effect on the number of malaria patients per 1,000 children under five in this specification. The estimated effect on under- five malaria deaths (column [5]) is just shy of significance. The relative magnitude of the coefficients estimated in column (5) implies that providing full spraying coverage has about half the effect of providing full net coverage. Overall, the estimated effects of the IRS campaign in the HMIS are surprisingly weak given the design of the spraying rollout. As figure 1.11 illustrates, the IRS campaign is generally directly focused around health facilities. Trying to maximize the impact of the campaign, programs generally concentrate their efforts to the (catchment) areas directly surrounding facilities reporting a high malaria caseload in the given period. Given the stochasticity of local malaria incidence, particularly high- incidence years in a particular area are likely to be followed by more moderate years, so that the simple difference- in-difference model estimated above should lead to an overestimation of the true effect. The fact that the effects are weak even in the HMIS is thus rather puzzling, but in our view consistent with the rather weak evidence on spraying emerging from the DHS analysis.

Effects of Large-Scale Health Interventions in Developing Countries

53

1.6 Conclusions 1.6.1

Future Directions for Research

As mentioned in the introduction, the current chapter is part of a larger project in which the authors hope to use the Zambia malaria initiative to better understand the economic effects of malaria. It is worth stepping back for a moment to see why the Zambian experience is particularly useful in this regard. Discussion in the policy community regarding the relative priority of improving health in developing countries often points to the economic benefits accruing from better health as an important secondary justification complementary to the direct humanitarian and health benefits associated with related programs. The question of how disease affects economic growth is well established in the literature. Much of the discussion of the economic effects of malaria among policymakers, for example, cites estimates from the work of Gallup and Sachs (2001). Trying to estimate the effect of health on economic outcomes runs into serious identification problems, however. Omitted factors that affect health may affect income directly, or health may respond directly to improvements in income. The standard solution for such an identification problem is to find instrumental variables that directly affect health. These instruments could be some purely exogenous factor or possibly the result of some discontinuous response of health inputs to local conditions. For example, Acemoglu and Johnson (2007) use variations in technological progress in controlling different diseases during the post– World War II period to instrument for health changes at the national level. Even though malaria is viewed as one of the most economically important diseases, most recent studies of its economic effects have had to rely on data from episodes of malaria eradication in South Asia and Latin America that took place half a century ago (see Lucas 2010b; Bleakley 2010). As discussed above, the large scaling up of resources devoted to malaria control in Zambia was not primarily the result of factors on the ground in the country itself. Rather, developments on the world stage, including the development of new technology and a new alignment of priorities within the development community, led to Zambia being chosen as a test case for the possibility of rapidly scaled-up malaria control. Thus, at the aggregate level, the timing of the reduction in malaria in Zambia may be viewed as largely exogenous. Of course, Zambia was chosen as the first country to receive such intensive support and treatment because it was viewed as having the institutional capacity to succeed, and the same factors that were expected to lead to success against malaria might have been expected to have independent economic effects, so the identification is not perfect. Nonetheless, the suddenness with which resources were applied suggests that reasonable identification may be possible. Beyond the inferences that can be drawn from developments at the national level, our hope is that additional identification can be achieved

54

Nava Ashraf, Günther Fink, and David N. Weil

by looking at the manner in which the campaign against malaria was rolled out within Zambia. Some of the issues that we hope to investigate in later work include: Fertility A substantial literature discusses the link between changes in mortality and changes in fertility. In particular, it is often argued that declining infant and child mortality initially leads to a rise in the total fertility rate, as actual deaths fall short of expectations, but that in the long run TFR declines because of reduced uncertainty. In the case of malaria, there is an additional set of considerations, because the disease works to lower fecundity directly (Lucas 2010). Figures from the DHS early release show that TFR in Zambia rose from 5.9 to 6.2 over the period 2002– 2007 (6.9 to 7.5 in rural areas, while urban TFR was flat at 4.0). The period between DHS surveys corresponds well to the period of rollout of the malaria program. Productivity One way in which malaria affects economic outcomes is by directly lowering the labor input of workers, both through absenteeism and reduced physical capacity due to anemia. There are some cases where we hope to observe directly the productivity effects of malaria control. We are working on obtaining data from Zambia Sugar, the country’s largest producer, which is located in the Mazabuka district. Zambia Sugar undertook a private eradication effort that predated the national effort by several years. Malaria morbidity has traditionally been quite high among cane cutters, who must work in swampy conditions. We hope to obtain data on changes in absenteeism over time (this data was actually collected once already, but was subsequently lost). Education Many economists have stressed malaria’s effect on educational attainment as an important channel through which the disease affects economic outcomes. By matching data from Zambia’s educational statistics system on grade progress and dropout rates to data on the rollout of the antimalaria initiative, we hope to investigate how much health improvements have led to an increase in educational attainment. 1.6.2

Sustainability and Further Progress

The progress already made against malaria and other sources of premature mortality in Zambia represents a major humanitarian success. One issue raised by progress so far is whether it will be possible to finish the job, and reduce malaria incidence to near zero. The history of antimalaria campaigns in the years after World War II contains several episodes in which malaria was substantially eliminated. Eradication is also the NMCC’s even-

Effects of Large-Scale Health Interventions in Developing Countries

55

tual goal, even if it is not possible in the short term. There are currently some discussions on pilot projects in a few districts to do a massive test and treat campaign to reduce parasitemia to near zero. Zambia was chosen as a test country for scale-up for multiple reasons. One important reason was that the institutional capacity, good governance, and political will existed to make use of the resources. However, it is also the case that the climate is favorable to eventual complete eradication because of the cold winters that ensure mosquito populations are reduced to near- zero for at least some period annually. Until malaria is completely eliminated in the country, however, a serious concern remains regarding the sustainability of gains achieved so far. As discussed above, Zambia has already been through an episode in which significant progress against malaria was followed by a resurgence of the disease. Similarly, in Zambia’s neighbor, Zimbabwe, malaria was almost completely eradicated, but the political environment led to the end of an effective malaria control regime and the disease has subsequently returned to epidemic levels. In the current Zambian environment, disease vectors remain present and a significant number of humans continue to host the disease. This means that unlike places where eradication has been complete, there is always the potential for a rapid resurgence, which could be all the more devastating as cohorts with lower acquired immunity age through the population. Maintaining the low current level of malaria mortality and morbidity will thus require continued application of inputs at near the current level. The life span of a bed net averages three years if properly treated, so maintaining a ratio of one net for every two persons will require the distribution of approximately two million nets per year. Since indoor spraying must be repeated annually, there is only limited scope for a reduction in spending and effort devoted to malaria control.

References Acemoglu, Daron, and Simon Johnson. 2007. “Disease and Development: The Effect of Life Expectancy on Economic Growth.” Journal of Political Economy 115 (6): 925– 85. Bleakley, Hoyt. 2010. “Malaria Eradication in the Americas: A Retrospective Analysis of Childhood Exposure.” American Economic Journal: Applied Economics 2 (2): 1– 45. Chanda, Pascalina, Felix Masiye, Bona M. Chitah, Naawa Sipilanyambe, Moonga Hawela, Patrick Banda, and Tuoyo Okorosobo. 2007. “A Cost-Effectiveness Analysis of Artemether-Lumefantrine for Treatment of Uncomplicated Malaria in Zambia.” Malaria Journal 6 (21). doi:10.1186/1475-2875-6-21. Frost, Laura J., and Michael R. Reich. 2008. How Do Good Health Technologies Get to Poor People in Poor Countries? Cambridge, MA: Harvard University Press.

56

Nava Ashraf, Günther Fink, and David N. Weil

Gallup, John L., and Jeffrey D. Sachs. 2001. “The Economic Burden of Malaria.” American Journal of Tropical Medicine and Hygiene 64 (1, 2)S: 85– 96. Government of Zambia. Estimates of Revenue and Expenditure: Activity-Based Budget 2008. Lusaka. Government of Zambia, Central Statistical Office. 2003. 2000 Census of Population and Housing. Lusaka. http://www.zamstats.gov.zm/media/sum_rpt.pdf. Government of Zambia, Central Statistical Office, et al. 2003. Zambia Demographic and Health Survey 2001–2002 Lusaka. February. ———. 2009. Zambia Demographic and Health Survey 2007. Lusaka. http://www.measuredhs.com/pubs/pdf/FR211/FR211[revised- 05– 12– 2009].pdf. Government of Zambia, Ministry of Health. 2008 National Malaria Control Action Plan: Actions for Scale-up for Impact on Malaria in Zambia. Lusaka. http://nmcc .org.zm/files/2008ZambiaNMCPActionPlan- 5– 20.pdf. Government of Zambia, Ministry of Health. 2005. National Health Strategic Plan 2006–2010. Lusaka. https://www.healthresearchweb.org/files/NHSP_2006.pdf. ———. 2006. A Roadmap for Impact on Malaria in Zambia: A 5-Year Strategic Plan, 2006–2010. Lusaka. Government of Zambia, Ministry of Health, et al. 2006. Zambia National Malaria Indicator Survey 2006. Lusaka. http://www.nmcc.org.zm/files/2006_Zambia _Malaria_Indicator_Survey.pdf. ———. 2008. Guidelines on the Distribution and Utilization of Insecticide-Treated Nets for Malaria Prevention and Control. Lusaka. November. ———. 2008. Health Facilities in Zambia: A Listing of Health Facilities According to Level and Location for 2008. Lusaka. Government of Zambia, Ministry of Health. 2009. Zambia National Malaria Indicator Survey 2008. Lusaka. Hamer, Davidson H., Micky Ndhlovu, Dejan Zurovac, Matthew Fox, Kojo YeboahAntwi, Pascalina Chanda, Naawa Sipilinyambe, et al. 2007. “Improved Diagnostic Testing and Malaria Treatment Practices in Zambia.” Journal of the American Medical Association 297 (20): 2227– 31. Lubell, Yoel, Hugh Reyburn, Hilda Mbakilwa, Rose Mwangi, Semkini Chonya, Christopher J. M. Whitty, and Anne Mills. 2008. “The Impact of Response to the Results of Diagnostic Tests for Malaria: Cost-Benefit Analysis.” British Medical Journal 336 (7637): 202– 05. Lucas, Adrienne. 2010a. “The Impact of Malaria Eradication on Fertility.” Working Paper, Wellesley College. ———. 2010b. “Malaria Eradication and Educational Attainment: Evidence from Paraguay and Sri Lanka.” American Economic Journal: Applied Economics 2 (2): 46– 71. Packard, Randall M. 2007. The Making of a Tropical Disease: A Short History of Malaria. Baltimore: Johns Hopkins University Press. Rolland, Estelle, Francesco Checchi, Loretxu Pinoges, Suna Balkan, Jean-Paul Guthmann, and Philippe J. Guerin. 2006. “Operational Response to Malaria Epidemics: Are Rapid Diagnostic Tests Cost-Effective?” Tropical Medicine and International Health 11 (4): 398– 408. Sipilanyambe, Naawa, Jonathan L. Simon, Pascalina Chanda, Peter Olumese, Robert W. Snow, and Davidson H. Hamer. 2008. “From Chloroquine to ArtemetherLumefantrine: The Process of Drug Policy Change in Zambia.” Malaria Journal 7 (25). doi:10.1186/1475-2875-7-25. Steketee, Richard W., Naawa Sipilanyambe, John Chimumbwa, James J. Banda, Abdirahuman Mohamed, John Miller, Surprotik Basu, et al. 2008. “National Malaria Control and Scaling Up for Impact: The Zambia Experience through 2006.” American Journal of Tropical Medicine and Hygiene 79 (1): 45– 52.

Effects of Large-Scale Health Interventions in Developing Countries

57

World Health Organization. 2007. “Malaria Rapid Diagnostic Tests: Assessing Cost-Effectiveness of Diagnosis.” http://www.wpro.who.int/sites/rdt/using_rdts /assessing_cost_effectiveness.htm. Zurovac, Dejan, Mickey Ndhlovu, Nawa Sipilanyambe, Pascalina Chanda, Davidson H. Hamer, Jon L. Simon, and Robert W. Snow. 2007. “Paediatric Malaria Case-Management with Artemether-Lumefantrine in Zambia: A Repeat CrossSectional Study.” Malaria Journal 6 (31). http://www.malariajournal.com/content /6/1/31.

2

Prevention of Mother-to-Child Transmission of HIV and Reproductive Behavior in Zambia Nicholas Wilson

2.1

Introduction

The rapid expansion of prevention of mother- to-child transmission of HIV (PMTCT) in sub-Saharan Africa in the past decade is one of the great humanitarian successes of this era. At the turn of the twenty- first century, PMTCT was virtually unavailable for the vast majority of women in subSaharan Africa, the region of the world most affected by the HIV/AIDS pandemic. In 2009, 54 percent of HIV- positive pregnant women in the region received PMTCT (WHO 2010). This scale-up surely has saved millions of lives. Five percent of adults age fifteen to forty- nine in sub-Saharan Africa are HIV positive (UNAIDS 2010). In the absence of PMTCT, a HIV- positive woman will transmit the virus in utero, during childbirth, or through breastfeeding with a cumulative probability of as much as 45 percent (Dabis and Ekpini 2002). A HIVNicholas Wilson is assistant professor of economics at Williams College. I would like to thank Quamrul Ashraf, William Dow, Lucie Schmidt, Lara Shore-Sheppard, Jeffrey Stringer, Anand Swamy, Waly Wane, Tara Watson, David Weil, seminar participants at the University of California, Berkeley, NBER Africa Project Zanzibar Conference, University of California, Santa Cruz, Williams College, and an anonymous referee for many excellent comments. Special thanks to Elisa Pepe for tremendous support throughout this project. Madeleine Watson and Wentao Xiong provided superb research assistance. The NBER Africa Project provided generous financial and institutional support. This research would not be possible without the assistance of the Network of Zambian People Living with HIV/AIDS (NZP+). All errors are my own. The findings, interpretations, and conclusions expressed in this chapter are those of the author and do not necessarily represent the views of the aforementioned individuals or the agencies that employ them. For acknowledgments, sources of research support, and disclosure of the author’s material financial relationships, if any, please see http://www .nber.org/chapters/c13372.ack.

59

60

Nicholas Wilson

positive infant will die with probability between 25 and 50 percent by age one (Spira et al. 1999; Taha et al. 1999; Dabis et al. 2001; Brahmbhatt et al. 2001). The PMTCT can reduce the cumulative transmission probability from a HIV- positive woman to her newborn child to as little as 3 percent (Dabis and Ekpini 2002; Canning 2006). This chapter documents the expansion of PMTCT in Zambia over the period 2001– 2007 and examines changes in reproductive behavior associated with the local introduction of PMTCT. In doing so, it provides some of the first evidence on reproductive behavior in the context of the widespread availability of the single- most effective HIV prevention intervention we know of in practice today, an intervention that dramatically reduces child mortality risk. Although the current analysis is mostly descriptive and does not attempt to identify the causal effects of PMTCT availability, it provides suggestive evidence that the rapid scale-up of PMTCT has generated large changes in reproductive behavior. Between 2001 and the end of 2007, the number of health facilities in Zambia offering PMTCT increased from fewer than six to nearly 600. By the end of this period, more than 40 percent of health facilities offered PMTCT. Because the expansion occurred with greater intensity in urban areas than in rural areas, individual- level coverage rates increased even more rapidly. Although it is clear that PMTCT changes the incentives that women and couples face in making decisions about reproductive behavior, a priori the sign of the fertility response to PMTCT is ambiguous. For example, the standard quantity- quality model following Becker and Lewis (1973) illuminates a mechanism by which PMTCT reduces fertility and one by which PMTCT increases fertility. First, PMTCT reduces the shadow price of child quantity. Because each birth is more likely to survive into older ages, the expected number of births to achieve a desired number of surviving children has fallen, in turn reducing the shadow price of child quantity and leading to an increase in fertility. Second, PMTCT simultaneously reduces the shadow price of child quality. Because each child is more likely to survive into older ages, household investments in children’s human capital made at a young age are more likely to realize a larger return. That is, the shadow price of child quality has fallen, inducing a reduction in fertility due to the quantity- quality trade- off embedded in this model. I use newly assembled data on the expansion of HIV/AIDS services in Zambia to examine changes in reproductive knowledge and behavior associated with the local introduction of PMTCT. A primarily descriptive analysis of conditional means yields three main findings. First, the local introduction of PMTCT was associated with an increase in knowledge of motherto-child transmission (MTCT) and an increase in knowledge of prevention of mother- to-child transmission (PMTCT). Second, the local introduction of PMTCT was associated with a decrease in child mortality and pregnancy rates. Third, the local introduction of PMTCT was associated with

Mother-to-Child Transmission of HIV and Reproductive Behavior

61

(proportionally) smaller reductions in breastfeeding rates. An analysis of the heterogeneity in these changes by the likelihood the respondent was HIV positive provides mixed support for a causal interpretation of these differences in conditional means. In addition, I uncover several significant differences by age and education level of the respondent. Future research will address the endogeneity of PMTCT expansion. The PMTCT largely was introduced at existing health clinics and PMTCT expansion occurred with greater intensity in urban areas. Thus, it is reasonable to believe that time- invariant as well as time- varying factors affecting reproductive behavior varied systematically with the intensity of PMTCT expansion. In the interim, the results of this analysis provide some of the first evidence on PMTCT expansion and reproductive behavior. The rest of the chapter is organized as follows. In section 2.2, I describe the existing literature in three related topics: (a) PMTCT and fertility, (b) HIV/AIDS and fertility, and (c) child mortality risk and fertility. Section 2.3 provides a few clinical facts about PMTCT and discusses some possible behavioral responses. Section 2.4 describes the data collected for this project. Section 2.5 discusses PMTCT scale-up in Zambia. Section 2.6 presents five sets of outcomes related to PMTCT expansion: (a) knowledge of mother- to-child transmission (MTCT), (b) knowledge of prevention of mother- to-child transmission (PMTCT), (c) child mortality, (d) pregnancy, and (e) breastfeeding. Section 2.7 discusses the implications of these findings as well as an agenda for future research on PMTCT and reproductive outcomes. 2.2 2.2.1

Existing Literature PMTCT and Fertility

There is no existing economic literature on the behavioral effects of PMTCT. Moreover, there is only a nascent literature in public health on the behavioral effects of PMTCT. However, a small number of studies examine fertility intentions in the context of PMTCT. The limited empirical evidence on fertility intentions in the context of PMTCT suggests that access to PMTCT may increase fertility among HIVpositive women with known HIV status. Cooper et al. (2007) conducted in-depth interviews with sixty- one HIV- positive men and women in Cape Town, South Africa. Providing information about PMTCT during the course of the interview increased the desire to have (more) children. However, female interviewees expressed additional concern about the availability of antiretroviral therapy for adults (ART) to ensure that they would be alive to take care of their children. Peltzer, Chao, and Dana (2009) investigate fertility intentions among a sample of women with known HIV status receiving postnatal care in Tembo District, South Africa. Among HIV- positive

62

Nicholas Wilson

women, knowledge of PMTCT was associated with increased desire for pregnancy. 2.2.2

HIV/AIDS and Fertility

A variety of studies examine the impact of HIV/AIDS on fertility in sub-Saharan Africa. These studies generally investigate the total effect of HIV/AIDS on fertility, rather than focusing on the child mortality risk channel. Because PMTCT only directly reduces child mortality risk without directly affecting adult mortality risk, it is unlikely that PMTCT simply will reverse the effects of HIV/AIDS. Nonetheless, these studies represent an important related literature, particularly because of the dearth of evidence on the fertility response to PMTCT. The initial economic analyses of the fertility response to the HIV/AIDS pandemic suggested that fertility in much of sub-Saharan Africa may have fallen in response to the HIV/AIDS pandemic. Young (2007) found that demand for children, and in turn fertility, fell in response to the HIV/AIDS pandemic. Similarly, Juhn, Kalemi-Ozcan, and Turan (2009) found that although community- level HIV prevalence had no effect on fertility, HIVpositive women in sub-Saharan Africa had fewer children. A second wave of economic research on this topic presents new evidence seemingly contradicting these initial findings. Kalemli-Ozcan and Turan (2011) revisited Young’s (2007) study and found that restricting the empirical analysis to the period for which HIV data are actually available suggests that HIV actually increased fertility.1 Fortson (2009) and Fink and Linnemayr (2009) present evidence suggesting that the HIV/AIDS pandemic has not affected fertility on average. However, Fink and Linnemayr (2009) also argue that fertility fell among more educated women in response to HIV/AIDS. More generally, Kalemli-Ozcan (2012) shows that the estimated relationship between HIV prevalence and fertility is very sensitive to the source of variation in HIV prevalence (e.g., cross- sectional versus time series) and suggests that HIV/AIDS has had little effect on fertility. In contrast, it appears that the majority of public health and medical studies on HIV/AIDS and fertility find lower fertility among HIV- positive women than among HIV- negative women. For example, Gray et al. (1998) found that pregnancy rates in Uganda were lower among HIV- positive women than among HIV- negative women, partly due to an increased likelihood of pregnancy loss and partly due to lower rates of conception. Likewise, Carpenter et al. (1997), Zaba and Gregson (1998), Glynn et al. (2000), Fabiani et al. (2006), Kongnyuy and Wiysonge (2008), and Chen and Walker (2010) present evidence indicating that fertility is lower among HIV- positive women than among HIV- negative women. In a study examining changes 1. In the absence of complete data on HIV prevalence at the start of the pandemic, Young (2007) assumed that HIV prevalence was zero from 1980 through 1998.

Mother-to-Child Transmission of HIV and Reproductive Behavior

63

in fertility among women receiving a HIV- positive test result in Malawi, Hoffman et al. (2008) found that pregnancy rates fell among HIV- positive women who learned of their HIV status. 2.2.3

Child Mortality Risk and Fertility

A broader economic literature examines the effect of child mortality on fertility. Using data from three different settings, Ben-Porath (1976) found that in each setting increased (realized) child mortality was associated with higher subsequent fertility, consistent with households engaging in replacement fertility. In contrast, Wolpin (1984) found evidence of only a small replacement fertility effect. However, Wolpin (1984) found a large negative effect of child mortality risk on fertility. Among more recent studies, Doepke (2005) examined the relationship between child mortality and fertility in a model following Barro and Becker (1989), as well as in several variants of this model. Doepke (2005) found that the existence of replacement fertility may produce a positive relationship between child mortality risk and fertility. However, for child mortality risk to have a positive effect on net fertility, households must practice precautionary fertility (i.e., “child hoarding”). Similarly, Angeles (2010) found that child mortality increased gross fertility, but had little effect on net fertility. In contrast, Soares (2005) shows that the existence of a quantity- quality trade- off for household investments in children yields the result that a reduction in child mortality risk reduces fertility. 2.3 2.3.1

Prevention of Mother-to-Child Transmission Background

Prevention of mother- to-child transmission of HIV (PMTCT) is the single- most effective HIV prevention intervention in practice today. When administered in accordance with World Health Organization (WHO) recommendations, PMTCT can reduce the cumulative probability of transmission from as much as 45 percent in the absence of PMTCT to as little as 3 percent (Dabis and Ekpini 2002; Canning 2006). In doing so, the availability of PMTCT has the potential to substantially reduce child mortality risk in high HIV prevalence environments. The WHO recommends “combination therapy” (i.e., a multiple- drug therapy) for HIV- positive mothers and infants (WHO 2006). The drugs in this combination therapy include azidothymidine (AZT) and nevirapine (NVP). In Zambia, as in much of sub-Saharan Africa (UNAIDS 2010), PMTCT consists of single- dose NVP (i.e., NVP without AZT) administered to a HIV- positive pregnant woman at diagnosis, at the onset of childbirth, and to her infant child during the first week or two of breastfeeding (Stringer et al. 2003, 2005).

64

Nicholas Wilson

By reducing child mortality risk, PMTCT changes the incentives that women and households face in their reproductive decisions. In a standard quantity- quality model following Becker and Lewis (1973), PMTCT reduces the shadow price of child quantity because it decreases the number of births required to achieve a desired number of surviving children. The reduction in the shadow price of child quantity should increase fertility. However, PMTCT simultaneously reduces the shadow price of child quality because household investments in a child’s human capital made at a young age are more likely to realize a return as that child is more likely to survive into older ages. Through the quantity- quality trade- off embedded in this model, this reduction in the shadow price of child quality should decrease fertility. The PMTCT may affect fertility through other channels as well, including reducing the need for precautionary or replacement fertility or by providing information to a HIV- positive woman about her HIV status.2 A change in fertility due to the availability of PMTCT may induce changes in other reproductive behaviors as well. For example, if PMTCT reduces fertility, then women may increase breastfeeding durations, possibly further reducing fertility due to increased lactational amenorrhea. Moreover, the availability of PMTCT may directly affect breastfeeding behavior. Motherto-child transmission of HIV through breastfeeding comprises roughly onehalf of the cumulative risk of mother- to-child transmission (Dabis and Ekpini 2002) and PMTCT reduces the risk of transmission associated with breastfeeding. 2.3.2

Scale-Up in Sub-Saharan Africa

In the early twenty- first century, in conjunction with global donor governments, many sub-Saharan African counties dramatically expanded access to PMTCT. Prior to the authorization of the United States President’s Emergency Plan for AIDS Relief (PEPFAR) in 2004 and the Global Fund to Fight  HIV/AIDS, Tuberculosis, and Malaria, PMTCT was virtually unavailable to the vast majority of the world’s population in this region. Between 2005 and 2009, the proportion of pregnant women living with HIV in this region who received PMTCT drugs increased from 15 percent (WHO 2007) to 54 percent (WHO 2010). Early scale-up has been greater in eastern and southern Africa, the area of sub-Saharan Africa with the highest HIV prevalence, than in western and central Africa. For example, in 2009, the fraction of pregnant women living with HIV who received PMTCT drugs was more than two- thirds in eastern and southern Africa and less than one- quarter in western and central Africa (WHO 2010). Coverage for infants to pregnant women living with HIV has increased 2. Personal communication with Dr. Jeffrey Stringer indicates that there do not appear to be any biochemical pathways linking NVP and fecundity.

Mother-to-Child Transmission of HIV and Reproductive Behavior

65

as well. Relative to coverage for women, scale-up for infants has lagged. In sub-Saharan Africa, the proportion of infants born to mothers living with HIV who received PMTCT drugs increased from 11 percent in 2005 (WHO 2007) to 35 percent in 2009 (WHO 2010). As one of the fourteen PEPFAR focus countries, Zambia has been among the leaders in PMTCT rollout. Between 2005 and 2009, PMTCT drug coverage for pregnant women living with HIV increased from 15 percent (WHO 2007) to nearly 70 percent (WHO 2010). By 2009, nearly 40 percent of infants born to mothers living with HIV received PMTCT drugs (WHO 2010). 2.4

Data

Despite the central role PMTCT has played in HIV/AIDS policy in subSaharan Africa, there has been little effort to document PMTCT expansion at a subnational level. Thus, in conjunction with the Network of Zambian People Living with HIV/AIDS (NZP+), I collected data documenting the expansion of PMTCT at the health- facility level from the beginning of HIV/AIDS services time in Zambia through June 2008.3 The Japanese International Cooperation (JICA) 2006 Health Facilities Census (HFC) formed the foundation for our data collection. The 2006 HFC enumerated every health facility in Zambia and recorded the precise GPS coordinates of each facility. Using this information on the universe of health facilities in Zambia, we collected information from each facility on the month and year (if any) it began offering PMTCT, as well as similar information for the other two main HIV/AIDS services (i.e., voluntary counseling and testing [VCT] and antiretroviral therapy for adults [ART]). This processes yielded a retrospective monthly health- facilities panel documenting the expansion of the three main HIV/AIDS services for an entire high HIV- prevalence country. To the best of my knowledge, these are the first such data that exist. Not only do they document HIV/AIDS service expansion at the health- facility level, but in conjunction with nationally representative household survey data on reproductive behavior they will provide evidence on the behavioral responses to HIV/AIDS service expansion (including PMTCT expansion). Data on reproductive behavior come from four nationally representative household surveys. These are the 2001 and 2007 Demographic Health Surveys (DHS) and the 2003 and 2005 Zambia Sexual Behavior Surveys (ZSBS). Several key variables I use in the empirical analysis are: (a) knowledge (yes/no) of mother- to-child transmission of HIV, (b) knowledge 3. Data collection continued past the middle of 2008. However, the panel only reliably captures PMTCT expansion through the middle of 2008. Data collection effectively began in June 2008 and facilities reporting no PMTCT at that time may have subsequently introduced it. Because our data collection process does not update expansion in real time, we are unable to track PMTCT expansion after June 2008 until we update the health facilities panel.

66

Nicholas Wilson

(yes/no) of prevention of mother- to-child transmission of HIV (PMTCT), (c) child death (yes/no) by age one, (d) pregnant (yes/no) at any point in the twelve months leading up to the interview date, and (e) breastfeeding (yes/no) for a child age zero to twenty- four months. Information on the location of the survey respondents allow me to calculate the distance from each survey respondent to each health facility. The 2007 DHS contains respondents’ GPS coordinates (intentionally coded with a small error component to ensure respondent confidentiality). I use administrative records on the locations of the primary sampling units in the 2001 DHS, the 2003 ZSBS, and the 2005 ZSBS. These tend to be statistical enumeration areas (SEAs), which are administrative units designed to capture approximately 1,000 residents. I calculate the centroid of each SEA and record its GPS coordinates as the location of the survey respondents who reside in that SEA. After calculating the distance from each survey respondent to each health facility, I am left with 7,683 adult females (i.e., ages fifteen to forty- nine) in the 2001 DHS, 2,296 adult females in the 2003 ZSBS, 2,072 adult females in the 2005 ZSBS, and 7,146 adult females in the 2007 DHS.4 2.5

Zambian Scale-Up

Before turning to an analysis of the association between local PMTCT introduction and reproductive behavior, I briefly describe the expansion of access to PMTCT as documented in the data collected for the current analysis. Between 2000 and the end of 2007, the number of health facilities in Zambia offering PMTCT increased from virtually zero to nearly six hundred. Figure 2.1 shows the number of health facilities that introduced PMTCT by the year of introduction. Although expansion was fairly steady over this period, the expansion occurred with greatest intensity in 2005. Facilities coverage rates increased from close to zero to more than 40 percent over this time period. Figure 2.2 shows the proportion of health facilities at the end of each calendar year offering PMTCT from 2000 through 2007. Individual coverage rates increased even more rapidly over this same time period. I calculate the individual coverage rate as the fraction of females age fifteen to forty- nine residing within twenty kilometers of a PMTCT site. Figure 2.3 shows the individual coverage rate at the end of each calendar year from 2000 through 2007. Because PMTCT expansion occurred with greater intensity in urban areas, the individual coverage rate tended to be higher than the facilities coverage rate during this time period. 4. The Zambia Central Statistical Office provided a digitized census map that I used to identify the location of each statistical enumeration area (SEA). However, this map is missing roughly 7 percent of the SEAs in Zambia and I am unable to calculate the location of approximately 7 percent of the 2001– 2005 survey respondents.

Fig. 2.1

The PMTCT expansion in Zambia, 2000–2007

Fig. 2.2

Facilities coverage rate for PMTCT, 2000–2007

68

Nicholas Wilson

Fig. 2.3

2.6 2.6.1

Individual coverage rate for PMTCT, 2001–2007

Results Knowledge of MTCT

I begin by examining the change in knowledge of mother- to-child transmission of HIV (MTCT) associated with the local introduction of PMTCT. By local introduction, I mean PMTCT was introduced at a health clinic within twenty kilometers of the respondent at least eleven months prior to the interview date.5 Unless most women were already informed about mother- to-child transmission of HIV, the local introduction of a service designed to prevent this transmission presumably should have increased knowledge of MTCT. If it did not, that might suggest that local availability does not translate into local access because of lack of information about local availability. Table 2.1 shows the proportion of female respondents who reported knowing of MTCT, disaggregated by (eventual) proximity to PMTCT. For respondents in (eventual) PMTCT locations, these sample means are further disaggregated by whether PMTCT has been available for at least eleven months. The sample means presented in table 2.1 suggest that PMTCT 5. In an analysis of the determinants of maternal health care usage in Zambia, Stekelenberg et al. (2004) found that willingness to visit a maternal health clinic fell rapidly when the clinic was more than a two- hour walk. I use a cut- off date of eleven months prior to the interview date because that is roughly when the representative conception occurred for pregnancies measured during the twelve months leading up to the interview date.

69

Mother-to-Child Transmission of HIV and Reproductive Behavior

may have increased knowledge of mother- to-child transmission. Although respondents in (eventual) PMTCT locations had higher knowledge of MTCT prior to local PMTCT introduction than did respondents in nonPMTCT locations, knowledge of MTCT increased by approximately 7 percentage points ( p- value = 0.00) among respondents in locations receiving PMTCT. This response may have been larger among women who were more likely to be HIV positive. Women who believed they were more likely to be HIV positive may have been more likely to be introduced to PMTCT (or the concern about MTCT) through a local health clinic. On the other hand, women who believed they were more likely to be HIV positive may have been more likely to have prior information about MTCT and hence would have had less scope for increasing their awareness. Table 2.2 presents information similar to that in table 2.1, except now I disaggregate the sample means by whether the respondent is in a demographic group with HIV prevalence above the median HIV prevalence by demographic group. For these purposes, I define demographic group as Table 2.1

Sample All adult females Observations

Knowledge of MTCT by access to prevention of mother-to-child transmission of HIV Greater than 20 km from eventual PMTCT site (1)

Within 20 km of eventual PMTCT site Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

0.84 3,992

0.86 6,063

0.93 7,596

0.07 —

0.00 —

Sources: Data on knowledge of mother-to-child transmission (MTCT) come from the 2001 and 2007 DHS survey rounds and the 2003 and 2005 ZSBS survey rounds.

Table 2.2

Sample HIV prevalence Below median Median or above

Knowledge of MTCT by HIV prevalence and access to prevention of mother-to-child transmission of HIV Greater than 20 km from eventual PMTCT site (1)

Within 20 km of eventual PMTCT site Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

0.83 0.86

0.84 0.89

0.91 0.95

0.07 0.06

0.00 0.00

Sources: Data on knowledge of mother-to-child transmission (MTCT) come from the 2001 and 2007 DHS survey rounds and the 2003 and 2005 ZSBS survey rounds.

70

Nicholas Wilson

the interaction of gender, five- year age group, and province of residence. Data on HIV prevalence come from a single cross- section, the 2007 Demographic Health Survey (DHS). This table shows that knowledge of MTCT was indeed higher among women in high HIV- prevalence demographic groups. Moreover, MTCT knowledge rates approaching the high nineties for women in high HIV- prevalence demographic groups suggest that the scope for increased knowledge was limited. Consistent with this observation, the increase in MTCT knowledge among respondents in (eventual) PMTCT locations associated with the local introduction of PMTCT was smaller among women in high HIV- prevalence demographic groups than among women in low HIV- prevalence demographic groups (i.e., 0.06 versus 0.07 percentage points). Social networks likely vary systematically across age groups, suggesting that social learning about MTCT through local PMTCT availability may vary by age as well. Table 2.3 explores this possibility, by further disaggregating knowledge about MTCT by the respondent’s age group. These results suggest that the increase in knowledge of MTCT among respondents in (eventual) PMTCT locations associated with the local introduction of PMTCT was greater among younger age groups. For example, the increase among women age fifteen to nineteen in (eventual) PMTCT locations was approximately 11 percentage points, or roughly twice that for women age twenty to twenty- nine. Again, high prior knowledge of MTCT among older respondents might explain the smaller increase among older age groups. Finally, I examine how the association between PMTCT availability and knowledge of MTCT varied by the education level of the respondent. Table 2.4 represents the results of this analysis. There do not appear to be particularly large differences by education level in the change in knowledge of MTCT associated with the local introduction of PMTCT. The proportion of women who reported knowing about MTCT increased by between 2 and Table 2.3

Knowledge of MTCT by age and access to prevention of mother-to-child transmission of HIV

Sample

Greater than 20 km from eventual PMTCT site (1)

Within 20 km of eventual PMTCT site Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

Age 15–19 20–29 30–39 40–49

0.73 0.85 0.87 0.88

0.77 0.88 0.90 0.88

0.88 0.94 0.96 0.95

0.11 0.05 0.06 0.06

0.00 0.00 0.00 0.00

Sources: Data on knowledge of mother-to-child transmission (MTCT) come from the 2001 and 2007 DHS survey rounds and the 2003 and 2005 ZSBS survey rounds.

Mother-to-Child Transmission of HIV and Reproductive Behavior Table 2.4

71

Knowledge of MTCT by education and access to prevention of mother-to-child transmission of HIV

Sample Education level Did not complete primary Completed primary Completed secondary

Greater than 20 km from eventual PMTCT site (1)

Within 20 km of eventual PMTCT site Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

0.81 0.91 0.89

0.82 0.90 0.96

0.90 0.94 0.98

0.08 0.04 0.02

0.00 0.00 0.09

Sources: Data on knowledge of mother-to-child transmission (MTCT) come from the 2001 and 2007 DHS survey rounds and the 2003 and 2005 ZSBS survey rounds.

8 percentage points in locations receiving PMTCT after the local introduction of PMTCT, regardless of the respondent’s education level. Although more educated women may be more knowledgeable about MTCT even in locations never receiving PMTCT, even these differences do not exhibit a consistent pattern of increased education associated with increased knowledge of MTCT. However, the small number of women in locations never receiving PMTCT who have completed secondary school means we should interpret these simple differences in means with caution. 2.6.2

Knowledge of PMTCT

Now I turn to the question of whether knowledge about prevention of mother- to-child transmission (PMTCT) increased in areas receiving PMTCT. Table 2.5 shows the proportion of female respondents who reported knowing of PMTCT, disaggregated by (eventual) proximity to PMTCT. In locations eventually receiving PMTCT, the change in knowledge of PMTCT associated with the local introduction of the service was nearly 50 percentage points ( p- value = 0.00). Much of this may be a secular change that was shared by individuals in locations greater than twenty kilometers from an eventual PMTCT site. However, the fact that only 25 percent of respondents in locations greater than twenty kilometers from an eventual PMTCT were aware of PMTCT also suggests that the local introduction of this service had a large impact on knowledge of the existence of an intervention aimed at reducing mother- to-child transmission of HIV. Table 2.6 allows average knowledge of PMTCT to vary by HIV prevalence in the respondent’s demographic group. The results suggest that the local introduction of PMTCT did not increase knowledge of its existence substantially more among women who were more likely to be HIV positive. The proportion of women aware of the existence of PMTCT increased by approximately 47 percentage points in low and high HIV- prevalence demographic groups alike.

72

Nicholas Wilson

Table 2.5

Sample All adult females Observations

Knowledge of PMTCT by access to prevention of mother-to-child transmission of HIV Greater than 20 km from eventual PMTCT site (1)

Within 20 km of eventual PMTCT site Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

0.21 3,833

0.11 5,989

0.59 7,287

0.48 —

0.00 —

Sources: Data on knowledge of prevention of mother-to-child transmission (MTCT) come from the 2007 DHS survey rounds and the 2003 and 2005 ZSBS survey rounds. Note: Knowledge of PMTCT in the 2001 DHS survey round is assumed to be zero.

Table 2.6

Sample HIV prevalence Below median Median or above

Knowledge of PMTCT by HIV prevalence and access to prevention of mother-to-child transmission of HIV Greater than 20 km from eventual PMTCT site (1)

Within 20 km of eventual PMTCT site Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

0.21 0.21

0.10 0.12

0.57 0.60

0.47 0.48

0.00 0.00

Sources: Data on knowledge of prevention of mother-to-child transmission (MTCT) come from the 2007 DHS survey rounds and the 2003 and 2005 ZSBS survey rounds. Note: Knowledge of PMTCT in the 2001 DHS survey round is assumed to be zero.

Partly because the change in knowledge of MTCT associated with local PMTCT varied by age, it is reasonable to believe that the change in knowledge of PMTCT did as well. Table 2.7 examines this possibility. The results indicate that the increase in knowledge of PMTCT associated with local PMTCT introduction was greater among women who were in age groups that were already more likely to know about PMTCT. For example, the proportion of women age fifteen to nineteen in areas eventually receiving PMTCT who were knowledgeable about PMTCT increased by 42 percentage points, whereas the increase among women age twenty to twenty- nine in the same areas was 50 percentage points. Table 2.8 examines whether the change in knowledge of PMTCT associated with local introduction of the service varied by the education level. As was the pattern with MTCT, there were only small- to-moderate differences by education level in the change in knowledge of PMTCT.

73

Mother-to-Child Transmission of HIV and Reproductive Behavior Table 2.7

Knowledge of PMTCT by age and access to prevention of motherto-child transmission of HIV

Sample

Greater than 20 km from eventual PMTCT site (1)

Within 20 km of eventual PMTCT site Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

Age 15–19 20–29 30–39 40–49

0.16 0.25 0.22 0.20

0.07 0.11 0.12 0.12

0.49 0.61 0.65 0.59

0.42 0.50 0.53 0.48

0.00 0.00 0.00 0.00

Sources: Data on knowledge of prevention of mother-to-child transmission (MTCT) come from the 2007 DHS survey rounds and the 2003 and 2005 ZSBS survey rounds. Note: Knowledge of PMTCT in the 2001 DHS survey round is assumed to be zero. Table 2.8

Knowledge of PMTCT by education and access to prevention of mother-to-child transmission of HIV

Sample

Greater than 20 km from eventual PMTCT site (1)

Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

0.27 0.32 0.36

0.12 0.11 0.15

0.59 0.67 0.66

0.46 0.56 0.50

0.00 0.00 0.00

Education level Did not complete primary Completed primary Completed secondary

Within 20 km of eventual PMTCT site

Sources: Data on knowledge of prevention of mother-to-child transmission (MTCT) come from the 2007 DHS survey rounds and the 2003 and 2005 ZSBS survey rounds. Note: Knowledge of PMTCT in the 2001 DHS survey round is assumed to be zero.

2.6.3

Child Mortality

The local introduction of PMTCT may have affected knowledge about MTCT and PMTCT, even if respondents accessed these services at low rates. Without data on use of PMTCT, the best information on access to PMTCT is information on child mortality. If respondents were receiving PMTCT, then presumably we should see a reduction in child mortality in these data. Table 2.9 reports under- age- one child mortality rates, disaggregated by (eventual) proximity to PMTCT. The local introduction of PMTCT was associated with a 1 percentage point (i.e., 10 percent) reduction in underage- one child mortality rates (significant at the 5 percent level). Table 2.10 disaggregates child mortality rates by whether the respondent

74

Nicholas Wilson

(i.e., the child’s mother) was more or less likely to be HIV positive. Child mortality appears to have fallen by 1 percentage point, invariant of the likelihood the mother was HIV positive. Table 2.11 examines whether the association between PMTCT availability and child mortality varied by the age of the respondent (i.e., the child’s Table 2.9

Sample All adult females Observations

Under-one-year child mortality by access to prevention of motherto-child transmission of HIV Greater than 20 km from eventual PMTCT site (1)

Within 20 km of eventual PMTCT site Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

0.11 4,077

0.10 3,826

0.09 8,167

–0.01 —

0.05 —

Sources: Child mortality data come from the 2001 and 2007 DHS survey rounds.

Table 2.10

Sample HIV prevalence Below median Median or above

Under-one-year child mortality by HIV prevalence and access to prevention of mother-to-child transmission of HIV Greater than 20 km from eventual PMTCT site (1)

Within 20 km of eventual PMTCT site Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

0.11 0.10

0.11 0.10

0.10 0.09

–0.01 –0.01

0.32 0.18

Source: Child mortality data come from the 2001 and 2007 DHS survey rounds. Table 2.11

Under-one-year child mortality by age and access to prevention of mother-to-child transmission of HIV

Sample

Greater than 20 km from eventual PMTCT site (1)

Within 20 km of eventual PMTCT site Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

Age 15–19 20–29 30–39 40–49

0.15 0.11 0.10 0.11

0.15 0.11 0.09 0.11

0.09 0.10 0.08 0.09

–0.05 –0.01 –0.01 –0.01

0.11 0.25 0.57 0.43

Sources: Child mortality data come from the 2001 and 2007 DHS survey rounds.

Mother-to-Child Transmission of HIV and Reproductive Behavior

Table 2.12

75

Under-one-year child mortality by education and access to prevention of motherto-child transmission of HIV

Sample Education level Did not complete primary Completed primary Completed secondary

Greater than 20 km from eventual PMTCT site (1)

Within 20 km of eventual PMTCT site Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

0.11 0.09 0.07

0.11 0.10 0.10

0.10 0.09 0.07

–0.01 –0.01 –0.03

0.27 0.32 0.23

Source: Child mortality data come from the 2001 and 2007 DHS survey rounds.

mother). Child mortality appears to have fallen the most for the youngest respondents. Child mortality fell by 5 percentage points for women age fifteen to nineteen. In contrast, it fell by 1 percentage point for all other age groups. Table 2.12 disaggregates the change in child mortality associated with local PMTCT availability by the education level of the mother. There is some evidence of a larger reduction in child mortality for more educated women. 2.6.4

Pregnancy

The availability of PMTCT changes the incentives that women and couples face in making decisions about reproductive behavior. In the standard Becker and Lewis (1973) model, PMTCT simultaneously reduces the shadow prices of child quantity and child quality. These price changes would affect fertility decisions. Similarly, PMTCT should reduce the need for replacement or precautionary fertility. Table 2.13 examines changes in pregnancy rates associated with the local introduction of PMTCT. Although pregnancy rates were already lower in locations eventually receiving PMTCT than in locations never receiving PMTCT, pregnancy rates fell by an additional 5 percentage points ( p- value = 0.00) in association with the local introduction of PMTCT. Presumably the reduction in pregnancy rates associated with the local introduction of PMTCT should have been larger among women who were more likely to be HIV positive. Table 2.14 explores this possibility by further disaggregating pregnancy rates by whether the respondent was in a demographic group with HIV prevalence above the median. Perhaps surprisingly, these results do not suggest that the response was larger among women who were more likely to be HIV positive. Table 2.15 allows pregnancy rates to vary by the age of the respondent. The results indicate the reduction in pregnancy rates associated with the

Table 2.13

Pregnancy rates by access to prevention of mother-to-child transmission of HIV

Sample

Greater than 20 km from eventual PMTCT site (1)

Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

0.37 4,073

0.32 6,125

0.28 7,641

–0.05 —

0.00 —

All adult females Observations

Within 20km of eventual PMTCT site

Sources: Pregnancy data come from the 2001 and 2007 DHS survey rounds and the 2003 and 2005 ZSBS survey rounds. Table 2.14

Pregnancy rates by HIV prevalence and access to prevention of motherto-child transmission of HIV

Sample

Greater than 20 km from eventual PMTCT site (1)

Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

0.37 0.38

0.32 0.32

0.27 0.28

–0.05 –0.04

0.00 0.00

HIV prevalence Below median Median or above

Within 20 km of eventual PMTCT site

Sources: Pregnancy data come from the 2001 and 2007 DHS survey rounds and the 2003 and 2005 ZSBS survey rounds. The HIV data come from 2007 DHS. Note: The HIV prevalence refers to HIV prevalence in the respondent’s demographic group as defined by the interaction of gender, five-year age group, and province of residence. Table 2.15

Pregnancy rates by age and access to prevention of mother-to-child transmission of HIV

Sample

Greater than 20 km from eventual PMTCT site (1)

Within 20 km of eventual PMTCT site Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

Age 5–19 20–29 30–39 40–49

0.26 0.49 0.41 0.15

0.23 0.44 0.35 0.10

0.18 0.38 0.31 0.07

–0.05 –0.05 –0.04 –0.04

0.00 0.00 0.01 0.00

Sources: Pregnancy data come from the 2001 and 2007 DHS survey rounds and the 2003 and 2005 ZSBS survey rounds.

Mother-to-Child Transmission of HIV and Reproductive Behavior Table 2.16

77

Pregnancy rates by education and access to prevention of mother-to-child transmission of HIV

Sample Education level Did not complete primary Completed primary Completed secondary

Greater than 20 km from eventual PMTCT site (1)

Within 20 km of eventual PMTCT site Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

0.39 0.33 0.33

0.37 0.28 0.20

0.34 0.25 0.19

–0.02 –0.03 –0.01

0.09 0.01 0.65

Sources: Pregnancy data come from the 2001 and 2007 DHS survey rounds and the 2003 and 2005 ZSBS survey rounds.

local introduction of PMTCT was concentrated among younger women (i.e., ages fifteen to nineteen and ages twenty to twenty- nine). These groups demonstrated roughly 5 percentage point reductions each in pregnancy rates ( p- value = 0.00 and 0.00), whereas older age groups demonstrated slightly smaller reductions in pregnancy rates. The change in pregnancy rates associated with the local introduction of PMTCT may have varied by the education level of the mother. If more educated women were better able to access this service, then they may have demonstrated a larger response. On the other hand, pregnancy rates were higher at lower levels of education so less educated women may mechanically have greater scope for reducing pregnancy rates. Table 2.16 provides evidence on whether the change in pregnancy associated with the local introduction of PMTCT varied by the education level of the respondent. The results indicate that women who had not completed primary school reduced their likelihood of being pregnancy by 2 percentage points, roughly twice as large as most educated women. Furthermore, the reduction for secondary school completers was not statistically significant at conventional levels ( p- value = 0.65). The simple difference in means was roughly the same for primary school completers and for women who had not completed primary school. 2.6.5

Breastfeeding

The PMTCT expansion likely affected breastfeeding behavior as well. There are at least three reasons to think that breastfeeding habits might have changed after the local introduction of PMTCT. First, PMTCT appears to have reduced pregnancy rates and breastfeeding is an important contraceptive mechanism in Zambia. Among sexually active females age fifteen to forty- nine in the 2007 DHS, approximately 4 percent report using the lacta-

78

Nicholas Wilson

Table 2.17

Sample All adult females Observations

Breastfeeding rates by access to prevention of mother-to-child transmission of HIV Greater than 20 km from eventual PMTCT site (1)

Within 20 km of eventual PMTCT site Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

0.84 1,511

0.83 1,745

0.77 2,264

–0.05 —

0.00 —

Source: Breastfeeding data come from the 2001 and 2007 DHS survey rounds.

tional amenorrhea method (Central Statistical Office et al. 2009). Because the denominator in this calculation includes many women who do not have infants of breastfeeding age, this understates the true prevalence of breastfeeding as a contraceptive method. Second, PMTCT reduces the shadow price of child quality, which should induce households to increase investment in children’s human capital (e.g., health and education). Because PMTCT increases the likelihood a newborn will survive into older ages, household investments in children’s human capital made at a young age (e.g., breastfeeding) are more likely to realize a greater return. Thus, PMTCT should increase breastfeeding rates as part of an overall increase in household investment in children’s human capital. Third, PMTCT directly reduces the expected cost of breastfeeding. Breastfeeding is one of the three stages at which a mother may transmit HIV to her child and comprises roughly one- half of the cumulative transmission probability (Dabis and Ekpini 2002). The PMTCT reduces the probability of transmission through breastfeeding and hence should increase breastfeeding rates among women receiving the service. Table 2.17 investigates the relationship between PMTCT availability and breastfeeding behavior among women with children ages birth to twentyfour months.6 A simple comparison of conditional means suggests that PMTCT availability was associated with a small decrease in breastfeeding. Breastfeeding rates were approximately 5 percentage points lower after the local introduction of PMTCT (significant at the 1 percent level). In table 2.18, I allow the association between PMTCT availability and breastfeeding to vary by the likelihood the respondent was HIV positive. Breastfeeding rates appear to have fallen by 5 percentage points (significant at the 1 percent level) invariant of the likelihood the respondent was HIV positive. Breastfeeding behavior and the change therein associated with PMTCT 6. Approximately twenty months is the median breastfeeding duration in Zambia (Central Statistical Office et al. 2009).

79

Mother-to-Child Transmission of HIV and Reproductive Behavior

expansion may vary by age of the mother. Differences by cohort in education levels or in exposure to particular reproductive health policies might generate differences in the breastfeeding response to PMTCT. Likewise, life cycle differences in labor supply might condition the responsiveness of breastfeeding to PMTCT. Table 2.19 explores these possibilities by disaggregating the change in breastfeeding by the age of the mother. Breastfeeding appears to have fallen more among women in the middle of the age distribution. For women ages twenty to twenty- nine and thirty to thirty- nine, breastfeeding rates fell by 5 to 7 percentage points (significant at the 1 percent level). In contrast, the simple difference in means suggests that breastfeeding rates fell by 2 to 3 percentage points for women ages fifteen to nineteen and forty to forty- nine, although these changes are not statistically significant for these groups ( p- values = 0.29 and 0.73, respectively). Finally, table 2.20 presents breastfeeding rates in PMTCT and nonPMTCT locations further disaggregated by the education level of the mother. Breastfeeding rates appear to have fallen more among women who

Table 2.18

Breastfeeding rates by HIV prevalence and access to prevention of mother-to-child transmission of HIV

Sample

Greater than 20 km from eventual PMTCT site (1)

Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

0.84 0.84

0.84 0.81

0.79 0.76

–0.05 –0.05

0.01 0.00

HIV prevalence Below median Median or above

Within 20 km of eventual PMTCT site

Source: Breastfeeding data come from the 2001 and 2007 DHS survey rounds.

Table 2.19

Breastfeeding rates by age and access to prevention of mother-to-child transmission of HIV

Sample

Greater than 20 km from eventual PMTCT site (1)

Within 20 km of eventual PMTCT site Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

Age 15–19 20–29 30–39 40–49

0.88 0.83 0.85 0.80

0.84 0.81 0.85 0.80

0.81 0.76 0.78 0.78

–0.04 –0.05 –0.07 –0.02

0.29 0.00 0.00 0.73

Source: Breastfeeding data come from the 2001 and 2007 DHS survey rounds.

80 Table 2.20

Nicholas Wilson Breastfeeding rates by education and access to prevention of mother-to-child transmission of HIV

Sample Education level Did not complete primary Completed primary Completed secondary

Greater than 20 km from eventual PMTCT site (1)

Within 20 km of eventual PMTCT site Before local introduction (2)

After local introduction (3)

Change (4)

p-value (5)

0.85 0.82 0.91

0.84 0.81 0.76

0.80 0.76 0.68

–0.04 –0.05 –0.08

0.02 0.02 0.23

Source: Breastfeeding data come from the 2001 and 2007 DHS survey rounds.

had completed more schooling, although the change for women who had completed secondary school was not statistically significant at conventional levels. 2.7

Discussion and Conclusion

Prevention of mother- to-child transmission of HIV (PMTCT) is the single- most effective HIV prevention intervention we know of in practice today. The PMTCT reduces the probability a HIV- positive mother transmits the virus to her child from as much as 45 percent in the absence of PMTCT to as little as 3 percent (Dabis and Ekpini 2002; Canning 2006). During the past ten years or so, the proportion of HIV- positive pregnant women in sub-Saharan Africa receiving PMTCT increased from virtually zero to more than one- half (WHO 2010). This chapter documents the rapid expansion of PMTCT in Zambia over the period 2000– 2007 and provides descriptive evidence on the association between PMTCT expansion and reproductive behavior. I use a newly assembled monthly health facilities panel identifying the expansion of access to the three main HIV/AIDS services, including PMTCT. In conjunction with four nationally representative household surveys spanning this time period, these data allow me to examine the change in reproductive knowledge and behavior associated with the local introduction of PMTCT. I find that local PMTCT introduction was associated with: (a) increased knowledge about mother- to-child transmission (MTCT) and about PMTCT, (b) reduced child mortality and pregnancy rates, and (c) a small reduction in breastfeeding rates. The reduction in pregnancy rates suggests that not only has PMTCT expansion been a humanitarian success, but it may have reduced total fertility rates, possibly promoting economic growth and development. These results are partly consistent with the standard Becker and Lewis

Mother-to-Child Transmission of HIV and Reproductive Behavior

81

(1973) model of fertility. In the Becker model, a fall in child mortality directly reduces the shadow price of child quality because household investments in children’s human capital are more likely to realize a return as the child is more likely to survive into older ages. Through the quantity- quality trade- off embedded in this model, fertility should fall in response to the decrease in the shadow price of child quality. It is somewhat puzzling that breastfeeding rates appear to have decreased, as a mortality- induced fertility reduction should be accompanied by an increase in children’s human capital investment. Future research is required to address the endogeneity of PMTCT expansion. For example, PMTCT expansion may have occurred with greater intensity in areas with fundamentally different time trends in fertility than the rest of Zambia, and the empirical methodology used in the current analysis does not address this concern. In addition, the data assembled for this project allow for the investigation of several other important research questions about PMTCT expansion and reproductive health, including the interaction between PMTCT and ART availability in the process of determining reproductive behavior.

References Angeles, Luis. 2010. “Demographic Transitions: Analyzing the Effects of Mortality on Fertility.” Journal of Population Economics 23:99– 120. Barro, Robert, and Gary Becker. 1989. “Fertility Choice in a Model of Economic Growth.” Econometrica 57 (2): 481– 501. Becker, Gary, and H. Gregg Lewis. 1973. “On the Interaction between the Quantity and Quality of Children.” Journal of Political Economy 81 (2, part 2): S279– 88. Ben-Porath, Yoram. 1976. “Fertility Response to Child Mortality: Micro Data from Israel.” Journal of Political Economy 84 (4, part 2): S163–S178. Brahmbhatt, Heena, Fred Wabwire-Mangen, Godfrey Kigozi, and Ronald Gray. 2001. “Associations of Maternal HIV and Child Survival in Rakai, Uganda.” Third Conference on Global Strategies for the Prevention of HIV Transmission from Mothers to Infants (abstract 59). Kampala, Uganda, September 9– 13. Canning, David. 2006. “The Economics of HIV/AIDS in Low-Income Countries: The Case for Prevention.” Journal of Economic Perspectives 20 (3): 121– 42. Carpenter, Lucy, Jessica Nakiyingi, Anthony Ruberantwari, Samuel Malamba, Anatoli Kamali, and James Whitworth. 1997. “Estimates of the Impact of HIV Infection on Fertility in a Rural Ugandan Population Cohort.” Health Transition Review 7 (S2): 113– 26. Central Statistical Office, Ministry of Health, Tropical Diseases Research Centre, University of Zambia, and Macro International Inc. 2009. Zambia Demographic and Health Survey 2007. Calverton, MD: CSO and Macro International Inc. Chen, W. J., and N. Walker. 2010. “Fertility of HIV-Infected Women: Insights from Demographic and Health Surveys.” Sexually Transmitted Infections 86 (Suppl. 2): ii22– 27.

82

Nicholas Wilson

Cooper, Diane, Jane Harries, Landon Myer, Phyllis Orner, and Hillary Bracken. 2007. “Life is Still Going On: Reproductive Intentions among HIV-Positive Women and Men in South Africa.” Social Science and Medicine 65:274– 83. Dabis, Francois, and Ehounou Rene Ekpini. 2002. “HIV- 1/AIDS and Maternal and Child Health in Africa.” Lancet 359:2097– 104. Dabis, Francois, Narcisse Elenga, Nicolas Meda, Valeriane Leroy, Ida Viho, Olivier Manigart, Laurence Dequae-Merchadou, Phillipe Msellati, and Issiaka Sombie. 2001. “18-Month Mortality and Perinatal Exposure to Zidovudine in West Africa.” AIDS 15:771– 79. Doepke, Matthias. 2005. “Child Mortality and Fertility Decline: Does the BarroBecker Model Fit the Facts?” Journal of Population Economics 18:337– 66. Fabiani, Massimo, Barbara Nattabi, Emingtone Ayella, Martin Ogwang, and Silvia Declich. 2006. “Differences in Fertility by HIV Serostatus and Adjusted HIV Prevalence Data from an Antenatal Clinic in Northern Uganda.” Tropical Medicine and International Health 11 (2): 182– 87. Fink, Gunther, and Sebastian Linnemayr. 2009. “HIV Does Matter for Fertility: Human Capital, Mortality, and Family Size.” Working Paper, Harvard University. Fortson, Jane. 2009. “HIV/AIDS and Fertility.” American Economic Journal: Applied Economics 1 (3): 170– 94. Glynn, Judith, Anne Buve, Michel Carael, Maina Kahindo, Isaac Macauley, Rosemary Musonda, Eva Jungmann, Francis Tembo, and Leopold Zekeng. 2000. “Decreased Fertility among HIV- 1-Infected Women Attending Antenatal Clinics in Three African Cities.” Journal of Acquired Immune Deficiency Syndromes 25 (4): 345– 52. Gray, R. H., M. J. Wawer, D. Serwadda, N. Sewankambo, C. Li, F. Wabwire-Mangen, L. Paxton, N. Kiwanuka, G. Kigozi, J. Konde-Lule, and T. C. Quinn. 1998. “Population-Based Study of Fertility in Women with HIV- 1 Infection in Uganda.” Lancet 351 (9096): 98– 103. Hoffman, Irving, Francis Martinson, Kimberly Powers, David Chilongozi, Emmie Msiska, Emma Kachipapa, Chimwemwe Mphande, Mina Hosseinipour, Harriet Chanza, Rob Stephenson, and Amy Tsui. 2008. “The Year-Long Effect of HIVPositive Test Results on Pregnancy Intentions, Contraceptive Use, and Pregnancy Incidence among Malawian Women.” Journal of Acquired Immune Deficiency Syndromes 47 (4): 477– 83. Joint United Nations Programme on HIV/AIDS (UNAIDS). 2010. UNAIDS Report on the Global AIDS Epidemic. Geneva: UNAIDS. http://www.unaids.org/global report/Global_report.htm. Juhn, Chinhui, Sebnem Kalemi-Ozcan, and Belgi Turan. 2009. “HIV and Fertility in Africa: First Evidence from Population Based Surveys.” NBER Working Paper no. 14248, Cambridge, MA. Kalemli-Ozcan, Sebnem. 2012. “AIDS, ‘Reversal’ of the Demographic Transition and Economic Development: Evidence from Africa.” Journal of Population Economics 25 (3): 871– 97. Kalemli-Ozcan, Sebnem, and Belgi Turan. 2011. “HIV and Fertility Revisited.” Journal of Development Economics 96 (1): 61– 65. Kongnyuy, Eugene, and Charles Wiysonge. 2008. “Association between Fertility and HIV Status: What Implications for HIV Estimates?” BMC Public Health 8:309. Peltzer, Karl, Li-Wei Chao, and Pelisa Dana. 2009. “Family Planning among HIV Positive and Negative Prevention of Mother to Child Transmission (PMTCT) Clients in a Resource Poor Setting in South Africa.” AIDS Behavior 13:973– 79. Soares, Rodrigo. 2005. “Mortality Reductions, Educational Attainment, and Fertility Choice.” American Economic Review 95 (3): 580– 601.

Mother-to-Child Transmission of HIV and Reproductive Behavior

83

Spira, Rosemary, Phillipe Lepgage, Phillipe Msellati, Phillipe Van d Perre, Valeraine Leroy, Arlette Simonon, Etienee Karita, and Francois Dabis. 1999. “Natural History of HIV Type 1 Infection in Children: A Five-Year Prospective Study in Rwanda.” Pediatrics 104:D1–D9. Stekelenburg, J., S. Kyanamina, M. Mukelabai, I. Wolffers, and J. van Roosmalen. 2004. “Waiting Too Long: Low Use of Maternal Health Services in Kalabo, Zambia.” Tropical Medicine and International Health 9 (3): 390– 98. Stringer, Elizabeth, Moses Sinkala, Jeffrey Stringer, Elizabeth Mzyece, Ida Makuka, Robert Goldenberg, Pascal Kwane, Martha Chilufya, and Sten Vermund. 2003. “Prevention of Mother- to-Child Transmission of HIV in Africa: Successes and Challenges in Scaling-Up a Nevirapine-Based Program in Lusaka, Zambia.” AIDS 17 (9): 1377– 82. Stringer, Jeffrey, Moses Sinkala, Courtney Maclean, Jens Levy, Chipepo Kankasa, Alain DeGroot, Elizabeth Stringer, Edward Acosta, Robert Goldenberg, and Sten Vermund. 2005. “Effectiveness of a City-Wide Program to Prevent Motherto-Child HIV Transmission in Lusaka, Zambia.” AIDS 19 (12): 1309– 15. Taha, Taha, Newton Kumwenda, Robin Broadhead, Donald Hoover, Diane Markakis, Len van der Hoeven, George Liomba, John Chiphangwi, and Paolo Miotti. 1999. “Mortality after the First Year of Life among Human Immunodeficiency Virus Type 1-Infected and Uninfected Children.” Pediatric Infectious Disease Journal 18:689– 94. World Health Organization (WHO). 2006. Antiretroviral Drugs for Treating Pregnant Women and Preventing HIV Infection in Tnfants: Towards Universal Access. Geneva: World Health Organization. ———. 2007. Toward Universal Access: Scaling Up Priority HIV/AIDS Interventions in the Health Sector. Geneva: World Health Organization. ———. 2010. Toward Universal Access: Scaling Up Priority HIV/AIDS Interventions in the Health Sector. Geneva: World Health Organization. Wolpin, Kenneth. 1984. “An Estimable Dynamic Stochastic Model of Fertility and Child Mortality.” Journal of Political Economy 92 (5): 852– 74. Young, Alwyn. 2007. “In Sorrow to Bring Forth Children.” Journal of Economic Growth 12:283– 327. Zaba, Basia, and Simon Gregson. 1998. “Measuring the Impact of HIV on Fertility in Africa.” AIDS 12 (S1): S41–S50.

3

Stimulating Demand for AIDS Prevention Lessons from the RESPECT Trial Damien de Walque, William H. Dow, Carol Medlin, and Rose Nathan

3.1

Introduction

Conventional approaches to HIV prevention have been important in educating populations about HIV risk factors and risk- reduction strategies, and many have been found to be cost effective, but by themselves they have had limited success in slowing the AIDS epidemic. Over the past decade the annual number of new infections has decreased by about one- fifth, but in 2009 there were still an estimated 2.6 million new HIV infections, and 1.8  million of these were in sub-Saharan Africa (UNAIDS 2010). HIV experts have highlighted the need for new combination approaches to complement existing prevention efforts (Coates, Richter, and Caceres 2008; Padian et al. 2011), including efforts to stimulate individual demand for prevention. Damien de Walque is a senior economist in the Development Research Group (Human Development and Public Services Team) at the World Bank. William H. Dow is the Henry J. Kaiser Professor of Health Economics at the University of California, Berkeley, and a research associate of the National Bureau of Economic Research. Carol Medlin is a director of health at the Children’s Investment Fund Foundation (CIFF). Rose Nathan is a senior research scientist at the Ifakara Health Institute. We have benefited greatly from the insights and efforts of the RESPECT Project Team and many others who contributed to the intellectual development and success of the RESPECT project. We are also grateful to Adam Wagstaff for useful comments. We acknowledge funding support for this chapter from the NBER Africa Project, and support for the overall RESPECT project from the William and Flora Hewlett Foundation (through the Population Reference Bureau), the World Bank Research Committee, and the Spanish Impact Evaluation Fund and the Knowledge for Change Program (KCP) managed by the World Bank. The findings, interpretations, and conclusions expressed in this chapter are entirely those of the authors, and do not represent the views of any of the authors’ institutions or funders. For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see http://www.nber.org/chapters/c13375.ack.

85

86

Damien de Walque, William H. Dow, Carol Medlin, and Rose Nathan

This chapter considers the use of economic incentives to encourage people to engage in behavioral change strategies that reduce risky sexual behaviors. Incentives have been widely used in a variety of related domains, but are only recently being explored as a possible HIV prevention strategy. One type of widely adopted incentive scheme in the health domain is as part of conditional cash transfer (CCT) programs such as the wellknown Oportunidades program in Mexico, which provides significant cash welfare support to households that engage in specific positive behaviors, including appropriate prenatal and well- child care. Lagarde, Haines, and Palmer (2007) and Fiszbein and Schady (2009) survey such programs in detail. Related programs are being explored in the sub-Saharan African context, but with conditionality focused on nonhealth behaviors such as schooling (which can then indirectly yield positive health consequences, e.g., Baird et al. [2010], [2012]). There is also a history of providing positive price incentives for using specific types of preventive health services, most relevant in this context being the use of cash incentives to promote receipt of HIV test results (Thornton 2008). The aforementioned incentive programs are all conditioned on observable input behaviors, which can be controlled by participants. A parallel set of “contingency management” programs have been designed for situations in which behavior changes are more complex and/or not easily observed, and in which conditionality is instead based on outcomes. This contingency management approach has been most widely adopted in the substance abuse field, for example, providing rewards to substance abuse treatment patients if they are negative on random drug tests; such programs have been found in developed country contexts to be feasible, acceptable, and effective (Stitzer 2006). In this chapter we discuss a variant of the above incentive programs, using outcome- based incentives adapted to the sexual domain (possibly within the context of a larger CCT program). After reviewing traditional HIV- prevention strategies as well as what is known from prior literature on incentives in the types of contexts discussed above (drawn partly from Medlin and de Walque [2008]), we discuss the theoretical pathways by which incentive programs could work to reduce risky sexual behavior. We then turn to a specific example in detail, elaborating on one such project in Tanzania (the RESPECT project, of which the authors are the principal investigators), which has implemented a novel randomized trial of outcomebased incentives to reduce risky sexual behavior.1 The RESPECT project was a one- year intervention in which participants were tested for curable 1. We are only aware of one other study that has attempted to similarly incent outcomes associated with risky sexual behaviors. Kohler and Thornton (2011) report on a trial in Malawi in which HIV- negative individuals were randomized to cash reward levels to be paid one year later conditional on remaining HIV negative. In contrast to the RESPECT study, which found significant reductions in STIs, the Malawi study found null results on HIV conversion (the main measured outcome), although there was low power to detect HIV changes in that study.

Stimulating Demand for AIDS Prevention

Fig. 3.1

87

The RESPECT randomized study design

sexually transmitted infections (STIs) at four- month intervals, and received cash rewards if STI negative. The results from the impact evaluation of this intervention are reported in de Walque et al. (2012). Figure 3.1 presents the basic randomized design of the RESPECT study’s 2,399 participants across two treatments arms (with different levels of cash rewards) and a control arm. The balance of the chapter focuses on lessons learned from the RESPECT trial and implications for such efforts in the future. 3.2

Traditional Approaches to HIV Prevention

The social, economic, and human costs of the AIDS epidemic in subSaharan Africa are staggering (Eiss and Glass 2007), and innovative solutions to stem the tide of the epidemic are desperately needed. Globally, an estimated 33 million people were living with HIV in 2009. That same year, an estimated 1.8 million people lost their lives to AIDS, and an estimated 2.6 million became newly infected. The global epicenter of the AIDS pandemic is in Africa, where an estimated 22.5 million people were living with HIV in 2009 including approximately 5 percent of adults, and an estimated 1.8 million new infections occurred during that year (UNAIDS 2010). Young people, ages fifteen to twenty- five, are at particularly high risk of new infection. The tragic reality is that many of these new infections could have been prevented. At its core, the global AIDS epidemic is fueled by risky sexual behavior. Over 80 percent of HIV infections occur through sexual contact with an infected partner (Askew and Berer 2003), and could have been avoided through the adoption of safer sexual behaviors including condom use, reduction in the number and concurrency of sexual partners, or abstinence. Since the early years of the epidemic and beyond, billions of dollars have

88

Damien de Walque, William H. Dow, Carol Medlin, and Rose Nathan

been invested in prevention programs, and a significant portion of these dollars have gone to what are known as “information, education, and communications” (IEC) investments. Unfortunately, IEC by itself has not been shown to have more than a minor impact on patterns of HIV transmission and the trajectory of the epidemic (Bertrand et al. 2006). Numerous studies have shown that information alone is typically insufficient to change risk behavior. Accurate information is indisputably a basic ingredient in informed policy discourse, and information, education, and communication campaigns in conjunction with condom promotion and distribution likely results in higher condom use and significantly lower sexually transmitted infection (STI) incidence (Bertozzi et al. 2006). Behavioral change will need to be an important part of prevention strategies going forward. But the dearth of specific guidance and recommendations from the literature about what can be done to improve the effectiveness of behavioral change interventions, especially as they relate to risky sexual behavior, has been one of the more poignant failures of global response to the epidemic. Certain other key AIDS- prevention strategies do not depend on the widespread adoption of safe sexual practices, but they nevertheless will depend on individual demand decisions to use available technologies. These include HIV and STI testing and treatment, male circumcision, and pharmacological prevention including drugs to stop mother- to-child transmission, vaginal microbicides, prophylactic antivirals, and potentially a future vaccine (Padian et al. 2011). Such clinical and technologically based approaches hold a great deal of promise; however, behavioral strategies must also be pursued. Current levels of investment in the development of new drugs, vaccines, and technologies have been at least partially fueled by the perception that behaviorally focused approaches cannot be made to work, or are politically and socially unviable. However, biomedical advances such as a cure for AIDS or the development of an effective AIDS vaccine are, at best, many years away or unattainable. Moreover, even once a clinic- based intervention, drug, or vaccine has been proven efficacious, changes in behavior are still needed to ensure access, uptake, and acceptability. Consequently, greater attention is now being given to a focus on behavioral change as an important component of prevention strategies and activities. In the Tanzanian context in which the RESPECT trial was fielded, table 3.1 shows that knowledge of AIDS and HIV prevention strategies is extremely high and accurate, with over 99 percent of respondents familiar with HIV and 85 percent identifying condom usage as a strategy to prevent HIV. Table 3.2 further shows that a large portion of young adults have been HIV tested (a national testing campaign occurred two years earlier and testing is standard during prenatal care), and people believe that HIV rates are quite high. The average respondent estimated an HIV- prevalence rate of 17 percent among people their age in their village, indicating a high level of perceived risk. Yet table 3.2 also shows 20 percent of men admitting to

Table 3.1

The RESPECT baseline knowledge of HIV and prevention methods

Heard of AIDS or HIV? Can die from AIDS Knowledge of prevention methods (ABC) AIDS can be prevented by: Just one partner Regular condom use Abstaining Witchcraft Other facts about HIV Healthy- looking person can be HIV infected Not being infected after having sex with HIV positive Child can be infected during pregnancy

Males (N = 1,175) (%)

Females (N = 1,191) (%)

99.8 94.3

99.9 92.5

97.0 85.6 91.3 1.9

97.4 84.2 90.9 4.1

97.2 50.2 81.6

95.6 42.7 88.9

Source: The RESPECT project baseline survey, conducted in 2009 after project recruitment and informed consent, but prior to receiving project counseling.

Table 3.2

The RESPECT baseline HIV test history, status beliefs, and risky behaviors

Ever tested for HIV? If yes, when last test? < 12 months ago 12– 23 months ago >= 2 years ago Received the HIV test results? Perceived HIV prevalence for same age group in community On scale 0– 10 what is your risk of being HIV positive? Number of partners in last 4 months 0 1 2 More than 2 Condom use during last sexual intercourse With spouse or union With other partner

Males (N = 1,175) (%)

Females (N =1,191) (%)

34.8 Males ( N = 409) 45.7 27.4 26.9 90.9 16.0 (N = 1,166) 2.12 (N = 1,175)

71.9 Females (N = 856) 45.0 28.6 26.4 92.3 18.4 (N = 1,175) 2.2 (N = 1,191)

12.2 68.1 15.6 4.2

10.4 86.2 2.7 0.7

15.1 61.3

13.8 49.5

Source: The RESPECT project baseline survey, conducted in 2009 after project recruitment and informed consent, but prior to receiving project counseling.

90

Damien de Walque, William H. Dow, Carol Medlin, and Rose Nathan

multiple partners in the past month (we suspect such statistics are underreported), and close to half of participants failed to use a condom during their last sexual intercourse with a nonmarital partner. These young people appear to understand their HIV risks and know how to behave to prevent transmission—yet they do not choose to act on that knowledge. Next, we review prior experience with incentivizing individuals so as to stimulate demand for prevention. 3.3 3.3.1

Literature Review on Incentives for Health Input Behaviors and Health Outcomes Conditional Cash Transfer Approaches

The CCT programs that provide cash to poor households in exchange for their active participation in educational and health care services have proven remarkably popular among developing country governments, sweeping the globe from Mexico to several other Latin American countries, including Columbia, Honduras, Jamaica, and Nicaragua, and much more recently, to Africa (Kakwani, Soares, and Son 2005; Nigenda and GonzalezRobledo 2005; Schuring 2005). The principle of conditionality—which may be applied differently in practice, but generally requires families to send their children to school or to receive a range of health care services, such as nutritional counseling, childhood vaccination programs, and so forth— distinguishes CCT programs from the more traditional social assistance programs that provide cash or vouchers directly to poor or otherwise distressed families with no strings attached. The CCT programs emphasize the use of market- oriented “demand- side” interventions as an instrument for longer- term human capital investments (Rawlings and Rubio 2005; Fiszbein and Schady 2009). In other words, the cash was intended to function as an incentive on credit- constrained (poor) families to invest in their children’s future, recognizing the powerful limitations that short- term financial constraints placed on poor families. Ideally, they are designed to complement, rather than replace, the more familiar “supply- side” investments, which channel resources directly toward schools, clinics, and service providers. The CCT programs that have received the most attention are those having an explicit orientation toward poverty alleviation, involving both educational and health components as part of a broader, long- term strategy of human capital investments. Such programs have been thoroughly evaluated, and overall the results of these programs have been extremely promising. The evaluations of such programs, some more rigorously designed than others, have generally shown positive impacts on health and education. Mexico’s program has been evaluated most thoroughly. Studies have found increases of 25– 60 percent in health care visits among children under three years of age, higher rates of nutritional monitoring, and higher immunization rates.

Stimulating Demand for AIDS Prevention

91

In addition, caloric intake increased by 7 percent, driven by higher expenditures on fruits, vegetables, and meats. Actual health impacts have also been observed, with 12 percent lower incidence of illness among children from birth to five years old, and a height increase of one centimeter among children ages six to thirty- six months (with the greatest effects in the poorest households with educated fathers; Lagarde, Haines, and Palmer [2007]). However, very little can be surmised about the relationship between the magnitude of the transfer and the behavioral effect it induces due to the lack of experimentation with this issue. Fernald, Gertler, and Hou (2008) and Fernald, Gertler, and Neufeld (2008) have instead used quasi- experimental variation in the magnitude of Mexico’s Oportunidades cash transfer to indicate that more cash is associated with better child anthropometric and cognitive outcomes (though worse adult blood pressure and obesity). Beyond the lack of information related to cash dose response, there is also an intense debate regarding the importance of the conditionality component of such programs. Some have argued that the conditionality component is inappropriate for the African context (Schubert and Slater 2006). Even in the Latin American context, it has been recognized that it is difficult to attribute health impact to the conditionality, per se, as the programmatic intervention has many component parts, and is not limited to the conditionality (Gertler 2004; Lagarde, Haines, and Palmer 2007). In the absence of more definitive research, related programs in several countries have been launched without conditionality, including some in Africa. Several more recent studies are now analyzing conditionality compared to randomized arms with unconditional cash payments, yielding sometimes surprising results. For example, Baird, McIntosh, and Ozler (2011) find that conditionality improves adolescent girl school outcomes in a CCT program in Malawi, but increased teen pregnancy and marriage rates relative to unconditional transfers. More will be known to help interpret these complex relationships as additional studies such as this are released in the near future. Overall, the evidence from an impressive collection of evaluations of CCTs as part of a broader poverty alleviation strategy demonstrates that such programs are feasible and effective. Such programs have demonstrated positive impacts on uptake rates of health (and education) services, and, in some cases, marked improvements in health outcomes (Sridhar and Duffield [2006] review nutritional outcomes, in particular). However, it is important to note that the evidence on how and whether the conditionality works remains weak. For STI/HIV prevention, this is the most critical piece of information that is needed to assess whether similar types of programs can be effective at reducing risky sexual behavior. To explore the role of conditionality in more detail we next turn to contingency management approaches, which typically use much smaller payments that would be unlikely to have income effects on behavior, and thus better isolate the role of the conditionality incentive per se.

92

3.3.2

Damien de Walque, William H. Dow, Carol Medlin, and Rose Nathan

Contingency Management Approaches

Similar to CCTs, contingency management (CM) relies on the mechanism of conditionality to elicit behaviors that are viewed to be in one’s longterm interests (or, those of society’s), and to discourage those behaviors that may be ultimately detrimental to one’s own health and well- being that may not be easily perceived or experienced in the short term. There are several key differences between CCT and CM, though. First, CCT programs typically have much larger cash rewards for complying with conditionality, hence exploit both price and income effects on behavior. Second, CCT programs have typically conditioned on easily monitored input behaviors (such as health care use), whereas CM has been used for behaviors that are harder to monitor directly (such as drug use), and hence CM conditions instead on the desired outcome (negative drug test). Third, while the CCT incentivized inputs (a prenatal care visit) require a “simple” behavioral response over which the individual has a high degree of control, CM incentives often require a “complex” behavioral change (over which the individual may have imperfect control; Kane et al. [2004]). This complexity may require changing multiple behaviors, reversing habit formation and addictive behaviors, and judging uncertainty (such as probability that a behavior will indeed cause a negative test). The CM applications span many areas of risky behaviors, including substance abuse, smoking, and overeating. It has been especially well studied by clinical psychologists as a therapeutic approach to encourage the practice of healthful behaviors and to discourage unhealthy behavioral practices, especially those that may be linked to addiction or other destructive behaviors that are deeply engrained and/or habit forming. The CM interventions provide “reinforcers” (e.g., incentives or rewards) contingent on an individual’s abstinence from a target drug or behavior. The reinforcement device, often cash payments, vouchers, or prizes, is contingent upon an objective measure of a predetermined therapeutic target. An “objective” measure often means a biochemical measure such as urine toxicology testing or the measurement of breath alcohol or carbon monoxide levels instead of self- reported compliance, which is not verifiable. The essential principles of CM, as outlined by Petry (2000), are to reinforce the treatment goals by (a) closely monitoring the target behavior; (b) providing tangible, positive reinforcement of the target behavior; and (c) removing the positive reinforcement when the target behavior does not occur. The CM techniques have been developed and tested in the context of clinical trials and settings, and are a clinically accepted tool in fields such as substance abuse, but have rarely (if ever) been implemented on a large scale in the manner of CCT programs. As with CCTs, CM interventions have been tied to participation and the

Stimulating Demand for AIDS Prevention

93

uptake of services in several domains, although risk behaviors are the important determinant for participant selection, rather than income constraints. The CM has been shown to improve drug abuse outcomes (Rawson et al. 2002) and uptake rates of counseling sessions (Petry et al. 2001); attendance at weight loss sessions; attendance at HIV drop-in center activities (Petry, Martin, and Finocche 2001) and antiviral medication adherence (Rosen et al. 2007); and attendance in smoking cessation clinics (e.g., see Higgins et al. 1994; Petry 2000; Emont and Cummings 1992). Of particular interest, however, is the use of CM to elicit a complex behavioral change—usually, to discourage an unhealthy behavior by positively reinforcing the cessation of that activity (e.g., drug or alcohol abuse, smoking, or overeating). The conceptual basis of CM and CCTs is thus largely similar, although advocates of CM impose no a priori assumptions about the effectiveness of the use of cash as the incentive or reinforcement device, and have experimented with a variety of reward mechanisms, including vouchers and prizes.2 In addition, many CM studies are designed to explore effect differences due to variations in the value of the conditionality (known as the “dose- response” curve), the frequency of monitoring and payments, and the length of time that the elicited behavior change is sustained after the program has ended. The use of CM has been most intensively studied in relation to its efficacy in treating substance abuse. A landmark study by Higgins et al. (1994) demonstrated that incentives delivered contingent on submitting cocainefree urine specimens significantly improved treatment outcomes in ambulatory cocaine- dependent patients. Over 50 percent in the treatment condition achieved at least two months of cocaine abstinence versus only 15 percent of the controls. Silverman et al. (1996) showed that 47 percent of cocaineabusing methadone patients assigned to the CM group achieved more than seven weeks of continuous abstinence, compared to only 6 percent of patients in the control group who achieved more than two weeks of abstinence. Similar results have been found for treating opioid dependency (Petry 2000). While CM has also been shown efficacious in treating alcohol abuse, the studies are fewer in number due to the difficulties associated with objectively verifying abstinence. Breath, urine, and blood tests can detect alcohol use only up to four to eight hours, which means that effective monitoring would have to take place two or three times a day (Stitzer and Petry 2006). Financial incentives to discourage smoking have also been extensively studied. Donatelle et al. (2000) used social support and financial incentives to induce high- risk pregnant smokers to quit during their pregnancies. They provided in the amount of $50 per month for each month of absti2. The findings of studies reviewed for this chapter suggested that cash is typically preferred by research subjects, and in some studies it has been shown to have a greater behavioral effect than the equivalent noncash reward (Kamb et al. 1998; Deren et al. 1994; Vandrey, Bigelow, and Stitzer 2007), although the findings are hardly conclusive.

94

Damien de Walque, William H. Dow, Carol Medlin, and Rose Nathan

nence (up to a maximum ten- month period, which included two months of postpartum). Lab- verified abstinence was required, and the biochemically confirmed quit rates within the treatment group were higher both at eight months and two months postpartum. Stitzer and Bigelow (1983) experimented with different levels of cash payment, providing a payment of $1, $5, or $10 per day for ten days to the three treatment groups (the control group received no cash). The study found that CO levels decreased in an orderly fashion as pay increased. However, another study by Windsor, Lowe, and Bartlett (1988), which provided cash payments of $25 at six weeks and six months as a reward for abstinence, found no difference in cessation rates between the control and treatment groups. Other earlier studies experimenting with prizes, vouchers, and in-kind gifts of free nicotine patches showed mixed results, but even positive results disappeared after six months. More recently, Volpp et al. (2009) found that while an incentive program’s effects on smoking cessation also declined after withdrawal of the incentives, significant effects did still remain three to six months later. The use of financial incentives to treat obesity has also gained in popularity, but the evidence regarding efficacy is decidedly more mixed (see e.g., Follick, Folwer, and Brown 1984; Jeffery, Thompson, and Wing 1978; and Jeffery et al. 1984.) For example, Volpp et al. (2008) found significant weight loss from a lottery- based incentive program, but they were not sustained four months after the program’s end; similarly, John et al. (2011) found matched commitment contracts led to significant weight loss after thirty- six weeks, but again it was not sustained during a thirty- two- week postincentive period. A recent systematic review of randomized controlled trials of treatments for obesity (Paul-Ebhohimhen and Avenell 2008) showed no significant effect of the use of financial incentives on weight loss or maintenance at twelve months and eighteen months. However, further subanalysis indicated that large transfers (greater than 1.2 percent of personal disposable income) had greater impact, as did rewards for behavioral change rather than weight loss, per se, and rewards based on group performance rather than individual results. The CM literature, overall, offers useful insights into aspects of the conditionality that appear to elicit the desired behavior change. This is an important area of inquiry that has not been sufficiently explored within CCT programs. However, unlike CCT programs, studies of CM have remained largely experimental and have not been brought to scale (Petry 2000; Kane et al. 2004). Furthermore, the small sample sizes of study groups—most typically involving groups of 20 to 100, and rarely more than 500—have made it difficult to detect effects that are statistically significant, much less estimate effect sizes accurately. Also, factorial designs with several treatment arms are common which—in combination with already small sample sizes—has led to even more constraints on power (Kane et al. 2004).

Stimulating Demand for AIDS Prevention

3.4

95

Theoretical Pathways for Incentive Effects on Risky Sexual Behaviors

As indicated in the above literature review, there are a variety of theoretical pathways via which incentives could influence risky sexual behaviors. In the STI domain, such behaviors might include sexual behavior (abstinence, fewer partners, less risky partners, condom use, pressure spouse/partner to reduce risky behaviors) as well as testing and treatment behaviors (regular STI testing, STI treatment, and encouraging partner to do the same). In this section we focus on behavioral changes induced particularly by incentives such as those employed in the RESPECT study: cash rewards conditional on testing negative for STIs. 3.4.1

Neoclassical Price Effect

Neoclassical economics predicts that the incentives will influence behavior in part via a price effect. Conditioning the reward on STI status increases the implicit “price” of risky sex, since there is now a potential loss of cash associated with risky behaviors. This theory assumes rational decision making in the sexual domain (Philipson and Posner 1995). The idea that individuals make trade- offs between price and the riskiness of sex is consistent with Gertler, Shah, and Bertozzi (2005), who find Mexican sex workers charge higher prices for sex without condoms, and Robinson and Yeh (2011) who find that sex workers charge more for anal sex. This, of course, does not indicate that individuals are perfectly rational in such decisions, but lends credence to the idea that people do respond to sex prices. However, this price effect may be muted by the fact that not all risky behaviors will result in a positive STI test, so the expected loss may be lower than the reward value. In the RESPECT trial, approximately 10 percent of individuals tested positive at each time point, thus a person of average risk who mixes with average risk partners could have an expected loss of only one- tenth of the reward amount. For example, the RESPECT study’s higher cash reward amount of $20 might yield only a $2 “price” of risky sex during a fourmonth period, which by itself could be a weak spur to behavioral change. 3.4.2

Neoclassical Income Effects

To the extent that health is a normal good, the rewards may change behavior through income effects, particularly with increasing value of cumulative repeated rewards. In many CM applications, though, the reward amounts are sufficiently small as to preclude neoclassical income effects. In the RESPECT study, the rewards over one year can be as high as 25 percent of mean annual earnings, which is a substantial amount. For some lowerincome women this could indeed ameliorate immediate economic pressures to engage in transactional sex, although there is mixed evidence on the size and even sign of the income effect on risky sexual behaviors. For men in

96

Damien de Walque, William H. Dow, Carol Medlin, and Rose Nathan

particular, it is often hypothesized that higher income will lead to more transactional sex, which over time would mute the incentive effects on male sexual behaviors. 3.4.3

Systematic Cognitive Errors

Some individuals may not be able to accurately perform the expected value calculations discussed above. Limited numeracy, availability heuristics, and bounded rationality may make some people particularly prone to systematic overestimation of small STI probabilities (see e.g., Kahneman and Tversky 1979). Thus they may behave as if the expected loss is substantially higher than it truly is. 3.4.4

High Discounting

In a society with a generalized AIDS epidemic (Tanzania has an estimated 6 percent adult prevalence rate), the expected cost of an AIDS diagnosis might be considered far larger than the modest cash rewards offered. For individuals who are present focused and heavily discount the future though, the prospect of an AIDS diagnosis many years in the future may not be considered a high cost. But if the price of risky sex would be incurred within months instead (the RESPECT study tested and offered cash rewards every four months), then high discounters may perceive an increased (discounted) price of risky sex, and thus behaviorally respond to this shortening of the time horizon. This, of course, depends on the extent of high discounting; studies in this context (including measures in the RESPECT study) have found extremely high rates of discounting, implying that even a time frame of months may result in substantial discounting of the potential reward value. And a reward in the time frame of months may have little impact on those risky sexual behaviors that may be driven by strong hyperbolic discounting (similar to the concept of compulsive immediate gratification used in developmental psychology) as discussed in the behavioral economics literature (O’Donoghue and Rabin 2001). 3.4.5

Other “Nudges”

A variety of other behavioral economics and psychological hypotheses have been proposed regarding the operation of incentives, now sometimes referred to as “nudges” following the popularization of the term by Thaler and Sunstein (2008). For example, some argue that introducing explicit monetary incentives into the sexual decision- making process may alter the frame within which people assess costs and benefits, resulting in unpredictable deviations from neoclassical theory. Others suggest that the incentives provide individuals with an excuse for deviating from social norms in order to act on underlying preferences for less risky behavior. Several such theories would predict a discontinuity of the dose- response relationship at zero: the first positive reward amount should have much larger behavioral

Stimulating Demand for AIDS Prevention

97

effects than subsequent amounts. Designing studies with multiple reward amounts (RESPECT used $20, $10, $0) will be particularly important for testing such hypotheses. In addition to the above discussed pathways for behavioral change in response to the incentives, it is also useful to consider potential long- term effects of time- limited incentives. Two competing hypotheses are of particular interest: Learning For behaviors that individuals may not have tried until encouraged to by the incentives (e.g., use of condoms), it is possible that the incentives will induce learning (and reinforcement) that could result in permanent positive behavior changes even after withdrawal of the incentives. Reduced Intrinsic Motivation Alternatively, psychologists have emphasized the potentially pernicious effects of extrinsic monetary incentives in destroying the intrinsic desire to engage in positive behaviors. Cameron, Banko, and Pierce (2001) reviews the literature on the possible destruction of intrinsic incentives and concludes that while this might occur for some high- interest tasks, in general, incentives do not have pervasive adverse effects. 3.5

The RESPECT Study: Design Considerations and Lessons Learned

In this section we highlight major considerations in designing a sexual behavior incentive study, providing rationale for choices made in designing RESPECT, so as assist others in designing similar studies in the future. Table 3.3 provides 2009 baseline summary statistics on key variables discussed below. Further details of the RESPECT study are described elsewhere (de Walque et al. 2012). 3.5.1

Ethical Considerations

Ethical concerns permeated all aspects of project planning, as referred to in many of the subsequent sections. The use of incentives in the sexual arena raises many potential controversial issues, so it is also natural that there will be disagreement regarding ethical considerations, both across individuals and across cultural settings, thus it is essential to have careful oversight and scrutiny of the research design and protocols. The project underwent several rounds of ethical review by several different committees and agencies, including the IRBs of the University of California, Berkeley and San Francisco, the Ifakara Health Institute, and the National Institute for Medical Research in Tanzania (NIMR). In addition, the study benefited from a (largely Tanzanian) ethical advisory group that reviewed the study protocol.

98

Damien de Walque, William H. Dow, Carol Medlin, and Rose Nathan

Table 3.3 Variables Arm Control Low value High value Female Age Education None Primary Secondary Married Low SES Income Chlamydia Gonorrhea Trichomonas HSV2 Syphilis HIV N

The RESPECT baseline summary statistics, among full-year panel sample and those attriting by one-year endpoint Mean among followed

Mean among attrited

0.46 0.28 0.26 0.51 27.49

0.54 0.19 0.27 0.43 25.85

0.12 0.78 0.10 0.76 0.55 258.25 0.02 0.01 0.12 0.35 0.02 0.03 2,242

0.11 0.75 0.13 0.61 0.51 223.06 0.02 0.02 0.15 0.33 0.02 0.12 157

Proportion attrited 0.076* 0.064 0.049* 0.056* —*** 0.063 0.063 0.084 0.053*** 0.061 0.058 0.136 0.077 0.063 0.075 0.212*** 0.065

Source: The RESPECT project baseline survey, 2009. Notes: Low SES is an indicator for response < = 3 on a 1– 7 ladder of subjective social status in the community. Income is an individual’s annual earnings in thousands of Tanzanian shillings (in 2009 the exchange rate for dollars was US$1 ~= TSh 1,100– 1,300, yielding average earnings of about $200/year). ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

3.5.2

Target Population

A key decision point is whether to target the general population of some region, or target a particular high- risk group. The intervention may be on average more effective among high- risk groups, but potentially have a smaller aggregate impact. The RESPECT study chose to target a general population. A pilot study by Galarraga (in process) in Mexico City is alternatively exploring the possibility for targeting a group of male sex workers. One concern with targeting a group such as sex workers is the political constraint to scaling-up implementation due to stigma and/or the concern about the rewards attracting people to sex work; however, the income value of sex work is likely to swamp the value of incentive payments, making that concern less salient. In order to ease study logistics, RESPECT chose to implement the inter-

Stimulating Demand for AIDS Prevention

99

vention within the context of Ifakara Health and Demographic Surveillance Site in rural southwestern Tanzania. We chose to recruit from ten villages; eight were rural and two were in Ifakara town. Rural villages are generally lower income, so a given reward level is likely to be more salient. Due to the limited number of villages, study- arm randomization occurred at the individual level. Demographically, we chose to enroll both men and women, as both could hypothetically respond to the program. A priori, there were concerns that women might not have sufficient agency to respond to the incentive regimen, while there were concerns that men’s responses could be dulled if they used cash rewards to pay for new risky behaviors, thus it was of interest to study both genders. We also chose to enroll both single and married individuals, and married individuals were issued invitations for their spouse to join the study as well (in the same intervention arm). Regarding age group, we initially sought to enroll fifteen to twenty- fouryear- olds (a particularly high- risk age group), but due to ethical concerns we decided not to enroll minors, thus our final target range was eighteen to thirty- year- olds (plus all spouses ages sixteen or older). For recruitment we chose a random sample of village residents from the demographic surveillance site computer registry, and went to each of their houses to explain the study and invite them to participate by reporting to a study station set up in their village the following week. Of individuals who were reached in their houses for recruitment, approximately 70 percent came to the study station and enrolled. Thus the enrolled sample is highly representative of the general population of eighteen to thirty- year- olds in these villages. Table 3.3 further shows that only 6.5 percent of enrollees attrited or were lost to follow-up during the one- year intervention. Attriters were younger and more likely to be single, but overall means of those choosing to complete the study were quite similar to attriters, except for the fact that attriters were substantially more likely to be HIV positive at baseline (with 21 percent of HIV positives attriting, possibly due to confusion about continued study eligibility). With this exception, the study sample was highly representative of the underlying population. 3.5.3

Conditionality Design

A primary goal of the study was to reduce risky sexual behaviors. But because these were not directly verifiable, we instead opted to link cash payments to objective measures—STI test results—that could serve as (imperfect) proxies for risky sexual behavior. Ideally we wished to condition on only STIs, which have been incontrovertibly linked to solely risky sexual activity. We also needed to balance cost considerations with the imperative to have enough tests so as to sufficiently capture risky behavior. In addition, we needed to consider local knowledge of STIs, testing capacity, and availability of appropriate drugs with low rates of treatment failure. Because RESPECT

100

Damien de Walque, William H. Dow, Carol Medlin, and Rose Nathan

was designed as a “proof of concept” to detect whether or not individuals appeared to responded to this type of incentives related to sexual behavior, we chose to conduct an expansive set of tests. These were: chlamydia, gonorrhea, trichomonas, Mycoplasma genitalium, syphilis, and HSV2. Each had unique considerations: chlamydia and gonorrhea were well known among the target population, thus, although they ended up having relatively low prevalence overall, they were useful to include due to their salience in the target population. Trichomonas has the merit of fairly high prevalence (and presumed incidence), although additional complications are introduced by the fact that it is biologically more common among women than men. Mycoplasma genitalium is also a high- prevalence STI, but has the drawback that it is not well known (many clinicians in this setting are unaware of it), there are no rapid tests for it, and treatment failure rates are not well understood; for these reasons we opted not to condition rewards on this STI, but still measured it for the purposes of increasing statistical analysis power. For highly reliable measurement of these first four STIs, every four months we collected urine samples from men and vaginal swabs from women (performed by a local nurse after careful explanation and consent—acceptability of swabs did not turn out to be problematic), and conducted assays using nucleic acid amplification tests (NAAT) at the Ifakara Health Institute microbiology lab. The NAAT tests were chosen because of their high sensitivity, but they are also expensive and thus would not normally be possible to do if scaling-up the intervention; a more feasible design for scale would be to focus just on a subset of STIs with reliable rapid tests. Use of laboratory- based NAAT testing also required a two- to three- week period in order to process all of the tests, thus we had to ask participants to return to their village study station to pick up the tests. Over 90 percent of participants came back to collect their results three weeks later, even in the control group; this is likely due to a combination of the fact that this is a population accustomed to research studies, as well as the high level of interest people had in “knowing their status” and rechecking it regularly (for both HIV and STIs). This also had the added benefit that we had another contact with study participants to remind them of the study and also provide individual posttest counseling (pretest counseling was provided to all at the time of specimen collection). Individuals who tested positive were provided vouchers for themselves and up to five partners for free treatment at the local health clinic, and a system was put into place to ensure that the first- and secondline medicines were always available in those clinics. A critical point is that each of these four urine- based STIs is curable. Thus, enrollees who test positive for an STI can continue to participate in the intervention after they have been treated and cured of the infection; learning is encouraged through positive reinforcement, and mistakes can be corrected and overcome. In addition to these urine- based NAAT STI tests, we also measured syphilis and HSV2 using blood- based tests. Because of the low prevalence

Stimulating Demand for AIDS Prevention

101

of syphilis, and because HSV2 detects whether the individual has ever contracted HSV2, we elected to only measure these blood- based STI’s at baseline and twelve months, rather than at each four- month interval. Finally, we also measured HIV at baseline and one- year follow-up. We did not condition reward payments on HIV status, in part due to ethical concerns about “punishing” somebody at the very time when they learn that they have contracted a fatal disease. We were not powered to look at intervention arm effects on HIV, but for scientific purposes did want to measure baseline HIV to better characterize our population. Our HIV rates turned out to be lower (just under 4 percent) than anticipated based on regional Demographic and Health Survey data; given the reasonably high rates of HIV testing in the population, it may be that HIV- positive individuals selected themselves out of the sample. We decided to also test HIV at the twelvemonth endpoint due to the study population’s strong interest in checking their results again. Although over half of the population had received a prior HIV test, study enrollees expressed concerns about confidentiality of testing at local clinics, and strong desire for retesting by RESPECT study personnel. 3.5.4

Incentive Size and Frequency

A major design consideration is the size of the incentive. The CCT programs focused on uptake have typically relied on formulas to compensate participants on the opportunity costs, or time requirements, of complying with programmatic requirements such as attending antenatal care visits or taking the child to the clinic for his or her regular check-ups. However, in the case of CCT programs designed to encourage safer sexual practices and discourage risky ones, the purpose of the incentive is not primarily to compensate the individual for the opportunity costs of participating in the program, but rather to change the decision calculus of the individual regarding his or her sexual behavior. The goal would be to increase the immediate costs associated with risky behavior by increasing the possibility of future reward (cash). Several studies of the CM interventions have found a positive relationship between the magnitude of the reward and the impact on the target behavior (and laboratory- confirmed health impact; Sindelar, Elbel, and Petry 2007; Stitzer and Bigelow 1983), at least during the treatment period; the effect appears to weaken during follow up (Higgins et al. 2007). However, the magnitude of incentive required for shaping sexual behavior is unknown. Preliminary assessments can be obtained through focus group discussions, survey data, and discrete choice experiments, but these are no substitute for direct measures of impact achieved by randomly assigning different- sized payments to individuals participants. In the RESPECT study, we chose an incentive amount of approximately $20 at each four- month testing point. Mean earnings in the study population are approximately $200 per year, thus staying STI negative could increase income by about 30 percent, which is in a similar ballpark as Mexico’s

102

Damien de Walque, William H. Dow, Carol Medlin, and Rose Nathan

Oportunidades program. This amount was discussed extensively with local populations, some of whom argued for much higher amounts (which would not be able to be feasibly scaled-up), but others of whom agreed that $20 sounded reasonable. We then chose to include a second incentivized arm with a reward half that size, in order to test dose response. One concern about our ability to test this dose response, however, is that randomization is at the individual level, thus individuals in the $10 group may feel “cheated” and respond less robustly than if they were not aware of the higher group. In addition, we also chose to assess the acceptability and feasibility of lottery- based incentives. All enrollees were eligible for a $100 village and gender- specific lottery drawing at the four- month, eight- month, and twelve- month testing rounds (thus, on average, the chances of winning were approximately one in one hundred at each four- month time point). However, if individuals in the cash reward arms tested STI positive, they were made ineligible for the lottery in that round (control enrollees were eligible regardless of their STI status, as long as they had not attrited from the project). Thus, this equally reinforced incentives for both of the incentivized reward arms. The lottery proved popular, as our ex ante focus groups had predicted (the national lottery is also popular). Large crowds attended each drawing, providing yet another opportunity to communicate project messaging. The lottery had two limitations to consider, though. First, some winners were individuals known in the community to engage in risky sexual behavior, which cause disillusionment among some village members. Second, the incentive structure was confusing to some enrollees, violating the “KISS” principle (keep it simple stupid); this could be remedied though in a simpler lottery- only design. At the end of the study we queried individuals about the role of the cash rewards versus the lottery; in general, people indicated that their behaviors responded to both sets of incentives in roughly equal proportions, providing promising evidence for the potential of pursuing (lower cost) lottery designs in future interventions. This needs to be counterbalanced against the finding in table 3.4 that although respondents generally “liked” both the cash rewards and the lottery, they liked the former somewhat more. A separate but related consideration is the frequency and immediacy in which the cash incentive is paid out. Given the unplanned and spontaneous nature of many sexual encounters, the incentive needs to be offered frequently enough to keep it ever present in the minds of the target group, and as immediately as possible to reinforce the link between the payment and the target behavior. It seems reasonable to assume, and the CM literature has confirmed (Stitzer and Petry 2006), that larger cash payments given frequently have greater impact than smaller payments given less frequently, but it is nevertheless unclear how these two dimensions of magnitude and frequency interact. In the RESPECT study we chose to balance the frequency with budgetary imperatives by testing every four months, although given

103

Stimulating Demand for AIDS Prevention Table 3.4

Stated effects of incentives on motivation to change behavior How much does/would motivate behavior change? (percentage)

Arm

Award type

Value

N

Very much

Somewhat

A little

None

Low value High value Control Control Control Control

Actual Actual Hypothetical Hypothetical Hypothetical Hypothetical

Rewards: Eight-month survey 10,000 486 36 20,000 518 59 10,000 172 24 20,000 264 47 50,000 232 57 100,000 195 57

21 13 21 11 13 7

12 7 12 10 8 5

31 21 43 32 23 31

Low value High value Control Control Control Control

Actual Actual Hypothetical Hypothetical Hypothetical Hypothetical

Rewards: Twelve-month survey 10,000 150 57 20,000 160 79 10,000 254 51 20,000 239 65 50,000 279 70 100,000 258 77

27 13 20 19 19 14

7 4 9 4 3 2

9 5 19 12 9 8

Low value High value Control Control Control

Actual Actual Hypothetical Hypothetical Hypothetical

Lottery: Twelve-month survey 100,000 209 87 100,000 175 89 50,000 345 78 100,000 349 83 200,000 336 80

6 4 13 7 10

2 0 2 2 1

5 7 7 7 9

Notes: Control- arm respondents were randomly assigned a single hypothetical reward value and a single lottery value, about which the behavior change motivation question was asked. Respondents in the lowvalue and high- value arms were asked about their actual reward amounts. Low- value and high- value sample sizes are reduced at twelve months because a portion of these respondents were instead randomized to questions about other hypothetical arms (not shown).

the lack of guidance in the literature this was inevitably a somewhat arbitrary decision. Of additional interest is whether the cash incentive should be paid directly to individuals, to a couple, or some other social grouping. The CCT programs for poverty alleviation have tended to target households and communities, but the incentive is actually paid to the mothers of young children, rather than the fathers or the legal head of household. This design feature has been a nod to a series of findings that women are more likely to use the money on food for the family rather than on alcohol or other purchases (Rawlings and Rubio 2005; Schady and Rosero 2007). However, some studies of CM interventions have found that the effect size of incentive was greatest if the payment was made to a predesignated group rather than the individual (Jeffery et al. 1983). This may be due to the social support provided by the group, peer pressure, or some combination, and the question is whether such a mechanism can or should be applied to the sexual relationship. In the RESPECT study, we chose the clearest and most

104

Damien de Walque, William H. Dow, Carol Medlin, and Rose Nathan

transparent method: paying each individual separately. Data analysis indicates that many of the individuals shared their rewards by buying household goods, but many others (both men and women) kept the funds to themselves and did not even share with their spouse. 3.5.5

Skill Building to Aid Behavioral Change

In order to ensure that participants had the necessary tools to act on the incentives for sexual behavior change, we chose to offer monthly (sex segregated) group skill- building sessions (in addition to individual preand posttest counseling every four months). For this purpose we trained local facilitators to deliver a modified version of the Stepping Stones curriculum (Jewkes et al. 2008), with sessions on topics such as sexual health, condom use, relationship skills, decision- making skills, women’s roles and gender- based violence. The last of these was particularly important due to the concern that women testing STI positive may be subject to violence; for this purpose we also carefully monitored for violence events. Fortunately, reported violence rates during the intervention were consistently lower than preintervention levels (Krishnan et al. 2012). We chose to make these group sessions voluntarily, and 20– 30 percent of enrollees attended each month. 3.5.6

Self-Reported Degree of Behavioral Change Motivation from Incentives

To judge the overall perceived effectiveness of the incentive mechanisms used in RESPECT, table 3.4 presents enrollees’ responses to direct questions regarding how much the incentives motivated sexual behavior change. At both eight- month and twelve- month time points the low- value and highvalue respondents were asked how much the cash rewards motivated sexual behavior change. After eight months, 36 percent of low- value arm (TSh 10,000) respondents stated “very much” and 31 percent stated “none.” In the high- value arm (TSh 20,000), the percent stating “very much” was substantially higher at 59 percent, with only 21 percent reporting “none.” The control- arm respondents were also queried, but as part of this question each was randomized to a single hypothetical reward amount (TSh 10,000/ 20,000/50,000/100,000) and asked how much that amount would motivate behavior change. Table 3.4 reports again a consistent dose- response relationship, with higher amounts related to higher self- reported behavioral change motivation. Interestingly, those actually enrolled in the reward arms reported higher levels of motivation than did the control respondents when reporting about the hypothetical reward scenario. These motivation questions were repeated again at the twelve- month intervention end point, as reported in the middle panel of table 3.4 (the questions were not asked prior to the eight- month survey). The above doseresponse patterns were replicated. A further finding is that reported motivation levels are uniformly higher at twelve months than they were at eight

Stimulating Demand for AIDS Prevention

105

months, suggesting that there may have been further learning and behavior change by participants during this interval, which is an argument for repeated rounds of incentivized behavior over an extended period of time. By this end point, 79 percent of the high- value arm enrollees reported “very much” motivation, with only 5 percent reporting no motivation. Finally, the bottom panel of table 3.4 reports on parallel questions that were asked at twelve months in regard to the lottery incentives. We first note that 89 percent of high- value arm respondents reported “very much” motivation from the lottery, which is even higher than for the cash rewards. This is despite the fact that the respondents had already participated in two rounds of the lottery, after which they should have been aware that the odds of winning the actual 100,000 lottery were approximately 1 percent, thus the expected value of the lottery was far lower than the certain cash rewards. Second, we note that the responses by control individuals (asked to imagine a single hypothetical lottery amount in which eligibility was conditioned on negative STI results) were quite similar to those of the actual high- value and low- value arm respondents; perhaps the fact that the control- arm enrollees had already been actually eligible for the lotteries (but regardless of STI status) had made them better informed about how the lottery could affect their motivation. Third, it is evident from table 3.4 that there is not a dose- response relationship between the different hypothetical lottery amounts and the degree of projected behavior change motivation. This is despite the fact that these control respondents had all been “anchored” by the actual lottery amount of TSh 100,000, and thus might be particularly expected to react negatively to the lower hypothetical amount of 50,000. One potential explanation for this lack of a dose- response relationship is that the lottery amounts tested were all sufficiently high (TSh100,000 is equal to approximately 40 percent of mean annual income in the study population) that virtually all respondents would be motivated by the prospect of winning. Indeed, in-depth focus group responses suggested that many individuals in this context failed to calculate the relevant expected value (few have completed schooling beyond the primary level), and thus were likely making cognitive errors that led to greater salience of the lottery than would be rationally expected. Although we must be cautious in not overinterpreting these self- reported “motivation” questions, and should await publication of further results from the more objective STI endpoints, this evidence does suggest that a lottery- based incentive system may be acceptable within this community and possibly be more cost effective than a system of certain rewards. 3.6

Discussion

Cash incentives have been shown to be effective at shaping behavior in a variety of health domains, from improving the uptake of health and edu-

106

Damien de Walque, William H. Dow, Carol Medlin, and Rose Nathan

cational services among the poor, to discouraging unhealthy or risky behaviors, such as substance abuse, smoking, and overeating. These successes inevitably raise the question of whether they can also be applied in areas of sexual and reproductive health beyond contraception (Mauldon 2003; Weeden et al. 1986), and in particular to HIV prevention (Haug and Sorenson 2006). A review of countries’ experiences with CCT programs for poverty alleviation and CM provides useful insights into how to design, implement, and evaluate a CCT- based STI/HIV prevention intervention. The particular experience of the RESPECT study finds that such a program can be designed to be efficacious and acceptable. Tables 3.5A and 3.5B report enrollee responses to questions regarding perceived success of the trial and which dimensions of the trial they liked and did not like. Although such responses must be interpreted cautiously in light of possible social desirability bias (not wanting to offend the interviewer)—indeed there is some variation in responses across questions— Table 3.5A

Enrollees’ attitudes toward elements of RESPECT, after one-year trial

What aspects of RESPECT did you like: HIV testing Very much Somewhat A little None STI testing Very much Somewhat A little None Cash rewards Very much Somewhat A little None Lottery Very much Somewhat A little None Free STI treatment Very much Somewhat A little None Individual counseling Very much Somewhat A little None Group counseling Very much Somewhat A little None

High value

Low value

Control

96.2 2.7 1.0 0.0 98.5 1.4 0.2 0.0 93.2 5.6 1.2 0.0 74.5 18.6 5.8 1.0 96.4 3.1 0.5 0.0 88.0 9.1 2.6 0.3 58.6 21.0 17.1 3.3

95.3 3.6 1.0 0.0 97.2 2.1 0.3 0.3 87.7 8.9 2.8 0.7 69.9 20.6 8.1 1.5 96.8 2.1 0.8 0.3 83.6 12.8 3.1 0.5 54.3 21.1 20.9 3.7

92.4 5.1 2.4 0.0 96.8 3.0 0.2 0.0 60.1 24.8 13.0 2.0 54.4 27.4 14.8 3.5 95.0 3.6 1.2 0.2 82.4 13.7 3.8 0.1 50.0 20.4 23.1 6.7

107

Stimulating Demand for AIDS Prevention Table 3.5B

Enrollees’ perceptions of success of RESPECT, after one-year trial High value

Low value

Control

At end of one- year study did RESPECT: Reduce STIs in village Very much Somewhat A little None

91.6 7.2 0.7 0.5

90.3 7.3 1.6 0.8

86.2 10.9 2.0 0.9

How did RESPECT study affect: Your life Greatly improved Somewhat improved Not much or worse Well- being of Greatly improved community Somewhat improved Not much or worse

74.8 25.2 0.0 82.1 17.4 0.5

74.8 24.2 1.0 85.0 14.5 1.0

76.2 22.8 1.0 83.3 16.0 1.0

Notes: All items were drawn from the twelve- month RESPECT survey (at end point of trial) except the final two questions, which were drawn from the twenty- four- month survey (one year after end of the trial).

overall, the study appears to be highly acceptable and valued by the study population. This is perhaps not surprising given the substantial amount of cash delivered by the project, but even control group enrollees (who did not receive cash beyond minimal inconvenience fees) responded quite positively about the trial. But could a trial such as this be brought to scale effectively and cost effectively? The CCT programs for poverty alleviation have already been brought to scale in several different settings, and therefore have broken the credibility barrier by proving that it can be done. Of course, this has not eliminated concerns about whether incentive programs will be equally effective in other settings, particularly those where initial health infrastructural investments are very low, as in many countries in sub-Saharan Africa (Lagarde, Haines, and Palmer 2007). Furthermore, the cost effectiveness of CCT programs when compared to traditional supply- side investments in health and education has not been sufficiently explored (Lagarde, Haines, and Palmer 2007). Cost analyses for STI- based incentive programs will need to account for not only the cash payments and monitoring infrastructure, but also STI- testing costs. Continued improvements in rapid diagnostic testing will help to ameliorate the latter costs over time, but they will still be central considerations. The RESPECT study design was not created for the purposes of scale-up. It will be important to experiment with alternative designs to test effectiveness of cheaper approaches. For example, one promising option could be to use a lottery to regularly choose random “winners” from a defined population (such as a subvillage), who would then be tested for STIs or HIV, and would only

108

Damien de Walque, William H. Dow, Carol Medlin, and Rose Nathan

receive their lottery payout if negative. An ongoing project in Lesotho is further exploring variants of this idea; future studies of this type would be a high priority. However, these cost concerns must be balanced against the benefits of an effective behavioral intervention, and—again—too little is yet known to draw generalizable conclusions at this stage. Certainly the epidemiological context is highly relevant to this discussion, as the costs can be more easily justified in settings where the rate of disease transmission is very high. De Janvry and Sadoulet (2004) have raised concerns about inefficiencies in the design of CCT programs, pointing out that large- scale CCT programs fail to distinguish between those families who would have attended prenatal clinics and sent their children to school regardless of the incentive and those who require the incentive to induce the desired behavior. They conclude that CCT programs can be made more efficient if they do a better job of targeting the group that needs the incentive to be induced to alter their behaviors. This will pose a continuing challenge for STI incentive program design. Future cost- effectiveness and cost- benefit studies of CCTs in different target groups will be needed to inform these types of implementation decisions.

References Askew, I., and M. Berer. 2003. “The Contribution of Sexual and Reproductive Health Services to the Fight Against HIV/AIDS: A Review.” Reproductive Health Matters 11 (22): 51– 73. Baird, S., E. Chirwa, C. McIntosh, and B. Özler. 2010. “The Short-Term Impacts of a Schooling Conditional Cash Transfer Program on the Sexual Behavior of Young Women.” Health Economics 19 (suppl.): 55– 68. Baird, S. R. Garfein, C. McIntosh, and B. Özler. 2012. “Effect of a Cash Transfer Programme for Schooling on Prevalence of HIV and Herpes Simplex Type 2 in Malawi: A Cluster Randomised Trial.” Lancet 379 (9823): 1320– 29. Baird, S., C. McIntosh, and B. Özler. 2011. “Cash or Condition? Evidence from a Cash Transfer Experiment.” Quarterly Journal of Economics 126(4): 1709– 53. Bertozzi, S., N. S. Padian, J. Wegbreit, L. M. DeMaria, B. Feldman, H. Gayle, J. Gold, R. Grant, and M. T. Isbell. 2006. “HIV/AIDS Prevention and Treatment.” In Disease Control Priorities in Developing Countries, edited by Jamison, Breman, Meashem, and Mills. New York: World Bank. Bertrand, J. R., K. O’Reilly, J. Denison, R. Anhang, and M. Sweat. 2006. “Systematic Review of the Effectiveness of Mass Communication Programs to Change HIV/AIDS-Related Behaviors in Developing Countries.” Health Education and Research 21 (4): 567– 97. Cameron, J., K. M. Banko, and W. D. Pierce. 2001. “Pervasive Negative Effects of Rewards on Intrinsic Motivation: The Myth Continues.” Behavioral Analyst 24 (1): 1– 44. Coates, T. J., L. Richter, and C. Caceres. 2008. “Behavioural Strategies to Reduce HIV Transmission: How To Make Them Work Better.” Lancet 372:669– 84.

Stimulating Demand for AIDS Prevention

109

De Janvry, A., and E. Sadoulet. 2004. “Conditional Cash Transfer Programs: Are They Really Magic Bullets?” Working Paper, Department of Agriculture and Resource Economics, University of California, Berkeley. Deren, S., R. Stephens, W. R. Davis, T. E. Feucht, and S. Tortu. 1994. “The Impact of Providing Incentives for Attendance at AIDS Prevention Sessions.” Public Health Reports 109 (4): 548– 54. de Walque, D., W. H. Dow, R. Nathan, R. Abdul, F. Abilahi, E. Gong, Z. Isdahl, et al. 2012. “Incentivising Safe Sex: A Randomised Trial of Conditional Cash Transfers for HIV and Sexually Transmitted Infection Prevention in Rural Tanzania.” BMJ Open 2012 (2): e000747. doi: 10.1136/bmjopen- 2011– 000747. Donatelle, R. J., S. L. Prows, D. Champeau, and D. Hudson. 2000. “Randomised Controlled Trial Using Social Support and Financial Incentives for High Risk Pregnant Smokers: Significant Other Supporter (SOS) Program.” Tobacco Control 9:67– 69. Eiss, R., and R. Glass. 2007. “Bridging the Coverage Gap in Global Health.” Journal of the American Medical Association 298 (16): 1940– 42. Emont, S., and K. Cummings. 1992. “Using a Low-Cost, Prize-Drawing Incentive to Improve Recruitment Rate at a Work-Site Smoking Cessation Clinic.” Journal of Occupational Medicine 34:771– 74. Fernald, L. C. H., P. J. Gertler, and X. Hou. 2008. “Cash Component of Conditional Cash Transfer Program is Associated with Higher Body Mass Index and Blood Pressure in Adults.” Journal of Nutrition 138:2250– 57. Fernald, L. C. H., P. J. Gertler, and L. M. Neufeld. 2008. “Role of Cash in Conditional Cash Transfer Programmes for Child Health, Growth, and Development: An Analysis of Mexico’s Oportunidades.” Lancet 371:827– 37. Fiszbein, A., and N. R. Schady. 2009. “Conditional Cash Transfers for Attacking Present and Future Poverty.” Policy Research Report, Washington, DC, World Bank. Follick, M. J., J. I. Folwer, and R. A. Brown. 1984. “Attrition in Worksite WeightLoss Interventions: The Effects of an Incentive Procedure.” Journal of Consultative Clinical Psychology 52:139– 40. Gertler, P. J. 2004. “Do Conditional Cash Transfers Improve Child Health? Evidence from PROGRESA’s Control Randomized Experiment.” American Economic Review 94 (2): 332– 41. Gertler, P. J., M. Shah, and S. M. Bertozzi. 2005. “Risky Business: The Market for Unprotected Commercial Sex.” Journal of Political Economy 113:518– 50. Haug, N., and J. Sorensen. 2006. “Contingency Management Interventions for HIVRelated Behaviors.” Current HIV/AIDS Reports 3 (4): 154– 59. Higgins, S. T., A. J. Budney, W. K. Bickel, F. E. Foerg, R. Donham, and G. J. Badger. 1994. “Incentives Improve Outcomes in Outpatient Behavioral Treatment of Cocaine Dependence.” Archives of General Psychiatry 51:568– 76. Higgins, S. T., S. H. Heil, R. Dantona, R. Donham, M. Matthews, and G. J. Badger. 2007. “Effects of Varying the Monetary Value of Voucher-Based Incentives on Abstinence Achieved during and following Treatment among Cocaine-Dependent Outpatients.” Addiction 102 (2): 271– 81. Jeffery, R. W., W. M. Bjornson-Benson, B. S. Rosenthal, C. L. Kurth, and M. M. Dunn. 1984. “Effectiveness of Monetary Contracts with Two Repayment Schedules of Weight Reduction in Men and Women from Self-Referred and Population Samples.” Behavioral Therapy 15:273– 79. Jeffery, R. W., W. M. Gerber, B. S. Rosenthal, and R. A. Lindquist. 1983. “Monetary Contracts in Weight Control: The Effectiveness of Group and Individual Contracts of Varying Size.” Journal of Consulting and Clinical Psychology 51:242– 48.

110

Damien de Walque, William H. Dow, Carol Medlin, and Rose Nathan

Jeffery, R. W., P. D. Thompson, and R. R. Wing. 1978. “Effects on Weight Reduction of Strong Monetary Contracts for Calorie Restriction or Weight Loss.” Behavioral Research and Therapy 16 (5): 363– 69. Jewkes, R., M. Nduna, J. Levin, N. Jama, K. Dunkle, A. Puren, and N. Duwury. 2008. “Impact of Stepping Stones on Incidence of HIV and HSV- 2 and Sexual Behaviour in Rural South Africa: Cluster Randomised Controlled Trial.” BMJ 7:337a506. John, L. K., G. Loewenstein, A. B. Troxel, L. Norton, J. E. Fassbender, and K. Volpp. 2011. “Financial Incentives for Extended Weight Loss: A Randomized, Controlled Trial.” Journal of General Internal Medicine 26 (6): 621– 26. Joint United Nations Programme on HIV/AIDS (UNAIDS). 2010. Global Report: UNAIDS Report on the Global AIDS Epidemic 2010. Geneva: UNAIDS. http:// www.unaids.org/globalreport/Global_report.htm. Kahneman, D., and A. Tversky. 1979. “Prospect Theory: An Analysis of Decision under Risk.” Econometrica 47:263– 91. Kakwani, N., F. Soares, and H. Son. 2005. “Conditional Cash Transfers in African Countries.” International Poverty Centre Working Paper no. 9, United Nations Development Programme. Kamb, M. L., F. Rhodes, T. Hoxworth, J. Rogers, A. Lentz, C. Kent, R. MacGowen, and T. A. Peterman. 1998. “What About Money? Effect of Small Monetary Incentives on Enrollment, Retention, and Motivation to Change Behavior in an HIV/STD Prevention Counseling Intervention. The Project RESPECT Study Group.” Sexually Transmitted Infections 74:253– 55. Kane, R. L., P. E. Johnson, R. J. Town, and M. Butler. 2004. “A Structured Review of the Effect of Economic Incentives on Consumers’ Preventive Behavior.” American Journal of Preventive Medicine 27 (4): 327– 52. Kohler, H. P., and R. L. Thornton. 2011. “Conditional Cash Transfers and HIV/ AIDS Prevention: Unconditionally Promising?” World Bank Economic Review 26 (2): 165– 90. doi: 10.1093/wber/lhr041. Krishnan, S., D. Vohra, D. de Walque, C. Medlin, R. Nathan, and W. Dow. 2012. “Tanzanian Couples’ Perspectives on Gender Equity, Relationship Power, and Intimate Partner Violence: Findings from the RESPECT Study.” AIDS Research and Treatment 2012:1– 9. Lagarde, M., A. Haines, and N. Palmer. 2007. “Conditional Cash Transfers for Improving Uptake of Health Interventions in Low- and Middle-Income Countries: A Systematic Review.” Journal of American Medicine 298 (16): 1900– 10. Mauldon, J. G. 2003. “Providing Subsidies and Incentives for Norplant, Sterilization and Other Contraception: Allowing Economic Theory to Inform Ethical Analysis.” Journal of Law, Medicine, and Ethics 31:351– 64. Medlin, C., and D. de Walque. 2008. “Potential Applications of Conditional Cash Transfers for Prevention of Sexually Transmitted Infections and HIV in SubSaharan Africa.” Policy Research Working Paper no. 4673, Washington, DC, World Bank. Nigenda, G., and L. M. Gonzalez-Robledo. 2005. “Lessons Offered by Latin American Cash Transfer Programmes, Mexico’s Oportunidades and Nicaragua’s SPN: Implications for African Countries.” Working Paper, DFID Health Systems Resource Centre, Department for International Development. http://www.eldis .org/fulltext/verypoor/5_ningenda.pdf. O’Donoghue, T., and M. Rabin. 2001. “Risky Behavior among Youths: Some Issues from Behavioral Economics” In Risky Behavior among Youths, edited by J. Gruber, 29– 68. Chicago: University of Chicago Press.

Stimulating Demand for AIDS Prevention

111

Padian, N. S., S. I. McCoy, S. S. Karim, N. Hasen, J. Kim, M. Bartos, K. Katabira, S. M. Bertozzi, B. Schwartländer, and M. S. Cohen. 2011. “HIV Prevention Transformed: The New Prevention Research Agenda.” Lancet 378 (9787): 269– 78. Paul-Ebhohimhen, V., and A. Avenell. 2008. “Systematic Review of the Use of Financial Incentive in Treatments for Obesity and Overweight.” Obesity Reviews 9 (4): 355– 67. Petry, N. M. 2000. “A Comprehensive Guide to the Application of Contingency Management Procedures in Clinical Settings.” Drug and Alcohol Dependence 58:9– 25. Petry, N. M., B. Martin, and C. Finocche. 2001. “Contingency Management in Group Treatment: A Demonstration Project in an HIV Drop-In Center.” Journal of Substance Abuse Treatment 21:89– 96. Petry, N. M., I. Petrakis, L. Trevisan, G. Wiredu, N. N. Boutros, B. Martin, and T. R. Kosten. 2001. “Contingency Management Interventions: From Research to Practice.” American Journal of Psychiatry 158 (5): 694– 702. Philipson, T., and R. A. Posner. 1995. “The Microeconomics of the AIDS Epidemic in Africa.” Population and Development Review 21 (4): 835– 48. Rawlings, L., and G. Rubio. 2005. “Evaluating the Impact of Conditional Cash Transfer Programs.” World Bank Research Observer 20 (1): 29– 55. Rawson, R. A., A. Huber, M. McCann, S. Shoptaw, D. Farabee, C. Reiber, and W. Ling. 2002. “A Comparison of Contingency Management and CognitiveBehavioral Approaches during Methadone Maintenance Treatment for Cocaine Dependence.” Archives of General Psychiatry 59 (9): 817– 24. Robinson, J., and E. Yeh. 2011. “Transactional Sex as a Response to Risk in Western Kenya.” American Economic Journal: Applied Economics 3 (1): 35– 64. Rosen, M. I., K. Dieckhaus, T. J. McMahon, B. Valdes, N. M. Petry, J. Cramer, and B. Rounsaville. 2007. “Improved Adherence with Contingency Management.” AIDS Patient Care STDS 21 (1): 30– 40. Schady, N., and J. Rosero. 2007. “Are Cash Transfers Made to Women Spent Like Other Sources of Income?” World Bank Policy Research Working Paper no. 4282, Washington, DC, World Bank. Schubert, B., and R. Slater. 2006. “Social Cash Transfers in Low Income African Countries: Conditional or Unconditional?” Development Policy Review 24 (5): 571– 78. Schuring E. 2005. “Conditional Cash Transfers: A New Perspective for Madagascar?” A World Bank Report, Washington, DC, World Bank. Silverman, K., S. T. Higgins, R. K. Brooner, I. D. Montoya, E. J. Cone, C. R. Schuster, and K. L. Preston. 1996. “Sustained Cocaine Abstinence in MethodoneMaintenance Patients through Voucher-Based Reinforcement Therapy.” Archives of General Psychiatry 53:409– 15. Sindelar, J., B. Elbel, and N. M. Petry. 2007. “What Do We Get For Our Money? Cost Effectiveness of Adding Contingency Management.” Addiction 102 (2): 309– 16. Sridhar, D., and A. Duffield. 2006. “A Review of the Impact of Cash Transfer Programmes on Child Nutritional Status and Some Implications for Save the Children UK Programmes.” Save the Children. http://www.savethechildren.org.uk/sites /default/files/docs/cash_transfer_prog_nutrition_1.pdf. Stitzer, M. 2006. “Contingency Management and the Addictions.” Addiction 101:1536– 37. Stitzer, M., and G. E. Bigelow. 1983. “Contingent Payment for Carbon Monoxide Reduction: Effects of Pay Amount.” Behavioral Therapy 14:647– 56.

112

Damien de Walque, William H. Dow, Carol Medlin, and Rose Nathan

Stitzer, M., and N. Petry. 2006. “Contingency Management for Treatment of Substance Abuse.” Annual Review of Clinical Psychology 2:411– 34. Thaler, R., and C. Sunstein. 2008. Nudge: Improving Decisions about Health, Wealth, and Happiness. New Haven, CT: Yale University Press. Thornton, R. 2008. “The Demand for and Impact of Learning HIV Status.” American Economic Review 98:1829– 63. Vandrey, R., G. E. Bigelow, and M. L. Stitzer. 2007. “Contingency Management in Cocaine Abusers: A Dose-Effect Comparison of Goods-Based versus CashBased Incentives.” Experimental Clinical Psychopharmacology 15 (4): 338– 43. Volpp, K., L. John, A. B. Troxel, L. Norton, J. Fassbender, and G. Lowenstein. 2008. “Financial Incentive-Based Approaches for Weight Loss: A Randomized Trial.” Journal of the American Medical Association 300 (22): 2631– 37. Volpp, K., A. B. Troxel, M. V. Pauly, H. Glick, A. Puig, D. Asch, R. Galvin, J. Zhu, F. Wan, J. DeGuzman, E. Corbett, J. Weiner, and J. Audrain-McGovern. 2009. “A Randomized Controlled Trial of Financial Incentives for Smoking Cessation.” New England Journal of Medicine 360:699– 709. Weeden, D., A. Bennett, D. Lauro, and M. Viravaidya. 1986. “An Incentives Program to Increase Contraceptive Prevalence in Rural Thailand.” International Family Planning Perspectives 12 (1): 11– 16. Windsor, R. A., J. B. Lowe, and E. E. Bartlett. 1988. “The Effectiveness of a Worksite Self-Help Smoking Cessation Program: A Randomized Trial.” Journal of Behavioral Medicine 11:407– 21.

4

Alternative Cash Transfer Delivery Mechanisms Impacts on Routine Preventative Health Clinic Visits in Burkina Faso Richard Akresh, Damien de Walque, and Harounan Kazianga

4.1

Introduction

Poor health is widespread among children in low- income countries. Often, such ill health is coupled with poor access to health care, either because of supply- side or demand- side constraints. These poor health outcomes negatively affect physical growth and cognitive development, with potential long- term consequences. For example, children who are in poor health are Richard Akresh is associate professor of economics at the University of Illinois at UrbanaChampaign and a research associate of the National Bureau of Economic Research. Damien de Walque is a senior economist in the Development Research Group (Human Development and Public Services Team) at the World Bank. Harounan Kazianga is associate professor of economics at Oklahoma State University. These data were collected for a project evaluating social protection strategies in Burkina Faso, which greatly benefited from the support of Marie-Claire Damiba, Seydou Kabré, and Victorine Yameogo from the Secrétariat Permanent du Comité National de Lutte contre le SIDA et les Infections Sexuellement Transmissibles (SP-CNLS-IST) in Burkina Faso and Hans Binswanger, Nono Ayivi-Guedehoussou, Ousmane Haidara, Timothy Johnston, Mead Over, and Tshiya Subayi-Cuppen at the World Bank. Data collection was supervised by Robert Ouedraogo, Jean-Pierre Sawadogo, Bambio Yiriyibin, and Pam Zahonogo from the University of Ouagadougou, Department of Economics. The project is funded by the NBER Africa Project and the following World Bank trust fund grants: Strategic Impact Evaluation Fund (SIEF), Bank-Netherlands Partnership Program (BNPP), Gender Action Plan (GAP), Knowledge for Change Program (KCP), WB-DFID Evaluation of the Community Response to HIV and AIDS, and Luxembourg Poverty Reduction Partnership (LPRP). The authors would also like to thank Pascaline Dupas and Adam Wagstaff as well as participants at the NBER Africa workshop in Zanzibar for helpful comments on an earlier draft. Finally, the authors thank Emilie Bagby, German Caruso, Igor Cunha, Christine Jachetta, Moussa Kone, Marleen Marra, and Nga Thi Viet Nguyen for their research assistance. The findings, interpretations, and conclusions expressed in this chapter are entirely those of the authors. They do not necessarily represent the views of the World Bank, its executive directors, or the countries they represent. For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see http://www.nber.org/chapters/c13377.ack.

113

114

Richard Akresh, Damien de Walque, and Harounan Kazianga

less likely to enroll in school, or enter school later, and when they enroll they tend to perform worse than children in good health (Moock and Leslie 1986; Brooker, Hall, and Bundy 1999; Glewwe, Jacoby, and King 2001). This in turn affects productivity later in life (Dasgupta 1993; Strauss and Thomas 1995; Schultz 2005). It is estimated that illness incidence and other health risks prevent approximately 200 million children in low- income countries from reaching their full potential in cognitive development (GranthamMcGregor et  al. 2007). Obviously, identifying mechanisms that could improve health outcomes among children will have large payoffs, not only at the individual level, but also in term of overall economic progress as well. Conditional cash transfer (CCT) programs are now one of the most popular government welfare interventions in developing countries.1 These programs transfer resources to poor households conditional on the family taking measures to increase the health and human capital of their children (predominantly enrolling their children in school and taking them for regular health care visits). Unconditional cash transfer (UCT) programs do not impose these conditionality constraints. The CCT programs represent a “top- down” approach in which outside organizations decide what is best for poor children and provide incentives to their parents to achieve these objectives. In contrast, UCT programs assume that, once a budget constraint is relaxed, parents are in a better position to make appropriate decisions regarding their child’s human capital. The CCT programs are also more costly per recipient to administer than UCT programs because of the costs associated with monitoring conditions. Unconditional cash transfers act through increased income, so that health outcomes should improve as long as the income elasticity of demand for health is positive and marginal productivity of health care is positive (Parker and Wong 1998). Mounting evidence shows both types of transfers improve child health outcomes (for CCTs in Colombia [Attanasio et al. 2005], Ecuador [Paxson and Schady 2010], Nicaragua [Maluccio and Flores 2005; Macours, Schady, and Vakis 2012], and for UCTs in South Africa [Case and Deaton 1998; Case, Hosegood, and Lund 2005; Duflo 2003]), but the evidence on which mechanisms—conditionality or income effects—are playing a role in driving the impacts is more limited.2 In this chapter, we present evidence of the impacts of a unique cash trans1. Numerous countries in Latin America, as well as in Asia, have implemented such programs (Fiszbein and Schady 2009). In Africa, several CCT pilot programs (in South Africa and Kenya) have been implemented but focus exclusively on orphans and HIV households and have not yet been rigorously evaluated, while other pilot programs (in Malawi and Morocco) are currently being evaluated (see Baird, McIntosh, and Özler [2011] for information about the Malawi evaluation) but focus mainly on educational outcomes or adolescent children. 2. However, not all impact studies show consistently positive results for these programs, and differences tend to depend on the specific details about the cash transfer program, the age range of the child studied, the evaluation method, and whether the evaluations were measuring short- or medium- term impacts. Even for outcomes that were directly conditioned upon, such as preventative health clinic visits, evaluations for Chile (Galasso 2006), Ecuador (Paxson and Schady 2010), and Nicaragua (Maluccio and Flores 2005) show no significant change in the

Alternative Cash Transfer Delivery Mechanisms

115

fer pilot program in rural Burkina Faso, the Nahouri Cash Transfers Pilot Project (NCTPP), on the take-up of routine preventative health clinic visits. The NCTPP incorporated a random experimental design to evaluate the relative effectiveness of the following four cash transfer programs targeting poor households in rural Burkina Faso: conditional cash transfers given to fathers, conditional cash transfers given to mothers, unconditional cash transfers given to fathers, and unconditional cash transfers given to mothers. Our evaluation focuses on health utilization for children from birth to fiftynine months old, in particular their visits to health clinics for routine preventative care. Even by African standards, child health outcomes in Burkina Faso are considered to be poor. In 2003, 38.7 percent of children under fifty- nine months were two standard deviations below the reference population for height- for- age, 18.6 percent for weight- for- height, and 37.7 percent for weight- for- age (INSD/ORC Macro 2004). Similarly, 38 percent of children in that age group experienced some form of fever or respiratory infection in the two weeks preceding the national Demographic and Health Survey. Among children who had been ill, only 30 percent received any care from a health professional (INSD/ORC Macro 2004). We find that children under age five in families that received conditional cash transfers have 0.43 more visits to the health clinic for routine preventative care during the previous year compared to children in control households, a 49 percent increase compared to the mean in the control group. We find similar magnitude beneficial impacts for children in families where the mother or father received conditional cash transfers, indicating that at least when the transfers are conditional, the gender of the cash transfer recipient is not a critical factor influencing the frequency of routine health facility visits. In contrast, we do not find beneficial effects from unconditional cash transfers that are given to either mothers or fathers. This highlights the point that at least for this particular outcome of routine preventative health clinic visits, conditionality is more important than transfer recipient gender in increasing health clinic utilization for children. Our experimental design has several distinctive features differentiating it from other studies. First, we randomly test conditional and unconditional cash transfers in the same environment, which allows us to determine whether the key factor driving impacts is an income effect or conditionality.3 Second, unlike many cash transfer programs that pay the transfers to mothers number of visits to the health clinic for preventative reasons, while there were positive impacts on routine health check-ups in Colombia (Attanasio et al. 2005), Honduras (Morris et al. 2004), and Jamaica (Levy and Ohls 2010). 3. A recent study by de Brauw and Hoddinott (2011) makes use of administrative errors in Mexico to try to untangle how much of any observed program impact is due to the program’s income transfers and how much is due to the conditionality forcing households to undertake specific behaviors in order to receive the funds. Baird, McIntosh and Özler (2011) also report results from a randomized experiment in Malawi comparing conditional and unconditional cash transfers, but their target population are adolescent girls and they mainly focus on education outcomes.

116

Richard Akresh, Damien de Walque, and Harounan Kazianga

with the hope mothers will spend the money more wisely, we randomly selected mothers or fathers to receive the payments. Hence we contribute to the long- standing question on the effects of resources controlled by the mother or the father on a child’s human capital. The remainder of the chapter is organized as follows. Section 4.2 describes the cash transfer pilot program as well as its experimental design and presents some descriptive statistics from the household survey. Section 4.3 discusses our empirical identification strategy, and section 4.4 presents main results. Section 4.5 concludes. 4.2 4.2.1

Burkina Faso Nahouri Cash Transfer Pilot Project Program Design

The pilot cash transfer program was conducted in Nahouri province in southern Burkina Faso, located approximately 100 miles from the capital, Ouagadougou. The seventy- five villages in Nahouri province that each have a primary school were randomly allocated to the following five groups (see figure 4.1, panel A): (a) conditional cash transfers given to the father (CCTF), (b) conditional cash transfers given to the mother (CCTM), (c) unconditional cash transfers given to the father (UCTF), (d) unconditional cash transfers given to the mother (UCTM), and (e) a control group. There were fifteen villages in each treatment arm and in the control group. Based on our experience with previous randomized program evaluations in Burkina Faso, as well as qualitative focus groups we conducted in Nahouri province, we believed that transparency in the randomization process was critical for maintaining a household’s participation in the surveys and to guarantee the local authorities’ support. Therefore, the village- level treatment randomization took place during a public meeting in the Nahouri provincial capital in which each village was represented by their local leaders. In the next stage, in each of the sixty villages that were randomly selected to receive a cash transfer program, all poor eligible households were present for a participatory lottery to randomly assign those households to either receive or not receive the particular type of transfer allocated to that village. To determine whether a household was eligible (based on their poverty status) to take part in the lottery, immediately prior to the baseline survey, we conducted an extended household census in every village to collect information from each household about household living structure (cement or mud brick walls, metal or straw roof, flooring, access to latrine), household asset ownership (plow, cart, draft animals, motorcycle, radio), whether the head of household ever attended school, whether the household grows cotton, and whether there was a weekly market in the village. We combined this information with a Burkina Faso nationally representative household survey to calculate a predicted consumption level for each household and

Alternative Cash Transfer Delivery Mechanisms

Fig. 4.1

117

Summary of treatment and control group randomization plan

compare that with the national poverty line to determine whether a household is considered poor and subsequently is eligible to receive the cash transfer. Given the government transfer program’s limited budget, in consultation with village leaders, we decided that randomization was the fairest way to determine which poor eligible households receive a transfer, and everyone was aware that not all poor households in a given village would be receiving a transfer during the pilot program.

118

Richard Akresh, Damien de Walque, and Harounan Kazianga

We conducted three rounds of surveys (baseline, one- year follow-up, and two- year follow-up) in June 2008, June 2009, and June 2010, interviewing all poor households eligible to receive a transfer in each of the treatment villages and who had been randomly selected to receive the transfer. In each of these four groups of fifteen villages, we interviewed approximately 540 poor households randomly selected to receive transfers. The control group consisted of 615 randomly selected poor households that did not receive a cash transfer in the fifteen control villages where no households received cash transfers. Among households randomly assigned to a CCT scheme, for their children under age six, the conditions imposed were quarterly visits to the local health clinic for child growth monitoring, while for children age seven to fifteen, the conditions were school enrollment with an attendance rate above 90 percent each quarter. In the villages randomly assigned to conditional cash transfers, satisfaction of the conditions was assessed using a family booklet in which the school teachers and the health workers signed and stamped the booklet for a given child to confirm school attendance and health visits, respectively. Further, in the CCT villages, local village committees, which had received specific training, randomly selected 20 percent of the booklets and verified the information reported in those with data from the school attendance registers and the health center visits register. For families randomly assigned to a UCT program, the mother or father received a quarterly stipend for each child without conditions. For each child under age six, in the CCT and UCT programs, the mother or father would receive 1,000 FCFA per quarter, for a total of 4,000 FCFA per year. Using the exchange rate during the 2008 baseline of 415 FCFA = 1 USD, the annual transfer was worth approximately $9.64, which is 9 percent of household per capita expenditures. For each child age seven to ten (or in grades 1 to 4), the mother or father would receive 2,000 FCFA per quarter (8,000 FCFA per year), while for each child age eleven to fifteen (or in grades 5 or higher), the mother or father would receive 4,000 FCFA per quarter (16,000 FCFA per year).4 There was no household transfer cap amount, so that each age- eligible child could participate irrespective of the composition of the family. The transfers were made in cash during a quarterly visit conducted by the program staff. Cash transfers were disbursed during the academic school year 2008/09 (October 2008 to June 2009) and 2009/10 (October 2009 to June 2010). The program design assumes that each of the treatment groups would receive equal amounts of resources per capita over the two- year transfer program period if households randomly allocated to the conditional cash transfers fully satisfied the conditions. 4. To minimize child fostering in response to the program introduction and reduce any potential sample attrition (see Akresh [2009] for evidence on the relationship between income shocks and child fostering), eligibility for cash transfers was based only on the children present in the household at the time of the baseline survey.

Alternative Cash Transfer Delivery Mechanisms Table 4.1

119

Summary statistics of Burkina Faso Nahouri Cash Transfer Pilot Project (NCTPP) evaluation data Mean (1)

Variable

Standard deviation (2)

Household characteristics Household size 7.023 Number of children under sixty months 1.474 Number of children age five to fifteen years 2.317 Proportion either parent ever enrolled in school 0.185 Household expenditures per capita (in FCFA) 46,257

3.294 0.952 1.772 0.388 66,105

Child characteristics Child age (in months) 30.74 Child gender (1 = female) 0.494 Routine preventative health clinic visits 1.025

16.85 0.500 1.823

Number of households Number of children

1,618 2,559

Source: Nahouri Cash Transfer Pilot Project (NCTPP) evaluation data from 2010. Notes: Household characteristics are based on the 1,618 households that were eligible to receive cash transfers (treatment and control groups) and that have children under sixty months old. Child characteristics are based on the 2,559 children under sixty months old in these households. Per capita household expenditures are measured in FCFA (415 FCFA = USD 1).

Meetings were also organized between the central government and the provincial health and education authorities to ensure the full participation of school administrators (principals and teachers) and health administrators (doctors and nurses) in the program. This included ensuring their role in verifying the cash transfer conditionality. However, no other changes were made and no additional government funds were provided for the supply of education or health services. 4.2.2

Descriptive Statistics

Households in the Nahouri region are predominantly subsistence farmers growing sorghum and groundnuts and have mean annual per capita expenditures of approximately 111 USD. Summary statistics in table 4.1 show that, on average, there are 7.0 members in each household, of whom 1.5 are children under sixty months and 2.3 are children age five to fifteen. Focusing on the children under sixty months old that are used in this chapter’s analysis, 49.4 percent of them are female and their average age is 30.7 months old.5 Only 18.5 percent of these children’s parents (father or mother) have ever been enrolled in school. On average, children have 1.03 routine 5. Note that of the poor eligible surveyed households in the treatment and control villages, only 1,618 households are used in the health analysis for this chapter because the remaining households did not have any children under sixty months old.

120

Richard Akresh, Damien de Walque, and Harounan Kazianga

preventative health clinic visits per year, although younger children (birth to twenty- three months) have more visits (1.43) than older (twenty- four to fifty- nine months) children (0.80).6 The health care system in rural Burkina Faso is relatively underdeveloped. In the seventy- five survey villages, not every village has a health clinic, but the median distance to the nearest health clinic is only two kilometers (average distance is 2.85 kilometers). This is below the Burkina Faso national average distance to a basic health care center, which is 7.5 kilometers (Ministry of Health 2009). In addition, there are no statistically significant differences across treatment and control groups in the average distance to a health clinic. Access to basic health care remains poor in Burkina Faso. In 2007, only 36 percent of the population had access to basic health care (Ministry of Heath 2010). The available data also suggest that the health system is understaffed. In 2009, the ratios of total population to health care professionals indicate there are 22,522 people per physician, 94,564 per pharmacist, and 2,892 people per nurse (Ministry of Health 2010). The poor access to health services translates into lower utilization of health care, especially preventive health care. In rural Burkina Faso, it is estimated that only 60 percent of baby deliveries are attended by a trained health care provider, with large variations across regions (INSD/ORC Macro 2010). Noticeable exceptions include immunization coverage (more than 80 percent of children in rural areas) and utilization of prenatal care (around 94 percent of pregnant women in rural areas). In table 4.2, we use baseline data to test the balance of the randomization experiment. We first present the mean of the specific variable measured at the baseline for the control group and each of the four treatment arms. In the subsequent columns, we present p- values from a Wald test comparing the treatment arm with the control group. Finally, in the last column (under the control group heading), we present the p- value for an F-test of the joint test that the means of the five groups are equal. Results show balance across characteristics for children, parents, and measures of supplyside health clinic quality. The main dependent variable used in this chapter, routine preventative health clinic visits, was only collected in the final survey round, so we are unable to examine whether the variable exhibits balance in the baseline round (preintervention) across the treatment and control groups.7 Household attrition was very low between the baseline and one- year 6. The American Academy of Pediatrics recommends the following schedule for preventive childhood health care visits: birth, two months, four months, six months, nine months, twelve months, fifteen months, eighteen months, twenty- four months, thirty- six months, forty- eight months, and sixty months. 7. However, we are able to test baseline balance for other health outcomes for which we do have three rounds of data, such as the probability the child was sick during the past month, health clinic utilization, and illness duration, and all of these exhibit balance across randomized treatment and control groups in the baseline.

0.507 29.745 0.144 0.155 1.056 6.535 0.593 0.207 0.200 0.570 0.383 0.028 0.578 0.215 0.278 0.498 2.180 9.089

Mean 0.191 0.413 0.342 0.420 0.122 0.016** 0.451 0.108 0.449 0.943 0.545 0.443 0.745 0.907 0.147 0.536 0.097* 0.347

Child and parent characteristics 0.525 0.488 0.993 0.454 0.095* 29.450 0.177 29.070 0.732 0.162 0.855 0.125 0.806 0.179 0.501 0.184 0.564 1.064 0.496 1.121 0.157 6.422 0.276 6.897 0.761 0.576 0.378 0.579 0.480 0.214 0.391 0.239 0.802 0.210 0.971 0.181 0.969 0.370 0.217 0.553 0.915 0.487 0.464 0.271 0.537 0.073 0.754 0.114 0.971 0.505 0.564 0.609 0.963 0.262 0.437 0.221 0.737 0.308 0.926 0.222 0.849 0.414 0.465 0.544 0.939 2.209 0.709 2.508 0.980 9.215 0.798 9.835

p-value

UCTF p-value

Mean

CCTM Mean

p-value

CCTF

Baseline means and randomization balance

Child is female Child age in months Head of household is female Proportion either parent ever enrolled Number of wives of household head Household size Marital status = monogamous Marital status = polygamous Marital status = single Ethnic group = Kassena Ethnic group = Nankana/Farfarse Ethnic group = Mossi Ethnic subgroup = Nakomse Religion = Muslim Religion = Christian Religion = Animist Number of wives of HH head’s father Number of children of HH head’s father

Table 4.2

0.489 28.154 0.106 0.200 1.141 6.569 0.622 0.226 0.152 0.711 0.144 0.089 0.585 0.257 0.270 0.459 2.174 8.955

Mean 0.974 0.839 0.108 0.298 0.078* 0.125 0.554 0.215 0.105 0.307 0.119 0.561 0.919 0.546 0.563 0.835 0.979 0.839

p-value

UCTM

0.488 28.336 0.156 0.143 1.010 6.049 0.605 0.184 0.211 0.564 0.366 0.057 0.574 0.211 0.302 0.478 2.171 9.075

Mean

0.317 0.276 0.353 0.824 0.380 0.189 0.509 0.547 0.362 0.209 0.134 0.393 0.927 0.917 0.469 0.701 0.568 0.875 (continued)

p-value

Control

(continued)

9.53 5.091 3.467 0.867 0.933 0.733 0.667 1.200

Mean

Mean

p-value

CCTM

Village-level health clinic variables 0.593 10.57 0.885 0.871 4.455 0.576 0.406 3.286 0.352 0.184 0.929 0.370 0.371 0.929 0.370 0.796 0.929 0.309 0.231 0.571 0.565 0.321 1.214 0.383

p-value

CCTF

9.73 4.375 3.333 0.933 0.933 0.667 0.538 1.267

Mean 0.782 0.582 0.458 0.366 0.366 0.545 0.656 0.542

p-value

UCTF

12.08 4.125 3.500 0.692 0.769 0.462 0.636 1.692

Mean

0.425 0.515 0.574 0.089* 0.164 0.051* 0.370 0.287

p-value

UCTM

10.31 5.250 4.000 1.000 1.000 0.769 0.462 1.385

Mean

0.411 0.818 0.863 0.296 0.461 0.069* 0.648 0.127

p-value

Control

Source: Nahouri Cash Transfer Pilot Project (NCTPP) evaluation data from 2008. Notes: The treatment arms are abbreviated as CCTF (conditional cash transfers given to fathers), CCTM (conditional cash transfers given to mothers), UCTF (unconditional cash transfers given to fathers), and UCTM (unconditional cash transfers given to mothers). The p-values are from a Wald test comparing the treatment arm with the control group. In the last column (under the control group heading), the p-values are from an F-test of the joint test that the means of the five groups are equal. Robust standard errors, clustered at village level. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

Villages per health clinic Number of days a doctor is present Number of nurses Childhood nutritional/growth counseling offered Vaccines offered Nutritional supplements offered Health epidemics Source of funding for clinic (patient, gov’t., NGO)

Table 4.2

Alternative Cash Transfer Delivery Mechanisms

123

follow-up survey (1.42 percent), and increases slightly when comparing the baseline and two- year follow-up survey (4.62 percent). In table 4.3, we explore the relative differences between attritor and nonattritor households. Column (1) presents means of the household- level characteristics from the baseline survey for the sample of households that were followed from the baseline to the two- year follow-up survey (nonattritors). Column (2) presents means for the sample of attritor households, and column (3) presents the average difference in characteristics between attritors and nonattritors, as well as a test of whether the difference is statistically significant. Results suggest that attrition is not likely random, as attritors are more likely to come from smaller households, with fewer adults, are younger, have fewer wives, and are more likely to be female- headed households. However, what is more relevant for our analysis is whether attrition differs across treatment and control groups. In columns (4) to (7) of table 4.3, we present differencein-differences regressions for each characteristic comparing the difference between attritors and nonattritors in each specific treatment arm with the same difference between attritors and nonattritors in the control group, and we find few significant differences. 4.3

Empirical Identification Strategy

The key question we want to answer is whether cash transfers increased the frequency of routine preventative visits to health facilities by young children in recipient households. The experimental design provides a strong identification strategy and allows us to attribute any differences in outcome indicators (frequency of preventative visits to health clinics) between the treatment and control groups to the impact of the program. Because precise questions about the purpose of the visits to the health clinics were only included in the last round of data collection (round 3), we cannot use a difference- in-differences model. We rely on the program randomization to identify causal impacts of alternative cash transfer delivery mechanisms on routine health clinic visits. Since table 4.2 indicates that health behaviors and supply- side measures of health clinic quality were well balanced across the five study groups at baseline, this provides support for our identification strategy. We focus on the program effects on the treated households. We start by pooling the treatment arms to estimate two distinct specifications. First, we consider households that were either randomly selected to receive conditional cash transfers or randomly selected to receive unconditional cash transfers (see panel B of figure 4.1). This approach combines into one group the conditional cash transfers given to fathers or mothers and into a second group the unconditional cash transfers given to fathers or mothers. With this specification, we focus on whether the conditionality matters and we ignore the intrahousehold allocation aspects of the design. Formally, the regression is specified as follows:

Single

Monogamous

Polygamous

HH head gender

HH head educated

HH head age

HH head literate

Household size

Number of adults

No. children < 60 months

46,036 (1,165) 1.095 (0.021) 3.123 (0.033) 6.551 (0.062) 0.157 (0.007) 45.25 (0.286) 0.157 (0.007) 1.136 (0.007) 0.218 (0.008) 0.593 (0.010) 0.189 (0.008)

Nonattritors (n = 2,622) (1) 50,228 (4,003) 1.118 (0.070) 2.575 (0.119) 5.323 (0.214) 0.189 (0.035) 42.23 (1.329) 0.142 (0.031) 1.189 (0.037) 0.102 (0.027) 0.042 (0.661) 0.236 (0.038)

Attritors (n = 127) (2) 4,192 (5,365) 0.023 (0.097) –0.548*** (0.153) –1.228*** (0.284) 0.032 (0.033) –3.02** (1.330) –0.015 (0.033) 0.053* (0.032) –0.116*** (0.037) 0.069 (0.045) 0.036 (0.047)

Mean difference (3)

Relative differences between attriting and nonattriting households

HH expenditures per capita (in FCFA)

Variables

Table 4.3

–4,240 (9,396) –0.083 (0.163) –0.005 (0.328) 0.524 (0.475) 0.007 (0.078) 0.122 (2.986) –0.078 (0.071) 0.092 (0.106) 0.049 (0.062) –0.137 (0.098) 0.088 (0.101)

CCT diff. in diff. (4) –10,172 (9,239) –0.064 (0.189) 0.156 (0.319) 0.478 (0.518) –0.046 (0.094) 1.521 (3.209) –0.106 (0.075) 0.085 (0.095) 0.016 (0.064) –0.160 (0.099) 0.144 (0.094)

UCT diff. in diff. (5) –7,525 (10,111) –0.169 (0.161) –0.014 (0.331) 0.309 (0.471) –0.001 (0.079) –1.935 (3.014) –0.020 (0.060) 0.173* (0.093) 0.028 (0.063) –0.171* (0.090) 0.143 (0.094)

CTF diff. in diff. (6)

–10,589 (8,669) 0.062 (0.180) 0.170 (0.320) 0.848* (0.501) –0.031 (0.093) 3.821 (2.982) –0.181** (0.086) –0.025 (0.113) 0.053 (0.063) –0.112 (0.115) 0.059 (0.109)

CTM diff. in diff. (7)

0.554 (0.010) 0.334 (0.009) 0.231 (0.008) 0.272 (0.009) 0.484 (0.010) 2.247 (0.038) 9.253 (0.139)

0.591 (0.044) 0.260 (0.039) 0.260 (0.039) 0.378 (0.043) 0.354 (0.043) 2.214 (0.165) 8.714 (0.531)

0.037 (0.045) –0.074* (0.043) 0.029 (0.038) 0.106*** (0.041) –0.130*** (0.045) –0.033 (0.177) –0.539 (0.641)

–0.063 (0.153) 0.186 (0.145) 0.124 (0.107) 0.022 (0.090) –0.129 (0.095) –0.550 (0.334) –0.208 (1.510)

–0.015 (0.152) 0.065 (0.140) 0.116 (0.111) –0.021 (0.128) –0.100 (0.134) –0.246 (0.475) 1.556 (1.316)

–0.033 (0.148) 0.092 (0.146) 0.123 (0.104) 0.078 (0.082) –0.180* (0.098) –0.369 (0.392) 0.995 (1.343)

–0.009 (0.161) 0.135 (0.144) 0.113 (0.115) –0.102 (0.117) –0.018 (0.111) –0.450 (0.370) –0.1000 (1.721)

Source: Nahouri Cash Transfer Pilot Project (NCTPP) data from 2008. Notes: Column (1) presents means of household-level characteristics from the baseline survey for the sample of households that were followed from the baseline to the two-year follow-up survey (nonattritors). Column (2) presents means for the sample of attritor households, and column (3) presents the average difference in characteristics between attritors and nonattritors. The treatment arms are abbreviated as CCT (conditional cash transfer), UCT (unconditional cash transfer), CTF (cash transfer given to fathers), and CTM (cash transfer given to mothers). Columns (4)–(7) test for differential impacts of attrition between treatment and control groups. Specifically, for each characteristic, we estimate difference-in-differences regressions comparing attritors and nonattritors for the treatment (CCT, UCT, CTF, CTM) and control groups. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

No. kids of HH head’s father

No. wives of HH head’s father

Religion = Animist

Religion = Christian

Religion = Muslim

Ethnic group = Nankana

Ethnic group = Kassena

126

(1)

Richard Akresh, Damien de Walque, and Harounan Kazianga

yih = b0 + a1CCTh + a2UCTh + b1X ih + ´ih,

where yih is the number of routine preventative health clinic visits for child i in household h, CCTh and UCTh are treatment indicators that take the value one if a child lives in a household h that was randomly selected to receive, respectively, conditional or unconditional cash transfers and is zero otherwise, Xih is a vector of child characteristics (gender and age) and ε is an error term.8 The estimated impact for conditional cash transfers is given by a1 and for unconditional cash transfers by a2. Second, we consider households that were either randomly assigned to receive the stipends via the mother or via the father (panel C of figure 4.1). This approach combines into one group the conditional and unconditional cash transfers given to fathers and into a second group the conditional and unconditional cash transfers given to mothers. When estimating this specification, we focus on intrahousehold allocation (whether paying the stipends to the mother or the father leads to different outcomes), and we ignore the role that conditionality might play. We estimate the following regression: (2)

yih = b0 + a1CTFh + a2CTMh + b1X ih + ´ih,

where CTFh indicates a household h that is randomly selected to receive cash transfer given to fathers, CTMh indicates a household h that is randomly selected to receive cash transfer given to mothers, and all of the other variables are as defined previously. In equation (2), a1 and a2 represent the impact of cash transfers to fathers and to mothers, respectively. It is plausible that mothers and fathers react differently not only to the transfers, but also to the conditionality. In order to test this hypothesis, we allow the four treatment groups to enter separately in the regression. The resulting specification is: (3) yih = b0 + a1CCTFh + a2CCTMh + a3UCTFh + a4UCTMh + b1X ih + ´ih, where the four treatment groups are defined as binary variables and are represented by CCTFh (conditional cash transfers to fathers), CCTMh (conditional cash transfers to mothers), UCTFh (unconditional cash transfers to fathers), and UCTMh (unconditional cash transfers to mothers). All of the other variables are defined as before. The estimated impact of each treatment scheme is given by the associated coefficient a1, a2, a3, and a4. 4.4

Empirical Results

In table 4.4, we present results from estimating the three equations discussed above for routine preventative health clinic visits. Each regression includes controls for the child’s gender and year of birth cohort. All regres8. Correlation among the error terms of children in a given village experiencing the same shocks might bias the ordinary least squares standard errors downward, so in all regressions we cluster the standard errors at the village level.

127

Alternative Cash Transfer Delivery Mechanisms Table 4.4

Impact of cash transfers on routine preventative health clinic visits

Dependent variable: Routine health clinic visits Conditional cash transfer (CCT) Unconditional cash transfer (UCT)

(1)

(2)

0.431** [0.205] –0.079 [0.195]

Cash transfer fathers (CTF)

0.070 [0.209] 0.235 [0.201]

Cash transfer mothers (CTM) Conditional cash transfer fathers (CCTF)

0.415 [0.258] 0.446** [0.223] –0.216 [0.200] 0.046 [0.231]

Conditional cash transfer mothers (CCTM) Unconditional cash transfer fathers (UCTF) Unconditional cash transfer mothers (UCTM) Child age & gender fixed effects? Number of observations P-value testing equality between CCT and UCT P-value testing equality between fathers and mothers P-value testing equality between CCTF and CCTM P-value testing equality between CCTF and UCTF P-value testing equality between CCTF and UCTM P-value testing equality between CCTM and UCTF P-value testing equality between CCTM and UCTM P-value testing equality between UCTF and UCTM

(3)

Yes

Yes

Yes

2,559

2,559

2,559

0.002 0.346 0.901 0.008 0.156 0.001 0.076 0.190

Source: Nahouri Cash Transfer Pilot Project (NCTPP) evaluation data from 2010. Notes: All regressions are restricted to children under sixty months old and include child age fixed effects and child gender dummies. Robust standard errors in brackets, clustered at the village level. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

sions are restricted to children under sixty months old. Column (1) focuses on the comparison between conditional and unconditional cash transfers, while column (2) focuses on the comparison between transfers given to fathers or mothers. Finally, in column (3), we examine the impact of all four treatment groups separately. We find that conditional cash transfers have a larger impact on routine health clinic visits than unconditional cash transfers. Young children less than sixty months old in households that receive conditional cash transfers have 0.43 more routine preventative health clinic visits in the preceding year compared to comparable children in the control villages. This represents an

128

Richard Akresh, Damien de Walque, and Harounan Kazianga

increase of 42 percent compared to the average number of routine visits. We find no significant impact of unconditional cash transfers on these routine health visits, and relative to unconditional cash transfers, the impact on routine health clinic visits is significantly larger for conditional transfers. We are able to strongly reject equality between the coefficients on the conditional and unconditional variables with a p- value of 0.002. Turning to evaluate the impact of cash transfers given to fathers compared to mothers, in column (2), we estimate a regression with the same dependent variable but with independent variables indicating if a child was in a household where the cash transfers were given to fathers or to mothers. Overall, children in households that received cash transfers (either to the father or mother) receive more routine preventative health clinic visits than children in control households, although neither coefficient is statistically significant. Although the point estimate is larger for children in households where the mother received the transfer (0.235), we cannot reject equality between that coefficient and the coefficient for cash transfer to fathers (0.070) with a p- value of the test of equality of 0.35. The above results where the cash transfers are grouped by the presence or absence of conditionality (column [1]) and the transfer recipient’s gender (column [2]) potentially hide the fact that there could be significant differences between how conditionality interacts with a recipient’s gender. To explore this further, in column (3), we estimate equation (3) in which we include four indicator variables corresponding to each of the four different treatments (conditional cash transfers to fathers, conditional cash transfers to mothers, unconditional cash transfers to fathers, and unconditional cash transfers to mothers). Results highlight that conditionality is critical for improving take-up of routine preventative health clinic visits. Children living in households where the cash transfers were conditional and given to mothers have 0.446 more preventative health clinic visits and the coefficient is statistically significant at the 5 percent level. For children living in households where the cash transfers were conditional and given to fathers, the results indicate they have 0.415 more routine health clinic visits, although the coefficient is not statistically significant at standard levels. In testing the equality of these two coefficients, we are not able to reject equality, with the p- value equal to 0.901. These results for conditional cash transfers contrast with those for unconditional cash transfers. Children living in households that received unconditional cash transfers given to fathers or living in households that received unconditional cash transfers given to mothers show no significant impact on the take-up of routine preventative health clinic visits. The point estimate is positive, but small for unconditional cash transfers to mothers and is actually negative for fathers, indicating children in these two treatment arms are no better off compared to comparable children in the control households. Formally testing equality between the coefficients shows that the impact of conditional cash transfers to mothers is larger than both unconditional cash transfers to mothers or fathers. The impact of condi-

Alternative Cash Transfer Delivery Mechanisms

129

tional cash transfers to fathers is larger than unconditional cash transfers to fathers and mothers, although we can only reject equality with a p- value of 0.156 for the test with mothers. In the next three tables, we examine if there is a heterogeneous impact of cash transfers based on child gender (table 4.5), child age (table 4.6), or the baseline poverty status of the household (table 4.7). In table 4.5, impacts of

Table 4.5

Impact of cash transfers on routine preventative health clinic visits by gender

Dependent variable: Routine health clinic visits CCT UCT

Boys (1)

(2)

(3)

0.389 [0.242] –0.045 [0.229]

Cash transfer fathers (CTF)

(4)

(5)

(6)

0.477** [0.222] –0.106 [0.207] 0.010 [0.232] 0.283 [0.240]

Cash transfer mothers (CTM) CCT-father

0.136 [0.235] 0.190 [0.210] 0.244 [0.290] 0.505* [0.269] –0.169 [0.234] 0.075 [0.272]

CCT-mother UCT-father UCT-mother Child age & gender fixed effects?

Girls

0.580* [0.301] 0.384* [0.223] –0.260 [0.203] 0.023 [0.251]

Yes

Yes

Yes

Yes

Yes

Yes

Number of observations

1,296

1,296

1,296

1,263

1,263

1,263

P-value testing CCT = UCT P-value testing CTF = CTM P-value testing CCTF = CCTM P-value testing CCTF = UCTF P-value testing CCTF = UCTM P-value testing CCTM = UCTF P-value testing CCTM = UCTM P-value testing UCTF = UCTM

0.019

0.002 0.140

0.790 0.354 0.100 0.547 0.003 0.102 0.281

0.497 0.003 0.076 0.001 0.128 0.190

Source: Nahouri Cash Transfer Pilot Project (NCTPP) evaluation data from 2010. Notes: All regressions are restricted to children under sixty months old and include child age fixed effects and child gender dummies. The treatment arms are abbreviated as CCT (conditional cash transfer), UCT (unconditional cash transfer), CTF (cash transfer given to fathers), CTM (cash transfer given to mothers), CCT-father (conditional cash transfer given to fathers), CCT-mother (conditional cash transfer given to mothers), UCT-father (unconditional cash transfer given to fathers), and UCT-mother (unconditional cash transfer given to mothers). Robust standard errors in brackets, clustered at the village level. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

130

Richard Akresh, Damien de Walque, and Harounan Kazianga

alternative cash transfers are fairly consistent across boys and girls. Conditional cash transfers have a larger impact than unconditional cash transfers for both boys and girls, and we are able to reject equality in both regressions ( p- value equal to 0.019 for boys and 0.002 for girls). We find no statistically significant impacts on routine health clinic visits when the cash transfers are grouped by mother or father (columns [2] and [5]). We find suggestive evidence that conditional cash transfers to fathers have a larger impact on routine health clinic visits for girls (an additional 0.580 visits), while conditional cash transfers to mothers have a larger impact on health clinic visits for boys (an additional 0.505 visits). Conditional cash transfers to mothers still have a significant impact for girls, and we are not able to reject equality between the impact of conditional cash transfers to fathers or mothers (for either boys or girls). Unconditional cash transfers to fathers or mothers show no significant positive impact for either boys or girls. Table 4.6 presents separate results for younger (ages birth to twenty- three months) and older (ages twenty- four to fifty- nine months) children. All of the impacts of cash transfers on routine preventative health clinic visits discussed previously are being driven by older children. This might be explained by the fact that parents in Burkina Faso are less likely to bring older children for routine preventative care visits. Indeed, as noted previously, on average, younger children have more visits (1.43) than older children (0.80), so that, while routine visits for young children might already be frequent for young children in the absence of the transfers, the transfers might have had a relatively larger impact on the less frequent visits for older children. We find no statistically significant impact of cash transfers on routine health clinic visits for younger children. To put this in perspective, this means that young children (birth to twenty- three months) in control households are as equally likely to have gone to a health clinic for a routine visit as a young child in a treatment household. This contrasts with older children for whom there are large impacts of the cash transfers on the number of routine health clinic visits. Conditional cash transfers yield larger impacts on increasing the number of routine preventative health clinic visits compared to unconditional cash transfers, with an increase of 0.592 visits for the older children, representing an increase of 74 percent over the mean number of visits. For older children, cash transfers to mothers or to fathers show increases in routine health clinic visits, although only the coefficient on the mother’s variable is statistically significant. Finally, conditional cash transfers to fathers or mothers show larger impacts for the older children, increasing the number of health clinic visits by 0.630 and 0.558 visits, respectively. This represents increases of 79 and 70 percent, respectively, for cash transfers to fathers or mothers that are conditional, compared to the average number of routine health clinic visits children of this age receive. In table 4.7, we do not find any significant differences between the impacts

Alternative Cash Transfer Delivery Mechanisms Table 4.6

131

Impact of cash transfers on routine preventative health clinic visits by child age

Dependent variable: Routine health clinic visits CCT

Young children (birth to 23 months) (1)

(2)

(3)

0.098 [0.498] –0.516 [0.485]

UCT Cash transfer fathers (CTF)

Older children (24–59 months) (4)

(6)

0.592*** [0.203] 0.143 [0.155] –0.406 [0.488] –0.086 [0.498]

Cash transfer mothers (CTM)

(5)

CCT-father

0.307 [0.192] 0.391** [0.172] –0.030 [0.492] 0.198 [0.554] –0.704 [0.501] –0.347 [0.506]

CCT-mother UCT-father UCT-mother

0.630** [0.308] 0.558** [0.218] 0.033 [0.151] 0.244 [0.206]

Child age & gender fixed effects?

Yes

Yes

Yes

Yes

Yes

Yes

Number of observations

897

897

897

1,662

1,662

1,662

P-value testing CCT = UCT P-value testing CTF = CTM P-value testing CCTF = CCTM P-value testing CCTF = UCTF P-value testing CCTF = UCTM P-value testing CCTM = UCTF P-value testing CCTM = UCTM P-value testing UCTF = UCTM

0.013

0.024 0.196

0.671 0.515 0.010 0.236 0.016 0.144 0.211

0.832 0.051 0.247 0.013 0.216 0.280

Source: Nahouri Cash Transfer Pilot Project (NCTPP) evaluation data from 2010. Notes: All regressions are restricted to children under sixty months old and include child age fixed effects and child gender dummies. The treatment arms are abbreviated as CCT (conditional cash transfer), UCT (unconditional cash transfer), CTF (cash transfer given to fathers), CTM (cash transfer given to mothers), CCT-father (conditional cash transfer given to fathers), CCT-mother (conditional cash transfer given to mothers), UCT-father (unconditional cash transfer given to fathers), and UCT-mother (unconditional cash transfer given to mothers). Robust standard errors in brackets, clustered at the village level. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

for cash transfers given to extremely poor or less poor households. As discussed in section 4.2.1, recall that all households that were eligible to receive the cash transfers were below the estimated national poverty line. The further distinction in table 4.7 is comparing extremely poor and less poor households, which might explain the absence of differential impacts

Table 4.7

Impact of cash transfers on routine preventative health clinic visits by baseline poverty status

Dependent variable: Routine health clinic visits CCT UCT

Poor households (1)

(2)

Nonpoor households (3)

0.450* (0.239) –0.132 (0.215)

Cash transfer fathers (CTF)

(5)

(6)

0.389* (0.234) 0.068 (0.222) 0.015 (0.233) 0.204 (0.228)

Cash transfer mothers (CTM) CCT-father

0.199 (0.240) 0.281 (0.223)

Yes

Yes

0.453 (0.302) 0.447* (0.270) –0.300 (0.207) 0.029 (0.264) Yes

Number of observations

1,731

1,731

1,731

P-value testing CCT = UCT P-value testing CTF = CTM P-value testing CCTF = CCTM P-value testing CCTF = UCTF P-value testing CCTF = UCTM P-value testing CCTM = UCTF P-value testing CCTM = UCTM P-value testing UCTF = UCTM

0.005

CCT-mother UCT-father UCT-mother Child age & gender fixed effects?

(4)

Yes

Yes

0.324 (0.298) 0.434* (0.260) 0.053 (0.274) 0.081 (0.240) Yes

828

828

828

0.102 0.364

0.685 0.985 0.006 0.182 0.002 0.142 0.145

0.715 0.387 0.387 0.173 0.148 0.913

Source: Nahouri Cash Transfer Pilot Project (NCTPP) evaluation data from 2010. Notes: All regressions are restricted to children under sixty months old and include child age fixed effects and child gender dummies. All households that were eligible to receive the cash transfers were below the estimated national poverty line. The further distinction in this table is comparing extremely poor and less poor households. Extremely poor households are defined as being below the median household per capita expenditure level in the baseline survey. The treatment arms are abbreviated as CCT (conditional cash transfer), UCT (unconditional cash transfer), CTF (cash transfer given to fathers), CTM (cash transfer given to mothers), CCT-father (conditional cash transfer given to fathers), CCT-mother (conditional cash transfer given to mothers), UCT-father (unconditional cash transfer given to fathers), and UCT-mother (unconditional cash transfer given to mothers). Robust standard errors in brackets, clustered at the village level. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

Alternative Cash Transfer Delivery Mechanisms

133

by poverty status.9 Conditional cash transfers increase the number of routine preventative health care visits for children in those households and the households’ poverty status does not alter this relationship (columns [1] and [4]). There are no significant improvements on health clinic visits for the regressions when we group transfers to mothers or transfers to fathers. As before, conditionality is driving our observed results and the gender of the recipient does not matter, and the estimated impacts are similar between extremely poor and less poor households. 4.5

Conclusions

This chapter presents evidence of the health impacts from a cash transfer pilot program in rural Burkina Faso, the Nahouri Cash Transfer Pilot Program (NCTPP). Our evaluation focuses on routine preventative health clinic visits for children ages birth to five years old. The NCTPP incorporated a random experimental design to evaluate the relative effectiveness of the following modalities for delivering cash transfers: conditional versus unconditional and transfers to mothers versus fathers. Families under the conditional cash transfer schemes were required to obtain child growth monitoring at local health clinics for all young children. Our results indicate that children in families that received conditional cash transfers had an additional 0.43 routine preventative health clinic visits during the previous year compared to children in the control households. These results contrast with those for unconditional cash transfers given to either fathers or mothers, which showed no beneficial impacts. As long as the transfers were conditional, we did not find any significant difference between whether the money was given to fathers or to mothers, indicating that at least for routine health clinic visits, the cash transfer recipient’s gender is not a critical factor influencing outcomes. One limitation of our analysis is that the health clinic visits are self- reported by the parents. In future work, we will investigate more objectively measured child health outcomes such as anthropometrics and infant mortality. From a policy perspective, conditional cash transfers appear to have stronger beneficial impacts on increasing the number of routine health care visits for children. This finding should be balanced with the additional administrative and logistical costs implied by imposing and verifying the conditionality. For the outcome analyzed in this chapter—routine preventative health clinic visits—the gender of the recipient does not seem to affect critically the impact of the cash transfers on those health outcomes. Such a conclusion, however, needs to be included in a larger analysis of other health and schooling outcomes as well as other age ranges for the children. 9. To make this additional distinction in table 4.7, extremely poor households are defined as being below the median household per capita expenditure level in the baseline survey.

134

Richard Akresh, Damien de Walque, and Harounan Kazianga

Indeed, the NCTPP is a broad social protection pilot program that covers all children until age fifteen in the eligible households and aims to improve both their schooling and health outcomes. It should be emphasized that while we focus on children under age five in this analysis, those children could have benefited from the (larger) transfers received for their older siblings. It is not clear that by only implementing the part of the intervention targeted for young children (birth to fifty- nine months), similar health results would necessarily be obtained.

References Akresh, Richard. 2009. “Flexibility of Household Structure: Child Fostering Decisions in Burkina Faso.” Journal of Human Resources 44 (4): 976– 97. Attanasio, Orazio, Luis Carlos Gomez, Patricia Heredia, and Marcos VeraHernandez. 2005. “The Short-Term Impact of a Conditional Cash Subsidy on Child Health and Nutrition in Colombia.” Report Summary, Familias 03, London, Institute for Fiscal Studies. Baird, Sarah, Craig McIntosh, and Berk Özler. 2011. “Cash or Condition? Evidence from a Cash Transfer Experiment.” Quarterly Journal of Economics 126 (4): 1709– 53. Brooker, Simon, Andrew Hall, and Donald A. P. Bundy. 1999. “Short Stature and the Age of Enrollment in Primary School: Studies in Two African Countries.” Social Science and Medicine 48 (5): 675– 82. Case, Anne, and Angus Deaton. 1998. “Large Cash Transfers to the Elderly in South Africa.” Economic Journal 108 (450): 1330– 61. Case, Anne, Victoria Hosegood, and Frances Lund. 2005. “The Reach and Impact of Child Support Grants: Evidence from KwaZulu-Natal.” Development Southern Africa 22 (4): 467– 82. Dasgupta, Partha. 1993. An Inquiry into Well-Being and Destitution. Oxford: Clarendon Press. de Brauw, Alan, and John Hoddinott. 2011. “Must Conditional Cash Transfer Programs be Conditioned to be Effective? The Impact of Conditioning Transfers on School Enrollment in Mexico.” Journal of Development Economics 96 (2): 359– 70. Duflo, Esther 2003. “Grandmothers and Granddaughters: Old-Age Pensions and Intrahousehold Allocation in South Africa.” World Bank Economic Review 17 (1): 1– 25. Fiszbein, Ariel, and Norbert Schady 2009. Conditional Cash Transfers: Reducing Present and Future Poverty. World Bank Policy Research Report. Washington, DC: World Bank. Galasso, Emanuela. 2006. “With Their Effort and One Opportunity: Alleviating Extreme Poverty in Chile.” Unpublished manuscript, World Bank. Glewwe, Paul, Hannan Jacoby, and Elizabeth King. 2001. “Early Childhood Nutrition and Academic Achievement: A Longitudinal Analysis.” Journal of Public Economics 81 (3): 345– 68. Grantham-McGregor, Sally, Yin Bun Cheung, Santiago Cueto, Paul Glewwe, Linda Richter, and Barbara Strupp. 2007. “Development Potential in the First 5 Years for Children in Developing Countries.” Lancet 369 (9555): 60– 70.

Alternative Cash Transfer Delivery Mechanisms

135

INSD/ORC Macro. 2004. “Enquête Démographique et de Santé 2003.” Ouagadougou : Ministère de l’Économie et du Développement, Burkina Faso. ———. 2010. “Enquête Démographique et de Santé 2010: Rapport Préliminaire.” Ouagadougou : Ministère de l’Économie et du Développement, Burkina Faso. Levy, Dan, and James Ohls. 2010. “Evaluation of Jamaica’s PATH Conditional Cash Transfer Program” Journal of Development Effectiveness 2 (4): 421– 41. Macours, Karen, Norbert Schady, and Reno Vakis. 2012. “Cash Transfers, Behavioral Changes, and Cognitive Development in Early Childhood: Evidence from a Randomized Experiment.” American Economic Journal: Applied Economics 4 (2): 247– 73. Maluccio, John, and Rafael Flores. 2005. “Impact Evaluation of the Pilot Phase of the Nicaraguan Red de Proteción Social.” Food Consumption and Nutrition Division Discussion Paper no. 141, International Food and Policy Research Institute. Ministry of Health, Burkina Faso. 2010. Tableaux de Bord de la Sante. Ouagadougou, Burkina Faso. Moock, Peter R., and Joanne Leslie. 1986. “Childhood Malnutrition and Schooling in the Terai Region of Nepal.” Journal of Development Economics 20 (1): 33– 52. Morris, Saul, Rafael Flores, Pedro Olinto, and Juan Manuel Medina. 2004. “Monetary Incentives in Primary Health Care and Effects on Use and Coverage of Preventive Health Care Interventions in Rural Honduras: Cluster Randomized Trial.” Lancet 364 (9450): 2030– 37. Parker, Susan W., and Rebeca Wong. 1998. “Household Income and Health Care Expenditures in Mexico.” Health Policy 40 (3): 237– 55. Paxson, Christina, and Norbert Schady. 2010. “Does Money Matter? The Effects of Cash Transfers on Child Development in Rural Ecuador.” Economic Development and Cultural Change 59 (1): 187– 229. Schultz, T. Paul. 2005. “Productive Benefits of Health: Evidence from Low Income Countries.” Economic Growth Center Working Paper no. 903, Yale University. Strauss, John, and Duncan Thomas. 1995. “Human Resources: Empirical Modeling of Household and Family Decisions.” In Handbook of Development Economics, edited by T. N. Srinivasan and J. Behrman. Amsterdam: North Holland.

5

Girl Power Cash Transfers and Adolescent Welfare: Evidence from a ClusterRandomized Experiment in Malawi Sarah Baird, Ephraim Chirwa, Jacobus de Hoop, and Berk Özler

5.1

Introduction

Adolescent girls are a key demographic target group to successfully break the cycle of poverty in developing countries (Levine et al. 2008). In Malawi, the focus of this chapter, the population of fifteen to nineteen- year- old girls is forecast to grow by 66.9 percent from 2005 to 2020, making its projected growth rate the third highest in the world (Warhurst, Molyneux, and Jackson 2010).1 Interventions that help adolescent girls reach their full potential not only bring immediate benefits to their own lives, but also longer- term benefits to their offspring and communities at large (Lloyd 2009; Duflo 2012). This message is the focus of organizations that favor social interSarah Baird is associate professor of Global Health and Economics at George Washington University. Ephraim Chirwa is professor of economics at Chancellor College, University of Malawi. Jacobus de Hoop is an impact evaluation specialist at the International Labour Organization. Berk Özler is a senior economist in the Development Research Group, Poverty and Inequality Team at the World Bank. Raka Banerjee, Diana Boss, Josefine Durazo, Fernando Galeana, Nicola Hedge, Cara Janusz, Amanda Moderson-Kox, Lucie Tafara Moore, James Mwera, and Erin Shedd provided excellent research assistance on the SIHR project. We are grateful to participants at the National Bureau of Economic Research (NBER) Africa Project Research Conference for valuable comments. We gratefully acknowledge funding from the Global Development Network, the Bill and Melinda Gates Foundation, NBER Africa Project, World Bank Research Support Budget Grant, 3ie Open Window (Round 2) as well as several trust funds at the World Bank: Knowledge for Change Trust Fund (TF090932), World Development Report 2007 Small Grants Fund (TF055926), Spanish Impact Evaluation Fund (TF092384), and Gender Action Plan Trust Fund (TF092029). For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see http://www.nber.org /chapters/c13380.ack. 1. As of 2005, there were 849.6 million girls age ten to twenty- four in the world, comprising 13 percent of the global population. Between 2005 and 2020, the population of girls age ten to twenty- four years is forecast to grow by 5.1 percent (Warhurst, Molyneux, and Jackson 2010).

139

140

Sarah Baird, Ephraim Chirwa, Jacobus de Hoop, and Berk Özler

ventions targeted at young women, such as the Coalition for Adolescent Girls founded by the United Nations Foundation and the Nike Foundation.2 This chapter examines whether a cash transfer program targeted at adolescent girls in Malawi helped empower its recipients in the short run, that is, during and immediately after the two- year intervention. The Zomba Cash Transfer Program (ZCTP) was a randomized intervention that provided initially never- married females ages thirteen to twenty- two with financial support in the form of monthly cash transfers for two academic years. The intervention had two treatment arms, one where cash was given conditional on regular school attendance, the Conditional Cash Transfer (CCT) arm, and one that transferred cash unconditionally, the Unconditional Cash Transfer (UCT) arm. The program was not only targeted to families of eligible young females as described above, but also, unlike in almost all CCT programs, part of the monthly transfer was given directly to the girl. Giving transfers to girls directly can make the transfers “stick” to them, potentially leading to larger impacts (see Muralidharan and Prakash [2013] for strong enrollment effects of a program that gave girls bicycles in India). Furthermore, as in BRAC’s Empowerment and Livelihoods for Adolescents (ELA) program in Uganda, girls who had already dropped out of school were part of the target population under ZCTP, allowing for the identification of impacts separately on this potentially vulnerable group. The target population and the experimental variation in treatment make the ZCTP an ideal intervention within which to evaluate the impact of cash transfers on the empowerment of adolescent girls.3 However, it is important to note that the notion of empowerment adopted here for this group of young and largely never- married females at the end of the two- year intervention is different than what we would consider if the target population were mostly married adult women. As is powerfully summarized in Duflo (2012), many academics and policymakers are interested in women’s decision- making power within their households, their bargaining power within their marriages, and their voice and political power within their communities. As the study population here are initially never- married adolescent girls, the vast majority of whom still lived with their parents (or another guardian) at the end of the two- year program, outcomes pertaining to bargaining power within marriage, investments in own children, or involvement in local politics are too early to measure. Potential program impacts on these outcomes are the focus of future work using longer- term follow-up data. Here, we focus on the empowerment of adolescent girls and 2. See, for example, “The Girl Effect,” a YouTube sensation that advocates for “the powerful social change brought about when girls have the opportunity to participate.” http://www.you tube.com/user/girleffect. 3. We discuss the details of these two different treatment arms in more detail in section 5.2.

Cash Transfers and Adolescent Welfare

141

summarize program impacts for a wide and rich set of outcomes during and immediately following the completion of the cash transfer experiment. Empowerment, as a concept, is hard to pin down. Kabeer (1999, 435) defines empowerment as “the process by which those who have been denied the ability to make strategic life choices acquire such . . . ability.” Essentially, Kabeer (1999) argues that two elements, resources and agency,4 determine an individual’s ability to exercise choice. Resources can broadly be defined as access and future claims to physical and human resources that are instrumental in making important choices in life. Agency is “people’s actual capacity to define their own life- choices and to pursue their own goals” (Kabeer 1999, 435). Agency includes both internal cognitive processes such as reflection and analysis and the social processes of bargaining, negotiation, manipulation, norms, and conventions. Ultimately, improved ability to exercise choice (as a result of enhanced resources and agency) can affect dayto-day functioning of the individual and her family members (for instance, in terms of health status, nutritional intake, and time use). While the literature approaches empowerment from multiple angles, it can largely be interpreted within the framework of Kabeer (1999) as investigating the impact of enhanced resources or agency on subsequent functioning. In economics, the empirical literature largely focuses on policies or programs that increase a woman’s bargaining power within marriage, both improving own outcomes as well as those of her children. For example, policies such as the old- age pension in South Africa (Duflo 2003) and extended alimony rights in Brazil (Rangel 2006) had beneficial impacts on the health and education of the female children of beneficiaries. In addition, there is some evidence that female- targeted interventions improve outcomes for women: a switch from a tax credit to a direct payment to the mother for child benefits was associated with an increase in the consumption of women’s and children’s clothing (Lundberg, Pollak, and Wales 1997); a savings product in the Philippines improved women’s influence on household decisions (Karlan, Ashraf, and Yin 2007); a microcredit program in Bangladesh increased the female beneficiaries’ financial resources and mobility (Pitt, Khandker, and Cartwright 2006); a community- level education program for women improved employment and empowerment outcomes (Kandpal, Baylis, and Arends-Kuenning 2013); and Oportunidades, a Mexican CCT scheme, empowered women by encouraging them to negotiate better care from health care providers (Barber and Gertler 2010).5 More relevant for 4. Sen (1999) refers to this as capabilities. 5. While giving women more power is on many occasions both efficiency and equity enhancing, Duflo (2012) notes that women and men have different preferences and women’s preferences are not always more benevolent than men’s and their decisions are not always more favorable to well- being and development. For example, girls age birth to five years benefited from old- age pensions given to women but not to men, and there was no effect among boys for

142

Sarah Baird, Ephraim Chirwa, Jacobus de Hoop, and Berk Özler

the target population of adolescent girls under examination here, Bandiera et al. (2015) find that an intervention that combined vocational training with information on health and risky behaviors led to increases in incomegenerating activities and decreases in sexual activity and pregnancies among females ages fourteen to twenty in Uganda. Within the theoretical economics literature, the discourse on female empowerment focuses on shifts in the balance of power within married couples in favor of the woman. A shift in the balance of power can take different forms, such as an increase in the woman’s education, an increase in her earning capacity, or her improved access to birth control technologies. Such shifts can result in the reallocation of resources within the household toward the woman (Chiappori, Iyigun, and Weiss 2009; Chiappori and Oreffice 2008), both increasing her welfare and perhaps leading to a reduction in total fertility and even child mortality rates (Iyigun and Walsh 2007; Eswaran 2002). Interestingly, the theoretical empowerment literature suggests that the determinants of empowerment within a relationship are to a large extent shaped earlier in life. Recent research finds positive impacts on adult life outcomes from child sponsorship programs (Wydick, Glewwe, and Rutledge 2013) and a potential mechanism may be higher levels of self- esteem, educational, and occupational aspirations (Glewwe, Ross and Wydick 2013). Hence, theory and some empirical evidence suggests that influencing the resources and agency of females at a young age may lead to improved outcomes in the future—privately and socially. Our study adds to this literature by summarizing the short- term effects of a two- year cash transfer intervention targeted explicitly at never- married females on a broad set of outcomes related to empowerment. Our discussion first focuses on empowerment impacts of CCTs, before turning to a brief comparison of them with UCTs. The reason for this approach is that the CCT experiment was conducted in two strata—girls who were in school at baseline (baseline schoolgirls) and girls who had already dropped out of school at baseline (baseline dropouts). The UCT intervention, on the other hand, was only conducted among baseline schoolgirls. Previous studies evaluating the impact of the ZCTP focused mostly on baseline schoolgirls, as this group allowed an experimental comparison of impacts between the CCT and the UCT arms of the intervention (see Baird, McIntosh, and Özler 2011). However, CCTs had large and statistically significant impacts on a number of outcomes among baseline dropouts as well—a group that is often left out of programs that are school based (Bandiera et al. 2015). Thus, we first present findings on the impacts of CCTs for both baseline schoolgirls and baseline dropouts, before turning either recipient (Duflo 2003). Under the same program, schooling for children age thirteen to seventeen increased more when the eligible recipient was male (Edmonds 2006). Ashraf, Field, and Lee (2014) show that women who are empowered to take charge of birth control decisions through concealable contraceptives in Zambia report a lower subjective well- being.

Cash Transfers and Adolescent Welfare

143

to a comparison of CCT and UCT impacts among baseline schoolgirls only.6 We examine impacts while the program was ongoing (Round 2) and immediately after the program ended (Round 3). When examining the effects of CCTs on female empowerment, we focus on baseline schoolgirls and baseline dropouts separately throughout the analysis for a number of reasons. First, the schooling condition works differently on these two groups—for baseline dropouts it brings them back into school, while for baseline schoolgirls it prevents them from dropping out. Second, as described in section 5.3 below, these groups look vastly different across a host of baseline characteristics and thus are best viewed as separate populations. Finally, baseline dropouts are a group that is often ignored in the analysis of CCT programs, even though the size of this population is nonnegligible. Thus, we feel that providing results separately for this group may provide the reader with some important insights. The remainder of this chapter proceeds as follows. Section 5.2 describes the cash transfer intervention and the experimental design of this study. Section 5.3 discusses the estimation strategy. Section 5.4 presents the main results for the CCT arm and focuses on program impacts on the dimensions of resources, agency, and functioning, with section 5.5 discussing the results for the UCT arm and comparing the two interventions. Section 5.6 concludes. 5.2 5.2.1

Research Setting and Design7 Location

Malawi, the setting for this research project, is a small and poor country in southern Africa. Eighty- one percent of its population of 15.3 million lived in rural areas in 2009, with most people relying on subsistence farming. The country is poor even by African standards: Malawi’s 2008 gross national income (GNI) per capita figure of $760 (purchasing power parity [PPP], current international $) is less than 40 percent of the sub-Saharan African average of $1,973 (World Development Indicators Database 2010). 5.2.2

Sample

Zomba district in the southern region was chosen as the site for this study. Zomba district is divided into 550 enumeration areas (EAs), which are defined by the National Statistical Office of Malawi and contain an average of 250 households spanning several villages. Fifty of these EAs lie in Zomba 6. The decision to leave the discussion of the impact of UCTs until after the discussion of CCTs is not a reflection on the relative importance of these two interventions. It is simply for ease of exposition—allowing for an explicit focus on baseline dropouts. 7. This section draws heavily from section 2 in Baird, McIntosh, and Özler (2011), which provides more detail on the study design and the intervention.

144

Sarah Baird, Ephraim Chirwa, Jacobus de Hoop, and Berk Özler

city, while the rest are in seven traditional authorities. Prior to the start of the experiment, 176 EAs were selected from three different strata: Zomba city (urban, 29 EAs), near rural (within a 16 km radius of Zomba city, 119 EAs), and far rural (28 EAs). In these 176 EAs, each dwelling was visited to obtain a full listing of never- married females, age thirteen to twenty- two.8 The target population was then divided into two main groups: baseline dropouts and baseline schoolgirls. In each selected EA, 100 percent of all eligible baseline dropouts and 14– 100 percent of all eligible baseline schoolgirls were randomly sampled to participate in the study, where the percentage depended on the core respondent’s age at baseline and the strata she lived in (urban, near rural, and far rural). This sampling procedure resulted in a total study sample of 3,796 women with an average of 5.1 baseline dropouts and 16.5 baseline schoolgirls per EA. 5.2.3

Research Design

Treatment status was assigned at the EA level and the sample of 176 EAs was randomly divided into two equally sized groups: eighty- eight treatment EAs and eighty- eight control EAs. In the eighty- eight treatment EAs, all baseline dropouts were offered conditional cash transfers. The eightyeight treatment EAs were then randomly assigned to one of three groups to determine the treatment status of baseline schoolgirls: in forty- six EAs baseline schoolgirls received transfer offers conditional on regular school attendance (CCT arm), while in twenty- seven EAs they received offers for unconditional cash transfers (UCT arm). In the remaining fifteen EAs no baseline schoolgirls received any transfer offers.9 5.2.4

CCT Intervention

After the random selection of EAs and individuals into the treatment group, the local non- governmental organization (NGO) retained to implement the cash transfers held meetings in each treatment EA between December 2007 and early January 2008 to invite the selected individuals to participate in the program. At these meetings, the program beneficiary and her parents/guardians were made an offer that specified the monthly transfer amounts being offered to the beneficiary and to her parents, the condition to regularly attend school, and the duration of the program. It was possible for more than one eligible girl from a household to be invited to participate in the program. The offer to participate in the program consisted of a transfer to the 8. The target population of thirteen to twenty- two- year- old, never- married females was selected for a variety of reasons. For details, we refer the reader to Baird, McIntosh, and Özler (2011). 9. Girls who live in treatment EAs but do not receive transfers allow for the measurement of spillover effects within treatment EAs.

Cash Transfers and Adolescent Welfare

145

parents, a transfer directly to the girl, and payment of school fees for girls attending secondary school. Transfer amounts to the parents were varied randomly across EAs between $4, $6, $8, and $10 per month, so that each parent within an EA received the same offer. Within each EA, a lottery was held to determine the transfer amount to the young female program beneficiaries, which was equal to $1, $2, $3, $4, or $5 per month. The lottery was held publicly to ensure that the process was transparent.10 Secondary school fees were paid in full directly to the schools.11 This chapter focuses on the average effect of these transfers and does not delve into elasticities of the outcomes under investigation with respect to the transfer amounts. Monthly school attendance of all the conditional cash transfer recipients was checked and payment for the following month was withheld for any student whose attendance was below 80 percent of the number of days that the school was in session for the previous month. However, participants were never removed from the program for failing to meet the monthly 80 percent attendance rate, meaning that if they subsequently had satisfactory attendance their payments would resume. Offers to everyone, identical to the previous ones they received, and regardless of their schooling status during the first year of the program in 2008, were renewed between December 2008 and January 2009 for the second and final year of the intervention, which ended at the end of 2009. 5.2.5

UCT Intervention

In the UCT EAs, the offers were identical with one crucial difference: there was no requirement to attend school to receive the monthly cash transfers. Other design aspects of the intervention were kept identical so as to be able to isolate the effect of imposing a schooling conditionality on primary outcomes of interest.12 Attendance was never checked for recipients in the UCT arm and they received their payments by simply presenting at the transfer locations each month. 10. The lottery was held among those selected to participate in the program. Hence, each girl present at the lottery was offered at least $1 per month individually and $4 per month for her parents. Girls not selected for the program were never contacted by the implementing NGO and were not present at the lottery. 11. Primary schools are free in Malawi, but student have to pay nonnegligible school fees at the secondary level. The program paid these school fees for students in the conditional treatment arm upon confirmation of enrollment for each term. Private secondary school fees were also paid up to a maximum equal to the average school fee for public secondary schools in the study sample. 12. For households with girls eligible to attend secondary schools at baseline, the total transfer amount was adjusted upward by an amount equal to the average annual secondary school fees paid in the conditional treatment arm. This additional amount ensured that the average transfer amounts offered in the CCT and UCT arms were identical and the only difference between the two groups was the “conditionality” of the transfers on satisfactory school attendance.

146

Sarah Baird, Ephraim Chirwa, Jacobus de Hoop, and Berk Özler

5.2.6

Data

The data used in this chapter were collected in three household survey rounds. Baseline data, or Round 1, was collected between October 2007 and January 2008, before the offers to participate in the program took place. First follow-up data collection, or Round 2, was conducted approximately twelve months later—between October 2008 and February 2009. The second follow-up (Round 3) data collection was conducted between February and June 2010—after the completion of the two- year intervention at the end of 2009. The intervention period coincided with the 2008 and 2009 school years.13 The annual household survey consisted of a multitopic questionnaire administered to the households in which the sampled respondents resided. It consisted of two parts: one that was administered to the head of the household and the other administered to the core respondent, that is, the sampled girl from our target population. The former collected information on the household roster, dwelling characteristics, household assets and durables, shocks, and consumption. The survey administered to the core respondent provides detailed information about her family background, schooling status, health, dating patterns, sexual behavior, fertility, and marriage. In addition to the household survey, biological data on HIV and Herpes Simplex Virus- 2 (HSV- 2) were collected eighteen months after baseline (approximately six months after the Round 2 household survey).14 Finally, the entire sample was given three achievement tests (Mathematics, English Reading Comprehension, and Raven’s Colored Matrices), conducted at home, in Round 3. 5.3 5.3.1

Estimation Strategy Attrition and Balance

Before turning to our overall estimation strategy, this subsection first examines two potential sources of bias: (a) differential attrition and (b) imbalance in baseline characteristics between treatment and control. Table 5.1 investigates attrition by regressing a binary indicator that takes on the value of one if a respondent was surveyed in all three rounds on the treatment indicator. Column (1) shows that among baseline dropouts the attrition rate in the control group was 15.7 percent with no significant difference between treatment and control. Column (2) shows that among baseline schoolgirls, the attrition rate among the control group was even lower at 10.7 percent, 13. At the time of the intervention, the Malawian school year corresponded with the calendar year. 14. See Baird et al. (2012) for more details.

Cash Transfers and Adolescent Welfare Table 5.1

147

Analysis of attrition Dependent variable: = 1 if surveyed in all three rounds

Conditional treatment

Dropouts (1)

Schoolgirls (2)

0.005 (0.028)

0.021 (0.030) 0.030 (0.024)

0.843 889

0.893 2,284 0.797

Unconditional treatment Mean in the control group Number of observations Prob. > F(conditional = unconditional)

Notes: Regressions are OLS models with robust standard errors clustered at the EA level. All regressions are weighted to make them representative of the target population in the study EAs. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

with again no significant differences between either treatment arm and the control group, nor between the two treatment arms ( p = 0.797). These findings suggest that the results we present in this chapter are unlikely to be biased due to differential attrition between the study arms. Table 5.2 investigates the balance of the experiment by regressing baseline covariates (including parental, personal, and household characteristics) that the literature suggests are correlated with outcomes of interest on treatment indicators. Column (1) presents the mean value of each of the baseline characteristics in the control group among baseline dropouts, while column (2) presents the coefficient on the difference between treatment and control for baseline dropouts. We observe no violations of balance among this group. Column (3) presents the mean in the control for baseline schoolgirls, with column (3) indicating the CCT difference with the control group, column (4) the UCT difference with the control group, and column (5) the p- value for the difference between the two treatment arms. First of all we observe that, on average, baseline schoolgirls come from better socioeconomic backgrounds than baseline dropouts. Baseline schoolgirls have completed more schooling, come from households with better access to durable goods, are more likely to have parents who are still alive, and less likely to be sexually active. There is one violation in baseline balance for the CCT baseline schoolgirls, and one for the UCT arm, with these two variables also being different between the two treatment arms: girls in the CCT arm are approximately half a year younger than those in the UCT arm and, hence, have a similarly low grade attainment at baseline. As described in the subsection

Balance of baseline characteristics

–0.037 0.004 –0.403 –0.015 –0.195 0.081 –0.050 –0.035 –0.029 0.063

17.579 0.302 6.223 –0.772 6.128 0.181 0.699 0.120

Treatment difference (2)

0.791 0.642

Control mean (1)

Dropouts

Notes: Observations are weighted to make results representative of all study EAs. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

Parental characteristics Respondent’s mother is alive Respondent’s father is alive Personal characteristics Respondent’s age Respondent never had sex Highest grade attended by respondent Household characteristics Asset index (first principal component of 15 durable goods) Household size Geographic strata (used for block randomization) Respondent lives in Zomba city Respondent lives in rural area within 16 km from Zomba city Respondent lives in rural area more than 16 km from Zomba city Observations

Table 5.2

0.347 0.563 0.090

0.581 6.432

15.252 0.797 7.479

0.842 0.705

Control mean (3)

0.112 –0.122 0.010

0.402 –0.049

–0.299* –0.001 –0.233

–0.040 0.008

Conditional treatment difference (4)

0.070 –0.031 –0.039

0.639 0.230

0.173 –0.023 0.417**

–0.006 0.054

Unconditional treatment difference (5)

Schoolgirls

0.807 0.577 0.458

0.623 0.202

0.007 0.582 0.004

0.360 0.288

P-value (conditionalunconditional) (6)

Cash Transfers and Adolescent Welfare

149

below, we control for the baseline values of these variables in our analysis of program impacts. 5.3.2

Specification

We analyze the intention- to-treat (ITT) effects of the intervention separately on Round 2 and Round 3 indicators using cross- sectional regressions. This approach allows us to investigate whether there were any empowering impacts of the program while the young women were still participating in the cash transfer program, as well as whether such impacts were still present immediately after the program ended. The regression- adjusted ITT impact of the program is estimated with ordinary least squares (OLS) using the following linear regression model: Yi = TiC gC + TiU gU + X i b + ´i,

(1)

where Yi is the empowerment outcome for individual i in Round 2 or Round 3; Xi is a vector that contains a set of baseline controls; TiC (TiU ) is a binary indicator that is equal to one if a girl was offered a CCT (UCT) and zero otherwise. For baseline dropouts, equation (1) excludes the UCT indicator. The standard errors εi are clustered at the EA level to account for the design effect of the EA- level treatment assignment. Age- and stratum- specific sampling weights are used to make the results representative of the target population in the study area. To make the results comparable across survey rounds, the analysis includes respondents if and only if they were interviewed in all three rounds. In choosing the covariates, Xi , included in this analysis, we follow the approach advocated by Bruhn and McKenzie (2009) and control for two types of variables: strata that were used for block randomization in the trial and baseline characteristics that are predictive of the outcome. In this study, these covariates include dummy variables for age and geographic strata along with an index of household assets, highest grade attained, and sexual activity status—all measured at baseline. 5.4 5.4.1

Impacts of CCTs Resources

We first investigate to what extent the ZCTP influenced the physical resources available to the core respondent. As explained above, the monthly transfers consisted of two components, one component paid to the parents (or the guardian) of the core respondent and one component paid directly to the core respondent herself. Table 5.3 examines whether the cash transfers translated into higher monthly expenditures by respondents on themselves. Columns (1) and (2) present the impacts in Round 2 (during the intervention) while columns (3) and (4) present impacts in Round 3 (after the intervention ended).

150 Table 5.3

Sarah Baird, Ephraim Chirwa, Jacobus de Hoop, and Berk Özler Program impacts on expenditure by the respondents on themselves (in USD) over the past thirty days Round 2

Conditional treatment

Dropouts (1)

Schoolgirls (2)

Dropouts (3)

Schoolgirls (4)

1.530*** (0.573)

1.799*** (0.497) 1.434*** (0.528) 2.263 2,087 0.619

0.334 (0.544)

0.788** (0.324) –0.229 (0.885) 2.488 2,086 0.276

Unconditional treatment Mean in the control group Number of observations Prob. > F(conditional = unconditional)

Round 3

3.593 750

3.740 749

Notes: Regressions are OLS models with robust standard errors clustered at the EA level. All regressions are weighted to make them representative of the target population in the study EAs. Baseline values of the following variables are included as controls in the regression analyses: age dummies, strata dummies, household asset index, highest grade attended, and an indicator for never had sex. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

Baseline dropouts spent US$1.5 per month more than the control group while the program was ongoing, an increase of approximately 42 percent (column [1]). This increase amounts to approximately half of the mean amount of US$3 per month transferred directly to the respondents. The increase in expenditures was similar for baseline schoolgirls at US$1.8 per month, an increase of approximately 80 percent over the control group (column [2]). The impact of the program remained after the program ended for baseline schoolgirls, although the magnitude had declined by over 50 percent with no significant effect among baseline dropouts. The results suggest that the cash transfer program led to an increase in the beneficiaries’ control of cash resources during the program, and that these impacts declined or disappeared shortly after the end of the program. Table 5.4 shows that the direct transfers to the beneficiaries most likely drove this increase in personal consumption. Panel A shows that respondents had little influence on the way the component paid to the household was spent. Roughly 90 percent of the respondents indicated that the decision on how to spend these funds was made by someone else. Panel B, on the other hand, shows that approximately 80 percent of the young female beneficiaries had control over how to spend the transfers made directly to them.15 15. This finding accords with findings from focus group interviews at the design stage of ZCTP, during which females eligible for the program indicated that they would be able to “keep their transfers” if these were physically given to them.

Cash Transfers and Adolescent Welfare Table 5.4

151

Who decides how the transfer is spent (in percentages)? Dropouts (1)

Conditional schoolgirl (2)

Unconditional schoolgirl (3)

A. Transfer to household Father Mother Self Other

10.09 55.05 13.76 21.10

11.02 59.70 6.77 22.52

7.64 64.23 4.96 23.18

B. Transfer to respondent Father Mother Self Other

3.37 11.66 80.98 3.99

2.77 13.70 77.51 6.02

0.81 7.08 86.01 6.10

Number of observations

326

448

253

Notes: These results are from Round 2 only. Observations are weighted to make results representative of all study EAs. These results are for treatment girls only.

5.4.2

Agency

Schooling Next, we turn to the impact of the intervention on social patterns that can be detrimental to the development of adolescent girls. We first investigate how the intervention affected schooling outcomes. Baird, McIntosh, and Özler (2011) show that the CCT program had a strong effect on school enrollment among baseline schoolgirls. Among this group the intervention significantly increased the average number of terms enrolled by 0.54 according to teacher reports (over a base of 4.79 terms in the control out of a maximum of six during the two academic years the program ran). Baseline dropouts experienced an increase of 2.35 terms (compared to 1.02 in the control group, using self- reported enrollment data—see table 5A.1, column [1], appendix)—an impact that more than tripled the number of terms enrolled.16 An important question is whether the program empowered respondents by providing them with additional knowledge and skills to make important life decisions. Baird, McIntosh, and Özler (2011) provide a first indication that the intervention indeed resulted in increased skills. They show that among baseline schoolgirls, there were significant improvements in math, 16. We did not collect teacher reports of enrollment or attendance for this stratum in Round 3. Hence, these findings should be treated with some caution due to the fact that the underlying data for school enrollment are self- reported. Please see Baird and Özler (2012) for more on the reliability of self- reported data on school participation. However, significant improvements in learning presented in table 5A.1 support the finding of a significant increase in the reenrollment rate for this group.

152

Sarah Baird, Ephraim Chirwa, Jacobus de Hoop, and Berk Özler

Table 5.5.

Program impacts on participation in health training over the past twelve months Any Round 2

Conditional treatment

Dropouts (1)

Schoolgirls (2)

Dropouts (3)

Schoolgirls (4)

0.152*** (0.041)

–0.041 (0.029) 0.012 (0.035) 0.879 2,087 0.215

0.025 (0.034)

0.071** (0.033) 0.068** (0.035) 0.775 2,086 0.939

Unconditional treatment Mean in the control group Number of observations Prob. > F(conditional = unconditional)

Round 3

0.547 750

0.696 749

Notes: Regressions are OLS models with robust standard errors clustered at the EA level. All regressions are weighted to make them representative of the target population in the study EAs. Baseline values of the following variables are included as controls in the regression analyses: age dummies, strata dummies, household asset index, highest grade attended, and an indicator for never had sex. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

English and cognitive test scores during Round 3.17 Among baseline dropouts there were also modest, but statistically significant, improvements across the three tests with impacts ranging from 0.13 standard deviations in English Reading Comprehension ( p < 0.10) to 0.16 standard deviations in mathematics ( p < 0.05) (table 5A.1, columns [2]–[4], appendix). Table 5.5 adds to this evidence by investigating whether the program increased participation in any form of health training (including nutrition, personal hygiene, food hygiene, sexual education, and HIV/AIDS) and shows that both groups of CCT beneficiaries saw increases in the probability of participating in health training during the past twelve months. Fertility and Marriage Having shown that the ZCTP improved the physical resources available to respondents and increased their school participation and learning, we now turn to other important outcomes, such as childbearing and marriage, which may be influenced by the intervention either through an income effect or an effect of the condition to regularly attend school. We first investigate the impact of the intervention on respondents’ fertility decisions, one of the 17. For more details on the specifics of these achievement tests, see Baird, McIntosh, and Özler (2011). These tests were only conducted in Round 3.

Cash Transfers and Adolescent Welfare

153

prime outcomes in the theoretical empowerment literature (see, for instance, Eswaran [2002] or Iyigun and Walsh [2007]). Baird et al. (2010) shows that the conditional cash transfers significantly reduced pregnancy among treated baseline dropouts during the first year of the program, with no significant impact among baseline schoolgirls.18 Baseline dropouts were 5.1 percentage points less likely to ever have been pregnant (61 percent in the control group) in Round 2 and 8.2 percentage points (78 percent in the control group) in Round 3 (table 5A.1, column [5], appendix). In table 5.6, we investigate whether the program also impacted preferred fertility timing or desired lifetime fertility. Looking at the impact of the intervention on the number of months the respondent would like to wait before having a child, we observe significant increases for baseline dropouts in both rounds and for baseline schoolgirls in Round 3 (columns [1]–[4]).19 The CCT effects on the ideal number of children are generally negative but only significant among baseline schoolgirls at the end of the intervention, who, on average, want 0.184 less children over their lifetimes. The evidence suggests a desire to delay childbearing as a result of the intervention rather than lowering the ideal number of children. Turning now to marriage, the CCT program led to a significant decline in marriage rates among baseline dropouts. At baseline all of the respondents were never married, while 28 percent of the baseline dropouts in the control group had gotten married by Round 2. In the treated group the marriage rate was 11 percentage points lower (Baird et al. 2010). This difference persisted after the program ended—baseline dropouts were 12.6 percentage points less likely to be married by Round 3—from a mean of 55.1 percent in the control group ( table 5A.1, column [6], appendix). Among baseline schoolgirls, on the other hand, there was no evidence of a significant impact of the CCT intervention on marriage during or after the program (Baird, McIntosh, and Özler 2011). It may be puzzling to the reader as to why a large effect on marriage is found among baseline dropouts but not baseline schoolgirls. As discussed in detail in Baird, McIntosh, and Özler (2011), the cash transfer effect on marriage works through two channels in Malawi, where marriage and schooling are exclusive: an income effect and a schooling effect. In the case of CCTs, the income effect is zero for those who drop out of school during the 18. For an explanation of why a significant effect on childbearing is observed among baseline dropouts but not baseline schoolgirls, please refer to Baird, McIntosh, and Özler (2011, 1735– 40). 19. This variable is missing for respondents who want zero children or do not want any more children. We find no impact of the intervention on whether or not this variable is missing for baseline dropouts. However, in Round 3, there is a significant and positive coefficient on this relationship for baseline schoolgirls. This result suggests that the CCT intervention may have also decreased the number of respondents wanting any additional children among baseline schoolgirls. The significant decline in the ideal number of children in this same group (table 5.6, column [8]) supports this interpretation.

729

52.056 2,047 0.616

4.934 (3.110) 2.790 (3.765) 79.092

Schoolgirls (2)

699

48.657

3.658* (2.093)

Dropouts (3)

1,994 0.100

9.056*** (3.158) 2.138 (3.130) 70.207

Schoolgirls (4)

Round 3

749

2.919

–0.064 (0.101)

Dropouts (5)

2,087 0.186

–0.112 (0.104) 0.084 (0.122) 2.813

Schoolgirls (6)

Round 2

747

2.942

0.006 (0.086)

Dropouts (7)

2,080 0.119

–0.184* (0.099) 0.025 (0.103) 2.909

Schoolgirls (8)

Round 3

Ideal number of children

Notes: Regressions are OLS models with robust standard errors clustered at the EA level. All regressions are weighted to make them representative of the target population in the study EAs. Baseline values of the following variables are included as controls in the regression analyses: age dummies, strata dummies, household asset index, highest grade attended, and an indicator for never had sex. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

Number of observations Prob. > F(conditional = unconditional)

Mean in the control group

Unconditional treatment

7.383** (2.992)

Dropouts (1)

Round 2

Months before next child

Program impact on preferred fertility timing or desired lifetime fertility

Conditional treatment

Table 5.6

Cash Transfers and Adolescent Welfare

155

program because they cease to receive payments. Hence, the primary channel through which CCTs can delay marriage is through increased school enrollment. In this experiment, the marginal effect of CCTs on school enrollment was, while significant, relatively small among baseline schoolgirls, meaning that the knock-on effect on marriage was undetectable. Baseline dropouts, on the other hand, experienced a massive surge in their odds of reenrollment, which translated into delayed marriage and pregnancy. The CCT intervention also had a significant impact on baseline dropouts viewing education as an important characteristic of a future spouse (results not shown). We come back to this issue in section 5.5, when we contrast the effects of UCTs on these outcomes, where the channels of impacts are different yet again. Overall, the results presented in this section indicate that CCTs had a strong impact on agency, as they caused beneficiaries to increase their school participation and learning, and, among baseline dropouts, beneficiaries postponed marriage and pregnancy. These changes may well affect the well- being of these respondents later in life as well as their bargaining power in future relationships. Moreover, as the next section will show, the intervention also had a substantial impact on the contemporary functioning of respondents. 5.4.3

Functioning

Position in the Household Finally, we examine how the intervention affected two areas of functioning within the household: (a) self- perceived position within the household, and (b) nutrition and health. Table 5.7 first examines how the intervention affected respondents’ answer to the question “Would you say your household cares more about your education now compared to twelve months ago?” Baseline dropouts and baseline schoolgirls in the CCT arm are significantly more likely to agree with the statement than those in the control group during the program (columns [1] and [2]), with no lasting effects once the program is over. The pattern is the same when we analyze responses to the question “Would you say your household cares more about your health now compared to twelve months ago?” These findings suggest that the intervention may have improved the standing of the school- age beneficiaries within their households by making them more of an asset to their families during the program period. Nutrition and Health Above, we showed that beneficiaries perceived their households to care more about their health while the program was ongoing. We now look at whether this perception is accompanied by tangible improvements in

733

0.161 2,077 0.034

0.075** (0.034) 0.165*** (0.036) 0.524

Schoolgirls (2)

728

0.141

0.017 (0.027)

Dropouts (3)

2,076 0.961

0.032 (0.033) 0.034 (0.036) 0.416

Schoolgirls (4)

Round 3

732

0.242

0.067** (0.034)

Dropouts (5)

2,075 0.686

0.158*** (0.045) 0.183*** (0.043) 0.255

Schoolgirls (6)

Round 2

728

0.201

0.016 (0.030)

Dropouts (7)

2,072 0.262

0.026 (0.029) 0.081* (0.043) 0.288

Schoolgirls (8)

Round 3

Notes: Regressions are OLS models with robust standard errors clustered at the EA level. All regressions are weighted to make them representative of the target population in the study EAs. Baseline values of the following variables are included as controls in the regression analyses: age dummies, strata dummies, household asset index, highest grade attended, and an indicator for never had sex. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

Number of observations Prob. > F(conditional = unconditional)

Mean in the control group

Unconditional treatment

0.334*** (0.038)

Dropouts (1)

Round 2

= 1 if household cares more about respondent’s health than twelve months ago

Dependent variable = 1 if household cares more about respondent’s education than twelve months ago

Program impact on self-perceived standing in the household

Conditional treatment

Table 5.7

Cash Transfers and Adolescent Welfare

157

investments in nutrition and health, as well as measurable health outcomes. Kabeer (1999) argues that nutrition and health outcomes are prime examples of “universally valued functionings,” and as such they are highly relevant empowerment outcomes. In table 5.8, we first look at the impact of the intervention on the intake of three sources of protein: meat, eggs, and fish (columns [1]–[4]). The outcome variable counts the number of days respondents ate any of these three items over the seven days prior to the interview.20 We find that CCTs caused a modest increase (approximately 10 percent) in the consumption of protein rich food items among both baseline dropouts and baseline schoolgirls during the program (columns [1] and [2]) and that this effect persisted among baseline schoolgirls after the program ended, with no such lasting effect among baseline dropouts (columns [3] and [4]). We then investigate whether the intervention affected the probability that respondents usually sleep under a bed net (columns [5]–[8]). We observe a significant improvement among baseline schoolgirls: they are 7.8 (8.5 percent) percentage points more likely to sleep under a bed net in Round 2 (Round 3)—representing a large increase compared to the control group mean of 49.3 (65.9 percent). Given the high prevalence of malaria parasitemia in this area, which is a frequent cause of school absenteeism, it makes sense for families to devote resources to preventive health measures in an effort to minimize the probability of missing transfer payments due to noncompliance. These findings are also consistent with the effects of CCTs on school attendance being highest during term 1, when the number of malaria cases reaches its peak in Malawi (Baird, McIntosh, and Özler 2011). We observe no similar improvements among baseline dropouts during or after the program. This is somewhat surprising, but it is worth noting that baseline dropouts come from significantly poorer households than baseline schoolgirls. There is also evidence that the intervention improved the mental health of its beneficiaries—at least during the two- year period while it was ongoing. Using the General Health Questionnaire 12, a screening instrument widely used in clinical settings to detect individuals who are likely to suffer from psychological distress, Baird, de Hoop, and Özler (2013) show that baseline schoolgirls in the CCT arm were approximately 6 percentage points (or 17 percent) less likely to be suffering from psychological distress than those in the control group during the program. These effects had become smaller and statistically insignificant soon after the program ended. No similar effects were detected among baseline dropouts. The authors suggest that the significant changes in the daily life of baseline dropouts due to reenrollment in school compared with the relative lack of such changes among baseline 20. This variable thus takes values from 0 to 21:0 if the respondent ate none of these food items and 21 if the respondent ate all of the ingredients every day during the seven days prior to the interview.

750

Number of observations Prob. > F(conditional = unconditional)

2,087 0.827

0.489** (0.202) 0.434** (0.184) 3.954

Schoolgirls (2)

749

4.076

0.120 (0.188)

Dropouts (3)

2,086 0.255

0.530*** (0.179) 0.282* (0.165) 4.084

Schoolgirls (4)

Round 3

746

0.584

–0.022 (0.040)

Dropouts (5)

2,081 0.204

0.078* (0.040) –0.015 (0.066) 0.493

Schoolgirls (6)

Round 2

748

0.724

0.028 (0.032)

Dropouts (7)

2,084 0.228

0.085* (0.044) 0.009 (0.055) 0.659

Schoolgirls (8)

Round 3

= 1 if respondent usually sleeps under a bed net

Notes: Regressions are OLS models with robust standard errors clustered at the EA level. All regressions are weighted to make them representative of the target population in the study EAs. Baseline values of the following variables are included as controls in the regression analyses: age dummies, strata dummies, household asset index, highest grade attended, and an indicator for never had sex. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

3.662

Mean in the control group

Unconditional treatment

0.340* (0.189)

Dropouts (1)

Round 2

Number of times respondent ate protein-rich food in past seven days (out of twenty-one)

Program impact on nutrition and health

Conditional treatment

Table 5.8

Cash Transfers and Adolescent Welfare

159

schoolgirls may partly explain the differential impact of the intervention on the psychological well- being of adolescents in these two strata. Together, the presented results suggest that the impact of the intervention differed substantially between baseline schoolgirls and baseline dropouts. We observe stronger effects on health and nutrition among baseline schoolgirls and larger changes in marriage, pregnancy, and school enrollment among baseline dropouts. A likely explanation for this finding lies in the baseline differences between these two strata and the actions needed to be taken by the beneficiaries and their households to ensure compliance with the condition to attend school regularly. As we saw earlier, pregnancy and marriage are much bigger hazards to school enrollment among baseline dropouts, so their households would have encouraged beneficiaries not to get married or pregnant, to spend more time attending school and less time on household chores and labor. In the households of baseline schoolgirls, the potential payoff from following this strategy was limited because the children were already spending a relatively large share of their time on schooling—that is, most of the transfers to baseline schoolgirls were inframarginal. These households are more likely to decrease the probability of noncompliance by investing in the health of their eligible children to minimize school absences due to illness. 5.5

Impacts of UCTs

Previous analysis contrasted the effects observed in the CCT and the UCT arms on schooling outcomes, marriage, pregnancy, and mental health among baseline schoolgirls (Baird, McIntosh, and Özler 2011; Baird, de Hoop, and Özler 2013). These papers report some interesting findings, which suggest that attaching conditions to cash transfers have costs and benefits. For example, while the UCT program significantly reduced the dropout rate among its beneficiaries, this impact was only 43 percent as large as the impact of the CCT arm at the end of the two- year program. Moreover, as confirmed by differential impacts on attendance and test scores favoring the CCT arm, CCTs were found to be more cost effective than UCTs in raising enrollment rates. However, as discussed above, while the CCT program had no impact on marriage and pregnancy among baseline schoolgirls, the UCT treatment led to a significant reduction on both marriage and pregnancy by the end of the program. Furthermore, while both programs caused significant declines in psychological distress while the program was ongoing, the effects in the UCT arm were significantly larger than those in the CCT arm. Baird, McIntosh, and Özler (2011) provide an explanation for the differential effects of CCTs and UCTs on marriage and pregnancy. While CCTs only had an indirect effect on these outcomes through their effect on increased schooling participation, UCTs had a direct effect on marriage and pregnancy through an income effect. This difference between the two study

160

Sarah Baird, Ephraim Chirwa, Jacobus de Hoop, and Berk Özler

arms was due almost entirely to the effect of UCTs among girls who dropped out of school during the two- year program—as this large group of girls still received regular transfer payments. The authors argue that schooling conditions, while effective in improving school participation and learning, may undermine the social protection aspect of cash transfer programs by denying support to noncompliant households. Baird, de Hoop, and Özler (2013) exploit the random variation in amounts that were transferred separately to the girls and their parents (guardians) in each study arm and find that CCT and UCT effects on mental health were similar at the lowest amounts given to the households. However, increasing transfers to parents conditional on school attendance by the beneficiary led to significant increases in psychological distress of the adolescent girls—no such gradient was found in the UCT arm or for the transfers to the girls. The authors speculate that the burden of becoming the main source of income for their families may have become too much for these school- aged children to shoulder when a large monthly transfer was conditional on their actions. The new results presented in this chapter provide some additional insights to these earlier findings. With respect to resources, like the CCT arm, the UCT arm significantly increased the personal consumption of the respondent while the program was ongoing, but this impact appears to have dissipated faster than that of the CCT arm (table 5.3). The UCT recipients also report controlling the transfer that was directly transferred to them (86 percent), with very few controlling the household- level transfer (5 percent). Thus, these results suggest that the cash transfer program led to an increase in the control of resources for the UCT beneficiaries during the program, but that the effect did not last once the program ended. In terms of agency, unlike the CCT arm, we find no impacts of UCTs on either the number of months before having their next child or the ideal number of children (table 5.6), and these effects differ from the CCT arm ( p = 0.100 and p = 0.119, respectively). Given the significantly larger delays in fertility among this group, it will be interesting to see whether the pregnancy and marriage rates in the UCT arm quickly catch up with the CCT arm in the near future. Finally, turning to functionings, respondents in the UCT arm report similar or larger effects than the CCT arm on their self- perceived standing within their households (table 5.7), similar increases in the consumption of food items rich in protein, but a lower likelihood of sleeping under bed nets (table 5.8). Income effects likely explain the improvements in their position within the household and increases in their personal and food consumption, while the lack of incentives to attend school is consistent with the lack of effect in preventive health investments, such as bed nets. These findings make clear that the impacts of CCT and UCT programs are likely to differ—at least in the short run—and that one is not clearly preferable to the other. The choice between these two approaches may depend on the aims of the intervention, the target population, and relative weights

Cash Transfers and Adolescent Welfare

161

the policymaker or the social planner assigns to various outcomes. This is, of course, only true under the assumption that there is a legitimate reason to attach conditions to cash transfer programs, such as market failures, externalities, or political economy reasons. Furthermore, it is not clear how these short- term impacts might translate into longer- term outcomes with respect to women’s empowerment. It will be interesting to observe whether there are longer- term impacts on a broad range of outcomes, such as subjective welfare, bargaining power within marriages, fertility choices, early childhood development of own children, labor force participation, voice and political participation, and so forth, and whether and how these differ between the experimental study arms. 5.6

Concluding Discussion

Adolescent girls in developing countries are considered to be an important target group for policymakers. Targeted interventions for this group may not only affect their welfare directly, but they also have the potential to bring benefits to future generations. This chapter investigates whether one such intervention, the Zomba Cash Transfer Program in Malawi, helped empower adolescent girls in the short run. Summarizing evidence from multiple papers examining the impacts of this program on a broad range of outcomes and providing some new analysis here, this chapter suggests that the answer is a clear “yes.” The program effectively increased access to financial resources, increased schooling outcomes, decreased teen pregnancies and early marriages, improved health, and generally enabled beneficiaries to improve their agency within their households. The intervention provided its beneficiaries with financial support conditional on attending school in the CCT arm, and unconditionally in the UCT arm. Furthermore, the CCT arm was targeted to two distinct groups of school- age girls: those that were in school at baseline and those that had already dropped out before the intervention began. The latter group, albeit small at approximately 15 percent of the eligible population, saw substantial increases in schooling outcomes, as well as large declines in early marriage and pregnancy rates. As many interventions targeting school- age populations are school based, these findings point to the importance of conducting population- based interventions to avoid exclusion of this particularly vulnerable group of young people. The CCT program changed some common socioeconomic patterns that affect young women in Malawi, as it induced beneficiaries to delay childbearing and marriage. There is some evidence that these changed socioeconomic patterns are accompanied with changed marital and fertility preferences, suggesting that empowering adolescent women may not only increase their bargaining power within future relationships, but it may also affect the type of relationship they enter into in the first place. The experiment also

162

Sarah Baird, Ephraim Chirwa, Jacobus de Hoop, and Berk Özler

revealed contrasting findings between the CCT and the UCT arms. While eligible girls in both arms experienced benefits, the domains in which they experienced these benefits and the strength of the effects differed between the two groups. The findings teach us that while there may be good reasons to implement CCT programs, there are serious trade- offs associated with attaching conditions to cash transfer programs. The design choice will depend on the target population and the goals at hand. Overall, the results presented here indicate that cash transfers targeted at adolescent girls and young women can empower them in significant ways in the short run—at least in this or similar settings. It can alter social patterns that cause suboptimal investments in the human capital of young women and it can improve both their standing within the household and their day- to-day functioning. While these short- run impacts are promising, the ultimate impact of the ZCTP will depend on whether this relatively short (two- year) cash transfer program—introduced at a particularly important period of transition from adolescence to adulthood—can have long- lasting effects on the lives of this cohort of young females and their future families.

Appendix Table 5A.1

Additional program impacts on baseline dropouts in Round 3

Conditional treatment Mean in the control group Number of observations

Number terms enrolled (out of 6) (1)

English test score (standardized) (2)

2.348*** (0.163) 1.021 749

0.131* (0.070) 0.000 729

Math test score (standardized) (3) 0.164** (0.066) 0.000 729

Cognitive test score (standardized) (4) 0.142** (0.071) 0.000 729

Ever pregnant (5)

Ever married (6)

–0.082*** (0.027) 0.780 749

–0.126*** (0.036) 0.551 749

Notes: Regressions are OLS models with robust standard errors clustered at the EA level. All regressions are weighted to make them representative of the target population in the study EAs. Baseline values of the following variables are included as controls in the regression analyses: age dummies, strata dummies, household asset index, highest grade attended, and an indicator for never had sex. *** Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

Cash Transfers and Adolescent Welfare

163

References Ashraf, N., E. Field, and J. Lee. 2014. “Household Bargaining and Excess Fertility: An Experimental Study in Zambia.” American Economic Review 104 (7): 2210–37. Baird, S., E. Chirwa, C. McIntosh, and B. Özler. 2010. “The Short-Term Impacts of a Schooling Conditional Cash Transfer Program on the Sexual Behavior of Young Women.” Health Economics 19:55– 68. Baird, S., J. de Hoop, and B. Özler. 2013. “Income Shocks and Adolescent Mental Health.” Journal of Human Resources 48 (2): 370– 403. Baird, S., R. Garfein, C. McIntosh, and B. Özler. 2012. “Effect of a Cash Transfer Programme for Schooling on Prevalence of HIV and Herpes Simplex Type 2 in Malawi: A Cluster Randomised Trial.” Lancet 379 (9823): 1320– 29. Baird, S., C. McIntosh, and B. Özler. 2011. “Cash or Condition? Evidence from a Cash Transfer Experiment.” Quarterly Journal of Economics 126 (4): 1709– 53. Baird, S., and B. Özler. 2012. “Examining the Reliability of Self-Reported Data on School Participation.” Journal of Development Economics 98 (1): 89– 93. Bandiera, O., N. Buehren, R. Burgess, M. Goldstein, S. Gulesci, I. Rasul, and M. Sulaiman. 2015. “Women’s Empowerment in Action: Evidence from a Randomized Control Trial in Africa.” Working Paper, University College London. http://www.ucl.ac.uk/~uctpimr/research/ELA.pdf. Barber, S. L., and P. J. Gertler. 2010. “Empowering Women: How Mexico’s Conditional Cash Transfer Programme Raised Prenatal Care Quality and Birth Weight.” Journal of Development Effectiveness 2 (1): 51– 73. Bruhn, M., and D. McKenzie. 2009. “In Pursuit of Balance: Randomization in Practice in Development Field Experiments.” American Economic Journal: Applied Economics 1 (4): 200– 32. Chiappori, P., M. Iyigun, and Y. Weiss. 2009. “Investment in Schooling and the Marriage Market.” American Economic Review 99 (5): 1689– 713. Chiappori, P., and S. Oreffice. 2008. “Birth Control and Female Empowerment: An Equilibrium Analysis.” Journal of Political Economy 116 (1): 113– 40. Duflo, E. 2003. “Grandmothers and Granddaughters: Old-Age Pensions and Intrahousehold Allocation in South Africa.” World Bank Economic Review 17 (1): 1– 25. ———. 2012. “Women Empowerment and Economic Development.” Journal of Economic Literature 50 (4): 1051– 79. Edmonds, E. V. 2006. “Child Labor and Schooling Responses to Anticipated Income in South Africa.” Journal of Development Economics 81 (2): 386– 414. Eswaran, M. 2002. “The Empowerment of Women, Fertility, and Child Mortality: Towards a Theoretical Analysis.” Journal of Population Economics 15:433– 54. Glewwe, P., P. H. Ross, and B. Wydick. 2013. “Developing Aspirations: The Impact of Child Sponsorship on Self-Esteem and Life Expectations.” Conference Paper, Young Lives, University of Oxford. http://www.younglives.org.uk/files/others /inequalities- conference- papers/wydick- glewwe- ross- impactchildsponsorship self- esteem- lifeexpectations. Iyigun, M., and R. P. Walsh. 2007. “Endogenous Gender Power, Household Labor Supply and the Demographic Transition.” Journal of Development Economics 82:138– 55. Kabeer, N. 1999. “Resources, Agency, Achievements: Reflections on the Measurement of Women’s Empowerment.” Development and Change 30:435– 64. Kandpal, E., K. Baylis, and M. Arends-Kuenning. 2013. “Measuring the Effect of a Community-Level Program on Women’s Empowerment Outcomes: Evidence from India.” Policy Research Working Paper no. 6399, Washington, DC, World Bank.

164

Sarah Baird, Ephraim Chirwa, Jacobus de Hoop, and Berk Özler

Karlan, D., N. Ashraf, and W. Yin. 2007. “Female Empowerment: Impact of a Commitment Savings Product in the Philippines.” World Development 38 (3): 333– 44. Levine, R., C. B. Lloyd, M. Greene, and C. Grown. 2008. Girls Count: A Global Investment and Action Agenda (reprint, 2009). Washington, DC: Center for Global Development. Lloyd, C. B. 2009. New Lessons: The Power of Educating Adolescent Girls. New York: The Population Council. Lundberg, S., R. A. Pollak, and T. J. Wales. 1997. “Husbands and Wives Pool Their Resources? Evidence from the United Kingdom Child Benefit.” Journal of Human Resources 32 (3): 463– 80. Muralidharan, K., and N. Prakash. 2013. “Cycling to School: Increasing Secondary School Enrollment for Girls in India.” NBER Working Paper no. 19305, Cambridge, MA. Pitt, M. M., S. R. Khandker, and J. Cartwright. 2006. “Empowering Women with Micro Finance: Evidence from Bangladesh.” Economic Development and Cultural Change 54 (4): 791– 831. Rangel, M. A. 2006. “Alimony Rights and Intrahousehold Allocation of Resources: Evidence from Brazil.” Economic Journal 116 (513): 627– 58. Sen, A. 1999. Development as Freedom. Oxford: Oxford University Press. Warhurst, A., E. Molyneux, and R. Jackson. 2010. “Girls Discovered: Global Maps of Adolescent Girls.” Accessed July 2010. http://www.girlsdiscovered.org/. Wydick, B., P. Glewwe, and L. Rutledge. 2013. “Does International Child Sponsorship Work? A Six Country Study on Impacts on Adult Life Outcomes.” Journal of Political Economy 121 (2): 393– 426. World Development Indicators Database. 2009. Accessed January 2010.

6

Comparing Economic and Social Interventions to Reduce Intimate Partner Violence Evidence from Central and Southern Africa Radha Iyengar and Giulia Ferrari

6.1

Introduction

Empowerment of women within households and reduction in domestic violence remains a major issue around the world, including Africa. Despite this, there is a lack of broad evidence and little consensus among scholars or practitioners as to what programs or policies are effective. In particular, the debate remains as to whether economic conditions, such as wage rates or labor market opportunities, affect bargaining power and reduction in violence or whether specific gender- based programs are required. This chapter describes an impact evaluation of a financial skills and negotiation- training program in conjunction with microfinancing in Burundi compared to data from a previously published study on gender- based training for women receiving microfinancing in South Africa. The Burundi program coupled discussion groups for both women and Radha Iyengar is a senior economist at RAND. Giulia Ferrari is a research fellow at the London School of Hygiene and Tropical Medicine. Previously circulated as “Discussion Sessions Coupled with Microfinancing May Enhance the Role of Women in Household Decision-Making in Burudi.” The authors are grateful to Jeannie Annan, Oriana Bandiera, Shawn Cole, Erica Field, Michael Kremer, Emily Oster, and participants at the NBER Africa Project conferences and Harvard Development Seminars for helpful comments. The authors also thank Bersebeh Beyene and Gabrielle Cole as well as members of the IRC Burundi staff for assistance in data collection. We thank Professor Charlotte Watts (London School of Hygiene and Tropical Medicine) for providing us the raw data from the IMAGE study. The authors gratefully acknowledge financial support from the NBER Africa Project and the Centre for Economic Performance. Iyengar also acknowledges financial support from the Robert Wood Johnson Foundation. Any remaining errors are our own. For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see http://www.nber.org/chapters/c13379.ack.

165

166

Radha Iyengar and Giulia Ferrari

men with participation and financing for women in local savings and loan associations. The discussion groups focused on financial decision making within households with the aim of increasing the role of women in household decision making and in reducing domestic violence toward women. The program was designed to provide women with access to economic resources through village savings and loan associations (VSLA). The International Refugee Committee (IRC) randomly selected half of the members in each of the twenty- five VSLA groups to participate in a set of six discussion sessions where it shared with women and their spouses progressive attitudes about the role of gender in household decision making regarding finances. The evaluation utilized focus groups to investigate whether the discussion sessions were effective at increasing the role of women in decisions regarding household purchases and concomitantly in reducing violence against women. In this study, we compared the Burundi program to the well- known program in South Africa—Microfinance for Gender Equity (IMAGE). The South African study was designed to increase access to resources and reduce violence for poor women (Kim et al. 2009). The program included both microfinancing and a ten- session group course for the women. In South Africa, the study focused on general life skills and specific gender issues like fertility and sexually transmitted diseases, and the program was purposefully targeted at women and intended specifically to reduce violence in the household. Both studies show that discussion sessions in conjunction with microcredit participation improves financial decision- making authority for women, reduces exposure to violence, reduces acceptance of violence, and increases consumption of household goods relative to luxury goods, such as alcohol. The evidence from both studies suggests that discussion groups may be a useful approach for empowering women when applied in combination with the improved economic access provided by microcredit participation. In the following sections we present the program and evaluation designs, the sources of data, and the related timeline and outputs. 6.2

Background

Programs to reduce gender- based violence have spanned a range of countries in southern, central, and eastern Africa with varying degrees of success. In part, this is because these programs have been largely divorced from theories on underlying causes of intimate partner violence. Theories on the relationship between resources and violence are abundant in several disciplines including psychology, sociology, and economics. Despite this, there is limited empirical evidence to distinguish between these models. In psychology, there are a range of theories and explanations for vio-

Economic and Social Interventions to Reduce Intimate Partner Violence

167

lence. Broadly speaking, there have been two types of theories.1 The first characterizes violence as due to a lack of control during escalating arguments. Such violence programs focus on anger management programs and more detailed cognitive behavioral therapy as a means to reduce violence (Dutton and Corvo 2006). If increased resources reduce conflict within the household, then regardless of who the resources are provided to, violence should reduce. On the other hand, if female resources increase conflict then the chance of escalation to violence increases. An alternative theory is that violence is a strategically chosen systematic means to exercise control. Strategic violence for the purposes of control might therefore increase as women have increased access to resources. Work by sociologists and criminologists largely focuses on social and contextual causes of violence, but parallels closely the psychology theories. In these fields, there are two prominent sets of theories: “exposure” and “backlash.” The exposure theory focuses on the amount of time spent together. This is similar to the “violence is due to lack of control” described in the psychology literature. Programs that increase income- generating activities by women or generally increase separation will reduce violence by reducing the time partners spend together. Similarly, increased unemployment by men may increase violence by increasing the time partners spend together (Dugan, Nagin, and Rosenfeld 1999). The other theory is “male backlash,” related to the strategic violence theory cited in psychology. Aizer (2010) gives a detailed description of this literature, but the basic concept is that increased financial independence by women increases repression by men in other areas of interaction. As a result, increased access to resources for women will increase violence. In contrast, work by economists typically focuses on modeling household interactions. While neoclassical models of unitary household decision making (e.g., Becker 1965) are still used occasionally, data from a wide range of settings have rejected several features of the unitary model. This has led to a large literature that models household decisions as the result of bargaining among household members (e.g., Browning and Chiappori 1998). These models have helped frame findings from developing countries that show that increases in the female share of household income, interpreted as providing the woman more power within the household, induce an allocation of resources that better reflects her preferences (Duflo 2003; Rangel 2005). This allocation tends to feature greater investment in education, housing, and nutrition for children (Strauss and Thomas 1995; Duflo 2003). Many now see women’s empowerment as key to improving the welfare of women 1. There is a rich and detailed psychology literature on both the motivation and effects of intimate partner violence, a full review of which is beyond the scope of this chapter. For a more detailed treatment, see Johnson and Ferraro (2000).

168

Radha Iyengar and Giulia Ferrari

and children. To date, however, there is little evidence that externally induced “empowerment” is effective. While experimental evidence does suggest that legal control of a new asset empowers women (Ashraf, Karlan, and Yin 2006), this empowerment effect is short lived. Typically these models do not include violence, though adaptations by Aizer (2010) and Pollak (2005) present results with the man’s utility increasing in violence and the woman’s decreasing in violence. In such models, increased resources increase women’s bargaining power, suggesting that violence should decrease. However, this is only true if the increased resources imply an increased outside opportunity. Empirical evidence distinguishing the theories largely comes from economists who have employed structural methods or used panel data to overcome the problem posed by endogenous wages. Bowlus and Seitz (2006) use structural methods to estimate a negative impact of female employment on abuse. Tauchen, Witte, and Long (1985) and Farmer and Tiefenthaler (1997) utilized panel data on victims of domestic violence to examine the impact of changes in a woman’s income over time on violence. In all cases, there is limited ability to distinguish between resources affecting violence and reverse causality of violence affecting resources and unemployment. Recent work by Aizer (2010) uses demand shocks in female- dominant sectors to identify the effect of increased wages on violence. Aizer’s findings suggest increased resources reduce violence and are thus inconsistent with the backlash/strategic control models that predict that as women’s wages increase, violence against them increases. In addition, Aizer finds that violence reduction occurs during nonworking hours, inconsistent with the exposure/lack of control models as well. Aizer’s work provides important insight into settings where outside options improve and women have substantial economic and social freedom. In many settings across the world, and particularly in sub-Saharan Africa, women have few outside options. Women often have few legal rights and there is extreme social pressure to stay in marriages that are often extremely abusive. There is even more limited evidence on the role of resources and negotiations in these settings. The only existing evidence on the impact of women’s economic status on domestic violence comes from an experiment in South Africa and Burundi VSLA interventions described below. 6.3

Experimental Evidence from South Africa and Burundi

Among the most prominent of these experiments to reduce intimate partner violence was the program with Microfinance for Gender Equity (IMAGE) in South Africa, a cooperative study between the London School of Hygiene and Tropical Medicine (LSHTM), Wits University in Johannesburg, and the microfinance non- governmental organization (NGO) Small Enterprise Foundation (SEF). When this study was first implemented, SEF had been working in the Limpopo Province of South Africa for nine years.

Economic and Social Interventions to Reduce Intimate Partner Violence

169

Limpopo is one of the poorer provinces in South Africa, with 50 percent of its population earning 800 Rands or less a month in 2001.2 The IMAGE pilot was introduced in the peri- urban area around the mining town of Burgersfort, Sekhukhuneland. The researchers from LSHTM and Wits University structured a ten- session curriculum on life skills, health, and gender training to be administered to the women receiving microloans. The sessions were structured as discussion forums for adult learning facilitated by a group of social workers specially trained for the task. The ten sessions were offered to women fortnightly at repayment meetings. The training entailed the discussion of gender roles and self- awareness, as well as communication on difficult issues, especially around HIV, and within the household.3 The training was bundled into a package that also entailed participation into the pro- poor group- lending program that SEF runs in the province, based on the Grameen model of group- lending schemes, where participants form groups of five individuals that are jointly liable for the repayment of their individual loans. The framework followed by the IMAGE researchers is illustrated schematically in figure 6.1. As shown in the diagram, there is a common risk environment that determines a number of structural (in public health terms), cultural, and socioeconomic conditions that determine the vulnerability of the women. The aim of the IMAGE program was to break this vicious cycle by breaking the cycle of poverty the women are trapped in and alter the women’s perceptions of gender norms, thereby increasing their empowerment not just by giving them access to increased income, but also by altering their attitudes to violence and their decision- making skills. In 2001, the IMAGE study was designed as a pilot study of the introduction of a microfinance and training (mf plus) product in a new market. It was geared toward understanding the intervention’s efficacy, in view of possibly expanding operations in the area. To our knowledge it was the first randomized controlled trial of an mf plus. It compared the full package of microfinance and life skills and gender training versus no program participation.4 This is in line with the multifaceted nature of the risk that the study 2. Equivalent to USD 93 a month at 2001 exchange rates (calculated September 2001; source: http://www.oanda.com/currency/converter/ ). 3. This could also include an intervention with life- skills training only, although the difference here is that these types of discussion forums typically do not enjoy the regularity in attendance that microfinance program- based sessions typically do. This, as noted above, is possibly one of the reasons why life- skill training programs are often coupled to microfinance programs. 4. A subsequent cross- sectional study compared the two initial groups at follow-up with a third group of women in similar villages that only had access to microfinance and found that the latter group tended to fare better on most economic outcomes, while the full intervention group fares comparatively better on all of the other empowerment and violence- reduction outcomes (Kim et al. 2009). Further, the group exposed to full treatment seemed to do better on some longer- term economic outcomes, a finding that we corroborate from the Burundi study we present here. Further investigations of these trends may be useful and here we also analyzed the data from Kim et al. (2009).

170

Fig. 6.1

Radha Iyengar and Giulia Ferrari

Theoretical framework of the IMAGE study (South Africa)

Source: Based on figure from IMAGE study results (Kim et al. 2009). Notes: A subsequent cross-sectional study compared the two initial groups at follow-up with a third group of women in similar villages that only had access to microfinance (not depicted here). Adapted from IMAGE study.

hypothesized the individuals to be faced with and does not allow us to disentangle the effects of each component, but only to observe their joint effect. The IMAGE evaluation randomly assigned villages to treatment (microloans plus discussion sessions) and control. The study villages were first assigned to three different groups on the basis of their size and accessibility—large and accessible, two pairs of medium and accessible, and small and inaccessible—to reflect the typology of villages in the area, because it was hypothesized that villages that were larger and closer to main roads would have a more dynamic market than smaller villages, or villages that were not as close to main roads. Village characteristics were measured during field reconnaissance visits due to lack of census data on these villages at the time the pilot started. Randomization happened at the level of the cluster defined by the village pair, and individual villages were randomized either to treatment or to control by means of a lottery. Each of the three clusters contained two villages, and the lottery randomly assigned these to immediate treatment or to deferred treatment. The women joined in groups of five in each village, generating between a minimum of nine to a maximum of thirty groups per village. Within both intervention and control villages, the pool of individuals eligible to join the program was identified by means of a participatory rural appraisal technique called participatory wealth ranking (PWR), devised by the collaborating NGO, and whose consistency with statistical methods

Economic and Social Interventions to Reduce Intimate Partner Violence

171

has been tested and discussed elsewhere (Hargreaves et al. 2007). The program was designed to measure the effectiveness of the intervention, and it compared treated individuals in the villages assigned to treatment versus non- treatment- eligible individuals in control villages; this implies that we would find some statistically significant differences at baseline for sociodemographic characteristics that could bias the estimates, and that we therefore controlled for. Finally, because the level of treatment is the village, we clustered the errors at the village level. Program evaluation happened at two points in time: the baseline survey was collected in 2001– 2002, and the follow-up survey in 2003– 2004, so that each individual would be interviewed two years after the baseline interview. Interviewers received one month of training prior to going to the field. In contrast to previous efforts to enhance women’s empowerment, the program in Burundi did not to focus on women’s empowerment explicitly, fearing backlash in the home and community and increasing women’s vulnerability to violence in the short term. Instead, the program approached the issues of empowerment and gender- based violence subtly by encouraging discussion among partners to analyze how men and women relate to one another within the privacy of their homes in negotiating access to and control over household resources. While the courses did not explicitly deal with gender issues, the hypothesis driving the program was that encouraging husbands and wives to discuss household decisions and to respect women’s opinions may improve women’s decision- making power in the home. The courses were conceived to help facilitate a household atmosphere where women (and their opinions) would be more valued and violence against them becomes a less acceptable way of solving issues and conflicts. This change in attitudes and respect could reduce vulnerability to violence within the household. The Burundi program, run by the IRC, established the pilot VSLA program in the Makamba province of Burundi. The pilot project involved twenty- five groups across the Makamba Province in Burundi: seven in Nyanza-Lac, six in Kibago, six in Kayagoro, and six in Mabanda. In addition to implementing the VSLA methodology according to the guidelines and principles developed by CARE International, researchers from London School of Economics (with input from IRC) developed a six- course discussion group series that addressed household decision making along gender lines, the respective roles of women and men, and the use of violence against women in the home more broadly. Half of the participants in the VSLAs were invited to attend these discussions with their spouses. The IRC staff members from the gender- based violence program were trained to facilitate these discussion groups. The basic logic model is presented below. This impact evaluation is among the few detailed studies IRC has ongoing in postconflict countries. The IRC has partnered with academic evaluators in

172

Radha Iyengar and Giulia Ferrari

Liberia, Congo, and Cote d’Ivoire to assess the net difference its work makes for people and investigate what works best to accomplish IRC objectives.5 The Burundi evaluation is relatively unique because it used randomization more narrowly than other evaluation designs that are based on a complete random assignment of units (communities, villages, individuals) into treatment and control. Such broad randomization, although very valuable for testing overall effectiveness, does not provide insights into what parts of a program work and how existing programs may be specifically enhanced. In Burundi, the VSLAs were already formed at the time the evaluation began, thereby limiting the random assignment to units into which individuals had already self- selected. Thus, although this evaluation cannot assess the effectiveness of the VSLA program overall—which would require a control group that does not receive VSLA—it can assess the degree to which an important additional variation of the program design is more or less effective at influencing women’s empowerment, defined here as their ability to not only access economic resources, but also participate in controlling them. The cluster unit of randomization was the VSLA but randomization occurred at the individual level, with half of each VSLA’s members selected randomly into the treatment—that is, the discussion groups—through a lottery, held in each VSLA.6 Slips were drawn from a hat, and those with “winning” slips were the ones who entered the discussion groups with spouses. Those selected were invited to attend a six- session course on household decision making with their spouses.7 The program was evaluated at different points in time, and both quantitative and qualitative data were analyzed statistically, in order to fully capture the complexity of the programmatic impact. After the initial formation of groups through the community- based facilitators, the IRC conducted a 5. This project was approved by Harvard University Human Subjects (Application Number: F15660– 101). 6. The VSLA groups initially formed through members of the community designated as community- based facilitators (CBF).The IRC identified CBFs during community mobilization on the VSLA approach. The IRC was able to reach four communes and eight zones. After having explained the VSLA approach and the role of CBFs, community members elected two or three people. In each commune, the IRC invited four CBFs (for a total of sixteen CBFs) to a meeting where a transparent selection process was conducted to identify the eight CBFs. The IRC chose two individuals that fulfilled all or the majority of the criteria in each commune. At the end of the process the IRC had retained eight CBFs, with four women and four men as facilitators. Each commune had one female and one male facilitator. The CBFs were responsible for training groups in the VSLA methodology. 7. All participants were informed that due to space constraints, only half of the members would be able to attend. In each discussion group, individuals drew numbers from a bag or hat. Those who drew a “winning” number were invited to attend the groups. Others were informed that they would not participate this time, but would hopefully be able to participate in the next round. The lottery was conducted this way due to concerns that choosing half of the discussion groups would result in insufficient statistical power to detect an effect.

Economic and Social Interventions to Reduce Intimate Partner Violence

173

baseline survey of all participants to determine attitudes and assess comparability of treatment (discussion group attendees) and control groups. During the course of the discussion groups, IRC- designed monitoring tools were used to test the comprehension and retention of discussion group material. These tools can also be used to improve the quality of how the discussion groups are designed and facilitated along the way and during the second phase of implementation. After the conclusion of the discussion groups, the IRC conducted a second survey to determine direct immediate effect of discussion groups on attitudes. At this stage, the IRC also conducted four focus groups, with both treated and nontreated men and women, to contextualize and enrich the quantitative findings from the postdiscussion focus group survey. After the VSLA groups had completed their one- year cycle, and savings plus interest had been distributed to all participants, the IRC conducted a final survey of the short- run effects of VSLA participation and attendance in facilitated discussion groups on reported outcomes. In theory, participation by both men and women can also open up opportunities for dialogue over economic decisions from more equalized positions of power, which is often a critical barrier to economic self- reliance among women. The discussion sessions were conceived to improve attitudes toward women’s empowerment, thereby decreasing their vulnerability to violence in the home. We formulated and tested four hypotheses to investigate whether and how women’s empowerment was increased and their vulnerability reduced as a consequence of the intervention. 6.3.1

Sampling and Design

Burundi The sample of treatment and control participants was drawn from the VSLA groups initially formed through members of the community designated as community- based facilitators (CBFs). In order to determine the sample size necessary to detect a significant change in the outcome measures, we conducted a power analysis of a one- tailed test of treatment = control against the two- sided alternative treatment ≠ control. To conduct a power analysis to determine feasibility, we used previous related work by Kim et al. (2007). Kim and colleagues provided microfinance and sexual health and empowerment counseling to women in South Africa, and found that average effect sizes among treatment- group women revealed a reduction of almost half relative to their control group counterparts. With such a large effect, the pilot study sample of 500 would be sufficient to detect statistically significant change. To determine if such a distribution was applicable to the Burundi population, we compared the results from the baseline survey to the South African sample. The baseline survey used the Hurt, Insult, Threaten, and Scream

174

Radha Iyengar and Giulia Ferrari

(HITS) screening tool.8 This tool was designed as a “paper- and- pencil” instrument for identifying both physical and verbal abuse. It includes four items: physical abuse (such as hitting or punching), insults, threats, and screaming. The four items are scored on a Likert five- point scale.9 Baseline results indicate that the distribution of violence among respondents in Burundi is similar to that of respondents in the South African sample analyzed by Kim et al. Applying the same distribution (mean and standard deviation) of the population in Burundi would imply that the minimum effect size the current pilot could significantly detect is a 27 percent change in outcome values. This is significantly smaller than the economic well- being and attitude effects detected in Kim et al. (2009).10 To increase power for analysis, the sample was randomly drawn from each of the twenty- five groups so that the probability of being chosen for any respondent was 50 percent conditional on being in their VSLA group. Because of a small number of absences, the overall probability of any given VSLA member being chosen to participate in the discussion groups was 48 percent. Absences were random and so this slight divergence does not significantly affect the comparability of the control and treatment groups. South Africa For the study of the IMAGE program, no prior estimates of impact for similar studies existed for all outcomes(Hargreaves et al. 2007), hence the protocol for the study published expected outcomes and interval estimates for such outcomes, discussing the sensitivity of results to changes in key statistical parameters. In particular, because the power for the estimation of impact in randomized trials at the village level is influenced by the number of villages included in the study, the number of individuals in each village, and the intraclass correlation coefficient within each village, the IMAGE 8. Sherin et al. (1998). The HITS tool is used globally now in China, Saudi Arabia, the Middle East, Africa, Europe, and South and North America. It has been validated for women in Spanish, and for partner violence with males. In the United States, the HITS tool is used or has been recommended by Kaiser Permanente Group of Northern California, the New Jersey Hospital Association, the Alaska Department of Health and Human Services, Parkland Hospital in Dallas, the Department of OB/GYN at USF in Tampa, the CDC, and others. It has been translated into multiple languages, including Mandarin Chinese and Arabic. 9. It has been validated against the CTS in a study of 160 female patients in an urban/ suburban family practice setting and ninety- nine self- identified abused women. The HITS scores were strongly correlated with the CTS, with sensitivity and specificity of 96 percent and 91 percent, respectively. Positive predictive and negative predictive values in the family practice setting were 87 percent and 97 percent, respectively. 10. We show that the randomization into the different treatment groups was successful, and that participants do not differ significantly along any identifiable socioeconomic dimension. This is an important step in the evaluation design, because it tells us whether the groups generated via the randomization process are indeed good counterfactuals for one another. As the analyses below illustrate, the groups do not differ in any statistically significant way on average and thus do constitute good counterfactuals for one another. In turn, this allows us to attribute any statistically significant difference in the outcomes to the intervention.

Economic and Social Interventions to Reduce Intimate Partner Violence

175

protocol presented expected estimates and relative precision taking these elements into account.11 However, because virtually no data was available on either the intraclass correlation coefficient or on various outcome variables, the protocol reported a range of such estimates for different values of both baseline prevalence rates and intraclass correlation coefficients for the key outcomes it focused on.12 6.3.2

Distinguishing between Theories

The evaluation design allows for distinguishing the effect of increased resources in addition to discussion sessions. Broadly the studies test whether increased access to resources results in increased control of resources, and the extent to which improved information on the benefits of cooperative household decision making, relative to sole male, sometimes violent decision making, reduced violence. In both studies, the underlying hypothesis is that the program acts by both improving the woman’s economic status and her ability to negotiate her role within the household. In turn, this should imply that the resources she brings into the household remain under her control to a greater extent because she has learned to better negotiate her role; her demands for increased autonomy are less likely to be perceived as threatening by the man because the woman is bringing in resources of her own, and therefore may be seen as more of an equal by her partner. Compared to a situation where a woman receives microfinancing and participates in discussion sessions, the joint administration of these two services should support the woman in improving her status in the household in a nonthreatening and rather constructive manner, so that she may play a more autonomous and constructive role in household decision making and thereby reduce her exposure to violence. If increased resources reduce conflict in the household, then microfinance programs alone may be sufficient to reduce violence. However, if who receives resources matter, this may be due to either the economic bargaining model by economists or the backlash/strategic violence theory by sociologists and psychologists. If increased resources to women reduce violence, then this is supportive of the bargaining model. However, changes in resource levels should be salient only if women have access and decisionmaking power over the use of these resources. Thus the question also arises about whether changes in resource control are required; this may be better accomplished in single- sex or mixed- sex interventions. 11. Except for HIV data. 12. We report the baseline values for a number of relevant socioeconomic dimensions and demographic characteristics that show the extent to which the randomization was successful. Women in the two groups are not statistically different on a number of accounts that will be shown below, and belong to the same socioeconomic milieu according to the participatory wealth- ranking exercise carried out by the microfinance NGO to identify the group of poorest individuals in the village (Hargreaves et al. 2007; Simanowitz and Nkuna 1998).

176

Radha Iyengar and Giulia Ferrari

The two investigations in South Africa and Burundi explore different facets of women’s empowerment and reduction in domestic violence in sub-Saharan Africa in ways that reflect the intrinsic differences in study design, as well as the different socioeconomic contexts. The IMAGE study mostly focused on the dimension of empowerment related to autonomy— that is, the ability of individuals to think for themselves, independently of what others around them say. Thus, it encouraged the women to think for themselves and see themselves as more self- interested individuals compared with the role of women who were not in the program. The Burundi VSLA intervention, in contrast, focused more on the aspect of empowerment that has to do with positive relations with others—that is, meaningful connections with significant others that are mutually enriching and constructive. These intrinsic differences explain the exclusive focus on women we find in IMAGE, and the inclusion of clients of both genders in the Burundi VSLA, respectively. Further, both programs were designed to respond to the local environment they were introduced in, to enhance their respective salience in relation to the local context and, by so doing, enhance their chance of successfully achieving their stated targets. Thus, IMAGE was geared toward South African women in peri- urban areas who have a tradition of joining women’s groups, both for economic purposes—as the presence of numerous women’s stokvels (locally initiated rotating savings associations) indicated—and for political and social purposes (Bozzoli 1990). The program in Burundi owed its structure instead to the fact that IRC preferred to entertain a dialogue with both genders in an effort to offset previous failed attempts that had exclusively focused on women, and in order to reduce the likelihood of rejection on the part of the population of an intervention that only focused on women, given the very conservative nature of the local culture in Burundi (see table 6.1). 6.4 6.4.1

Methods Quantitative Data

Burundi The evaluation relies on four sources of data: (a) a baseline survey conducted in January 2008; (b) a post- discussion- group survey conducted in July 2008; (c) qualitative monitoring to complement the discussion group findings in January 2009; and (d) a final survey conducted in April 2009. These, together with the methods used for data analysis, are briefly described. The Survey Instrument. The three survey waves collected data on household consumption, decision making and conflict resolution, gender roles, attitudes toward violence, exposure to violence, and women’s rights. The first wave of the survey also included a household roster, while the second included sections on asset ownership and income, VSLA loans and savings,

Economic and Social Interventions to Reduce Intimate Partner Violence Table 6.1.

177

Summary of comparative measures

Question

Burundi VSLA

South Africa IMAGE

Household roster (relation, age) Education Displacement Assets

Y Y Y Land ownership

Consumption Who decides on a variety of issues How disputes get resolved Attitudes toward women’s roles and rights Controlling behavior Violence levels Response to violence Knowledge of & communication on HIV Networks & community participation

Y (past 2 weeks) Y Y

Y Y N/A LO + index of hh durables + heads of cattle & other domestic animals N Y N

Y N HITS tool (past 2 weeks) N

Roles only Y WHO tool (past 12 mos.) Y

N/A N

Y Y

and wealth and well- being. The measure of exposure to violence used in this survey is the Hurt, Insult, Threaten, and Scream (HITS) instrument (Sherin et al. 1998). The HITS was chosen due to its proven applicability in a variety of settings, and because it allows for a rapid appraisal of past experiences of violence. Its measurement regards the two weeks prior to the interview. The surveys were conducted by twelve interviewers, four of which were males. Each interview lasted approximately thirty minutes. Analysis. The data were collected in Excel spreadsheets and imported into Stata, a statistical package widely used for econometric analysis. We performed regression analysis on the data using a “difference- in-differences” approach. This method allowed us to compare the magnitude and statistical significance of the relative change in the outcomes of interest experienced by the relevant groups as a result of the treatment, compared to the initial situation. The rigorous randomization design allowed us to attribute the observed changes to the intervention. South Africa The Survey Instrument. The two waves of the IMAGE panel contained data on sociodemographics, group membership, community participation, household dynamics, economic well- being and shortages, HIV/AIDS awareness and communication, societal norms on gender roles, decision making in the household, intimate partner violence including controlling behavior, responses on experiences of abuse, and questions on loan performance for microfinance clients. Importantly, the tool that measures

178

Radha Iyengar and Giulia Ferrari

exposure to violence in the IMAGE study—based on the World Health Organization (WHO) indicators of domestic violence as found in the WHO multicountry study (García-Moreno et al. 2005)—measured incidences over a period of twelve months prior to the interview. This implies that comparisons between the HITS results from the Burundi intervention and the tool used for IMAGE are not directly comparable. The measures of exposure to violence used in the two evaluations differ in terms of the time span they cover. The women were also administered a household questionnaire that included a household roster, questions on the type of the two most significant sources of income, characteristics of the dwelling, household assets, credits and savings, perception of own wealth, and food security. Questionnaires were in total about twenty- five pages long, and took forty minutes to one hour to administer. Interviewers were all females, and during the monthlong training prior to the first wave of survey data collection they learned interviewing techniques for sensitive issues, and studied the questionnaire in depth. Analysis. The data was entered in Access databases, and transferred into Stata. We used the “difference- in-differences” approach in an OLS model to measure impact. We clustered the errors at the village level, as this was the level at which treatment was administered, and control for village pair effect, as these are identifying geographical characteristics of relevance, as well as for a number of sociodemographic variables to correct for baseline imbalances. The model also includes a dummy variable equal to one to capture nonresponse in the outcome variable. 6.4.2

Qualitative Data

There were important differences in the way the qualitative data were collected in South Africa versus Burundi. The qualitative data from South Africa used in this study focused on the understanding of the women’s conception of subjective well- being (SWB) and was used to formulate initial hypotheses as to what aspects of SWB the objective measures of empowerment may relate to exposure to violence. For the Burundi project the qualitative data were the transcripts of the discussion sessions. The data from Burundi was collected in a manner that could be analyzed statistically by the text analysis program Alceste. Burundi The Focus Group Data Collection. Perceptions and customs around decision making within households, including daily and major household purchases, family planning issues, and women’s ability to negotiate sex, domestic violence, and the recent conflict between different ethnic groups were investigated in focus group discussions. We used verbal descriptions to obtain answers to questions.

Economic and Social Interventions to Reduce Intimate Partner Violence

179

We then analyzed the transcripts from focus group sessions to enrich and contextualize the interpretation of the quantitative data. In line with the underlying evaluation design, focus group participants were divided into focus group according to their treatment status and further separated by gender, so that a total of four focus groups were run: one with treated women, one with nontreated women, one with treated men, and one with nontreated men. One moderator supported by an interpreter conducted each focus group. A female moderator and interpreter conducted the focus groups with the women and a male moderator and interpreter those with men because it was thought this would favor a greater degree of understanding and trust during the sessions. At times, especially for the most delicate parts of the discussion, moderators and interpreters explicitly appealed to this form of trust and understanding to reassure participants that their thoughts would be comprehended, valued, and respected.13 The focus groups all had the same structure, and therefore produced information that may be compared across the different groups. The sessions opened with the moderator and interpreter briefly explaining the purpose of the focus group, introducing themselves, and requesting an informed oral consent of participants. The focus group interactions were mainly devoted to obtaining answers for eight key questions, some of which were further articulated as subquestions or themes: the market day, joint decision making, women and their ability to manage money, the ethnic conflict and marital relationships, family planning, and conflict and violence within the household. Analysis. The verbal material from the focus groups was analyzed with the aid of Alceste, a software suite for the analysis of the content of textual data. The software applies a statistical technique called correspondence analysis to identify themes (technically called classes) in the text; this is done by computing the relative frequency and cooccurrence of different parts of speech in the text. By identifying those elements in the text’s vocabulary that tend to occur together, the software defines the key themes woven in the text itself. This technique is useful in the analysis of text that reports responses to open questions, as it allows us to characterize the worldviews of respondents associated with the prompts they were given. This feature makes Alceste particularly apt for the analysis of the material from the focus groups conducted in Burundi, where the moderators only gave hints to introduce the themes they wanted to discuss, and let participants express their own personal views in their own manner on these themes. The software identified seven separate themes that we discuss below in conjunction with 13. In the following excerpt, the female facilitator introduces the part of the discussion on violence to the women in the discussion group: “Let us now talk about violence. I would like to remind you that you are free to talk according to your understanding and whatever you say will be confidential, you know we are almost the same age, so, feel free to express yourselves. Tell me, when a man is angry, what kind of reactions can he have?”

180

Radha Iyengar and Giulia Ferrari

the results from the quantitative econometric analysis of the survey data to more fully capture the impact of the intervention. 6.4.3

Comparison of Results

The results are presented separately for our own data from Burundi and our analysis of the Kim et al. data from South Africa. We provide the results in the immediate context of the hypotheses outlined above. The results are described around the three broad themes of financial decision making, attitudes to violence, and violence outcomes, integrating both quantitative and qualitative findings for Burundi. This integration allows us to provide a rigorous discussion of the hypotheses in light of results while contextualizing them clearly into the areas of relevance for policy making. Only quantitative analysis is provided for South Africa, as Kim et al. are currently conducting qualitative analyses of their data. The tools used in the two interventions to measure experiences of domestic violence differ somewhat, with the Burundi intervention relying on the HITS measure (Sherin et al. 1998), and the IMAGE intervention relying on the WHO methodology found in the WHO multicountry study on violence against women (Garcia-Moreno et al. 2005). However slightly different, both these measures capture a measure of physical assault—push and hit with a fist or object in the case of IMAGE, and physically hurt you in the case of the VSLA intervention in Burundi—as well as a measure of insult, though the IMAGE instrument only gages whether insults are administered in public, and is therefore likely to capture less instances. The two measures differ in that the HITS measure also captures instances of threat and cases when the woman has been screamed at, and hence in general focuses on aggressive behavior of the man toward the woman, broadly understood. The IMAGE tool looks explicitly at sexual violence, investigating whether the woman has been forced to have sex and/or has had sex for fear of the consequences had she refused to, and also at controlling behavior more generally. In both cases, the choice of questions is related to the context where the interventions were introduced, which in turn, as we have seen, determined the nature of the interventions themselves. In the case of IMAGE, the choice of asking explicit questions about sexual violence may be connected both to the widespread incidence of sexual violence itself, and by the fact that in South Africa this is an issue that is openly discussed in the media and by policymakers. In contrast, the choice of the HITS tool—whose efficacy in detecting instances of domestic abuse is documented (Sherin et al. 1998)— has rather to do with the overarching spirit of the intervention and the choice not to focus explicitly on domestic violence in order to not alienate men and the general population in the communities where it was introduced.

Economic and Social Interventions to Reduce Intimate Partner Violence

6.5

181

Results

We report on results from our analyses of the effects of programs that couple microfinancing with discussion sessions in Burundi and South Africa in the areas of decision making, attitudes to traditional norms and violence against women, and reported exposure to violence, both in the form of controlling behavior on the part of men and of more direct forms of violence. The results are shown so as to assess the degree to which the programs have met the targets they had originally defined. For each program we first provide an overview of results to convey the overall impact of the program, we then evaluate the success of the randomization in order to justify the econometric models, and then describe the results in relation to the hypotheses we formulated for the two programs. We focus on the specific indicators related to these hypotheses for each program and develop a comparative perspective. We discuss the common and discordant features of the results from the two programs. 6.5.1 Burundi The VSLA microfinancing program coupled with the discussion sessions in Burundi was targeted at reducing male control over all household decision making. Included in this was the goal of changed attitudes toward household violence. If successful in execution, both men and women who participated in the discussion sessions will have developed a more nuanced understanding of domestic violence without an explicit discussion or consideration of violence. In particular, women who participated in the discussion sessions would be able to describe the elements that constitute the cultural risk environment for domestic violence. “Cultural risk environment” means the set of criteria that identify an acceptable behavior for the woman in the household and that at the same time underscore her lesser stand in the negotiation of roles. In testing the impact of the Burundi program, we find that participation in the discussion groups is associated with increased decision making for women. In particular, 26 percent more women in the discussion groups report an increase in spending on their own earnings. There is no substantial change in decisions on how men’s income is spent. In addition, women report that increased decision- making authority over major household purchases also increases by nearly 14 percent. This change in decision making directly impacts household consumption, with women reporting more than an 11 percent increase in household consumption. Attitudes toward violence changed by 9 percent, with men reporting more often that violence was unacceptable, in particular, when the wife is perceived as neglecting the children and when the wife refuses sex. However, these changes in attitude are not reflected in substantial changes to violence exposure. The program appears to reduce violence by less than 1 percent.

182

Radha Iyengar and Giulia Ferrari

Verifying Randomization Before considering the initial reported attitudes of VSLA participants on gender issues, we asked respondents detailed information about their households, including information about displacement, education, and wealth. This information is important from a methodological standpoint because it provides important information to test that discussion- session participating and control communities are similar across a range of background variables that might shape the outcomes of interest or impact the efficacy of the program. In addition, the data offers a detailed picture of the VSLA participants, many of whom are recent returnees to post- ethnic- conflict Burundi. The individuals who selected into the VSLA program were not necessarily representative of all Burundians. About two- thirds of participants, and thus roughly as many respondents, are female (69 percent). The average age of participants is 37.9, with the youngest participants at age fifteen and the oldest at age eighty. On average, respondents had four children living at home. The maximum number of children living at home is twelve. Just short of two- thirds (61 percent) had young children (under five) in the household. Only 2.2 percent of respondents reported never being displaced due to the ethnic conflict. More than half of respondents were displaced from their homes but remained within Burundi, while 41 percent of respondents reported having to leave their homes and Burundi due to the ethnic conflict. A majority of participants own some land (72 percent). Among landowners, the average number of hectares owned by the household is 4.5 hectares. Approximately 61 percent of respondents had attended some primary school. Only 16 percent of respondents had attended secondary school. An important component in ascertaining the validity of an experiment is to compare the outcome variables of interest in the control and treatment groups to ensure that there are no systematic baseline differences. If randomization is successful, then on average there should be no statistically detectable difference between the control and treatment groups for baseline variables. Confirming this, we found that almost no outcome variable recorded a statistically significant baseline difference between average values recorded in the participants’ and nonparticipants’ groups, respectively (see table 6.2). The only statistical difference in characteristics prior to the discussion sessions was whether the husband decides how the money his wife earns is to be spent. The discussion- session participants reported an approximately 10 percent rate (that is, for discussion- session participants, more husbands decided how the money the wife earned is to be spent). This difference in one pretreatment outcome is not of concern given the large number of outcome variables tested. Statistically, there is a 5 percent chance that an outcome would appear significantly different, consistent with our baseline results. To ensure the groups appear similar on observable characteristics, we

184 184 184 184 184 184 119 119 246 206 243 206 244 205 173 138 241 189 220 180 128 194 110

How money is spent: spouse decides Spend money disagree: Spouse changes Daily hh purchases: Spouse decides Daily hh purchases disagree: Spouse changes Large hh purchases: Spouse decides Large hh purchases disagree: Spouse changes Alcohol/cigarettes: Spouse decides Alcohol/cigarettes disagree: Spouse changes When to visit family & friends: Spouse decides When to visit family & friends disagree: Spouse changes When to visit spouse’s family & friends: Spouse decides When to Visit Family & Friends Disagree: Spouse Changes How many kids: Spouse decides Have sex: Spouse decides When to visit family & friends disagree: Spouse changes

N

SD

2.34 4.05 2.20 4.23 2.02 4.14 1.76 3.52 2.52 4.20 2.53 4.42 2.81 1.89 4.26

B. Decision-making outcomes 2.09 1.24 209 4.272 1.231 161 2.313 1.088 199 4.311 1.177 158 2.053 1.264 201 4.200 1.326 159 1.705 1.000 140 3.667 1.577 104 2.672 1.296 202 4.349 1.155 155 2.691 1.305 184 4.567 0.922 142 2.602 1.433 97 1.938 1.318 142 4.555 0.982 82

Mean 1.20 1.81 1.35 2.23 6.58 0.32 6.48 0.31

N

1.21 1.32 1.04 1.16 1.15 1.33 1.02 1.55 1.21 1.22 1.21 1.07 1.40 1.24 1.26

0.46 1.05 0.79 1.05 2.54 0.47 2.72 0.47

SD

Nonparticipants

154 154 154 154 154 154 99 99

A. Violence outcomes 1.207 0.534 1.728 1.004 1.245 0.693 2.087 1.057 6.266 2.515 0.250 0.434 6.345 2.857 0.269 0.445

Mean

Participants

2.093 –1.646 –1.152 –0.619 –0.291 –0.439 0.513 –0.728 –1.322 –1.153 –1.261 –1.272 1.117 –0.361 –1.784

–0.217 0.744 1.301 1.223 1.152 1.38 0.344 0.712

t

–0.24 0.22 0.12 0.08 0.03 0.06 –0.06 0.15 0.16 0.15 0.16 0.14 –0.21 0.05 0.30

0.01 –0.08 –0.11 –0.14 –0.32 –0.07 –0.13 –0.04

Diff.

1.01 0.95 0.92 0.94 0.95 0.79 0.98 0.86

Chng. (%)

0.90 1.05 1.05 1.02 1.02 1.01 0.97 1.04 1.06 1.04 1.06 1.03 0.92 1.03 1.07 (continued)

Comparison

Comparison of baseline levels of decision making, attitudes, and violence between control and program participation groups (Burundi)

Woman has been physically hurt Woman has been insulted Woman has been threatened Woman has been screamed at Total HIT score Total HIT score greater than 5 Total HIT score females’ reports only Total HIT Score females’ only > 5

Table 6.2

(continued)

Mean

SD N

241 192 257 257 257 257 257

D. Consumption 13,379.3 15,372.22 5,646.2 6,140.84 1,360,000 1,630,000 0.9 1.23 1.0 — 1.0 0.50 1.0 0.15

200 150 226 224 224 224 224

C. Attitudes to women and violence 257 2.202 0.571 225 257 2.525 0.587 225 257 2.467 0.599 225 257 2.486 0.560 225 256 2.879 0.350 225 256 2.555 0.636 222 256 2.773 0.759 222 254 1.630 0.484 223 257 1.455 0.499 224 254 1.736 0.442 221 249 1.707 0.456 216 257 1.887 0.317 223 256 1.859 0.348 223 257 1.864 0.344 224 255 1.525 0.500 222

N

Participants

11,291.93 4,901.47 1,140,000 0.74 1.00 0.99 0.97

2.13 2.60 2.45 2.45 2.89 2.50 2.66 1.58 1.48 1.74 1.74 1.89 1.87 1.89 1.53

Mean

11,592.24 4,689.70 1,500,000 0.72 — 0.09 0.16

0.56 0.55 0.63 0.54 0.34 0.71 0.75 0.50 0.50 0.44 0.44 0.31 0.33 0.31 0.50

SD

Nonparticipants

–1.624 –1.272 –1.524 –1.661 — –1.261 –0.240

–1.418 1.441 –0.241 –0.658 0.46 –0.953 –1.611 –1.145 0.491 0.145 0.702 0.182 0.484 0.974 0.033

t

2,087.33 744.73 220,000. 0.15 — 0.04 0.00

0.07 –0.08 0.01 0.03 –0.01 0.06 0.11 0.05 –0.02 –0.01 –0.03 –0.00 –0.02 –0.03 –0.00

Diff.

Comparison

1.18 1.15 1.19 1.20 — 1.04 1.00

1.03 0.97 1.01 1.01 1.00 1.02 1.04 1.03 0.98 1.00 0.98 1.00 0.99 0.98 1.00

Chng. (%)

Notes: Outcome measures based on survey data collected with assistance of the International Rescue Committee (Burundi). Participants refer to individuals randomly selected to attend the program that consisted of a set of six discussion sessions. Nonparticipants refer to individuals who did not attend the program. Confidence intervals are based on village-cluster estimated standard errors. Number of respondents varies due to differential response rates to questions. Comparison columns report mean difference between participant and nonparticipant groups. Also reported are the t-statistics testing if this difference is distinguishable from zero.

Weekly consumption (value) Weekly self-production (value) Yearly consumption (value) Rights for women to own and inherit land Women reps. in reconstruction programs Women reps. in local community meetings Increased penalties for spousal abuse

Women should do as men say Wife should give money she earns to husband Okay for husband to abandon wife if he wants Woman’s job to gather water, even if unsafe Women cannot manage money Women should have sex when husband wants Women should have as many kids as husband wants Okay to beat wife if goes out w/out telling husband Okay to beat wife if neglects kids Okay to beat wife if argues w/husband Okay to beat wife if refuses sex Okay to beat wife if burns food Okay to beat wife if does something annoying Okay to beat wife for any reason Never okay to beat wife

Table 6.2

Economic and Social Interventions to Reduce Intimate Partner Violence

185

also compared sociodemographic variables. The two groups are also similar along sociodemographic lines, with no variable recording a statistically significant difference between the two groups (see table 6.3). In both groups, approximately 60 percent of respondents are females. The two groups also do not differ significantly in terms of amount of land owned: both spouse and respondent own on average two half hectares of land, and the majority of spouses do not own land jointly. Decision-Making Authority The first objective of the Burundi program was to improve women’s participation in decision making. Women’s participation in decision making in the household is generally considered as an important step in their emancipation. If she learns to take part in the management of household matters, and if her husband learns that it is useful to listen to her, this is interpreted to indicate a greater appreciation of the woman’s input in the household and, in turn, to lead to a reduced likelihood that she is subjected to violence. The intervention aimed to improve decision- making dynamics in this direction, encouraging both men and women to make increasingly more decisions jointly. We identified several areas that may be critical to women’s empowerment: income/asset- related decision- making authority, fertility decisionmaking authority, safety, and political rights. Based on this objective, hypothesis B-H1 was that female members of discussion sessions are more likely to be involved in increasing areas of household decision making. Men who participate in the discussion sessions will be more likely than those that do not to believe that women are capable of making decisions in a broader set of areas. Related to this, hypothesis B-H2 was that members of discussion sessions would become more likely to apply negotiation skills during conflict to reduce the risk of arguments escalating to violence Our results show a statistically significant impact on three of the eight decision- making (DM) indicators: the husband’s role in deciding how the wife’s money is spent, husbands deciding unilaterally on large purchases for the household, and husbands deciding unilaterally on how many children to have. In addition, all three indicators show a similar trend, with women reporting having become more able to cooperate. In all three cases, the percentage increase in cooperative behavior among women with respect to the baseline group is one hundredfold the increase reported by men in absolute value. Tables 6.4 and 6.5 report full results for the battery of decision- making indicators at baseline and follow-up; table 6.5 also reports the percent change in each indicator, relative to the baseline levels. The results suggest that joint decision making on large purchases improves markedly in the group of treated individuals—with an increase in about 15 percent for women relative to baseline decision- making authority. This same trend is observed

257 256 174 137 254 239 175 115 129 257

Males Age Respondents’ level of schooling Spouse’s level of schooling Displaced Displaced outside Burundi Respondent half hectares Spouse’s half hectares Use others’ land Respondent & spouse own land jointly

0.401 39 0.897 0.912 0.996 0.402 2.061 1.935 0.744 0.412

Mean 0.491 12.395 0.305 0.284 0.063 0.491 1.617 1.427 0.438 0.493

SD 226 226 148 116 222 207 149 93 109 226

N 0.336 37 0.892 0.905 0.995 0.391 1.935 1.949 0.761 0.345

Mean

SD 0.473 13.209 0.312 0.294 0.067 0.489 1.666 1.726 0.428 0.476

Nonparticipants

–1.468 –1.31 –0.135 –0.198 –0.095 –0.223 –0.688 0.063 0.307 –1.524

t

0.065 1.534 0.005 0.007 0.001 0.011 0.126 –0.014 –0.017 0.067

Diff.

Comparison

0.16 0.04 0.01 0.01 0.00 0.03 0.06 –0.01 –0.02 0.16

Change (%)

Notes: Outcome measures based on baseline survey data collected with assistance of the International Rescue Committee (Burundi). Participants refer to individuals randomly selected to attend the program that consisted of a set of six discussion sessions. Nonparticipants refer to individuals who did not attend the program. Confidence intervals are based on village-cluster estimated standard errors. Number of respondents varies due to differential response rates to questions. Comparison columns report mean difference between participant and nonparticipant groups. Also reported are the t-statistics testing if this difference is distinguishable from zero.

N

Participants

Comparison of baseline levels of sociodemographic characteristics between control and program participation groups (Burundi)

Variable

Table 6.3

Table 6.4.

Comparison of preprogram decision making and dispute resolution behavior (Burundi) Nonparticipants Females

Males

Participants Females

Males

A. Decision-making authority over household decisions How money you earn is spent 2.276*** 0.0790 –0.247** (0.128) (0.0826) (0.117) Major household purchases 2.054*** 0.136* 0.0186 (0.0897) (0.0786) (0.0858) Daily household purchases 2.213*** –0.0712 0.112 (0.0892) (0.0621) (0.0919) Purchases of alcohol and cigarettes 1.725*** –0.151* –0.0449 (0.0752) (0.0812) (0.0863) Visits your family or friends 2.393*** 0.121 0.145 (0.109) (0.0919) (0.0977) Visits your spouse’s family or friends 2.442*** 0.114* 0.131 (0.0988) (0.0625) (0.0810) How many children to have 2.542*** 0.331*** –0.137 (0.146) (0.111) (0.102) When to have sex 1.649*** 0.556*** 0.0212 (0.0978) (0.0738) (0.0913)

0.00247** (0.00118) –0.000184 (0.000866) –0.00113 (0.000925) 0.000459 (0.000870) –0.00152 (0.000992) –0.00134 (0.000812) 0.00138 (0.00102) –0.000201 (0.000913)

B. Dispute resolution over disagreements on household decisions (no dispute because spouse knows better) Disagree with spouse on how money 4.074*** 0.137 0.210** is spent (0.136) (0.0953) (0.101) Disagree with spouse on major hh 4.159*** 0.150 0.0540 purchases (0.161) (0.0906) (0.126) Disagree with spouse on daily hh 4.175*** 0.259** 0.0539 purchases (0.135) (0.111) (0.122) Disagree with spouse on purchases 3.639*** –0.315** 0.123 of alcohol and cigarettes (0.115) (0.145) (0.156) Disagree with spouse on visit your 4.387*** 0.200*** 0.119 spouse’s family or friends (0.0996) (0.0607) (0.0936) Disagree with spouse on visit your 4.243*** 0.130 0.141 family or friends (0.113) (0.0919) (0.138) Disagree on having sex 4.144*** –0.293*** 0.275** (0.113) (0.0731) (0.107)

–0.00212** (0.00102) –0.000561 (0.00127) –0.000562 (0.00123) –0.00124 (0.00156) –0.00119 (0.000951) –0.00143 (0.00138) –0.00275** (0.00108)

Notes: Outcome variable in panel A is an indicator variable that is 1 if the decision was taken unilaterally by the respondent. Outcome variable in panel B is an indicator variable that is 1 if there was no discussion because the respondent believes the spouse knows better for each of the categories listed in the panel. Participants refer to individuals randomly selected to attend the program that consisted of a set of six discussion sessions. Nonparticipants refer to individuals who did not attend the program. Robust standard errors clustered at the village level are reported in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

Table 6.5.

Estimated effect of discussion sessions on decision making and dispute resolution outcomes (Burundi) Nonparticipants Females

Males

Participants Females

Males

A. Decision-making authority over household decisions ( = 1 if respondents decide alone) How money you earn is spent –0.00272 4.37e–05 0.602*** –0.00605*** (0.0817) (0.000822) (0.138) (0.00138) Percent change (relative to female baseline) –0.01 0.00 26.45*** –0.27*** Major hh purchases 0.106 –0.00104 0.293** –0.00293** (0.0942) (0.000938) (0.119) (0.00120) Percent change (relative to female baseline) 5.16 –0.05 14.26** –0.14*** Daily hh purchases 0.0818 –0.000803 0.0927 –0.000924 (0.0770) (0.000776) (0.115) (0.00116) Percent change (relative to female baseline) 3.70 –0.04 4.19 –0.04 Purchases of alcohol and cigarettes 0.0132 –0.000103 0.0156 –0.000158 (0.0860) (0.000866) (0.127) (0.00128) Percent change (relative to female baseline) 0.90 –0.01 0.90 –0.01 Visit your family or friends 0.232** –0.00232** 0.0836 –0.000772 (0.0961) (0.000962) (0.144) (0.00145) Percent change (relative to female baseline) 9.69** –0.10** 3.49 –0.32 Visit your spouse’s family or friends –0.0367 0.000351 0.0563 –0.000525 (0.0776) (0.000787) (0.0900) (0.000913) Percent change (relative to female baseline) –1.50 0.01 2.31 –0.02 How many children to have –0.147** 0.00149** 0.359*** –0.00359*** (0.0639) (0.000645) (0.108) (0.00109) Percent change (relative to female baseline) –5.78** 0.06** 14.12*** –0.14*** Having sex 0.0118 –0.000115 0.104 –0.00102 (0.0748) (0.000750) (0.0949) (0.000954) Percent change (relative to female baseline) 0.72 –0.01 6.31 –0.06 B. Dispute resolution over disagreements on household decisions ( = 1 if there was no dispute because respondent feels spouse knows better) Disagree with spouse on major hh purchases 0.183 0.00104 0.183 (0.135) (0.00152) (0.135) Percent change (relative to female baseline) 4.40 0.03 4.40 Disagree with spouse on how money is spent –0.00814 8.84e–05 –0.00253 (0.139) (0.00139) (0.147) Percent change (relative to female baseline) –0.20 0.00 –0.06 Disagree with spouse on daily hh purchases –0.213* 0.00216* 0.0474 (0.109) (0.00110) (0.119) Percent change (relative to female baseline) –5.10* 0.05* 1.14 Disagree with spouse on purchases of –0.0731 0.000747 –0.0404 alcohol and cigarettes (0.125) (0.00124) (0.209) Percent change (relative to female baseline) –2.01 0.02 –1.11 Disagree with spouse on visit your family –0.0341 0.000367 0.0658 or friends (0.108) (0.00109) (0.165) Percent change (relative to female baseline) –0.80 0.01 1.55 Disagree with spouse on visit your spouse’s 0.124 –0.00126 0.0154 family or friends (0.0891) (0.000902) (0.109)

–0.00184 (0.00135) –0.04 3.37e–05 (0.00148) 0.00 –0.000477 (0.00119) –0.01 0.000406 (0.00209) 0.01 –0.000667 (0.00165) –0.02 –0.000174 (0.00111) (continued)

Economic and Social Interventions to Reduce Intimate Partner Violence Table 6.5.

189

(continued) Nonparticipants

Percent change (relative to female baseline) Disagree on having sex Percent change (relative to female baseline)

Females

Males

2.83 –0.868*** (0.133) –20.95***

–0.03 0.00876*** (0.00134) 0.21***

Participants Females 0.35 –0.0710 (0.148) –1.71

Males 0.00 0.000688 (0.00149) 0.02

Notes: Outcome variable in panel A is an indicator variable that is 1 if the decision was taken unilaterally by the respondent for each of the categories listed in the panel. Outcome variable in panel B is an indicator variable that is 1 if there was no discussion because the respondent believes the spouse knows better for each of the categories listed in the panel. Participants refer to individuals randomly selected to attend the program that consisted of a set of six discussion sessions. Nonparticipants refer to individuals who did not attend the program. Robust standard errors clustered at the village level are reported in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

for decisions on when to visit one’s own or one’s spouse’s family; further, the management of disagreements in these two areas also shows similar patterns, although these results are not statically significant. The results in table 6.5 also suggest that negotiation skills targeted as nonviolent resolution of disagreements improve for women in the intervention group. Women are 4 percent more likely to discuss with their husbands when they disagree with both large and daily purchases. A similar trend in negotiation abilities is observed for resolving disagreements in relation to visits to either their own or their spouse’s family. Men in the treatment group report a very small reduction in their ability to negotiate. Although this effect is small in magnitude (less than 1 percent) it appears consistent across various indicators and may be due to heightened awareness of what negotiation entails. The results are most pronounced for decision making on sex and alcohol and cigarette purchases, which are also the outcomes that had the lowest levels of cooperation as well as the areas with the largest influence of men in decision making. Figure 6.2 shows the change in whether the respondent’s partner decides how to spend the respondent’s money: female VSLA members who were members of discussion groups reported a 26 percent increase in level of empowerment when compared to female VSLA members in the control group at baseline. Men who participated in the discussion sessions, however, reported a greater degree of control of household money when compared to female participants at baseline, that is, a lower tendency to cooperate. Though significant at the 95 percent level, this decrease in cooperation is one hundred

Decision making on economic issues: Who decides how to spend money? (Burundi)

Notes: Outcome measures based on survey data collected with assistance from the International Rescue Committee (Burundi). Participants refer to individuals randomly selected to attend the program that consisted of a set of six discussion sessions. Nonparticipants refer to individuals who did not attend the program. Confidence intervals are based on village-cluster estimated standard errors.

Fig. 6.2

Economic and Social Interventions to Reduce Intimate Partner Violence

191

times smaller than the increase reported by the women (– 0.27 percent for females in the control group at baseline and – 0.26 percent when compared to males in the control group at baseline). Females who participated in the discussion groups cooperate on major purchases on average 14.26 percent more than the women at baseline ( p < 0.05), reporting an average of 2.347. Men showed a decrease, instead, of 0.14 percentage points in their ability to share decision making on the initial 2.054 average recorded by women in the control group at baseline (see figure 6.3). Females who had participated in the discussion sessions recorded an increase of 0.36 points over female clients in control groups at baseline regarding the decision of how many children to have, recording a 14.12 percent increase from 2.542 ( p < 0.05). Men reported a decrease of 0.14 percentage points on the initial value reported by women in the control group (see figure 6.4). When considering the full range of the decision- making indicators, even those that did not change significantly showed trends similar to the ones illustrated above. In general, the women report an increase in joint decision, while men report a very small (often one hundredfold smaller) increase in their own role in household decision making. This is consistent with the evidence from the qualitative analysis. Two themes emerged in the focus groups related to the economic sphere of access to and management of resources. One theme specifically describes the role of the woman in the management of the household. Typically the activities included were cleaning, making the bed, fetching water and wood, and preparing meals. The role of women in this area was most often associated with her responsibility or duty. The related theme for men contains a very rich description of all activities revolving around the market, and is associated explicitly with men in their role as fathers. This is specifically in relation to the key role they seem to play in providing for their children’s school purchases. The sentences characteristic of this theme seem to report the husband is in a dominant position, that is, as the one who ultimately has the privilege to make decisions that revolve around the acquisition of resources in the market, possibly because of his more direct access to money. The focus on the woman’s duties in the household and her role in its management is associated with the discussion on the division of labor, and was largely concentrated among participants in the discussion sessions. In contrast, the discussion on men’s privileged access to the market was prevalent with both men and women who were not participants. This is consistent with some initial effect of the treatment in fostering a more diverse set of perceptions and ideas around the role of men and the household’s access to the market among treated individuals. This interpretation is supported by the quantitative results, where a greater degree of joint decision making is reported by both males and females, as opposed to the control groups.

Decision making on economic issues: Who decides on large household purchases? (Burundi)

Notes: Outcome measures based on survey data collected with assistance from the International Rescue Committee (Burundi). Participants refer to individuals randomly selected to attend the program that consisted of a set of six discussion sessions. Nonparticipants refer to individuals who did not attend the program. Confidence intervals are based on village-cluster estimated standard errors.

Fig. 6.3

Decision making on household issues: Who decides on having children? (Burundi)

Notes: Outcome measures based on survey data collected with assistance of the International Rescue Committee (Burundi). Participants refer to individuals randomly selected to attend the program that consisted of a set of six discussion sessions. Nonparticipants refer to individuals who did not attend the program. Confidence intervals are based on village-cluster estimated standard errors.

Fig. 6.4

194

Radha Iyengar and Giulia Ferrari

Attitudes toward Gender Norms and Violence A second objective of the Burundi program was to affect attitudes toward domestic violence by challenging traditional views of women. The nature of the challenge was in the economic and decision- making sphere, but theory suggests that increased decision- making authority may enhance perceptions of capability and reduce tolerance and acceptance of violence. The program aimed to improve attitudes in this direction by encouraging both men and women to consider why conflicts arise and why such resolutions are achieved via violence rather than negotiation. We identified several areas in which violence might be more or less tolerated: mobility, fertility, household behavior, and general social/political rights. Based on this objective, hypothesis B-H3 stated that members of the discussion groups would be more likely to think that abuse is never justified (see table 6.6). In general, the program has a positive and statistically significant impact in the reduction of the tolerance of violence in two out of the six areas measured: neglecting children and refusing to have sex. The impact is stronger than the time trend observed in the control group in the case of neglecting children, and approximately the same when it comes to the refusal to have sex, though the estimate for the control group is more precise than that for discussion- session participants. Moreover, the acceptance of wife beating in cases of child neglect records an increase in rejection of abuse among discussion session- participating women equal to 0.137 when compared to baseline females (23 percent increase), versus a negligible decrease in the rejection of violence on the part of discussion session- participating men (– 0.23 percent), also significant at the 90 percent level (see figures 6.5 and 6.6). Participants in discussion sessions are less likely to accept violence for indiscriminate reasons and, in particular, if women go out without saying, argue with their husband, burn food, or say something annoying. Female participants in the discussion sessions are less likely to accept violence against women if they say something annoying or for indiscriminate reasons. Unlike the cases of neglect of children and refusal to have sex, the change in these other dimensions is, however, not statistically significant with respect to levels of acceptance recorded by females at baseline for discussion- session participants. The control group recorded statistically significant reductions in the tolerance of violence if the wife goes out without letting her husband know, argues with him, neglects the kids, refuses sex, or burns food. In all these cases, trends are identical in both groups and for both genders across groups, with a change of approximately the same order of magnitude for women in the intervention and control groups in relation to arguments. Moreover, the control group recorded a statistically significant opposite shift in views between males and females: female clients see it as less acceptable that wives be beaten, while men see it as more acceptable. The

Percent change (relative to female baseline) = 1 If agree that it’s acceptable to beat one’s wife if she neglects kids

B. Postprogram = 1 If agree that it’s acceptable to beat one’s wife if she goes out without her husband’s permission

= 1 If agree that it’s never okay to beat wife

= 1 If agree that it’s acceptable to beat one’s wife for any reason

= 1 If agree that it’s acceptable to beat one’s wife if she’s annoying

= 1 If agree that it’s acceptable to beat one’s wife if she burns food

= 1 If agree that it’s acceptable to beat one’s wife if she argues { { = 1 If agree that it’s acceptable to beat one’s wife if she refuses sex

= 1 If agree that it’s acceptable to beat one’s wife if she neglects kids

0.187*** (0.0384) 12.68*** 0.141*** (0.0398)

1.475*** (0.0375) 1.390*** (0.0459) 1.724*** (0.0387) 1.668*** (0.0345) 0.203** (0.0797) 1.854*** (0.0199) 1.866*** (0.0239) 1.617*** (0.0421)

Females

–0.00191*** (0.000390) –0.13*** –0.00143*** (0.000404)

0.176*** (0.0281) 0.189*** (0.0302) 0.0478** (0.0223) 0.124*** (0.0268) –0.00739 (0.0173) 0.0540*** (0.0160) 0.0385* (0.0190) –0.0986** (0.0398)

Males

Nonparticipants

Estimated effect of discussion sessions on attitudes toward gender norms and violence (Burundi)

A. Baseline = 1 If agree that it’s acceptable to beat one’s wife if she goes out without her husband’s permission

Table 6.6

0.0223 (0.0380) 1.51 0.130** (0.0533)

0.0476 (0.0432) –0.0262 (0.0449) –0.00899 (0.0467) –0.0328 (0.0387) –0.0525 (0.0653) –0.0160 (0.0354) –0.0292 (0.0328) –0.00332 (0.0556)

Females

–0.000222 (0.000386) –0.02 –0.00132** (0.000538) (continued)

–0.000490 (0.000437) 0.000271 (0.000451) 0.000106 (0.000473) 0.000335 (0.000393) 0.000535 (0.000666) 0.000158 (0.000357) 0.000297 (0.000332) 3.49e–05 (0.000561)

Males

Participants

(continued)

10.14*** 0.127*** (0.0343) 7*** 0.101*** (0.0247) –0.06*** 0.0570** (0.0219) 3** 0.0325 (0.0259) 1.75 0.0265 (0.0232) 1.41 –0.0848** (0.0373) –5.40**

Females –0.10*** –0.00129*** (0.000345) 0*** –0.00102*** (0.000250) 4* –0.000580** (0.000223) –0.03** –0.000333 (0.000262) –0.02 –0.000265 (0.000236) –0.01 0.000863** (0.000377) 0.05**

Males

Nonparticipants

9.35 0.0275 (0.0478) 2 0.0661* (0.0341) –0.04* 0.0166 (0.0307) 0.89 0.0495 (0.0460) 2.67 0.0334 (0.0355) 1.77 –0.0185 (0.0566) –1.18

Females

–0.000169 (0.000312) –9.02296E–05 –0.000496 (0.000463) –0.03 –0.000346 (0.000361) –0.02 0.000182 (0.000571) 0.01

–0.09 –0.000295 (0.000483) –0.02 –0.000674* (0.000347)

Males

Participants

Notes: Each row presents the results from a separate regression with the dependent variables listed in each row. Percent changes in panel B are based on comparison to females in the control group prior to treatment. Participants refer to individuals randomly selected to attend the program that consisted of a set of six discussion sessions. Nonparticipants refer to individuals who did not attend the program. Robust standard errors clustered at the village level are reported in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

Percent change (relative to female baseline)

Percent change (relative to female baseline) = 1 If agree that it’s never okay to beat wife

Percent change (relative to female baseline) = 1 If agree that it’s acceptable to beat one’s wife for any reason

Percent change (relative to female baseline) = 1 If agree that it’s acceptable to beat one’s wife if she’s annoying

Percent change (relative to female baseline){6*** = 1 If agree that it’s acceptable to beat one’s wife if she burns food

Percent change (relative to female baseline) = 1 If agree that it’s acceptable to beat one’s wife if she argues . Percent change (relative to female baseline) = 1 If agree that it’s acceptable to beat one’s wife if she refuses sex

Table 6.6

Attitudes toward violence: Acceptability of abuse if wife refuses sex (Burundi)

Notes: Outcome measures based on survey data collected with assistance from the International Rescue Committee (Burundi). Participants refer to individuals randomly selected to attend the program that consisted of a set of six discussion sessions. Nonparticipants refer to individuals who did not attend the program. Confidence intervals are based on village-cluster estimated standard errors.

Fig. 6.5

Attitudes toward violence: Acceptability of abuse if wife neglects children (Burundi)

Notes: Outcome measures based on survey data collected with assistance from the International Rescue Committee (Burundi). Participants refer to individuals randomly selected to attend the program that consisted of a set of six discussion sessions. Nonparticipants refer to individuals who did not attend the program. Confidence intervals are based on village-cluster estimated standard errors.

Fig. 6.6

Economic and Social Interventions to Reduce Intimate Partner Violence

199

discussion- session participants recorded a similar trend; in both cases, the absolute difference in change is very large (about one hundredfold). Exposure to Violence The ultimate objective of the program was to reduce women’s exposure to violence. Rather than approach the norms that affect violence directly, the program in Burundi was based on the theory that improving women’s authority over household decisions could challenge the norms that enable violence, thus reducing violence. Based on this theory, B-H4 was that the program reduces the prevalence of domestic violence. Table 6.7 presents the impact of the program on reported violence. It appears that females that did not participate in the discussion sessions reported a reduction in the levels of violence in all areas except for the experience of threats. Males who did not participate in the discussion sessions reported an increase in the level of violence imposed on their partners with respect to the levels control women reported at baseline, and virtually unchanged from their own report at baseline (conditional average was 1.20 at baseline and is 1.19 at follow-up). Females in the control group reported a statistically significant reduction in the experience of violence at the end of the savings cycle, that is, when they receive their total savings back, suggesting that the actual access to the savings makes a difference. However, males in this group reported increasing levels of violence in time, and especially once they receive their savings, suggesting that increased access to material resources may induce men to inflict more violence. Across individuals in the discussion sessions, there is an overall reduction in violence for women. Similarly, men kept reporting higher levels of violence, just as their nonparticipating counterparts, however, the increase is much lower among those in the discussion session than the levels reported by men in the control group. Thus, while there may have been a secular trend of violence increasing (perhaps due to other environmental factors), the effect of this increase is subdued among discussion- session participants. Female participants exhibit a similar reduction in the exposure to violence both immediately after the discussion sessions and after receiving their savings. Treated men, like their untreated counterparts, still report stepping up the violence; in this case, however, the increase is much lower and no longer statistically significant. This seems to suggest that while increased access to resources does encourage violence among men, partaking in activities designed to increase their awareness of the importance of negotiation encourages them to resort to violence less frequently than they would otherwise do. It is critical to note that while table 6.7 shows an overall reduced exposure to violence for women in both the intervention and control group, the selfreports may be affected directly by the program. It is possible that women

200

Radha Iyengar and Giulia Ferrari

Table 6.7

Estimated effect of discussion sessions on violence levels (Burundi) Nonparticipants

Physically hurt Insult Threaten Scream tothit > 5 . Physically hurt Percent change (relative to female baseline) Insult Percent change (relative to female baseline) Threaten . Percent change (relative to female baseline) Scream Percent change (relative to female baseline) tothit > 5 Percent change (relative to female baseline)

Participants

Females

Males

Females

A. Baseline 1.185*** (0.0320) 1.821*** (0.0731) 1.350*** (0.0575) 2.144*** (0.0759) 0.304*** (0.0336)

0.011 (0.0185) –0.217*** (0.0415) –0.131*** (0.0392) –0.131 (0.0514) –0.0685*** (0.0180)

0.015 (0.0359) –0.044 (0.0656) –0.069 (0.0504) –0.082 (0.0734) –0.0424 (0.0341)

0.000 (0.000363) 0.000 (0.000663) 0.000676 (0.0392) 0.001 (0.000740) 0.000424 (0.000342)

0.001*** (0.000233) 0.001*** 0.00152** (0.000675) 0.001** 0.000 (0.000446) 0.000 0.002*** (0.000754) 0.001*** 0.000836** (0.000302) 0.003**

–0.0385 (0.0356) –0.032 –0.0573 (0.0912) –0.031 0.005 (0.0504) –0.0685 –0.019 (0.0907) –0.009 0.00999 (0.0450) 0.033

0.000399 (0.000360) 0.000 –0.00122 (0.00148) 0.001 –0.0000432 (0.000541) 0.000 0.000191 (0.000918) 8.90858E–05 –0.000273 (0.000454) –0.001

B. Postprogram –0.09*** (0.0232) –0.08*** –0.15*** (0.000663) –0.08*** –0.03 (0.0437) –0.02 –0.22*** (0.0743) –0.10*** –0.08 (0.0300) –0.28***

Males

Notes: Each row presents the results from a separate regression with the dependent variables listed in each row. Percent changes in panel B are based on comparison to females in the control group prior to treatment. Participants refer to individuals randomly selected to attend the program that consisted of a set of six discussion sessions. Nonparticipants refer to individuals who did not attend the program. Robust standard errors clustered at the village level are reported in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

report lower reductions in experienced violence after increased awareness about domestic violence because they are more likely to categorize events as violence relative to their less- aware counterparts (Abramsky et al. 2014). This has been consistently documented over a range of other studies. Men in the control group report increases in violence across the board, significant in all cases except for threatening behavior. Men in the intervention group instead report a more mixed picture, though the changes they report are never statistically significant. In the focus groups men show a finer

Economic and Social Interventions to Reduce Intimate Partner Violence

201

understanding of the issues around domestic violence; hence, these mixed results may be interpreted as further corroborating the hypothesis that they have developed a clearer understanding of domestic violence. The reported increase may indicate an ability to distinguish improvements in one area from a worsening situation in another. The analysis of focus group data revealed two themes of discussion related to household violence. The first may be labeled “modes of violence,” as it contains words that refer to the type of violence inflicted on the women. This is mostly reported to be of a physical nature, with effects both psychological and physical on the victim. The main violent acts are (in order of importance) beating, insulting, and threatening. An important form of physical violence that differs from other areas is burning and scorching, which both men and women report. This seems to be consistent with conflict- related dynamics previously identified in the academic literature. In general, the verbs characterizing this theme are verbs of active aggression: beat, burn, and hurt. The preposition “against” is also typical of this theme, further indicating an antagonistic interaction. The juxtaposition of the language of “psychological” versus “physical” suggests that violence is not only perceived as physical, and there is a sense of what is a cause of psychological violence.14 Words that refer to feelings associated to this theme are “anger” or “mood” in relation to the man and “frustration” and “abused” associated with the woman.15 Though the focus group data illustrated an enhanced ability to identify and categorize antagonistic physical abuse, it is particularly associated to the focus group run with men that had participated in the program. This is consistent with a greater awareness of domestic violence, and in particular the ability to recognize it and describe its different facets, as well as the implications it has for the victims. This, coupled with lower—though small in magnitude—reported threatening and insulting behavior in this group, suggests the participant group may have had some initial impact on the men’s perception of what is acceptable and, hence, on their behavior. A second issue identified in the focus group is the acceptability of some forms of physical violence. Thus violence associated with aggressive behavior on the part of the man, and in particular violence initiated because of 14. In the transcripts, the word “moral” is used to qualify some types of violence. This seems to be an inaccurate translation from the French “moral.” In French, this part of speech may be both a noun and an adjective, and the noun may be both masculine and feminine, both with slightly different meanings. In the masculine acceptation it has the same meaning as the English “mood.” This latter sense seems to be the one meant by the interviewees in this context. It will thus be substituted here by “psychological,” as this adjective best captures the meaning interviewees gave it. 15. In the transcript, the word “nervous” is also found and is highly correlated with this theme. This is the other example of inaccurate translation from the French into English. In this case the original French word “nerveux” means precisely irritable, irascible or, more commonly, angry.

202

Radha Iyengar and Giulia Ferrari

changes in his mood, is considered distinct relative to violence instigated by the wife’s behavior or the general social conditions. The nature of violence is most often categorized as reasonable versus unreasonable rather than existing or not. There is a general attitude that tolerates some motivations for violence as reasonable while others are considered inappropriate or mistreatment. The unreasonable and unpredictable nature of male violence is typically associated with the language of powerlessness by the women. Consistent with this, the attitudes described are aggressiveness on the part of the man, while submissiveness and politeness are associated to the woman. 6.5.2

South Africa

The microcredit program, coupled with the discussion sessions in South Africa, was targeted at identifying harmful gender norms with the objective of reducing violence. Included in the gender issues considered were cultural norms, domestic violence, sexuality, and HIV/AIDS status in addition to broader skills such as communication, conflict resolution, solidarity, and leadership. The discussion sessions in South Africa were only for women. If successful in execution, women who participated in the discussion sessions would develop a better understanding of domestic violence and an explicit set of skills to reduce their own risk and exposure to violence. Though not explicitly targeted at other spheres of influence, the more general skills could be adapted to a range of other circumstances to more broadly increase women’s decision- making authority at home. In testing the impact of the South Africa program, we find that participation in the discussion sessions reduced experience of violence in the previous year among women compared to women in the control group by 23 percent. There is no substantial change in decisions on household spending or purchases. Attitudes toward gender norms and violence did change with substantial (nearly 50 percent) increases in willingness to request condom use. There was no significant change in attitudes toward violence. Verifying Randomization The IMAGE program in South Africa used a village- cluster design to randomize. Assuming the clustering achieved randomization, we would expect no significant difference in preprogram outcome measures. Among the baseline measures of the outcomes we investigated (presented in table 6.8), only three record differences between the discussion participant and the nonparticipant groups at or below the 5 percent significance level. One of these is from the group of indicators that measures attitudes to social norms: women in the nonparticipant group revealed themselves as more progressive than women in the discussion sessions, disagreeing on average more with the proposition that women should do all household chores. They also reported a more progressive attitude of the partner in relation to the

178 178 178 178 178 178 178 178 176 178 178 178 178 178 178 178 178 178 178 178 178 178 92

Large purchases self, ask partner Small purchases household, ask partner Medium purchases household, ask partner Large purchases household, ask partner Visit family of birth, ask partner Visit friends in the village, ask partner Visit family or friends outside village, ask partner Join credit association, ask partner Partner encouraged to participate outside household Partner asks for advice Partner keeps from friends Partner restricts contact w\family Partner insists on knowing where she is Partner controls access to health care Partner boasts g-friends Partner threatened eviction Spend own money—Ask partner

N 0.593 0.302 0.251 0.22 0.208 0.208

B. Decision making 0.225 0.419 0.781 0.415 0.426 0.496 0.152 0.36 0.354 0.48 0.601 0.491 0.236 0.426 0.101 0.302 0.506 0.501 0.416 0.494 0.888 0.317 0.927 0.261 0.854 0.354 0.899 0.302 0.938 0.241 0.944 0.231 2.935 0.248

0.163 0.899 0.067 0.051 0.045 0.045 190 190 189 190 189 189 189 190 193 193 193 193 193 193 193 193 153

193 193 193 193 193 193 0.237 0.847 0.36 0.105 0.265 0.577 0.238 0.126 0.477 0.451 0.876 0.922 0.819 0.788 0.922 0.891 2.941

0.187 0.87 0.083 0.067 0.036 0.036

Mean

N

Mean

SD

Participants

Nonparticipants

0.426 0.361 0.481 0.308 0.442 0.495 0.427 0.333 0.501 0.499 0.331 0.268 0.386 0.41 0.268 0.312 0.236

0.609 0.337 0.276 0.251 0.187 0.187

SD

–0.275 –1.636 1.296 1.326 1.853 0.474 –0.048 –0.76 0.556 –0.679 0.357 0.17 0.918 2.991 0.601 1.855 –0.199

–0.378 0.856 –0.565 –0.687 0.421 0.421

t

–0.012 –0.066 0.066 0.046 0.089 0.024 –0.002 –0.025 0.029 –0.035 0.012 0.005 0.035 0.111 0.016 0.053 –0.006

–0.024 0.028 –0.015 –0.017 0.009 0.009

Diff.

Summary

1.054 1.085 0.844 0.694 0.747 0.959 1.009 1.249 0.943 1.084 0.986 0.995 0.959 0.876 0.983 0.944 1.002 (continued)

1.145 0.968 1.23 1.332 0.807 0.807

Chng. (%)

Comparison of baseline levels of decision making, attitudes, and violence between control and program participation groups (South Africa)

A. Violence Total violence (push, hit, force sex) Insulted by partner—Past year experience Pushed by partner—Past year experience Partner hit w\fist—Past year experience Had forced sex w\partner—Past year experience Had sex for fear of reprisal—Past year experience

Table 6.8

(continued)

N 1.663 1.614 1.835 1.804 1.823 1.690 1.522 1.5 1.434 1.428

Mean

SD

Mean

C. Attitudes toward women and violence 419 1.752 0.432 424 419 1.668 0.471 425 410 1.917 1.459 413 408 1.811 0.392 413 420 1.812 0.391 423 418 1.689 0.463 422 416 1.538 0.499 423 413 1.528 0.5 418 419 1.477 0.5 424 416 1.498 0.501 423

N

Participants

Nonparticipants

0.473 0.487 0.371 0.398 0.382 0.463 0.5 0.501 0.496 0.495

SD 2.852 1.64 1.099 0.269 –0.405 –0.018 0.464 0.802 1.264 2.027

t

0.089 0.054 0.082 0.007 –0.011 –0.001 0.016 0.028 0.043 0.07

Diff.

Summary

0.949 0.968 0.957 0.996 1.006 1.00 0.99 0.982 0.971 0.953

Chng. (%)

Notes: Outcome measures are based on survey data collected by Intervention with Microfinance for Gender Equity (IMAGE) in South Africa, a cooperative study between the London School of Hygiene and Tropical Medicine (LSHTM), Wits University in Johannesburg, and the microfinance NGO Small Enterprise Foundation (SEF). Results presented in Kim et al. (2009). Estimates are based on authors’ own calculations of the data from Kim et al. (2009). Participants refer to individuals randomly selected to attend a program consisting of a set of ten discussion sessions. Nonparticipants refer to individuals who did not attend the program. Comparison columns report mean difference between participant and nonparticipant groups. Also reported are the t-statistics testing if this difference is distinguishable from zero.

Women should do all chores If paid lobola, wife must obey Wife asks condom, disrespectful Wife asks condom, sleeps around Man has g-friends, must tolerate Wife must not divorce Okay to refuse sex if not want Okay to refuse sex if no condom Okay to refuse sex if angry for other g-friends Okay to refuse sex if worried about AIDS

Table 6.8

Economic and Social Interventions to Reduce Intimate Partner Violence

205

woman’s seeking health care for herself: the partners of women in the control group are on average reported as expecting to ask for permission less often than the partners of the women in the intervention group. However, women in the control group disagree more at baseline with the proposition that the wives are entitled to refuse sex if they are worried that their partner may have AIDS. As in the case of Burundi, given the large number of outcomes considered, it is not surprising to find a few cases of statistical differences at baseline. Table 6.9 presents the comparison of sociodemographic characteristics of individuals in the IMAGE study. Of those characteristics considered, parity, access to sanitation in the house, and access to electricity differ at the 5 percent level of significance between the participants and nonparticipants groups. We also consider the degree of connectedness as a proxy for baseline levels of entrepreneurship and initiative the women display. We measured connectedness as a count of the associations the women report being a member of the baseline. Connectedness does differ significantly between the two groups at the 1 percent level. To address baseline differences, we controlled for these variables in our regressions in order to correct for these baseline differences between the two groups.

Table 6.9

Comparison of sociodemographic characteristics between control and program participation groups (South Africa) Nonparticipants

Age Marital status Parity Connectedness Maximum schooling Total asset value Non-livestock value Livestock value Type of toilet Access to electricity Dwelling walls material Access to water

Participants

Comparison

N

Mean

SD

N

Mean

SD

t

Diff.

Chng. (%)

420 420 420 426 425 412 412 413 421 422

42.519 2.15 4.417 2.178 1.386 4,265.09 3,204.97 1,057.56 2.268 1.218

12.594 1.09 2.885 1.263 0.572 7,284.84 6,115.00 3,139.42 0.485 0.413

426 426 425 428 426 421 421 422 425 425

42.077 2.277 5.009 2.874 1.458 5,245.02 3,576.25 1,664.92 2.191 1.195

10.904 1.049 2.955 1.383 0.632 9,927.12 6,991.39 5,978.12 0.405 0.397

0.545 –1.727 –2.95 –7.671 –1.739 –1.627 –0.816 –1.843 2.533 0.816

0.442 –0.127 –0.593 –0.695 –0.072 –979.929 –371.283 –607.358 0.078 0.023

0.99 1.059 1.134 1.319 1.052 1.23 1.116 1.574 0.966 0.981

422 416

4.265 2.565

1.241 1.162

425 425

4.393 3.226

1.306 1.483

–1.457 –7.204

–0.128 –0.661

1.03 1.258

Notes: Outcome measures are based on survey data collected by Intervention with Microfinance for Gender Equity (IMAGE) in South Africa, a cooperative study between the London School of Hygiene and Tropical Medicine (LSHTM), Wits University in Johannesburg, and the microfinance NGO Small Enterprise Foundation (SEF). Results presented in Kim et al. (2009). Estimates are based on authors’ own calculations of the data from Kim et al. (2009). Participants refer to individuals randomly selected to attend a program consisting of a set of ten discussion sessions. Nonparticipants refer to individuals who did not attend the program. Comparison columns report mean difference between participant and nonparticipant groups. Also reported are the t-statistics testing if this difference is distinguishable from zero.

206

Radha Iyengar and Giulia Ferrari

Decision-Making Authority Although the IMAGE intervention was targeted at gender roles and violence, the authors posited that such a directed program might have broader impacts on the role of women in the household. The hypothesis in SA-H1 was that women participating in the program would be more likely to participate in decision making in the household relative to women in control groups. The results for decision making in South Africa are presented in table 6.10. The decision- making indicators often indicate a tendency toward increased decision- making authority among discussion- session participations, but nearly all indicators are insignificant. The only indicators that mark a statistically significant positive change are the ones capturing medium purchases for the home and the degree of controls the husband exerts over the money the respondent earns. Women appear to increase in decision- making authority relative to the nonparticipant group baseline level. However, there appears to be no substantial difference in the postprogram levels in the control and treatment group. Thus, while decision- making authority does appear to improve, it is ambiguous the extent to which such change is due to the IMAGE program rather than other environmental factors. Attitudes toward Gender Norms and Violence The primary target of the IMAGE study was to affect the set of gender norms and cultural practices that facilitate violence against women. Specifically, hypothesis SA-H2 was that women participating in the IMAGE program are more likely to exhibit gender norms that are more favorable for women. Table 6.11 presents the results of the IMAGE program on attitudes toward gender norms and shows that women in the treatment group are less likely to accept gender norms biased against women. In particular, treatment group participations are 12– 15 percent more likely to reject traditional roles for women (e.g., women do all the household chores, obey if husband paid bride price). Effects on attitudes toward fidelity and sex are much larger, showing a near 50 percent change. Subsequent to participation in discussion sessions, women become less tolerant of the husbands’ other girlfriends and are more open to the possibility of divorce, compared to control women at baseline, albeit none of the latter three changes reaches statistical significance. This is consistent with women also being less likely to think that it is acceptable for a woman to refuse to have sex with her partner if he does not want to use a condom. Areas specifically targeted by the program do show substantial changes after the program. There is a 63 percent decline among participants who believe women can refuse sex if her husband refuses to wear a condom and a 50 percent decline in women believing that requesting her husband to use a condom indicates the woman is having an affair.

Economic and Social Interventions to Reduce Intimate Partner Violence Table 6.10

207

Estimated effect of discussion sessions on decision making and dispute resolution outcomes (South Africa) Baseline

Does not ask husband’s permission for: Small purchases for herself Percent change relative to nonparticipant baseline Large purchases for own self, does not ask for husband’s permission Percent change relative to nonparticipant baseline Small purchases for the hh, does not ask for husband’s permission Percent change relative to nonparticipant baseline Medium purchases for the hh, does not ask for husband’s permission Percent change relative to nonparticipant baseline Large purchases for the hh Percent change relative to nonparticipant baseline Taking children to hospital Percent change relative to nonparticipant baseline Visit family of birth Percent change relative to nonparticipant baseline Visit friends Percent change relative to nonparticipant baseline Visits family and friends outside the husband’s permission Percent change relative to nonparticipant baseline

Nonparticipants 0.359*** (0.0626) 0.228*** (0.0503) 0.672*** (0.068) 0.351*** (0.0522) 0.228*** (0.0503) 0.520*** (0.118) 0.352*** (0.097) 0.491*** (0.0854) 0.260*** (0.0598)

Postprogram

Participants

Nonparticipants

Participants

–0.067 (0.0791) –19 –0.00302 (0.09) –1 0.0495 (0.0459) 7.40 –0.0921 (0.0615) –26.24 –0.0518 (0.0781) –23 –0.0785 (0.113) –15.10 –0.0986 (0.108) –28 –0.0317 (0.0869) –6.46 –0.00703 (0.0814) –2.70

0.200* (0.0856) 56 0.143 (0.0924) 63 0.123 (0.0688) 18.30 0.240** (0.0774) 68.38** 0.0837 (0.0891) 37 0.0618 (0.171) 11.90 –0.0776 (0.136) –22 0.16 (0.118) 32.59 0.0165 (0.0874) 6.30

0.174 (0.103) 48 0.0945 (0.156) 41 0.00542 (0.0807) 0.80 0.233** (0.0928) 66.38** 0.171 (0.13) 75 0.199 (0.195) 38.30 0.314 (0.177) 89 0.124 (0.143) 25.25 0.169 (0.143) 65.00

Notes: Each row presents the results from a separate regression with the dependent variables listed in each row. Percent changes in panel B are based on comparison to females in the control group prior to treatment. Outcome measures are based on survey data collected by Intervention with Microfinance for Gender Equity (IMAGE) in South Africa, a cooperative study between the London School of Hygiene and Tropical Medicine (LSHTM), Wits University in Johannesburg, and the microfinance NGO Small Enterprise Foundation (SEF). Results presented in Kim et al. (2009). Estimates are based on authors’ own calculations of the data from Kim et al. (2009). Participants refer to individuals randomly selected to attend a ten-session discussion group series. Nonparticipants refer to individuals not selected to attend the program. Robust standard errors clustered at the village level are reported in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

The measures of the acceptability of the husband beating his wife were only measured at follow-up for IMAGE. They depict a mixed picture, with women who participated in discussion sessions becoming less willing to accept that a man beats his wife because she does not want to have sex or because she may be unfaithful; however, treated women become more likely

Percent change relative to nonparticipant baseline

Percent change relative to nonparticipant baseline Married woman can refuse sex if husband will not wear a condom

Percent change relative to nonparticipant baseline Married woman can refuse sex if does not want it

Percent change relative to nonparticipant baseline Women should never divorce their husband

Percent change relative to nonparticipant baseline A man must have many girlfriends; wife tolerate

Percent change relative to nonparticipant baseline If wife asks for condom she is unfaithful

Percent change relative to nonparticipant baseline A wife is disrespectful if asks to use a condom

Percent change relative to nonparticipant baseline Wife must obey husband who paid lobola

Woman should do most hh chores

1.596*** (0.132)

1.494*** (0.0611)

1.711*** (0.0421)

1.684*** (0.0994)

2.124*** (0.156)

2.149*** (0.201)

1.615*** (0.0250)

1.747*** (0.0224)

Nonparticipant

Baseline

Estimated effect of discussion sessions on gender norms and violence (South Africa)

Percent who disagree that

Table 6.11

–0.0766* (0.0363) –4.38* –0.0599 (0.0461) –3.7 –0.0386 (0.265) –1.8 0.0367 (0.172) 1.7 –0.0153 (0.110) –0.9 0.0120 (0.0549) 0.70 –0.0223 (0.0759) –0.01493 –0.0199 (0.145) –1.2

Participant –0.133* (0.0562) –7.61* –0.0473 (0.0969) –2.9 1.427* (0.710) 66.4 1.287** (0.501) 60.6** 0.0250 (0.0901) 1.5 –0.0465 (0.0854) –2.72 –0.0450 (0.108) –0.03012 0.874* (0.453) 54.8*

Nonparticipant

0.210*** (0.0598) 12.02*** 0.212* (0.0951) 13.1* –1.242 (0.660) –57.8 –1.093* (0.463) –51.5* 0.265 (0.240) 15.7 0.150 (0.0886) 8.77 –0.141 (0.112) –0.09438 –1.003** (0.419) –62.8**

Participant

Postprogram

1.924*** (0.181)

1.724*** (0.0367)

2.634*** (0.195)

1.830*** (0.00972)

1.282*** (0.224)

1.395*** (0.106)

–0.0579 (0.0537) –4 –0.0491 (0.255) –4 0.0368* (0.0184) 2.0* –0.477 (0.291) –18.1 0.0150 (0.0284) 0.87 –0.502 (0.354) –26.09

0.146 (0.202) 10 1.017 (0.799) 79 –0.00277 (0.00192) –0.2 0.00632 (0.00654) 0.2 –0.00257 (0.00280) –0.15 0.00536 (0.00658) 0.28

–0.311 (0.192) –22 –1.172 (0.739) –91 0.000886 (0.00183) 0.0 –0.00740 (0.00934) –0.3 0.00120 (0.00282) 0.07 –0.00184 (0.00434) –0.10

Notes: Each row presents the results from a separate regression with the dependent variables listed in each row. Percent changes in panel B are based on comparison to females in the control group prior to treatment. Outcome measures are based on survey data collected by Intervention with Microfinance for Gender Equity (IMAGE) in South Africa, a cooperative study between the London School of Hygiene and Tropical Medicine (LSHTM), Wits University in Johannesburg, and the microfinance NGO Small Enterprise Foundation (SEF). Results presented in Kim et al. (2009). Estimates based on authors’ own calculation of the data from Kim et al. (2009). Participants refer to individuals randomly selected to attend a program consisting of a set of ten discussion sessions. Nonparticipants refer to individuals who did not attend the program. Robust standard errors clustered at the village level are reported in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

Percent change relative to nonparticipant baseline

Percent change relative to nonparticipant baseline Okay to beat wife if she disagrees in public

Percent change relative to nonparticipant baseline Okay to beat wife if she is unfaithful

Percent change relative to nonparticipant baseline Okay to beat wife if she asks to use a condom

Percent change relative to nonparticipant baseline Okay to beat wife if she refuses to have sex

Percent change relative to nonparticipant baseline She worries he might have AIDS, married woman can refuse sex

If she is angry because he has other g-friends, married woman can refuse sex

210

Radha Iyengar and Giulia Ferrari

to accept that a man beats his wife because she asks him to use a condom or because she disagrees publicly with him. Overall the evidence on attitudes toward violence indicates only limited changes, many of which are not statistically significant. Exposure to Violence The primary objective of the program was to reduce women’s exposure to domestic violence. In contrast to the Burundi study, the IMAGE program was specifically targeted at reducing violence. As a result, the intervention was expected to substantially reduce violence. The specific hypothesis stated in SA-H3 was that women in the IMAGE program would be more likely to experience a reduction in the exposure to violence. Overall, the results appear consistent with this hypothesis: the IMAGE intervention reduced violence among discussion- series participants by 38 percent relative to the control group experienced at baseline, conditional on baseline values of women’s parity, connectedness, and access to drinking water and sanitation. Compared to their own level at baseline, these women experienced a 34 percent reduction in the experience of violence (see figure 6.7). Breaking down the aggregate measure in its components, we observe that women in the control group also experienced a statistically significant reduc-

Fig. 6.7 Change in violence exposure over the previous twelve months (South Africa) Notes: Outcome measures are based on survey data collected by Intervention with Microfinance for Gender Equity (IMAGE) in South Africa, a cooperative study between the London School of Hygiene and Tropical Medicine (LSHTM), Wits University in Johannesburg, and the microfinance NGO Small Enterprise Foundation (SEF). Results have been presented in Kim et al. (2009). Estimates shown here are based on authors’ own calculations from the data in Kim et al. (2009). Participants refer to individuals randomly selected to attend a program consisting of a set of ten discussion sessions. Nonparticipants refer to individuals who did not attend the program. Confidence intervals are based on village-cluster estimated standard errors.

211

Economic and Social Interventions to Reduce Intimate Partner Violence

tion in the incidence of physical assault—both in the form of their partner hitting them with a fist or an object and of her pushing or shoving them. The third component of our aggregate measure of violence—experience of sexual assault in the form of forced sex—was also reduced, albeit not to a statistically significant extent. Finally, we included among the indicators of violence the measure of insult contained in the IMAGE survey that in fact here belongs to another group of indicators, and namely those aimed at detecting controlling behavior. The results are presented in table 6.12. Much of the violence reduction seems to come from declines in physical violence. To facilitate comparison to the Burundi results, we include a measure of “insults” from the South Africa violence measurement tool. This Table 6.12

Estimated effect of discussion sessions on violence and consumption levels (South Africa) Baseline

Insult Percent change relative to nonparticipant baseline Push Percent change relative to nonparticipant baseline Has been hit with a fist by partner Percent change relative to nonparticipant baseline Has had forced sex with partner Percent change relative to nonparticipant baseline Total violence Percent change relative to nonparticipant baseline

Postprogram

Nonparticipant

Participant

0.882*** (0.0237)

–0.0176 (0.0104)

0.0239 (0.0140)

0.0309* (0.0137)

0.0915*** (0.00684)

–2 0.0112 (0.00801)

2.71 0.00612 (0.0117)

3.50* –0.0466** (0.0168)

0.0782*** (0.00532)

12.2 0.0129* (0.00550) 16.50*

6.7 0.0157 (0.00950) 20

–51** –0.0371** (0.0121) –47**

0.0720*** (0.00799)

–0.00237 (0.00435)

0.0195* (0.00833)

–0.0223 (0.0164)

0.231*** (0.0189)

–3 0.0204 (0.0147)

27.08* 0.0337* (0.0161)

–31 –0.0844*** (0.0239)

14.59*

–37***

9

Nonparticipant

Participant

Notes: Each row presents the results from a separate regression with the dependent variables listed in each row. Percent changes in panel B are based on comparison to females in the control group prior to treatment. Outcome measures are based on survey data collected by Intervention with Microfinance for Gender Equity (IMAGE) in South Africa, a cooperation between the London School of Hygiene and Tropical Medicine (LSHTM), Wits University in Johannesburg, and the microfinance NGO Small Enterprise Foundation (SEF). Results presented in Kim et al. (2009). Estimates are based on authors’ own calculations. Participants refer to individuals randomly selected to attend a ten-session discussion group series. Nonparticipants refer to individuals not selected to attend the program. Robust standard errors clustered at the village level are reported in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

212

Radha Iyengar and Giulia Ferrari

is similar to the HITS indicator that was used in Burundi. In light of this, it is important to note that the question in the IMAGE questionnaire is worded slightly differently from the HITS questionnaire, asking as it does whether the respondent has been insulted or humiliated by her partner in public, rather than more generally insulted. Hence, for the same individual, the IMAGE question would elicit responses that are only a subset of the instances captured by the HITS indicator in Burundi. There is a marginal increase in the rate of insults among participants, although there was an increase among the control group as well. This may suggest that violence shifted from physical to verbal after the intervention. It is important to note, however, that like in the case of Burundi, this may be because of differences in how the respondent categorizes violence. 6.6

Discussion

The interventions we have studied here constitute two of the first randomized evaluations of the impact of microfinance products on domestic violence in sub-Saharan Africa. Concentrating only on female users, IMAGE explores the impact of the introduction of a microfinance and training product in a new market. The Burundi VSLA investigation explores the impact of training skills on a population that has self- selected for receiving microfinance services. Both programs sought to enhance women’s decisionmaking power, reduce acceptance of gendered social norms that facilitate domestic violence against women, and reduce exposure to such violence. The emphasis of the programs, however, differed in two key dimensions: first, the IMAGE program focused only on women, while the Burundi program included both women and men. Second, the IMAGE program specifically targeted gender norms and violence, while the Burundi program discussed household economic matters and issues in an effort to highlight and challenge traditional norms. We find that both interventions had impacts on a range of desired indicators. In South Africa, the IMAGE program is associated with a substantial decline in violence, and some changes in tolerance of gender- specific norms that facilitate violence. However, IMAGE appears to have had only limited impact on enhancing the role of women in decision- making authority in the household. In Burundi, on the other hand, there were substantial changes in household decision- making authority over purchases and even fertility decisions by women, but limited impact on decision making about sex. There were marginal changes in acceptability of violence. Violence was categorized in reasonable and unreasonable dimensions consistent with existing attitudes prevalent in Burundi. There were only marginal and often insignificant changes in exposure of women to domestic violence in Burundi. The findings suggest that discussion groups in conjunction with VSLA may empower women by increasing decision- making authority over household purchases. The evidence suggests a trend toward potentially important improvements

Economic and Social Interventions to Reduce Intimate Partner Violence

213

in reducing domestic violence. Beating one’s wife, together with insulting and threatening her, seem to be the most common forms of violence. While within the relatively brief study period physical violence did not significantly decrease, insults did, which may indicate reduced levels of violence and abuse within the household in the future. Together these studies suggest a few key take- home messages: First, programs that target violence and do not incur backlash from the community may reduce exposure to violence (as in the case of South Africa). Second, in areas where it is infeasible to introduce gender- specific programming on violence, programs on economic factors may improve decision- making authority and may aid in reducing violence, albeit to a lesser degree. Third, targeted programs tend to impact the areas in which they are targeted, whether that target is violence or economic decisions. Spillover effects to related areas appear to be insignificant. As a result, policymakers should be careful in assuming that limited programs will have broad effects across areas of decision making. The central message that emerges from both studies is that long- term (~5– 10 year) prospective studies are needed to assess the real efficacy of discussion group- based interventions. Of critical importance is an assessment of whether impacts are permanent or decay over time and if periodic reinforcement helps. The quantitative evidence indicates that in Burundi the greatest change in attitudes takes place in the management and access to resources, while in South Africa it is on violence. It should be noted that these results are not directly comparable because of the different subpopulations the two interventions compare; it is, however, of interest to note that in relation to the array of results both interventions measure, it is those they targeted most directly that record the greatest impact, at least in the short term. In both cases, longer- term evaluations are warranted. Programs and evaluations should be designed to serve the purpose of assessing whether these initial results are maintained and broader effects in other areas reinforce the set of intended impacts.

References Abramsky, T., K. Devries, L. Kiss, J. Nakuti, N. Kyegombe, E. Starmann, B. Cundill, L. Francisco, D. Kaye, and T. Musuya. 2014. “Findings from the Sasa! Study: A Cluster Randomized Controlled Trial to Assess the Impact of a Community Mobilization Intervention to Prevent Violence against Women and Reduce HIV Risk in Kampala, Uganda.” BMC Medicine 12 (1): 122. Aizer, Anna. 2010. “The Gender Wage Gap and Domestic Violence.” American Economic Review 100 (4): 1847– 59. DOI:10.1257/aer.100.4.1847. Ashraf, N., D. Karlan, and W. Yin. 2006. “Tying Odysseus to the Mast: Evidence from a Commitment Savings Product in the Philippines.” Quarterly Journal of Economics 121 (2): 673– 97.

214

Radha Iyengar and Giulia Ferrari

Becker, G. 1965. “A Theory of the Allocation of Time.” Economic Journal 75 (299): 493– 51. Bozzoli, Belinda, (with Mmanto Nkotsoe). 1990. The Women of Phokeng: Consciousness, Life Strategy, and Migrancy in South Africa. Portsmouth, NH: Heinemann Educational Books, Inc. Bowlus, A., and S. Seitz. 2006. “Domestic Violence, Employment, and Divorce.” International Economic Review 47 (4): 1113– 49. Browning, M., and P. A. Chiappori. 1998. “Efficient Intra-Household Allocations: A General Characterization and Empirical Tests.” Econometrica 66 (6): 1241– 78. Duflo, E. 2003. “Grandmothers and Granddaughters: Old Age Pension and IntraHousehold Allocation in South Africa.” World Bank Economic Review 42:1– 25. Dugan, L., D. S. Nagin, and R. Rosenfeld. 1999. “Explaining the Decline in Intimate Partner Homicide: The Effects of Changing Domesticity, Women’s Status, and Domestic Violence Resources.” Homicide Studies 3 (3): 187– 214. Dutton, Donald G., and Kenneth Corvo. 2006. “Transforming a Flawed Policy: A Call to Revive Psychology and Science in Domestic Violence Research and Practice.” Aggression and Violent Behavior 11:457– 83. Farmer, A., and J. Tiefenthaler. 1997. “An Economic Analysis of Domestic Violence.” Review of Social Economy 55 (3): 337– 58. Garcia-Moreno, C., H. A. F. M. Jansen, M. Ellsberg, L. Heise, and C. Watts. 2005. “WHO Multi-Country Study on Women’s Health and Domestic Violence against Women: Initial Results on Prevalence, Health Outcomes, and Women’s Responses.” Washington, DC: World Health Organization. Hargreaves, James R., L. A. Morison, J. S. S. Gear, M. B. Makhubele, J. D. H. Porter, J. Busza, C. Watts, J. C. Kim, and P. M. Pronyk. 2007. “ ‘Hearing the Voices of the Poor’: Assigning Poverty Lines on the Basis of Local Perceptions of Poverty. A Quantitative Analysis of Qualitative Data from Participatory Wealth Ranking in Rural South Africa.” World Development 35 (2): 212– 29. Johnson, Michael P., and Kathleen J. Ferraro. 2000. “Research on Domestic Violence in the 1990s: Making Distinctions.” Journal of Marriage and the Family 62:948– 63. Kim, J., G. Ferrari, T. Abramsky, C. Watts, J. Hargreaves, L. Morison, G. Phetla, J. Porter, and P. Prony. 2009. “Assessing the Incremental Effects of Combining Economic and Health Interventions: The IMAGE Study in South Africa.” Bulletin of the World Health Organization 87 (11): 824– 32. Kim, J., C. Watts, J. Hargreaves, L. Ndhlovu, G. Phetla, L. Morison, J. Busza, J. Porter, and P. Pronyk. 2007. “Understanding the Impact of a Microfinance-Based Intervention on Women’s Empowerment and the Reduction of Intimate Partner Violence in South Africa.” American Journal of Public Health 97 (10): 1794– 802. Pollak, R. 2005. “Bargaining Power in Marriage: Earnings, Wage Rates and Household Production.” NBER Working Paper no. 11239, Cambridge, MA. Rangel, M. 2005. “Alimony Rights and Intrahousehold Allocation of Resources: Evidence from Brazil.” Economic Journal 116 (July): 627– 58. Sherin, K., J. M. Sinacore, X.-Q. Li, R. E. Zitter, and A. Shakil. 1998. “HITS: A Short Domestic Violence Screening Tool for Use in Family Practice Settings.” Family Medicine 30 (7): 508– 12. Simanowitz, A., and B. Nkuna. 1998. Participatory Wealth Ranking Operational Manual. Tzaneen, South Africa: Small Enterprise Foundation. Strauss, J., and D. Thomas. 1995. “Empirical Modeling of Household and Family Decisions.” RAND Reprint Series Paper no. 95– 12, RAND Corporation. Tauchen, Helen V., Ann Dryden Witte, and Sharon K. Long. 1985. “Domestic Violence: a Non-Random Affair.” NBER Working Paper no. 1665, Cambridge, MA.

7

Family Ties, Inheritance Rights, and Successful Poverty Alleviation Evidence from Ghana Edward Kutsoati and Randall Morck

7.1

Introduction

In much of sub-Saharan Africa, the idea of a family extends beyond its conjugal members. A lineage, or extended family, is a far larger web of relationships in which all members have a common ancestor, either male or female. One’s relationship with members of one’s extended family may be as important as, and in some cases more important than, one’s relationship with one’s spouses and children. Historically, lineages are bastions of emotional and financial support (Plateau 1991). Lineages can pay for education and training, and their social safety nets can support risk taking and entrepreneurship. However, expectations of being supported by, and of having to support, members of one’s lineage can also deter human capital

Edward Kutsoati is associate professor of economics at Tufts University. Randall Morck holds the Stephen A. Jarislowsky Distinguished Chair in Finance in the School of Business at the University of Alberta and is a research associate of the National Bureau of Economic Research. We are deeply grateful to Kofi Awusabo-Asare for coordinating our household survey and his numerous suggestions along the way. We are also grateful to the management of the Social Security and National Insurance Trust (SSNIT) Fund of Ghana for providing access to individual records; the team of research assistants, Yvonne Adjakloe, Eugene Darteh, and Kobina Esia-Donkoh for their superb field work; William Angko and Alex Larbie-Mensah for the painful task of collecting the SSNIT data; and Greatjoy Ndlovu and Panos Rigopoulos for excellent assistance with the data analysis. We thank participants at the National Bureau of Economic Research (NBER)/Africa Project meetings in Cambridge and Zanzibar for their comments and gratefully acknowledge financial support from the NBER and the SSHRC. The views expressed here, however, are solely ours and do not reflect those of the NBER or the SSNIT. For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see http://www.nber.org/chapters/c13378.ack.

215

216

Edward Kutsoati and Randall Morck

accumulation, labor supply, entrepreneurship, and risk taking.1 The actual importance of extended families in any given context is thus an empirical question. This study explores how inheritance rules in two distinct Ghanaian systems of defining extended family membership interact with formal legal inheritance rules to affect asset accumulation during marriage and the economic situations of widows and their children. Until 1985, intestate inheritances were determined by traditional custom, and this depended on how one’s extended family was defined. People whose tribal customs are matrilineal define their lineages through their female bloodlines only: their mothers and maternal cousins, aunts, uncles, grandparents, and so on are their blood kin, but their fathers and paternal cousins, aunts, uncles, grandparents, and so on are not. People whose tribal customs are patrilineal analogously define blood kinship as flowing through their paternal, but not maternal, bloodlines. Under matrilineal lineage norms a man’s children are thus not his blood kin, and his heir should he die intestate (without a legal will) is his sister’s son, his nearest blood relative in the next generation. Under patrilineal norms his estate devolves to his children, who are considered his blood kin. Under both traditional norms widows have no inheritance rights, and are often left with no assets because of a traditional presumption that assets acquired during marriage belong to the husband. Rather than relying on their husbands’ estates, they must depend on their lineages’ social safety nets. The 1985 Intestate Succession (PNDC) Law 111 was enacted to alter perceived adverse effects of these traditional norms, especially on widows with husbands of matrilineal lineage. Our surveys of widows living in villages selected for matrilineal, patrilineal, or mixed- lineage norms reveal that a quarter of a century later, Law 111 is little used and traditional inheritance norms persist. This confirms previous work (FIDA 2007; Fenrich and Higgins 2005; Scholz and Gomez 2004). We link this to a dearth of information about the formal law, lack of access to the formal judicial system, and the continued social importance of overtly adhering to traditional norms. Most low- income Ghanaians die intestate, and while some profess to know of Law 111, remarkably few make use of it. Those who know of the law have, however, built up more family assets jointly, even after controlling education level. This effect is, however, least evident for widows with husbands from matrilineal lineage traditions—the very people the reforms focused on advancing. 1. African extended families are attracting new attention in both the theoretical and empirical economics literatures. Bertrand, Mullainathan, and Miller (2003) find older relatives becoming eligible for pension payments affect adult labor supply decisions in black South African homes. Chiteji and Hamilton (2002) find transfers from richer to poorer members of African American families deter wealth accumulation more than in white families. Hoff and Sen (2006) model extended families becoming poverty traps, and Alger and Weibull (2008, 2010) show that the expectation of financial assistance from family members can prevent the development of insurance markets.

Family Ties, Inheritance Rights, and Successful Poverty Alleviation

217

We also survey widows about the extent of support from their lineages, and their access to economic (money, education, and health care) and social support. Widows who acknowledge closer ties with members of either their own or their spouse’s lineage report more support, as do those very few who made use of Law 111 or who inherited via a legal will. Intriguingly, widows of matrilineal lineage also report better economic support, consistent with maternal lineage social safety nets being more effective. Our survey targets widows living in villages, and not connected to the formal sector of the economy. While these are most representative, the inheritance practices of people in the formal sector are also of interest. Due to the low survey response rates from middle- and high- income families, we therefore complement our survey analysis with individual- level data from the Social Security and National Insurance Trust (SSNIT), the sole pension annuity program for retirees. This program provides retirees with fixed pension annuities and should they die before the annuity expiration dates, survivor benefits are payable to selected heir(s). In a second attempt to improve the economic situations of widows and their children, the 1998 Children Act 560 mandates that 60 percent of this survivor benefit pass to the decedent’s children under age eighteen, with the decedent’s choice of beneficiaries governing only the disposition of the remaining 40 percent. Due to confidentiality rules, the SSNIT allowed access only to older files. Despite this working against finding significant effects, the data provide some evidence that Act 560 benefits the nuclear families of decedents, especially those of matrilineal lineage. Together these results indicate that these formal legal reforms have a very limited impact on most Ghanaians. Specifically, they are efficacious only for people connected to the formal economy. For most Ghanaians, living in villages and dependent on the traditional economy, the reforms are either irrelevant or of only indirect help. Our study complements a growing empirical literature on the economics of the family, and on the importance of inheritance rights in developing countries. Quisumbing and Ostuka (2001) link land inheritance rights to skills acquisition decisions in Sumatra; Quisumbing et al. (2001) report that improved women’s land rights in Ghana incentivize the cultivation of tree crops, such as cocoa. Lastarria-Cornhiel (1997) link privatization to the land rights of marginalized Africans. Hacker (2010) provides a broad literature review, and discusses gender- related inheritance issues in different parts of the world. Ellul, Pagano and Panunzi (2010), in a sample of 10,004 family and nonfamily businesses across thirty- eight countries, find that strict (traditional) inheritance laws interact with weak investor protection laws to impede investment in family businesses, but not in nonfamily businesses. Where inheritance norms allow (or require) business owners to bequeath more substantial proportions of their estates to noncontrolling heirs, investors are more reluctant to provide external capital.

218

Edward Kutsoati and Randall Morck

The chapter is organized as follows: section 7.2 provides a brief background on traditional inheritance rules in matrilineal and patrilineal lineages. Section 7.3 outlines the relevant features of Law 111 and Act 560. Section 7.4 describes our data, and section 7.5 summarizes our econometric methodology and empirical results. Section 7.6 concludes. 7.2

Traditional Inheritance Rules: A Background

The inheritance rights of spouses and children depend on the form of their marriage and on their lineage traditions. Marriages in Ghana can be monogynous or polygynous, and can be ordinance marriages (legally valid civil or Christian marriages) or customary marriages as prescribed by customary tribal traditions.2 The last is by far the most popular, with up to 80 percent of marriages in contemporary Ghana entered solely under the customary system (Awusabo-Asare 1990). In practice, almost all couples marry in a traditional ceremony recognizing the new bond between the two families. Subsequently, some follow up with an ordinance marriage in a church. These are usually wealthier couples. Not all traditional marriages can be recapitulated as ordinance marriages because traditional marriages can be polygynous, while ordinance marriages cannot. While the ordinance marriage may specify inheritance rules—for example, that equal thirds might go to the decedent’s spouse, children, and extended family—customary rules take precedence in marriages that are also entered in traditional ceremonies. Under customary rules, the corpse and all property of a person who dies without having written a will (an intestate decedent) passes to the family. One’s family is customarily defined as one’s lineage: “the extended group of lineal descent of a common ancestor or ancestress” (Kludze 1983, 60). The head of the lineage appoints a “successor” to assume the estate, rights, and obligations of the decedent on behalf of the lineage. Only a legal will overrides customary law, and few Ghanaians have legal wills. The applicable customary law varies across ethnic groups, and each tribal tradition is an intricate body of rules, obligations, and norms. However, Ghana’s customary legal regimes as regards inheritances can be meaningfully divided into two broad categories: matrilineal and patrilineal traditions.

2. Islamic marriage has a “special” status, with the Quran defining marriage and inheritance rules. These let a man marry up to four women, let only men inherit certain assets, and so forth. Islamic law shapes the customary traditions of Muslim tribes, which predominate in the far north. Consensual unions, with neither an ordinance marriage nor a marriage under tribal custom, provide no inheritance rights whatsoever. Deceased common- law spouses’ property reverts to their families. Intervivos transfers to a common- law spouse are subject to legal challenge.

Family Ties, Inheritance Rights, and Successful Poverty Alleviation

Fig. 7.1

219

Regional map of Ghana depicting the ten regions

Note: The matrilineal societies are found in the southwest regions (Ashanti, Central, and Western) and parts of the Northern region. Patrilineal groups are in the southeastern (Greater Accra and Volta) and the Upper regions.

7.2.1

Matrilineal Customary Inheritance Norms

The Akans (Ashanti, Central, and Western regions) and the Lobi, Tampolese, and Baga (Northern Ghana) all use variants of matrilineal customary law. The Akans, constituting about 48 percent of Ghana’s population and the largest tribe,3 are often considered an archetypical matrilineal culture. Under matrilineal tradition, a family’s controlling spirit passes from generation to generation only through female blood lines, from whom Akan children are believed to inherit their “flesh and blood,” that is, their source 3. The Akan tribe contains subgroups defined by their mostly mutually intelligible dialects. The largest groups are the Asante, Akuapem Twi, Akyem, Brong (in the Brong-Ahafo region), Fante, and Agona.

220

Edward Kutsoati and Randall Morck

of existence (Bleeker 1966). Family ties, traced only through female ancestors, define one’s extended family, lineage, or matriclan.4 In a matrilineal tribe, one is thus related by blood to one’s mother, full siblings, and half siblings by a common mother (uterine half siblings), but not to one’s father nor to any half siblings by a common father. Thus, children belong to their mother’s lineage, but not the father’s. A traditional Akan male thus feels blood kinship to his mother’s brother (wɔfa: pron. wə-fa), but at most a weak connection to his father’s brother. An Akan male does not consider his children to be his blood kin. His closest blood relative in the next generation is his sister’s son, and this maternal nephew (wɔfase: pron. wə-fa-si ) is his presumed heir if his brothers predecease him and he dies intestate. Because Akan traditional rules revert a married couple’s acquired property to the decedent’s matrilineal extended family (Awusabo-Asare 1990), a widow and her children can be left destitute by the husband’s death. She must thus look to her brothers for support, and her children must look to their maternal uncles for bequests. The expectation of inheriting a maternal uncle’s wealth is often said to blunt an Akan nephew’s incentives to acquire human capital or seek a job, and is captured neatly in an old Akan adage wɔfa wɔho nti me nye egyu ma (lit. “I have a rich uncle; I don’t need a job”). Note that a matrilineal definition of who is, and is not, in one’s family does not imply a matriarchal power structure over that family. Chiefs and tribal leaders in matrilineal tribes are almost always male, and the leaders of matrilineally defined extended families are almost always their highest status males.5 Figure 7.2 illustrates how a matrilineal controlling spirit flows from generation to generation. The members of a matriclan all share a common female ancestor, to whom their mothers are tied by female- to-female lines of descent (shown in black). 7.2.2

Patrilineal Customary Inheritance Norms

The main patrilineal societies in Ghana are the Ga tribe (in the Greater Accra region), the Ewe tribe (in the Volta region), and the Dagomba and Nanumba tribes in the Upper East region. In a patrilineal tribe, a family’s controlling spirit passes from generation to generation only through male blood lines, and these connections define one’s extended family, or patriclan.6 Under patrilineal custom, one’s extended family thus includes one’s chil4. One’s matriclan is precisely those with whom one shares identical mitochondrial DNA. 5. Matrilineal definitions of ethnicity are not unknown in the West. For example, one is Jewish by birth only if one’s mother is Jewish. A Jewish father does not count. As with the Akan, a matrilineal definition of family did not imply matriarchal control of ancient Hebrew tribes or kingdoms. Many American aboriginal cultures also use a matrilineal definition of blood kinship—the Cherokee, Gitksan, Haida, Hopi, Iroquois, Lenape, and Navajo, among others. 6. One’s patrician is precisely male relatives with Y chromosomes identical to that of one’s father, plus their immediate children of both genders.

Family Ties, Inheritance Rights, and Successful Poverty Alleviation

Fig. 7.2

221

Matrilineal definition of blood relatives

Note: A circle represents a female and a square represents a male. One’s lineage consists of all descendants (white) of all common female ancestors through female blood lines. Children of both genders belong to their mother’s, but not their father’s, lineage. One is thus related to one’s mother but not one’s father, and to all members of one’s mother’s lineage but not to members of one’s father’s lineage.

Fig. 7.3

Patrilineal definition of blood relatives

Note: A circle represents a female and a square represents a male. One’s lineage includes all descendants (white) of common male ancestors through male blood lines. Children of both genders belong to their father’s, but not their mother’s, lineage. One is thus related to all members of one’s father’s lineage, but not to members of one’s mother’s lineage.

dren as well as one’s father, siblings, half siblings by a common father, aunts and uncles, and so on. One’s sisters and half sisters by a common father are members of one’s lineage, but their children are not. This is because they belong to that sister’s or half sister’s husband’s family. Likewise, one’s grandchildren through a son belong to one’s family, but grandchildren though a daughter belongs to their father’s family, and are thus not one’s blood relatives. In a patrilineal society, children inherit their father’s estate and widows thus look to their children for support see Ollennu (1966) for details. Figure 7.3 distinguishes members of a common patriclan (in white) from persons

222

Edward Kutsoati and Randall Morck

normally considered relatives in Western societies (in gray), who do not count as blood relatives in a patrilineal culture. 7.2.3

Criticisms of Traditional Inheritance Norms

A key difference between figures 7.2 and 7.3 is that matrilineal cultures do not number a deceased man’s widow or children among his blood relatives. Western observers thus often see patrilineal traditions as more supportive of widows and children. However, patrilineal norms also appear superficially more familiar to Western observers, who may neither understand nor appreciate the support provided by brothers, maternal half brothers, and maternal uncles in matrilineal societies. A widow with a wealthy brother in a matrilineal tribe may be much better off than a widow in a patrilineal tribe whose poor husband left her children a meager estate. Which system better provides for widows and orphans on average is thus an empirical question, and may not even be subject to broad generalization. Some communities might apply a given set of traditional norms with more generosity to widows and children than others. Publicized cases of impoverished widows and children in matrilineal tribes, buttressed by survey evidence assembled by women’s advocacy groups and Christian organizations, repeatedly made the poverty of widows’ and their children a public policy issue in the decades subsequent to Ghana’s 1957 independence. Widow- headed households throughout Ghana, but most evidently in rural matrilineal homes, were highlighted as having extreme levels of poverty—due in part, at least, to traditional inheritance norms. Intestacy law reform attracted increasing debate, but actual reform was slow to come. One reason for this deadlock was the absence of a viable reform proposal, and another was doubtless the legislators’ fear of providing fodder for tribal chauvinists. The case for reform grew to encompass several arguments. The most direct was the case for conjugal (nuclear) families retaining all or most of a deceased spouse’s assets to shield widows and their children from poverty. But the case for reform went beyond such welfare considerations. Of at least equal importance were the incentives inheritance customs created for wealth accumulation by individuals and conjugal families. Especially in matrilineal tribes, a plausible case was made that the transfer of a conjugal family’s assets to the deceased man’s maternal nephew undermines the incentives of the husband and wife to acquire skills, exert effort, and accumulate assets, and to blunt the same incentives in maternal nephews. Another problem concerns the alienability of assets passed to a lineage. No individual person owns these assets, and the conditions under which they can be bought and sold are still murky. A lineage is a corporate entity, but often lacks necessary legal titles because of the difficulties of deeding an asset to multiple owners. For example, throughout Ghana lineages own land and other assets that have no value beyond their primary use. These

Family Ties, Inheritance Rights, and Successful Poverty Alleviation

223

assets cannot serve as collateral for a loan and improvements to them are the property of the lineage, not the individual who pays for the improvements. Individuals thus have scant incentive to add to the value of such assets. Traditional inheritance systems might thus explain, at least in part, the failure of many sub-Saharan countries to formalize titles to land and capital assets (De Soto 2000). Reforms that keep assets from reverting to lineages might ultimately spread clearer property rights and thus improve allocative efficiency. It is possible to contract around these problems, but lineages must solve a collective action problem to act in concert. Individuals can nullify traditional inheritance norms with a legal will. However, most Ghanaians die intestate.7 High illiteracy rates, a lack of access to the formal legal system, and the fear of retaliation by the extended family doubtless all play a role. Males in matrilineal households can attempt to protect their wives and children with intervivos transfers, but these can be undone—either legally or by social pressure. In fact, actual monetary transfers may also go in the opposite direction: for example, the child pays money to the father. La Ferrara (2006) finds Akan (matrilineal) sons transferring more money to their fathers than do otherwise similar sons in patrilineal cultures, especially if a paternal aunt’s son resides with, or lives in the same village, as the father. La Ferrara concludes that the increased transfers from Akan sons are partially attempts to influence their fathers to direct land gifts to them, rather than to the father’s nephew. 7.3

Legal Reforms

By the mid- 1970s, a case for comprehensive reform was widely acknowledged. For example, in 1979, the Constitution of the Third Republic of Ghana proclaimed in its Article 32 (Woodman 1985): §2 No spouse may be deprived of a reasonable provision out of the estate of a spouse, whether the estate be testate or intestate. §3 Parliament shall enact such laws as are necessary to ensure that every child, whether or not born in wedlock, shall be entitled to reasonable provision out of the estate of its parents. Parliament, of course, did no such thing, and the Constitution was abolished in a military coup later that year. The military junta reiterated the two pledges, but took no immediate judicial or legislative action. On June 14, 1985, the Provisional National Defence Council (PNDC), the ruling military junta, proclaimed four interrelated reforms that, in theory at least, radically reformed the ground rules for intestate inheritances. These 7. In our survey of widows, only 8 percent reported that their spouse had a legal will.

224

Edward Kutsoati and Randall Morck

were: the Intestate Succession Law (PNDC Law 111, 1985); the Customary Marriage and Divorce (Registration) Law (PNDC Law 112, 1985); the Administration of Estates Law (PNDC Law 113, 1985); and the Head of Family Accountability Law (PNDC Law 114, 1985). All four initiatives were justified in an accompanying memorandum as reflecting “the growing importance of the nuclear family” relative to the extended family. 7.3.1

The Intestate Succession Law (PNDC Law 111, 1985)

The most important of these for our purposes, and for rebalancing customary inheritance norms against the needs of surviving members of the conjugal family, is the Intestate Succession Law (PNDC Law 111, 1985)— hereinafter Law 111. Indeed, it has been characterized as the most radical legislative reform ever made in the private law of Ghana (Woodman 1985). We therefore pause to elaborate. Although Law 111 is phrased to be gender neutral, it was seen as a victory for women and so hailed by women advocacy groups. The law allows a widow and her children—hitherto completely denied rights to the nuclear family’s assets under matrilineal norms—to be the primary beneficiaries of the deceased husband. The writ of Law 111 is restricted in two ways. First, the law applies only to property not disposed of in a legal will. Because most Ghanaians die intestate, this restriction is not thought to be of paramount importance. More importantly, Law 111 does not apply to lineage property—a concept unfamiliar to most Western observers. Much land, and other sorts of property too, belongs to a lineage, and not to any individual. Such property assigned to the deceased husband for use during his life automatically reverts to the lineage upon his death, and in a matrilineal tribe most likely passes to one or more of his maternal nephews.8 Law 111 does apply only to selfacquired property—assets the deceased, or his nuclear family, purchased or created during his life. Because the husband is typically considered its sole owner, a conjugal family’s self- acquired property virtually always reverted to a deceased husband’s lineage. A woman’s role, in whatever form, was rarely recognized. The lawmakers explicitly referred to this issue in an accompanying memorandum, which explained the reforms thus: “It is the right that 8. Lineage property, which encompasses the extended family- owned assets, is distinct from tribal property. A typical example of tribal property is the communal land “owned,” in principle, by the paramount chieftaincy (called the stool) in trust. Individuals have rights to use the land for farming or for some other commercial activity by virtue of membership of the tribe, but only with the consent from the stool (i.e., the chief ). Tribal and lineage land are essentially inalienable because of a “tragedy of the anticommons” problem. All members of the currently living generation of the lineage or tribe are considered custodians for property that also belongs to all past and future generations, and thus cannot be sold without the explicit consent of all living lineage members plus countless deceased and unborn generations of the lineage—a condition prospective buyers can be certain is never satisfied. Sagas abound of foreigners thinking they have purchased such property when they have not.

Family Ties, Inheritance Rights, and Successful Poverty Alleviation

225

the husband with whom the woman has lived and whom she has probably served, is the person on whose property she must depend after his death.” Law 111 partitions a decedent’s assets into two categories: household chattels and residue assets. Household chattels include all household belongings in regular use: clothes, furniture, appliances, a family noncommercial vehicle, farm equipment, and household livestock. All household chattels automatically devolve to the conjugal family. Residue assets include business- related and investment assets: business properties, commercial vehicles, nonprimary residential properties, bank accounts, savings, and investments. Residue assets are distributed to members of the decedent’s conjugal family and extended family according to a set of formulae set forth in sections 5– 8 and 11 (Articles 1 and 2) of Law 111. Table 7.1 summarizes these. The first row of the table sets out a baseline case, where the decedent has surviving relatives in all relevant classes. In this example, section 5 stipulates that three- sixteenths of the residual goes to the spouse, nine- sixteenths to the surviving children, and one- fourth to the lineage, that last to be split equally between the parents and those entitled to inherit under the decedent’s traditional norms. The other rows modify the baseline formula in the absence of surviving heirs of one or more sorts. For example, the second line shows that if the decedent has no living spouse—because of either a divorce or the spouse’s prior death—what would have been the spouse’s share passes instead to the children. If the decedent has neither a living spouse nor living children, what would have been their shares passes to the decedent’s lineage—with the parents receiving three- fourths of the residue property and the remaining one- fourth distributed by the lineage in accordance with its traditional Table 7.1

Residue property distribution under Intestate Succession Law (PNDC Law 111, 1985) Share of residue Law 111 assigns to:

Living conjugal and extended family members If all survive No living spouse No living children No living spouse or children No living spouse, children, or parents No surviving known relatives

Spouse

Children

Parents

Lineagea

3/16 — 1/2 — — —

9/16 3/4 — — — —

1/8 1/8 1/4 3/4 — —

1/8 1/8 1/4 1/4 1 —

Stateb 0 0 0 0 0 1

Note: The decedent’s residue property (property not classified as household chattels or lineage property) is apportioned to relatives by one of the following formulas. Residue assets include business-related and investment assets: business properties, commercial vehicles, nonprimary residential properties, bank accounts, savings, and investments. The applicable formula depends on which of the decedent’s relatives survive. a To be distributed in accordance with the traditions of the lineage. b In trust for any person subsequently identified as sufficiently close to the deceased to be a legitimate heir.

226

Edward Kutsoati and Randall Morck

norms. In the rare event of the decedent having no known relatives of any kind, the residue property goes to the state in trust, and can subsequently be disbursed to one “who was maintained by the intestate or with whom the intestate was closely identified,” should such a person be found to exist. Thus, someone who lived with, or was related in some sufficiently close way to, the decedent can seek a court order to inherit a portion or all of the estate (see Woodman 1985). Over the more than a quarter of a century since Law 111 took effect, anecdotal evidence and reports by women’s advocacy and religious groups concur that the law is not widely followed. The most likely reasons for this are a lack of information about the law, the inaccessibility of the formal legal system to many people, and a very real fear of reprisals from the lineage for violating customary laws. Many families, especially in rural areas, know only the customary laws of their tribes. Moreover, government officials in these areas are often reluctant to enforce the formal law and apply sanctions when it is violated because these same officials are often also charged by their traditional communities with upholding customary laws. The formal law is a written body of knowledge, while customary law is passed along orally, and thus more accessible to illiterate people. If legislation from Accra conflicts with tribal custom, the latter usually wins out. A 2007 study by the Ghana office of the International Federation of Women Lawyers (FIDA 2007) shows about 40 percent of survey respondents interviewed in Accra (the capital and mainly patrilineal), Kumasi (the secondlargest city and predominantly matrilineal), and Koforidua (a mixture of inheritance systems) had either no or an erroneous knowledge of Law 111, with these responses much more frequent among people with little or no education. A mere 3 percent had a complete knowledge of the law.9 The FIDA study pursues the issue with the few respondents who knew of the law. Even these find actual use of Law 111 to be restricted by multiple barriers. Widows often lack the financial resources to mount a legal challenge, and are often overwhelmed and frustrated by a cumbersome legal procedure. The widow must petition for a Letter of Administration from the courts to gain standing—and this requires the approval of the head of the decedent’s lineage, typically a contending party. In addition, she must obtain competent legal guidance to execute this document precisely in accordance with the letter of Law 111, for any procedural error nullifies her case. Added to the expense of legal advice is the cost of the decedent’s funeral and burial rituals, which the widow must pay in their entirety should she contest the customary law. These costs are easily prohibitive given the importance of 9. Recent education drives and social awareness programs are actively working to inform people of their rights under the formal legal system. Prominent among them are: the Ministry of Women and Children, the Federacion Internacion de Abogadas ([FIDA], known in Ghana as the International Federation of Women Lawyers), the Women’s Initiative for SelfEmpowerment (WISE), and Women in Law and Development in Africa (WiLDAF).

Family Ties, Inheritance Rights, and Successful Poverty Alleviation

227

elaborate funerals in Ghanaian cultures. The community typically expects a grand funeral, and this is only financially possible with the support of the decedent’s lineage. Perhaps even more daunting than all of these financial costs are the social costs a widow risks by challenging traditional norms. The repercussions from overtly disregarding deeply rooted tribal custom can be devastating. This messy and divisive process, with all its attendant costs, conflicts, and adverse consequences, can readily be avoided if the decedent left a will. But the FIDA (2007) study reports that women are unlikely to press their husbands to write a will. Indeed, a majority of interviewed widows did not know if their spouses had a will, and never discussed writing a will with him. This response from one respondent, when asked why, captures the general sentiment: I could never ask [my husband] if he had a will or not. . . . If I asked, he may even think I am planning to kill him so I can take his assets; or accuse me of being a witch or something bad. He may even ask for a divorce. 7.3.2

Survivors’ (Pension) Benefits under the Children’s Act 560, 1998

A second comprehensive reform to Ghanaian inheritance laws developed in stages, and provides a wealth of government data pertaining to middleand high- income Ghanaian households, whose survey participation rates tend to be very low in any event. The Social Security and National Insurance Trust (SSNIT) runs the sole government- sponsored pension annuity program for retirees. Should the contributor die before his accumulated benefits are fully disbursed, 60 percent of the remaining benefits pass to the decedent’s children under age eighteen. Each contributor apportioned the remaining 40 percent to one or more chosen heirs. This reflects a sequence of reforms, but primarily the Social Security Act (PNDC Law 247, 1991) and the Children’s Act (Act 560, 1998), hereinafter Law 247 and Act 560, respectively. Law 247 designates the permissible choices open to members of patrilineal versus matrilineal cultures, and is summarized in table 7.2. Thus, Law 247 prohibits a member of a patrilineal culture from listing a sister’s son as an heir, and forbids a member of a matrilineal culture from listing a father’s father or father’s brother as an heir. But in both cases, one has the option of either adhering to the traditional norms of one’s lineage or bequeathing benefits to one’s conjugal family. A pension contributor’s choice of heirs is confidential, buried in government files, and available to interested parties only after the contributor’s death. This theoretically lets one defy customary inheritance norms by bequeathing one’s accrued pension wealth to one’s conjugal family, not one’s lineage, with no one to know until well after one is safely dead. The social security system dates back (at least) to 1946, when Chapter

228 Table 7.2

Edward Kutsoati and Randall Morck Schedule 45 of the Social Security Act (PNDC Law 247, 1991) Decedent’s traditional norm

Patrilineal

Matrilineal

Mother, father Husband Wife, son, daughter Brother, sister Father’s father Mother’s mother Father’s brother Mother’s brother Mother’s sister Sister’s son Sister’s daughter

Allowed Allowed Allowed Allowed Allowed Prohibited Allowed Prohibited Prohibited Prohibited Prohibited

Allowed Allowed Allowed Allowed Prohibited Allowed Prohibited Allowed Allowed Allowed Allowed

Note: Under Law 247, only certain persons are eligible to be listed as heirs to a deceased beneficiary’s SSNIT pension accruals. Different choice sets are offered to members of matrilineal versus patrilineal tribes.

30 of the Pension Ordinance of 1946 provided government pensions for certain public sector employees, a scheme that became known as CAP30. A more general social security system began with the military government of the time, the National Redemption Council, decreeing (NRCD 127, 1972) the expansion of a previous Parliamentary Act 279 to establish a Provident Fund to pay every formal sector worker a lump sum upon retirement. In 1991, another military government, the Provisional National Defence Council proclaimed the Social Security Act (PNDC Law 247, 1991), hereinafter Law 247, under which the Social Security and National Insurance Trust (SSNIT) is made the sole government- sponsored pension system. The system resembles standard Western pension systems in some ways, but deviates markedly from them in others. Our data cover much of the period when Law 247 was in effect.10 Like many government pension schemes, the SSNIT is a pay- as-you- go system: formal- sector workers’ current contributions fund benefits paid to pensioners. Law 247 requires that all employers contribute 17.5 percent of their base salaries to the fund. This appears as a 5 percent deduction from the employee’s monthly pay check, and is matched by a 12.5 percent employer contribution invisible to the employee. Act 766 now allows the self- employed to join the system by making the full 17.5 percent contribution. Also like many other systems, the SSNIT is a defined benefit system. The minimum annual pension benefit is 50 percent of the average of the contributor’s highest three annual salaries earned in the twenty years prior 10. A new National Pensions Act (Act 766) went into effect in January 2010 and expands the scheme under Law 247 to include voluntary contributions from self- employed persons and individuals in the informal sector. Social Security for the informal sector will be administered by the SSNIT Informal Sector Pension Fund.

Family Ties, Inheritance Rights, and Successful Poverty Alleviation

229

to retirement. The benefit rises by 1.5 percent of average of that average for each additional year of work and contributions to the fund. So, theoretically, anyone who retires at age sixty, having contributed for forty years to the fund, merits a pension equal to 80 percent of average earnings in the retiree’s three most prosperous years.11 The major difference between most other pension systems and the SSNIT it its fixed term: paid-up contributors are entitled to exactly 144 consecutive monthly benefit payments in retirement, no more and no fewer. Under Law 247, an individual becomes entitled to full- pension benefits after contributing to the scheme for 240 months and attaining the age of sixty, the mandatory retirement age.12 The retiree then receives a monthly pension for the next twelve years. When the retiree turns seventy- two, the benefit terminates and the retiree must rely on relatives or savings. Individuals may opt out of this default scenario and receive 25 percent of the payments’ present value as a lump sum upon retirement, and the other 75 percent as monthly payments over the next twelve years. If the contributor dies before age seventy- two, the present value of the remaining payments the contributor would have received, computed at the treasury rate over the same period, was then paid to the heirs the contributor chose from the options made available in table 7.2.13 The largest subsequent change to this system, the Children’s Act (Act 560, 1998) of 1998, mandates that the SSNIT pay 60 percent of such survivor benefits to the contributor’s minor children (age eighteen or younger). In theory, whenever the SSNIT receives a procedurally complete claim, with all necessary supporting documents (e.g., death and burial certificates), it should investigate the family to ascertain whether or not the decedent has minor children not listed as beneficiaries. In practice, the SSNIT lacks the resources to do this, and merely ascertains the validity of the submitted claim. The remaining 40 percent of the survivor benefit is then disbursed to the beneficiaries the decedent selected from the appropriate column in table 7.2, in the selected proportions. Another major difference is that Ghana’s government- sponsored pension program was, until very recently, restricted to formal- sector employees. Workers in the informal sector—subsistence agriculture, fishing, roadside stands, and so forth—are not covered.14 Some 80 percent of Ghanaians work 11. Early retirement, starting from age fifty- five, with reduced pension payments is permitted under Law 127. Individuals who retire before age sixty lose 7.5 percent of their age sixty retirement benefit for each year until their sixtieth birthday. People in high- risk occupations, such as mining, are exempted and can retire at fifty- five years with full pension. 12. If a contributor falls short of the required 240 months, the total contributions plus interest at half the T-Bill rate is refunded as a lump sum at retirement. 13. Benefits are discounted at the lower of the prevailing Treasury bill rate and ten. 14. Recent reforms (Act 766, 2008) mandate that the SSNIT organize the Informal Sector Social Security Fund, which actively encourages informal- sector workers to sign up and save for their retirement.

230

Edward Kutsoati and Randall Morck

in the informal sector—subsistence agriculture, fishing, roadside stands, petty trading, and the like (Heintz 2005). The SSNIT data thus pertain only to middle- and upper- income Ghanaians. We therefore use both SSNIT data and surveys of inheritance patterns among low- income Ghanaians to gain a broader picture of the current situation. 7.4 7.4.1

Data and Empirical Strategy Survey Data Descriptive Analysis

We surveyed 322 widows living in four villages in Southern Ghana: Bortianor and Ingleshi Amanfro are both predominantly patrilineal, and in the Greater Accra Region; Abura Dunkwa and Nyankomase Ahenkro are in the Central Region and most lineages there are matrilineal. Potential survey participants were identified with the help of a town or village council member, a town leader, or the traditional chief, whose approval was sought for our team of researchers to conduct the survey. Households were randomly chosen, and the questionnaires were administered to an adult person in the house, in private. Because visits to randomly selected, seemingly more affluent households in the urban areas generally yielded no responses, our final data consist mostly of very low- income respondents, though not the poorest of the poor living rough. About 50 percent of our respondents reported no formal education, and 25 percent reported the same for their spouses. The average is five years of formal education. Only 5.2 percent reported a bank account (single or joint), though 26 percent have other personal assets. Nonetheless, over half reported contributing directly to the conjugal family’s wealth. The survey data contain information to assemble a profile of each respondent’s age, education level, inheritance system, years married, children (by spouse), minor children, and marriage type (polygynous or monogynous). Our youngest respondent was age twenty and our oldest was age ninetythree. Consistent with the ethnic distribution in Ghana, about 48 percent reported their traditional custom as matrilineal. Respondents also provided information to let us assemble similar profiles of their spouses. Spouses’ profiles appear similar, though they are slightly more educated—about eight years of formal education. A key part of our analysis is to gauge how respondents’ information about the law shaped the ownership structure of the family assets (i.e., individually owned or jointly), which law was applied in the distribution of the assets when a spouse died, and the welfare of the family after the death. We therefore asked widows an additional set of questions. Nearly 47 percent report prior knowledge of the law at their marriage or before the death of their spouse, but only 3.2 percent of widows reported its use in their cases. Rather, just over 7.4 percent reported their deceased spouse having a will and over

Family Ties, Inheritance Rights, and Successful Poverty Alleviation

231

75 percent reported the estate having been distributed in accordance with traditional norms. The remaining 14 percent did not know which rule was employed, or reported using a religious- based principle to divide up the estate (e.g., Muslim). Almost 46 percent reported being dissatisfied with the distribution. To assess widows’ economic and social status, we asked respondents to compare their situations in the years immediately before versus after their spouse’s death. Because they virtually all dwell entirely in the informal economy, questions about monetary income do not capture their economic situations. We therefore asked them to rate their economic situations or opportunities, defined as access to financial services (formal or informal), health care for themselves and their children, and educational opportunities for their children. Access to financial services in this context means an expectation of being able to borrow money in a pinch from a financial institution, or from the head of the lineage or its more prosperous members. Health care and education typically entail small informal monetary outlays. Thus, obtaining needed medical care and clothing and provisioning children for school require cash outlays that, in a subsistence economy, typically require economic support from one’s lineage. We are also interested in whether or not widows, if they attain better financial situations than traditional law prescribes, encounter tension with their lineages, or those of their husbands. We therefore also ask each widow to compare the quality of her relationships with her spouse’s extended family, before versus after his death. Finally, we are interested in how divergent formal and customary laws affect welfare. We therefore asked widows questions about their economic positions to assess how well their traditional support networks performed. Here, quantitative assessment is tricky. Previous work (e.g., Awusabo-Asare 1990) finds that poor villagers, such as those we interviewed, tend to report their economic situation as very bad, leaving little variation to study. It is therefore necessary to introduce a norm to which they can compare themselves. However, a common norm is inappropriate because different lineages have markedly different capacities to help their less fortunate members. Some lineages include government officials, police, formal- sector workers, émigrés, or others, whose formal or informal income can be redistributed to needy relations. Some lineages belong to tribes with mineral rights, whose revenue streams can also be redistributed from chiefs to lineage heads and then to needy widows. However, most lineages have few or no such resources, and genuinely cannot provide more than minimal subsistence support. To extricate information about how well traditional lineage support systems function, we asked each widow to consider the financial capacity of the lineage in question to support people such as herself. With this benchmark in mind she was then asked if she received any lineage support, and if so whether this was less than, about in line with, or more than that capacity

232

Edward Kutsoati and Randall Morck

Table 7.3

Marriage patterns: Within versus cross-lineage marriages Spouses’ traditional lineage Matrilineal

Respondents’ traditional lineage

Matrilineal Patrilineal Total

145 93.5% 10 6.5% 155

88.4% 6.3% 48.1%

Patrilineal 19 11.4% 148 88.6% 167

11.6% 93.7% 51.9%

Total 164 50.9% 158 49.1% 322

Note: Numbers in each cell are numbers of respondents of horizontal lineage classification whose spouses are in vertical lineage classification. Percentages to the right of these numbers are percentages of respondents, percentages below each number are percentages of spouses.

allowed. A study of “mixed marriages,” in which one spouse has a patrilineal tradition and another has a matrilineal tradition, would have been especially instructive here. Unfortunately, interlineage marriages within our sample are extremely rare, as shown in table 7.3. Roughly nine in ten Ghanaians marry within their lineage tradition. Studying interlineage marriages would be interesting, but table 7.3 reveals our sample of these to be very small. Table 7.4 highlights differences in means by respondents’ inheritance options. Panel A compares widows who know of Law 111 to respondents who do not, while panel B compares by the deceased spouse’s lineage type. Unsurprisingly, panel A shows that knowledge of Law 111 correlates positively with education for both the respondents and spouses, and the educated are more likely to be in monogamous marriages. Confirming the importance of financial incentives, knowledge of Law 111 is also greater among respondents who report contributing more wealth to their conjugal families. Widows who had knowledge of the law were significantly more likely to have settled their deceased spouse’s estate under it. In fact, no one that indicated no knowledge of the law prior to their spouse’s death made use of the law in dealing with the estate, suggesting that proponents of the Law’s greater usage might consider more energetically distributing information at the time of a spouse’s death or serious illness. More surprisingly, though consistently with previous studies such as FIDA (2007), panel A shows that about 70 percent of respondents with a prior knowledge of Law 111 nonetheless settled their deceased spouses’ estates in accordance with traditional lineage or religious customs, a figure that is only slightly less than that for respondents with no prior knowledge of Law 111. Presumably, much or all of the estates passed to the lineage in both cases. Knowledge of the formal law correlates with a better economic situation in widowhood overall. However, it has little traction in explaining the relative economic and emotional support widows receive from their lineages versus those of their husbands. Widows knowledgeable of the law obtain insignificantly better financial support both from their own and spouses’

Table 7.4

Mean survey responses by knowledge of Law 111

A. Mean survey responses by widows’ knowledge of Law 111, which extends limited formal legal inheritance rights to conjugal families regardless of customary law Knowledge of Law 111 Yes Respondent profile Age (years) Education (yrs. of formal educ.) Monogamy at marriage (% of widows) Monogamy at spouse’s death (% of widows) Information on spouse Years married Spouse education (yrs. of formal educ.) Customary/Islamic marriage (%) No. of children (with spouse) —age < 18 (at spouse’s death) Ownership of assets Joint account w/spouse (% of widows) Percent contributed to family assets Personal assets (% of widows) Intervivos from spouse: As per ability or better (%) Distribution of assets Will (%) PNDC Law Customary/Islamic Percent dissatisfied with distribution Welfare after spouse’s death Economic situation (% worse off) Relationship with in-law (% worse) Lineage financial support Own: As per ability or better Spouse’s: As per ability or better Difference (own minus spouse’s) Lineage emotional support Own: As per ability or better Spouse’s: As per ability or better Difference (own minus spouse’s)

No

Difference

52.9 5.1 82.4 79.1

56.6 2.8 65.5 66.8

–3.7** 2.3*** 16.9*** 12.2**

23.8 8.0 82.6 4.6 1.8

23.9 5.5 85.0 4.8 1.9

0.1 2.5*** –2.4 –0.2 –0.1

4.7 27.6 29.7 34.2

5.8 18.4 23.8 27.1

–1.1 9.2*** 5.9 7.1*

12.8 6.1 67.1 36.2

2.9 0.6 87.3 47.9

9.8*** 5.5*** –20.2*** –11.7**

55.7 22.8

69.9 16.7

–14.2*** 6.1*

31.5 14.1 17.4***

31.2 12.1 19.1***

0.3 2.0

49.7 25.5 24.2***

52.6 27.2 25.4***

–3.9 –1.7

B. Mean survey responses by customary law applicable to deceased spouse. The customary law of the deceased husband’s lineage determines her inheritance rights Spouse inheritance custom Matrilineal Respondent profile Age (years) Education (yrs. of formal educ.) Monogamy at marriage (% of widows) Monogamy at spouse’s death (% of widows)

54.6 5.1 79.1 80.4

Patrilineal 55.1 2.8 68.1 65.3

Difference 0.5 2.3*** 11.0** 15.1*** (continued)

234 Table 7.4

Edward Kutsoati and Randall Morck (continued) Spouse inheritance custom Matrilineal

Information on spouse Years married Spouse education (yrs. of formal educ.) Customary/Islamic marriage (%) No. of children (with spouse) – age < 18 (at spouse’s death) Ownership of assets Joint account w/spouse (% of widows) Percent contributed to family assets Personal assets (% of widows) Intervivos from spouse: As per ability or better (%) Distribution of assets Will (%) PNDC Law Customary/Islamic Percent dissatisfied with distribution Welfare after spouse’s death Economic situation (% worse off) Relationship with in-laws (% worse) Lineage financial support Own: As per ability or better Spouse’s: As per ability or better Difference (own minus spouse’s) Lineage emotional support Own: As per ability or better Spouse’s: As per ability or better Difference (own minus spouse’s)

Patrilineal

Difference

24.5 7.5 85.1 4.2 2.1

23.2 5.8 82.6 5.3 1.6

1.3 1.7*** 2.5 –1.1 0.5***

3.9 26.1 34.4 28.4

6.6 19.4 19.3 32.3

–2.7 6.6*** 15.1** –3.9

9.7 5.2 70.7 45.1

5.4 1.2 85.4 40.1

4.3* 4.0** –14.7*** –5.0

50.3 27.7

75.4 12.0

–25.1*** 15.7***

31.6 13.5 18.1***

31.1 12.5 18.6***

0.5 1.0

58.1 31.0 27.1***

44.9 22.2 22.7***

13.2** 8.8*

***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

lineages than do widows unfamiliar with the formal law. Indeed, widows unfamiliar with the law actually report insignificantly better financial and emotional support from their spouse’s lineage, relative to their own. Across the board, widows’ own lineages provide more support. Panel B repeats the exercise, but partitions the data by respondents’ spouses’ inheritance tradition. Knowledge of the law is substantially greater among widows whose husbands were from a matrilineal tradition, suggesting women’s and various organizations have successfully reached more of those that the law was specifically intended to empower. However, panel B also reveals both matrilineal deceased spouses and their widows to be better educated, perhaps also partially explaining their better familiarity with the formal law. Matrilineal spouses’ widows are also more likely to have been in monogynous marriages; a situation that presumably improved their im-

Family Ties, Inheritance Rights, and Successful Poverty Alleviation

235

plicit bargaining power with their husband and his family. Matrilineal widows reported having fewer children with their deceased husband, although a slightly larger number of their children were less than eighteen years old at the time she became a widow. Widows of matrilineal spouses, those least likely to inherit a deceased spouse’s assets under customary law, nonetheless report having contributed more to the conjugal family’s wealth, and also having accumulated more personal assets. This might explain their small, though insignificantly, greater dissatisfaction with the distribution of those assets after their husbands’ deaths. It seems knowledge of Law 111 mitigates any adverse effects of husbands’ matriliny on their widows. Matrilineal men’s widows knowledgeable about Law 111 report less dissatisfaction with the distribution of the conjugal family’s assets (36 percent vs. 48 percent were dissatisfied). Perhaps knowledge of the law strengthens their bargaining power within the traditional inheritance process. Finally, widows of matrilineal men report insignificantly better financial support both from both their own and their deceased husbands’ lineages than do widows of patrilineal men. Widows of matrilineal men also report better emotional support from both lineages. Recall from table 7.4 that 93 percent of matrilineal men’s widows are themselves from matrilineal cultures, so strong support from the widow’s lineage is unsurprising. But the finding that their deceased husbands’ matrilineal lineages also support them, even though their children from such a marriage are not considered theirs, suggests that matrilineal lineages in general provide unexpectedly strong traditional safety nets for widows. 7.4.2

Pension Bequests Data Descriptive Analysis

To complement our survey data, which cover very low- income Ghanaians, we utilize official data on individuals’ bequest instructions regarding their Social Security and National Insurance Trust (SSNIT) benefits. These data pertain only to Ghanaians with employment in the formal sector— persons considered to be of middle to high socioeconomic status. Hardcopy records of each beneficiary’s instructions are retained by the SSNIT, and are considered confidential until the contributor’s death. Thereafter, the record is opened so that interested parties can learn of their rights, if any, to the deceased contributors remaining benefits. Because of this confidentiality requirement, we were only allowed access to the bequeathal instructions of deceased contributors whose residual pensions had been disbursed, and whose files were closed. Names, addresses, and other information that might identify contributors or their relatives were withheld. Our total sample of SSNIT data consists of records of 860 contributors who passed away between 1992 (when Law 247 came into effect) and 2006. The median age at death is fifty- four years (mean 52.5 years), which

236

Edward Kutsoati and Randall Morck

is lower than the mandatory retirement age of sixty years. About 70 percent of our contributors were married at death and only 10 percent are women, reflecting the overwhelming predominance of men in the formal sector and a corresponding slighter predominance of women in the far larger informal sector. The data the SSNIT made available also contain more observations from later years. We were told this is because many older files are incomplete, missing much critical information, and less readily accessible. Summary statistics for the variables we construct from these records are reported in table 7.5. Each record sets out the contributor’s pension bequeathal decisions and identifies tribal background, from which we can infer the traditional inheritance custom. Each record also lists the contributor’s marital status and most, though not all, provide the average of the Table 7.5

Summary statistics of pension bequest data

Sample Age at death All decedents Male decedents Female decedents Married (fraction) All decedents Male decedents No. heirs listed All decedents Married decedents Male decedents Married male decedents If claims adjusted by Act 560a Bequest to nuclear family (%) All decedents Male decedents Married male decedents Pre-Act 560 decedentsb Post-Act 560 decedents Act 560 audited decedents Benefits paid/claim (2006/07)c

All (1)

Matrilineal (2)

Patrilineal (3)

Difference (t-stat) (2) minus (3)

52.5 [860] 53.12 [761] 47.27 [99]

52.13 [421] 53.14 [357] 46.56 [64]

52.75 [439] 53.1 [404] 48.57 [35]

–0.62 (0.82) 0.02 (0.20) –2.01 (–0.83)

69.7 [600] 71.9

69.0 [290] 72.5

71.0 [310] 71.3

–2.0 (0.55) 1.3 (–0.39)

2.64 2.94 2.67 2.96 [547] 4.91 [250]

2.85 3.21 2.96 3.29 [259] 5.06 [123]

2.43 2.68 2.41 2.65 [288] 4.76 [127]

0.422 (3.25)*** 0.53 (3.19)*** 0.55 (3.79)*** 0.64 (3.63)*** 0.30 (–1.04)

58.4 58.2 73.3 51.8 [348] 62.9 [512] 0.547 4,500 [85]

59.8 60.5 74.1 55.7 [168] 62.5 [253] 0.564 5,279 [45]

57.1 56.2 72.7 48.2 [180] 0.633 [259] 0.530 3,622 [40]

2.7 (0.85) 4.3 (1.30) 1.4 (0.43) 7.5 (1.52) –0.008 (–0.21) 0.034 (0.575) 1,657 (–1.38)

Notes: Means of key variables for all files, files of matrilineal decedents, and files of patrilineal decedents. Numbers in square brackets are sample sizes. Final column contains difference between matrilineal and patrilineal mean, with t-statistic for the difference being significantly different from zero in parentheses. a Length of time (in years) from death of contributor to completion of disbursement of survivor benefits. b Act 560, passed in 1998, altered the permissible distribution of survivor benefits. c In Ghanaian cedis per claim. The exchange rate in 2006/2007 was approximately GH¢1 = US$1. ***Significance at the 1 percent level. **Significance at the 5 percent level. *Significance at the 10 percent level.

Family Ties, Inheritance Rights, and Successful Poverty Alleviation

237

contributor’s best three annual incomes from among the twenty years prior to the contributor’s death. The 46.5 percent of contributors reporting tribal affiliations that imply matrilineal inheritance norms aligns well with an estimated 48 percent for the national average. Thus, 46.5 percent of contributors chose from the list of permissible heirs under the matrilineal heading in table 7.2, and the remaining 53.5 percent selected heirs from the list under the patrilineal heading. Recall from section 7.3 that the purpose of this restriction is to restrict the contributor to leaving residual benefits to the conjugal family or the traditional lineage. Bequests of pension benefits to others—for example, persons not belonging to the conjugal family or traditional lineage—are proscribed. Thus, the SSNIT does not permit a contributor from a patrilineal tribe to list a maternal uncle as an heir. Should a contributor attempt this, the list would be rejected. In private conversations with SSNIT staff, we were told that the SSNIT cannot enforce this rule completely. In practice, a mislabeled maternal uncle might become a heir. The SSNIT officials informed us that they simply lack the resources to thoroughly investigate each list of beneficiaries, and in the absence of a challenge from other relatives and if the claims are procedurally valid, simply distribute remaining funds to the pension recipients’ selected beneficiaries without further investigation. The record of each contributor’s bequeathal decision lists the chosen beneficiaries, the fraction of the total benefits bequeathed to each, and the relationship of each to the contributor. Upon the death of a contributor, the SSNIT takes no action. Potential heirs must submit claims for survivors’ benefits after a qualified contributor dies. The SSNIT staff informed us in private conversations that substantial benefits go unclaimed because heirs are unaware the benefits exist, and because people who learn they are not listed beneficiaries often fail to inform those who are of their rights. Unsurprisingly, contributors who were married at death list more beneficiaries: 2.94 versus 1.94 for single contributors. Male contributors list more beneficiaries than do female contributors, and contributors from matrilineal tribes list a slightly larger average number of beneficiaries (2.85) than do those from patrilineal tribes (2.43). Finally, each record provides the total value of survivor benefits paid out. These can be substantial by Ghanaian standards: the median for the 319 records closed in the 2006– 2007 fiscal year, when the Ghanaian cedi was at GH¢1 = US$1, was ¢2,142; the mean was ¢4,500; and the standard deviation was ¢5,539.15 The bequests were thus typically four to over seven times more than Ghana’s gross domestic product (GDP) per capita, which then stood at only about US$500. Also recall from section 7.2 that the 1998 Children’s Act 560 altered the 15. Not all 319 passed in that year; some died earlier, but survivor- benefits were not disbursed until 2006.

238

Edward Kutsoati and Randall Morck

permitted distribution of survivors’ benefits by the SSNIT. Prior to this, the contributor’s list of beneficiaries determined the distribution of all remaining benefits, but afterward 60 percent of the total benefits must go to the contributor’s child(ren) under eighteen years of age, regardless of whether they are listed as beneficiaries or not, and the remaining 40 percent is distributed in accordance with the contributor’s list of beneficiaries. The SSNIT theoretically investigates each claim to uncover other minor children, including illegitimate children, though in practice resource constraints limit this. One result of this is an increase in the number of beneficiaries in records closed after 1998 because the SSNIT adds the names of minor children to these. Thus, files closed under Act 560 named an average of 4.91 beneficiaries each, while those closed prior to 1998 named an average of 2.61 beneficiaries each. This Act thus substantially shifted SSNIT survivor benefits away from what contributors initially intended and toward their own children. This proves useful in the econometric analysis below. 7.5

Methodology and Econometric Results

We now examine econometrically the impacts of these two reforms. Our goal is to estimate the extent to which tribal inheritance norms shape economic outcomes of those on the margins of Ghanaian society, and for the case of retirees, how the reforms influence private, end- of-life bequest decisions. We begin the analysis with individuals’ bequest decisions about their unexpired pensions. Because we are interested in the status of widows, we first focus on the 90 percent of our SSNIT records that are males. 7.5.1

Pension Bequest Decisions and the Children’s Act (Act 560, 1998)

The 1998 Child Act 560 sought to improve the status of widows and their children by instructing the SSNIT to adjust survivors’ benefits so that at least 60 percent of unexpired benefits passes to his minor children, regardless of the contributor’s instructions. Ideally, we would also like to investigate how Act 560 altered contributors’ bequest decisions. In 1998, the changes were widely publicized, and contributors were urged to alter their bequests to accord with the new rules. No doubt many contributors ignored this advice, and left their original instructions in place. Unfortunately, our data on closed SSNIT files include too few decedents whose initial instructions are dated after 1998 to allow statistically meaningful analysis. We therefore contrast the recorded bequest decisions of contributors who died after 1999, some of which were changed to reflect the new rules, against the decisions of contributors who died earlier, and who thus felt no pressure to alter their instructions to the SSNIT. We take 1999, rather than 1998, as our transition year to ensure that contributors had sufficient opportunity to react to the rule change.

Family Ties, Inheritance Rights, and Successful Poverty Alleviation

239

To determine whether the new rule altered the number of heirs contributors listed in their SSNIT files, we estimate regressions of the form (1)

No. of heirs = X i·*a + bo*matrilineal + b1*postAct560 + b2*postAct560*matrilineal + b3*matrilineal*X i + b4* ln (age)i + e,

where postAct560 is an indicator variable set to one if the contributor’s death occurred after 1999, and zero otherwise; log(age) is the logarithm of the age at which he died, and the Xi are individual characteristics including marital status and an indicator variable for the contributor being among the top 25 percent in total unexpired pension (Top25Pension) in the cohort who died in the same year. We interpret this indicator variable as a proxy for the contributor’s total wealth, which is unavailable. Finally, we control for age, the contributor’s age at death. Table 7.6 reports estimated parameters for equation (1). The 2.67 grand mean summarizes a statistically significant increase from a bit over 2.5 heirs per contributor prior to 1999 to just below three thereafter. The typical contributor dying after 1999 thus names more heirs as beneficiaries, regardless of customary law. The positive significant coefficient on ln(age) indicates that older men also name more heirs in the bequests decisions. While this might be explained by older men having had time to sire larger families,

Table 7.6 Sample

Number of heirs listed in SSNIT records All males

Matrilineal 0.086 Post-Act 560 0.100 Married 0.223 Top 25 0.127 Post-Act 560 * matrilineal –0.038 Married * matrilineal 0.122 Top 25 * matrilineal 0.017 log(age at death) 0.572 Intercept –1.842 R2 No. of obs.

Married males

(0.90) 0.120 (1.66*) 0.098 (3.40***) — (1.68*) 0.158 (–0.42) 0.087 (1.25) — (0.16) 0.433 (5.85***) 0.593 (–5.03***) –1.707 0.13 761

(1.26) (1.25)

Married females

–0.332 –0.577 — (1.54) –0.967 (0.74) 0.240 — (0.31) –0.009 (4.28***) 1.034 (–3.16***) –2.686 0.08 547

(–1.46) (–2.75***) (–0.47) (0.78) (–0.03) (3.04***) (–1.95*) 0.26 53

Note: Dependent variable is the logarithm of the number of heirs listed by the deceased male contributor to the SSNIT pension system. Reported values are OLS estimates, and numbers in parentheses are t-statistics. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

240

Edward Kutsoati and Randall Morck

it could also be due to older men being more conscientious about their legacies. We include all male contributor files in the above analysis because many males who declare themselves unmarried nonetheless report children. These may be widowers, men who interpret the question as pertaining only to formally recorded marriages, men with illegitimate children, or the SSNIT may not update this field when updating data about children. Reestimating the regressions using males who report themselves as married in their SSNIT files generates broadly similar results, though the post-Act 560 indicator variable now fails to attain significance in the smaller sample. For completeness, we also estimate the regression for female contributors (not shown) and for married female contributors. Like males, older female contributors have more heirs, but unlike their male coworkers, females list markedly fewer heirs after Act 560 than before it. The very small sample makes this result somewhat uncertain, despite its statistical significance. Conceding the considerable limitations of our data and methodology, we infer that the data are not inconsistent with a discernable difference between bequest decisions before and after the Act. The purpose of the Act was to induce contributors to provide more fully for their conjugal families and to divert bequests away from their lineages. To quantify its effectiveness, we calculate the total fractions of unexpired pension contributors’ bequests passed to different categories of relatives, implicitly assigning zero to unmentioned relations. Again, data limitations necessitate caveats. We need not have complete information about each contributor’s family. For example, no children listed as beneficiaries means the contributor made no provisions for children, not that he had none. He may have neglected to update his record as his family grew, or he might have deliberately omitted his children. We partition each contributor’s heirs into two groups: conjugal (or nuclear) family—sons, daughters, and surviving spouse(s)— and other lineage members. This partition highlights the difficulties we confront in drawing inferences from these data: only 72 percent of males are classified as married. This is far lower that the married fraction of the male population known from census records, and therefore suggests that many contributors likely do not update their SSNIT records. Because Ghanaian culture exerts huge social pressure on men to father children, most place marriage and raising a family among their highest life priorities. This is particularly so for men in the formal sector, whose economic positions make them highly marriageable. Finally, every individual, no matter how isolated, belongs to a lineage. The only conceivable exceptions would be orphaned foreigners from outside sub-Saharan Africa who adopt Ghanaian citizenship. Our regressions explain the fraction of residual benefits each contributor in our sample bequeathed to members of his nuclear family (as opposed to his lineage), which we denote percentNUC, are of the form

Family Ties, Inheritance Rights, and Successful Poverty Alleviation

(2)

241

percentNUC = X i*a + bo*matrilineal + b1*postAct560

+ b2*postAct560*matrilineal + b3*matrilineal*X i + b4*ln (age) + e, where the right- hand- side variables are defined as in regression (1). Because the dependent variable is bounded by the unit interval, with mass at both endpoints, table 7.7 reports tobit regressions. Unsurprisingly, married men bequeath more to conjugal families than do men listed as unmarried. Because of the problem of stale records, as mentioned above, we reestimate the tobits restricting the sample to men designated as married. The third column presents results for female contributors. Again, age is significant: older men leave more pension wealth to their nuclear families. As above, this may be because older men have longer to build larger families or because they grow more attached to their conjugal families. Participants whose deaths occur after the 1999 implementation of Act 560 bequeath 12.7 percent more of their pension wealth to their nuclear families. This too indicates that the Act had an effect: When the law mandated that contributors provide more to their nuclear families, they complied. In addition, the data show a secular time trend toward increasing pension allocations to nuclear families—perhaps because of an ongoing erosion of traditional values. A time trend also accords with a growing social advocacy Table 7.7

Fraction of unexpired pension bequeathed to nuclear family members

Sample Matrilineal Post-560 Married Top 25 Post-560 * matrilineal Married * matrilineal Top 25 * matrilineal Log(age at death) R2 Observations

All males 0.17 0.127 0.543 –0.069 –0.136 –0.146 0.034 0.508

(1.97)** (2.45)** (7.71)*** (–1.15) (–1.91)* (–1.77)* (0.43) (5.47)*** 0.18 761

Married males 0.001 0.098 — –0.055 –0.074 — 0.018 0.415

(0.02) (1.97)** — (–0.94) (–1.09) — (0.23) (4.59)*** 0.04 547

All females –0.209 –0.168 0.353 –0.215 0.267 –0.120 0.296 0.534

(–0.88) (–0.91) (1.97)** (–1.14) (1.18) (–0.58) (1.21) (2.20)** 0.11 99

Note: Marginal effects estimated from tobit regressions explaining fraction of benefits bequeathed to surviving spouse(s) and children, as opposed to lineage. Right-hand-side variables are as in table 7.5. Numbers in parentheses are robust z-statistics, adjusted for clustering by age of death. Pseudo R-squared may not represent variation explained by dependent variables. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

242

Edward Kutsoati and Randall Morck

role for the SSNIT. The SSNIT records office staff (in particular, female staff) shared with us stories about how, over time, they increasingly assertively reminded men of their “responsibilities to their nuclear family,” and to provide for their spouses and children when listing heirs in their SSNIT records. A time trend added to the regressions in table 7.7 is significant, but a jump is nonetheless discernible at 1999. We conclude that the Act had an effect. Intriguingly, the tobits reveal that the Act’s major effect was not its intended one: altering bequests by men from matrilineal tribes. These men bequeathed more to their conjugal families before the reform, and did not substantially increase these bequests to conjugal families after 1999. One possible explanation is that men from matrilineal tribes used pension bequests to circumvent their tribal inheritance customs all along. These results appear independent of the magnitude of unexpired benefits. The SSNIT bequest decision was, after all, deliberately held confidential until the contributor’s death, and could thus provide privacy from pressure to adhere to the traditional inheritance system. Indeed, the SSNIT was intended, in part at least, to provide a defense to men from matrilineal tribes who wished to provide for their children, but who feared the wrath of traditionally minded relatives. To further test if values might be genuinely changing, we gauge for married men’s generosity toward their nuclear families. We define “generosity” as bequeathing more than the mandatory minimum of 60 percent to his nuclear family. That is, generosity = 1 if percentNUC > 60 percent and zero otherwise. Table 7.8 presents logistic regressions, similar in form to equation (2), but with generosity on the left- hand side. Matrilineal and married males are more “generous” than patrilineal and unmarried males to their children, and all males grow more generous after Act 560. However, again, no discernible difference is evident in the Act’s effect on men from matrilineal versus patrilineal backgrounds. The generosity of female contributors’ pension bequeathals is remarkably different. Although both male and female contributors are more “generous” if they report more children, female formal- sector workers are decidedly less generous to their conjugal families if their tribal tradition is matrilineal, and grow even less generous after Act 560 takes effect. Policymakers may wish to consider education programs directed at matrilineal females if further steps to force contributors to provide for their own children are deemed desirable. Next, we estimate the likelihood that the SSNIT ascertains a decedent’s instructions to be in violation of Act 560. Also, in cases where the SSNIT discovers a violation of the Act, we explore the sizes of the adjustments it imposes.16 16. The Act gives SSNIT staff a degree of discretion where a violation is discovered. In such cases, the bequests to minor children can be raised to more than the minimum allowed by the Act.

Family Ties, Inheritance Rights, and Successful Poverty Alleviation Table 7.8

Generosity of pension bequests to nuclear family

Sample Matrilineal Post-560 Married Top 25 Post-560 * matrilineal Married * matrilineal Top 25 * matrilineal Log(age at death) Pseudo R2 Observations

243

All males 0.235 0.118 0.550 0.035 –0.105 –0.133 –0.118 0.558

(2.30)** (1.95)** (10.92)*** (0.48) (–1.20) (–1.30) (–.018) (5.61) 0.22 761

Married males 0.048 0.081 — 0.024 –0.025 — –0.02 0.427

(0.76) (1.47) — (0.37) (–0.31) — (–0.02) (4.71)*** 0.05 547

All females –0.361 –0.346 0.373 –0.278 0.414 –0.219 0.218 0.523

(–2.04)** (–2.39)** (2.09)** (–1.29) (2.70)*** (–0.88) (1.16) (2.41)** 0.14 99

Note: Marginal effects are estimated using logit regressions explaining an indicator variable set to one if bequest to surviving spouse(s) and children, as opposed to lineage, exceeds the mandatory minimum required by law for minors (60 percent). Right-hand-side variables are as in table 7.5. Numbers in parentheses are robust z-statistics, adjusted for clustering by age of death. Pseudo R-squared may not represent variation explained by dependent variables. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

We estimate the probability of the SSNIT finding a violation using the following probit regression: (3) Pr (Act560 violated = 1) = X i*a + b0*matrilineal + b1*matrilineal*X i + b2*ln (age)i + e, where Act560violated is an indicator variable set to one if the SSNIT decides that the contributor’s bequests decision violates Act 560, and to zero otherwise. Table 7.9 reports these results in its first column. The instructions of married decedents are about 19 percent more likely to violate Act 560, but older men are actually less likely to leave instructions that violate the Act. This may well reflect higher odds that older men have only adult children. Finally, high- income matrilineal males are more likely to leave bequests instructions the SSNIT must override. The second column of table 7.9 reports regressions explaining the magnitudes of the adjustments SSNIT staff make to bequests found to violate Act 560. These regressions have the form (4)

pct( paid ) − pct(beq) = X i*a + b0*matrilineal + b1*matrilineal*X i + b2*ln (age)i + e,

where [ pct(paid) – pct(beq)] is the percentage of unexpired pension ultimately paid to the decedent’s conjugal family minus the percentage bequeathed. The left- hand- side variable is always nonnegative, but can be zero,

244

Edward Kutsoati and Randall Morck

Table 7.9

Act 560 audits and bequest adjustments

Dependent variable Regression type Matrilineal Married Ln(age at death) Top 25 Married * matrilineal Ln(age) * matrilineal Top 25 * matrilineal McFadden’s R2 Sample (No. of obs.)

Pr(Act560 violated) Probit Marginal effect

t-stat

–0.035 0.189 –0.487 –0.006 0.010 0.002 0.182

(–0.04) (2.34)** (–2.76)*** (–0.08) (0.08) (0.01) (1.69)*

0.04 All male contributors who died after 1999 454

Percent paid minus percent bequeathed to nuclear family Tobit Marginal effect 1.324 –0.303 –0.104 0.0001 0.066 –0.393 0.016

z-stat (1.71)* (–4.46)*** (–0.71) (0.00) (0.69) (–1.96)** (0.18)

0.38 All male contributors who violated Act 560 rules 232

Note: The first column is estimated coefficients and z-statistics for a probit regression explaining which bequests the SSNIT files ultimately judge in violation of Act 560, which requires 60 percent of residual pension benefits be paid to a decedent’s conjugal family if he has one or more minor children at the time of his death. The sample of 454 files includes those of all males who died after 1999, where after the SSNIT gained authority to alter bequests. The second column presents estimated coefficients and z-statistics for a tobit regression explaining the adjustments to bequests to nuclear families in the 232 files the SSNIT deemed in violation of Act 560. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

so we employ tobit regressions. Larger adjustments are made to bequests by unmarried, younger matrilineal men. Our data do not fully clarify their circumstances. One interpretation is that, despite all of the above government policy initiatives and the SSNIT’s exhortations, many young men from matrilineal tribes still feel a lesser duty toward their wives’ children, or perhaps to their out- of-wedlock children. Alternatively, young matrilineal men might be less attentive to updating their SSNIT files to record marriages and new children. However, we have no a priori reason that they are less attentive than their patrilineal peers. And if they are less attentive, this also suggests less concern about their conjugal families compared to young men from patrilineal tribes. 7.5.2

Knowledge of Law 111 and Asset Acquisition Decisions

We now turn from relatively well- off formal- sector Ghanaian men back to our surveys of very poor informal- sector Ghanaian women. Here, we examine the correlation between knowledge of the Intestate Succession Law

Family Ties, Inheritance Rights, and Successful Poverty Alleviation

245

111 and decisions that affect asset accumulation within the marriage, inheritance outcomes, and the economic conditions of widows. As with the SSNIT files, data limitations prevent cut and dried assessments. We cannot compare responses before and after Law 111—the law was enacted many decades ago. Moreover, knowledge of the law is almost surely not randomly distributed across women. Social networks among Ghanaians are generally very strong, and a woman may well belong to many social groups—religious organizations, trade groups, and so on. Information about the law is likely disseminated through these groups, so more socially active women are more likely to learn of it. Access to the media—newspapers, radio, and TV—depends on living nearer an urban center, literacy, literate friends or relatives, access to electricity, and other factors that quite plausibly also correlate with widows’ economic situations. We can control for education and literacy, but concede the potential importance of left- out variables. To explore these issues, we estimate regressions of women’s knowledge of Law 111 and other factors on her fractional contribution to her nuclear family’s assets, whether or not she builds up own assets, and the likelihood of intervivos during marriage. These regressions are of the form (5)

Yi = b1*KnowLaw111 + b2*matri _ spouse + b2*matri _ spouse*KnowLaw111 + e

where KnowLaw111 is one if the respondent indicates she knew of Law 111 at the time of her husband’s death and zero otherwise, educ is her years of formal education,17 matri_spouse is one if her deceased husband was from a matrilineal tribe and zero otherwise, and X is a vector of control variables: the husband’s education level, the log of their years married, and dummies for the widow’s lineage system being matrilineal, their marriage being monogamous, and the widow having minor children at the time of her husband’s death. The dependent variable, Yi in equation (5), is one of the magnitude of intervivos transfers—that is, assets the widow received from her husband before his death; the fraction of the conjugal family’s assets she estimates to have been her contribution; and an indicator variable set to one if she had personal assets when her husband died, and to zero otherwise. To gauge the magnitude of intervivos transfers, widows were asked the following question: “Did you receive financial support from your spouse when he was alive?” The responses were (a) not all, (b) very little, (c) as expected, and (d) very much. To elicit widows’ estimates of their proportional con17. This variable is zero in many cases, generally low, so the efficiency gain from using the log of one plus years of formal education is small. Using the latter approach yields qualitatively similar results throughout.

246

Edward Kutsoati and Randall Morck

tribution to their conjugal families’ assets, we asked them the question: “In your estimation, what percentage of total family assets (since marriage) is your contribution?” Respondents were presented with six choices: (a) zero, (b) less than 20 percent, (c) 20 to 39 percent; (d) 40 to 59 percent, (e) 60 to 79 percent, and (f) 80 percent or more. For whether or not the widow had personal assets that were explicitly her property, as opposed to the property of her husband or a lineage, we asked a simple yes/no question. Because of the categorical nature of these responses, we employ ordered logit regressions to estimate (5) regressions on the magnitude of intervivos transfers and on the fraction of the conjugal family’s assets the widow estimates as her contribution. A probit regression suffices in regressions on her possession of personal assets. Table 7.10 summarizes the results of these regressions. Respondents who report knowledge of Law 111 also report making significantly higher contributions to their nuclear family’s assets, and are significantly more likely to receive intervivo transfers from their spouses. Note that both the widow’s education and that of her spouse are included as controls. However, having

Table 7.10

Economic/financial decisions within marriage Intervivo Husband’s gifts to wife while alive

Know law111 Matrilineal spouse Know Law × matrilineal spouse Controls Regression Sample Pseudo R-squared

1.228*** (3.76) 0.776* (1.92) –1.675*** (–3.62) Yes Ordered logit 297 0.042

Contribution Widow’s reported contribution to conjugal family assets 0.861** (2.44) –0.031 (–0.06) –0.204 (–0.44) Yes Ordered logit 292 0.045

Personal assets Widow had personal assets at husband’s death –0.043 (–0.56) 0.021 (0.21) 0.081 (0.75) Yes Probit 295 0.041

Note: Table reports the coefficients of ordered logit regressions explaining the widow’s financial situation at the time of her husband’s death. The first column summarizes ordered logit regressions explaining approximate fractions of the conjugal family’s assets that were intervivo transfers; that is, gifts from husband to wife during his life. The second column summarizes ordered logit regressions explaining the widow’s estimate of her approximate fractional contribution to the conjugal family’s wealth. The third column summarizes marginal effects in probit regression explaining the widow having acquired personal assets. Robust z-statistics are in parentheses. Control variables are: the log of the widow’s and spouse’s education levels, log of years married, widow’s lineage system, whether monogamous marriage, and whether widow had children under age eighteen at the time of spouse’s death. ***Significant at the 10 percent level. **Significant at the 5 percent level. *Significant at the 1 percent level.

Family Ties, Inheritance Rights, and Successful Poverty Alleviation

247

a matrilineal spouse statistically counteracts the intervivo effect entirely, perhaps indicating persistent tendency of matrilineal men to pass wealth to their lineages. Widows with matrilineal spouses contributed insignificantly smaller fractions of their conjugal families’ assets, though not significantly less than with a patrilineal spouse. 7.5.3

The Economic Status of Widows

Ultimately, the purpose of the reforms we study above is to improve the lot of widows and their children. Our final set of regressions therefore examines the economic outcomes and opportunities for widows in the informal economy. Based on survey responses in the four villages we study, we construct the following dependent variables. To assess widows’ economic status, we asked them to compare their access to financial services (formal or informal), health care, and education for their children in the years immediately after losing their spouse to the years immediately before. For each comparison, respondents were asked to choose between: things got much worse than before the death (quantified as 1), slightly worse (2), the same (2.5), slightly better (3), and much better (4). Each widow’s responses to the three questions were averaged, with equal weights, to construct a change in overall economic status. To assess the efficacy of widows’ support from traditional safety nets, we also inquired about the level of financial support they received from their own lineage, and from that of their deceased husband. Their responses to each of these two queries is quantified as follows: no support at all (1); worse support than the lineage’s economic situation could readily have allowed (2), a level of support roughly in accordance with what the lineage’s economic situation allowed (3), and more support than the lineage’s economic situation readily allowed (4). We control for the widow’s “closeness” to each extended family during the marriage by asking how often they “visited or were visited by” members of each lineage. Respondents were asked how often they saw their own or their spouse’s family members during the marriage. They were given a choice of four responses: never (quantified as 1); rarely, less than once per year (2); occasionally, about once every six months (3); and very often, at least once a month (4). Because all three variables are discrete, we run ordered logits to explain them in terms of other responses from our survey. These results are displayed in table 7.11. Unsurprisingly, a widow’s “closeness” to her own extended family correlates with their providing her greater financial support for her during widowhood. A similar effect is evident for a widow’s “closeness” to her deceased husband’s extended family, with an almost identical coefficient. Minor children elicit stronger financial support from her own extended family, but not from her husband’s lineage.

248

Edward Kutsoati and Randall Morck

Table 7.11

Widows’ financial support and changes in economic status Financial support from own lineage

Inherited via Law 111 or a will “Close” to own family Matrilineal lineage “Close” to spouse’s family Matrilineal spouse Child under 18 yrs. Education level Other controls Sample R-squared

0.016 (0.05) 0.358** (2.30) –0.497 (–1.16) 0.110 (0.79) 0.196 (1.92) 0.634** (2.03) 0.004 (0.15) Yes 296 0.044

Financial support from spouse’s lineage –0.0700 (–0.16) 0.115 (0.72) –0.078 (–0.18) 0.323** (2.24) –0.013 (–0.03) 0.269 (0.87) –0.047 (–1.43) Yes 297 0.046

Change in economic status after widowed

Change in relationship with in-laws (worsening)

1.049*** (2.64) 0.091 (0.70) 0.556* (1.76) –0.064 (–0.62) 0.329 (0.98) –0.913*** (–3.85) 0.091*** (2.96)

1.053*** (2.62) –0.120 (–0.92) 0.514 (1.35) –0.219** (–1.96) 0.565 (1.51) –0.654*** (–2.71) –0.0001 (0.01)

Yes 295 0.038

Yes 300 0.062

Note: Ordered logit regressions explain affirmative survey response to questions about the generosity of financial support from the widow’s own lineage (column [1]) and her deceased spouse’s lineage (column [2]) after her bereavement, taking into account the overall wealth of the lineage in question as well as the widow’s estimate of how her economic status changed, comparing her situation in the years before her husband died to her situation in the two to five years afterward. Control variables not shown are: spouse’s education level, log of years married, indicator variables for a monogamous marriage, and whether the widow lived with the spouse. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

The third column of table 7.11 shows that a widow’s economic status deteriorates less upon the death of her husband if her lineage is matrilineal, but more sharply if she has minor children to support. The first result is somewhat at odds with the presumption of many advocacy groups and government officials that matrilineal widows are less financially secure, though the latter suggests that more could be done to support widows with minor children. We inquired casually about what sorts of support matrilineal lineages provided, and were told of brothers’ foster parenting their nieces and nephews, and of the lineage overall providing economic situations widows deemed more appropriate. The last column in table 7.11 sheds some light on how legal reforms, well intended as they may be, can lead to tensions between a spouse or widow and her in-laws. While the results show that widows with closer ties to the in-laws during marriage and those raising younger children

Family Ties, Inheritance Rights, and Successful Poverty Alleviation

249

tend to experience improved relationship with the spouse’s lineage members after his death, the use of legal structures to inherit assets (e.g., will or Law 111) leads to a worsening relationship with the in-laws. Advocacy groups can, however, clearly take heart from the significant coefficient in the first row of the third column. The very few widows—about 11 percent of the sample for this regression—who either made use of Law 111 or inherited under a will report a significantly better shift in economic status around their husbands’ deaths. The coefficient is economically large: almost twice that associated with a widow being of matrilineal lineage. Finally, all else equal, widows are substantially better off if they are educated. 7.6

Conclusions

Traditional Ghanaian cultural norms, especially regarding customary inheritance rules, differ starkly across tribes and from familiar Western and Asian norms. Reformers saw some of these traditions deterring wealth accumulation among the country’s poorest citizens, and impoverishing widows and their children. These concerns were most voluble regarding the traditions of matrilineal tribes, in which men considered their sisters’ children, not their wives’ children, to be their nearest relatives in the next generation. These traditions were thought to leave widows and their children economically vulnerable. This latter result seems consistent with that traditionalist argument that brothers and other members of matrilineal lineages support widows, including foster parenting the widow’s children. Matrilineal traditionalists countered that their customs held brothers, maternal uncles, and the like accountable for supporting their widowed female relatives, and that this system works well. Ghana’s governments has enacted two major legal rules designed to nudge men of all tribal traditions toward making decisions that would better protect their own children, thus bringing Ghanaians into closer accord with global cultural norms. First, the 1986 Intestate Succession Law 111 changed the rules for dividing up the estate of individuals who die intestate. This reform lets widows file formal legal procedures to secure ownership of certain classes’ the conjugal family’s assets, and was a major departure from customary rules that assigned most assets to the decedent’s lineage (extended family) which, if the deceased belonged to a matrilineal tribe, does not include his widow and children. Our extensive survey of widows living in matrilineal and patrilineal traditional village societies shows that Law 111 is little used, even by women familiar with it. Law 111 is procedurally touchy, necessitating expensive legal representation that makes its use expensive beyond the means of people in the traditional economy. Its use also overtly challenges traditional law, a sensitive course of action for anyone who must reside in the traditional culture

250

Edward Kutsoati and Randall Morck

for the rest of her life. Those few widows who do use Law 111 are financially more secure, and evidence that they bear social costs is unpersuasive. Our survey evidence finds matrilineal widows to be more financially secure than patrilineal widows. This supports the position of matrilineal traditionalists that brothers, uncles, and other matrilineal lineage members support widows financially, acting as de facto foster parents for her children. However, our survey also supports the efforts of reformers. Widows who used the formal legal system to settle their husbands’ estates—either by executing a formal will or by applying Law 111—gain almost twice the financial advantage associated with a matrilineal custom. Further, we find evidence consistent with mere knowledge of the law providing an incentive for couples to build up family assets jointly, and to motivate intervivo transfers from husband to wife, irrespective of education level. But surprisingly, this is least evident among widows whose spouses have a matrilineal lineage tradition—the very people the reforms focused on advancing. These findings are consistent with knowledge of the law encouraging people to accumulate wealth. However, we cannot preclude the possibility that people who have accumulated assets are more concerned about losing them and therefore seek information about the law regarding inheritances. The survey data also show that lineages remain an important safety net. Widows who had close ties with their own relatives obtain more financial support from their lineages, and widows who had closer ties with their in-laws likewise report more financial support from his lineage. While formal legal rights might give widows more bargaining power, their scant usage suggests major obstacles impede access to the law for the villagers we survey. The second major reform we examine is the 1998 Children’s Act 560, mandating that the Social Security and National Insurance Trust (SSNIT), the sole pension administrator, allocate 60 percent of deceased contributors’ unexpired pensions to their minor children, regardless of their written instructions. The SSNIT reform affects only the estates of Ghanaians who worked in the formal sector and thus contributed to the pension system, while Law 111 sought to affect widows living at the margins of the Ghanaian economy. Regression analysis of SSNIT records is somewhat problematic. First, we could only access closed files—those in which all bequests had been fully paid out. As many files remain open because the beneficiaries are unaware of their rights, our data suffer from unknown sample selection problems. Second, conversations with SSNIT staff and management indicate that many contributors’ bequest instructions were made when they first filled out a SSNIT membership form and never updated. This “stale data” problem weighs against our finding significant effects of the legal reforms affecting SSNIT survivors’ pension rights. Nonetheless, we find some limited evidence consistent with the reforms benefiting decedents’ nuclear families. Males who died after Act 560 was

Family Ties, Inheritance Rights, and Successful Poverty Alleviation

251

implemented allocate more of their pensions to the nuclear families; indeed, they bequeath significantly more than the mandatory 60 percent minimum prescribed by the law. Also, when the SSNIT judges a bequest decision to violate Act 560, the adjustments it imposes are overtly favorable to nuclear families, especially those of decedents with matrilineal lineage traditions. While our analysis finds some evidence of successful reform, we feel their deeper message is a confirmation of the tenacity of African cultural traditions. Formal legislation adopted in Accra does not change the social, economic, and spiritual forces that constrain people’s decisions. Genuinely effective reforms appear to require intense efforts to promote social awareness and provide legal aid for those who would exercise their rights under the reforms, and even then can achieve only qualified success. Our findings complement recent work by Scholz and Gomez (2004). Their examination of formal- sector inheritance rights in Botswana, Ethiopia, Ghana, Nigeria, Rwanda, Senegal, South Africa, Swaziland, Zambia, and Zimbabwe leads them to conclude that traditional inheritance norms prevail regardless of the paper law. Our findings suggest that the absence of resources to exercise formal legal rights may well be more important than a lack of information about those rights. Absent formal social assistance or legal aid programs—which are tenuous to nonexistent in all of these countries—poor Africans rationally conclude that the economic security, however minimal, offered by one’s lineage is too valuable to sacrifice for uncertain, inaccessible, and often effectively unenforceable formal legal rights. If African governments wish to effect reforms to inheritance customs, their formal legal reforms need bolstering by awareness campaigns, meaningful rule of law, and a sensitivity to existing traditions. However, challenges to long- standing traditions are likely to meet sustained resistance in Africa, as in Asia and in the Western world. Our findings also suggest that governments seeking to counter cultural norms thought to deter development might consider designing formal legal rights to empower people within the context of their traditional cultures. Matrilineal widows’ economic welfare may well have been improved by the fact that they could challenge traditional inheritance norms, even if few actually did. The law sounds hollow without an accompanying “highlife” drumbeat.

References Alger, Ingela, and Jorgen Weibull. 2008. “Family Ties, Incentives and Development: A Model of Coerced Altruism.” In Arguments for a Better World: Essays in Honor of Amartya Sen Volume II: Society, Institutions, and Development, edited by K. Basu and R. Kanbur. Oxford: Oxford University Press. _____. 2010. “Kinship, Incentives and Evolution.” American Economic Review 100:1725– 58.

252

Edward Kutsoati and Randall Morck

Awusabo-Asare, K. 1990. “Matriliny and the New Intestate Succession Law of Ghana.” Canadian Journal of African Studies 24 (11): 1– 16. Bertrand, Marianne, Sendhil Mullainathan, and Douglas Miller. 2003. “Public Policy and Extended Families: Evidence from South Africa.” World Bank Economic Review 17 (1): 27– 50. Bleeker, S. 1966. The Ashanti of Ghana. London: Dennis Dobson. Chiteji, N. S., and Darrick Hamilton. 2002. “Family Connections and the BlackWhite Wealth Gap among Middle-Class Families.” Review of Black Political Economy 30 (1): 9– 27. De Soto, Hernando. 2000. The Mystery of Capital: Why Capitalism Triumphs in the West and Fails Everywhere Else. New York: Basic Books. Ellul, Andrew, Marco Pagano, and Fausto Panunzi. 2010. “Inheritance and Investment in Family Firms.” American Economic Review 100:2414– 50. Fenrich, Jeanmarie, and Tracy E. Higgins. 2001. “Promise Unfulfilled: Law, Culture, and Women’s Inheritance Rights in Ghana.” Fordham International Law Journal 25:259– 341. http://ir.lawnet.fordham.edu/ilj/vol25/iss2/1. FIDA. 2007. “The Impact of Intestate Succession Law on Women’s Inheritance: Accra, Kumasi, and Koforidua.” Research Report, International Federation of Women’s Lawyers. Hacker, Daphna. 2010. “The Gendered Dimensions of Inheritance: Empirical Food for Legal Thought.” Journal of Empirical Legal Studies 7 (2): 322– 54. Heintz, James. 2005. “Women, Work, and Poverty in Ghana.” Presented at the conference on the Informal Economy in Ghana: A Comparative Perspective, co-hosted by the Ghana Statistical Service and Women in Informal Employment: Globalizing and Organizing, and held at La Palm Royal Beach Hotel, Accra, Ghana, October 24. Hoff, Karla, and Arijit Sen. 2006. “The Kin System as a Poverty Trap.” In Poverty Traps, edited by Samuel Bowles, Steven N. Durlauf, and Karla Hoff. Princeton, NJ: Princeton University Press. Kludze, A. K. P. 1983. “Property Law and Rural Development in Ghana.” Rural Africana 17:57– 67. La Ferrara, Eliana. 2006. “Descent Rules and Strategic Transfers: Evidence from Matrilineal Groups in Ghana.” Journal of Development Economics 83:280– 301. Lastarria-Cornhiel, S. 1977. “Impact of Privatization on Gender and Property Rights in Africa.” World Development 25 (8): 1317– 33. Ollennu, N. A. 1966. The Law of Testate and Intestate Succession in Ghana. London: Sweet and Maxwell. Plateau, Jean-Phillipe. 1991. “Traditional Systems of Social Security and Hunger Insurance.’’ In Social Security in Developing Countries, edited by E. Ahmad, J. Dreze, J. Hills, and A. Sen, 112– 70. Oxford: Clarendon Press. Quisumbing, A., and K. Otsuka. 2001. “Land Inheritance and Schooling in Matrilineal Societies: Evidence from Sumatra.” World Development 29 (12): 2093– 2110. Quisumbing, Agnes R., Ellen Payongayong, J. B. Aidoo, and Keijiro Otsuka. 2001. “Women’s Land Rights in the Transition to Individualized Ownership: Implications for the Management of Tree Resources in Western Ghana.” Economic Development and Cultural Change 50 (1): 157– 82. Scholz, Birte, and Mayra Gomez. 2004. Bring Equality Home: Promoting and Protecting the Inheritance Rights of Women. Geneva: The Centre on Housing Rights and Evictions (COHRE) Series on Law and Practicing in Sub-Saharan Africa. Woodman, Gordon R. 1985. “Ghana Reforms the Law of Intestate Succession.” Journal of African Law 29 (2): 118– 28.

8

The Surprisingly Dire Situation of Children’s Education in Rural West Africa Results from the CREO Study in Guinea-Bissau (Comprehensive Review of Education Outcomes) Peter Boone, Ila Fazzio, Kameshwari Jandhyala, Chitra Jayanty, Gangadhar Jayanty, Simon Johnson, Vimala Ramachandran, Filipa Silva, and Zhaoguo Zhan

8.1

Introduction

Despite declining global poverty, there are many regions of the world where poverty remains widespread and chronic (Young 2012). Children grow up in these regions with poor health, and their prospects are harmed by poor education (Boone and Johnson 2009). The right of every child to Peter Boone is an associate of the Centre for Economic Performance and director of the Effective Intervention program at the London School of Economics. Ila Fazzio is the field research manager of Effective Intervention in Madrid. Kameshwari Jandhyala is a researcher with ERU Consultants. Chitra Jayanty is chief executive officer of Effective Intervention in Guinea-Bissau. Gangadhar Jayanty is advisor to the chief executive office at Effective Intervention in Guinea-Bissau. Simon Johnson is the Ronald A. Kurtz (1954) Professor of Entrepreneurship and Professor of Global Economics and Management at the MIT Sloan School of Management and a research associate of the National Bureau of Economic Research. Vimala Ramachandran is director of ERU Consultants Private Limited. Filipa Silva is education project coordinator at Effective Intervention in Guinea-Bissau. Zhaoguo Zhan is assistant professor of economics at Tsinghua University. We are grateful to the National Bureau of Economic Research (NBER), Cambridge, MA, USA and Effective Intervention, a UK charity, for financial support. Pratham, an Indian charity, kindly provided assistance with test design. Participants at the NBER African Successes conference in Accra, Ghana, and the CREO Study Report conference in Bissau, GuineaBissau, provided helpful comments. Alex Eble advised on early study design. For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see http://www.nber.org/chapters/c13376.ack. Authors’ roles: Peter Boone conceived the study, participated in the design of the survey, analyzed results, coauthored this chapter, and raised the financing for the survey. Ila Fazzio conceived the study, lead the design of the survey, lead the training and management of field work teams, analyzed results and coauthored this chapter. Kameshwari Jandhyala advised on study design, surveys, and implementation, and coauthored this chapter. Chitra Jayanty conceived the study, participated in the design of the survey, supervised the implementation of the project, and coauthored this chapter. Gangadhar Jayanty conceived the study, participated in implementation of the project, and coauthored this chapter. Filipa Silva participated in the design of the survey, trained and supervized the field work teams, and coauthored this

255

256

Peter Boone et al.

primary school education is one of the Millennium Development Goals for 2015, yet in many of the most extreme “pockets of poverty” little is being done to address these issues. However, in order to understand the extent of problems and potential solutions, we need a good understanding of the current conditions. Statistics from very poor regions are generally unsatisfactory for this purpose (Jerven 2013). Often statistics are biased or missing because authorities have incentives to paint a rosy picture, or they simply do not make the effort or allocate resources to measure. Even when good statistics are available, the measures usually cover service provision (such as the level of school enrolment), rather than the desired outcome (indicators of educational and skills levels achieved). In this chapter, we report on a unique survey that was conducted in 2010 in rural villages and schools of Guinea-Bissau, in West Africa. GuineaBissau is one of the poorest nations of the world, ranking 192 on income per capita, at $600 per person (World Bank 2012). The United Nations Education, Scientific and Cultural Organization (UNESCO) reports the national adult literacy rate is 54.2 percent, while 72 percent of Guinea-Bissau’s youth ages fifteen to twenty-four are literate. The net primary school enrolment ratio is 73 percent. These figures represent national averages, but our work suggests they mask extremely poor educational outcomes in rural regions.1 (See figure 8.1.) One goal of the CREO (Comprehensive Review of Education Outcomes) survey was to provide an accurate, comprehensive overview of children’s learning outcomes and the relation of these to school quality, parental care, and socioeconomic variables. We are not aware of any other similarly comprehensive, integrated surveys of schools, households, caregivers, and children in such poor regions of West Africa.2 Our analysis is based on surveys from 202 villages (approximately 20 percent of the population) with interviews from 3,968 households. We interviewed 8,782 parents, and analyzed numeracy and literacy tests for 9,947 children ages seven to seventeen. We chapter. Simon Johnson conceived the study, analyzed results, coauthored this chapter, and raised financing for the survey. Vimala Ramachandran advised on study design, surveys, and implementation, and coauthored this chapter. Zhaoguo Zhan carried out statistical analysis and coauthored this chapter. 1. The UNESCO defines literacy as “The ability to read and write with understanding a simple statement related to one’s daily life. It includes a continuum of reading and writing skills, and often includes basic arithmetic skills.” Our survey population represents approximately 25 percent of the country’s population. If youth literacy rates in urban areas were approximately 90 percent, then the UNESCO data could be consistent with our data. We have not surveyed urban areas; however, our experience suggests literacy rates would be substantially lower than 90 percent, but well above the rural rates reported here. 2. Demographic and Health Surveys (DHS) cover important socioeconomic variables. This survey allows us to identify children’s outcomes with their respective schools, villages, and parents, thus permitting us to examine richer correlations. Since this is a survey, we cannot make causal interpretations based on these correlations.

The Dire Situation of Children’s Education in Rural West Africa

Fig. 8.1

257

Map of Guinea-Bissau

asked children which schools they attended, and then visited 351 schools and interviewed 781 teachers. We checked whether schools were operating, measured attendance compared to enrolment, and surveyed school materials. The results of the survey present a bleak picture of educational resources and outcomes in rural Guinea-Bissau. According to the national curriculum, by third grade (children age nine) children should be able to read and comprehend a story, and complete the simple math test we used (see section 8.3.3). We found that 81 percent of ten-year-olds could not sum two singledigit numbers, and 91 percent could not read single words. Among the 1,169 ten-year-olds with test results, we found only one child who was capable of completing both the numeracy and literacy tests. Why are these results so poor? While survey responses should be taken with caution since they may not reflect actions, our “demand” indicators suggest parents have a strong desire to send their children to good schools. Over 98 percent of household heads reported they would be willing to pay an average 20 percent of monthly income for school fees for each of their school-aged children if good schooling would be provided. In our spot checks of schools, we found 72 percent of enrolled children at the school. Attendance rates are probably boosted by functioning school lunch programs. The fact that children attend schools despite poor outcomes, and the reported willingness to pay for schooling, suggest additional interventions that target attendance, such as conditional transfers implemented in Mexico, are probably not of primary importance to improve educational outcomes (Kremer 2003; Schultz 2004).

258

Peter Boone et al.

On the supply side, given the poor test results, it is surprising to learn that 86 percent of schools were open and roughly three-quarters of enrolled children were attending when we conducted spot checks. The teachers reported substantial experience and some training. However, very few schools have adequate text books, and many teachers had a poor grasp of spoken Portuguese. Guinea-Bissau has many local languages. The most widely spoken language is a mix of Portuguese, the national language, and local dialects called “Kriol.” The only spoken language with a written script is Portuguese. Our numeracy tests (apart from number recognition) were simple math questions written on paper, and instructions were given verbally in the local language, so children did not need to know Portuguese in order to succeed in the test. Instructions for literacy tests were given in the child’s local language, but children were tested on reading and comprehension of Portuguese. An important goal of our project was to find “examples of success” with the aim of using these to better understand how projects to improve education can be developed. We anticipated that we could find individual schools or villages where education levels were high, and lessons could be learned from these. Unfortunately, we found only six schools (with more than ten pupils) where average scores implied students could read a paragraph. For the numeracy tests, there were no schools where the average student could multiply or divide. These survey results provide a comprehensive picture of a dysfunctional school system. The reasons for the poor functioning are numerous. However, we believe the evidence points to a strong latent demand for education in these villages that remains unsatisfied due to the very poor quality of existing teaching and schools’ organization. While there are many steps that could be taken to make schools function better (Ramachandran 2003; Kremer and Holla 2009), it is important to understand which should be prioritized in regions such as Guinea-Bissau. The public sector cannot be relied on to provide regular services due to political instability, institutional capacity, and a political system that does not serve the very poor. Private-sector provision of schooling in these villages is not profitable enough to generate activities due to low-income levels. The difficult logistics of finding trained teachers, gaining finances, securing materials, and supervising performance makes it daunting for individual villages to improve schools on their own for a sustained period. Similar problems plague non-governmental organizations (NGOs) and foreign donors who attempt to improve school outcomes. To some extent, when faced with the choice of attempting to build and maintain effective schools, or waiting to learn whether the existing public schools might get better, many villages choose to wait and hope. The result is that one more generation of children is growing up without even basic numeracy and literacy skills.

The Dire Situation of Children’s Education in Rural West Africa

8.2 8.2.1

259

Methods Survey Location and Eligible Population

When the survey was conducted, the last published census in GuineaBissau dated from 1991. We compiled a list of 913 potentially eligible villages from the census based on demographic estimates of whether they were likely to have a population between 200 and 1,000 in 2010. We then randomly ordered these villages numbering them from 1 to 913.3 Field workers then visited each village and interviewed village leaders to learn the estimated true population of the village. Village and household-level surveys were completed from December 2009 until the end of the academic year in 2010. Some delays in implementation and data checking were caused by security concerns during and after the survey was completed. A village was eligible for the survey if the village leaders agreed to the interview and survey, and if the population was estimated to be between 200 and 1,000 during the field worker’s visit with a minimum of twenty households, and it was accessible by land during the dry season. From the list of 913 villages, we visited villages sequentially with a target of 200 villages. The teams visited 411 villages; however, 209 of these were ineligible due to the number of households, or population, being above or below eligibility thresholds. Ultimately, surveys covering 202 villages were completed and reported here in the analysis.4 Village elders were interviewed to complete baseline data on villages. There are no maps of households in villages. We selected households for interviews according to systematic sampling (Luman et al. 2007). This sampling methodology generates an average of twenty households per village, independent of the size of the village.5 3. The randomization was conducted in ACCESS, assigning each village a unique number using a random number generator. Villages were then ordered sequentially according to assigned numbers. The first 202 eligible villages from this ordering are reported here. 4. We planned to complete 200 villages before the end of the 2009/10 school year. The teams kept working to complete their last village once the 200 total had been achieved. 5. The system used to select houses was derived from the SystRS methodology, which was developed to be an easy but representative way of randomly selecting households when there is no sampling frame. An estimate of the number of households in the village is made, and then a “skipping interval” is calculated by dividing the total houses by twenty. To establish the “starting” point of the “random walk,” supervisors asked a senior member of the community to take them to two opposite edges of the village, preferably along the longest axis in the village (i.e., the two most distant houses in the village). These were the starting points for two interviewers. The interviewers then picked a random number between 0 and the “skipping interval” to determine the first house to visit. They then walked in a line counting houses until they reached this starting number. That would be the first house to interview. They then walked toward the village center in a serpentine way, enumerating houses according to the calculated skipping interval. If they did not enumerate a minimum of twenty houses using the interval system, they would walk back toward the edge and enumerate houses that had not been enumerated until they reached twenty.

260

Peter Boone et al.

A household was eligible for inclusion if they were in an eligible village, they were selected by systematic sampling, they had at least one eligible child, and the household head agreed to be interviewed. We conducted an initial interview with a household head to determine characteristics of the household, and the eligible children in the household. A child was eligible if she lived in an eligible house, she was age seven to seventeen years, she was available on the day of the survey visit, and she had at least one eligible caregiver who was available to interview on the day of the visit. We interviewed each child and conducted literacy and numeracy tests. We also interviewed all the child’s available caregivers. When we surveyed villages, we listed all Portuguese-language schools in the region. There are Koranic schools in these regions where Portuguese is not taught; however, we did not visit these as they do not teach the national curriculum, and nearly all such children were also enrolled in a Portugueselanguage school. We then visited all schools that children reported they were enrolled at (during our child interviews). We sought permission from head teachers to survey schools and conduct teacher interviews. We interviewed every teacher who was available on the day of interview if they taught any grades 1 to 4. We sought permission to interview from village leaders, head teachers, household heads, and each participant. All villages and schools agreed to be included in the study. Only one household head refused to permit interviews. In the remaining households, all children and caregivers agreed to be interviewed. 8.2.2

Main Outcome: Test Scores

The primary outcome measure for the survey is test score results from literacy and numeracy tests given to eligible children. In our initial field work we recognized that a wide range of outcomes was possible, but that a substantial fraction of the population would likely score poorly on common international tests, so we needed a test that was more sensitive at discriminating at poorer levels of education. We chose to modify the ASER tests developed by Pratham, a large Indian NGO that specializes in educational outcome tests in India, to suit the environment in Guinea Bissau. We conducted multiple field tests in villages prior to the survey. The literacy test was implemented as follows: The test starts at a mediumlevel task where children are asked to read a sentence. If they were able to read at least two of the four sentences offered, they would be asked about the meaning of some underlined words and also asked (in local language) two questions about the comprehension of these sentences. If a child was able to read sentences they would be offered to read a story, and tested on comprehension of the story. A child who was not able to read a sentence would be offered mono- and dissyllabic words to read. Children who were not able to correctly read four out of five words were then asked to recognize

The Dire Situation of Children’s Education in Rural West Africa

261

letters. A child got the highest possible mark (a score of eight) if they could read and comprehend a short story, while they received zero if they could not read monosyllabic words. The numeracy tests are designed similarly, starting with addition of two single-digit numbers without carrying forward. If a child was able to solve correctly at least one sum, they would be offered to solve subtraction with borrowing, followed by multiplication, simple division, and inexact division. A child got the highest possible mark if they could answer all questions correctly. If they could not complete the initial addition questions, they were asked if they could recognize two-digit and single-digit numbers. A child scored zero if they could not recognize single-digit numbers. The interviewers conducted the tests in languages that were convenient for the child. However, since local languages are not written, and the national curriculum teaches children to read in Portuguese, we tested children’s ability to read and comprehend Portuguese. We also conducted background surveys to help interpret the reasons behind test score results. We interviewed caregivers to learn their socioeconomic status as well as attitudes toward schooling for their children. We also conducted a survey to measure the quality of school facilities, and we interviewed teachers to learn their overall training and work conditions. Our school interviews also provided spot checks to learn if schools were functioning, and they measured attendance. 8.2.3

Data Collection and Management

An initial team of ten people and ten vehicles traveled and located eligible villages based on the 1991 census. They sought consent from village leaders, recorded the GPS location of the village, determined the estimated current size of the village, and recorded the schools that children resident in the village attended. The remaining survey data collection was managed by five teams of three people, including one supervisor for each team. These teams completed the household and child surveys, as well as implementing child numeracy and literacy tests. They first sought permission to conduct the survey from the head of the household. If permission was granted, and if there was an eligible child, they proceeded to complete the survey. They sought interviews with all caregivers for eligible children in the household. The field workers would typically stay in a village three days, and they were trained to repeatedly seek out interviews with caregivers and children until all eligible children and caregivers had been reached. When conducting tests, children were asked to step away from their friends and other onlookers to complete the test. All surveys and discussion were completed in a local language. For the numeracy tests, children were asked to answer written questions. If the child could sum two digits correctly, we proceeded to harder questions. Hence, to pass the numeracy tests, children did not need to speak any Portu-

262

Peter Boone et al.

guese.6 Instructions for literacy were stated in local languages; however, the words and stories used in the test were in Portuguese. The schools survey was completed by our survey supervisors during the village visits. Supervisors made several attempts to visit schools during hours when they should be open, and if schools were functioning, they sought the head teacher’s approval to conduct the survey. They examined the official enrollment data at the school, and they visited classrooms to count the number of children actually attending along with textbooks available. During each village survey, supervisors double-checked survey forms and errors were corrected at that time. Four data entry personnel conducted double-blind data entry. A supervisor was responsible for correcting errors found when checking double-blind entry. If needed, we attempted to revisit interviewees to correct any errors found during data entry. Due to political uncertainty before and after the survey, some data correction had to be delayed for security reasons and this slowed the completion of data analysis; however, such corrections were very minor. 8.2.4

Sample Size

The survey was designed to cover a national footprint and provide adequate power to learn the importance of covariates. We were limited by access to good quality vehicles and needed to complete the study within a school year. We were also limited by political upheaval, which delayed implementation of the survey due to security concerns. By surveying approximately 200 villages with 10,000 children tested, we have a well-powered data set to examine correlations between test outcomes and other variables. 8.3 8.3.1

Results Population Covered

The analysis is based on outcomes for the 202 surveyed villages, including 3,968 household surveys, 4,907 female caregiver surveys, and 3,875 male caregiver surveys. In total we found 19,776 children age birth to seventeen living in these households, of which 10,014 were age seven to seventeen and eligible for our literacy and numeracy tests. We managed to complete surveys, as well as numeracy and literacy tests, and so base the analysis on 9,947 of these children (table 8.1).

6. If a child could not add two single-digit numbers, then the test reverted to checking if they could recognize single- and double-digit numbers. They were required to respond verbally in Portuguese to this number recognition test. So a child that spoke no Portuguese, and could not add two single digits in a written test, would score zero. A child that could add single digits in a written test was presumed to be able to recognize numbers and received no verbal test.

263

The Dire Situation of Children’s Education in Rural West Africa

8.3.2

Characteristics of Villages and Households

Table 8.2 shows average population of villages is 458 people. The average household size is ten persons. Our survey of twenty households covers 40 percent of an average village. Fifty-four percent of the villages had a school in the village, and when there was no school, the average walking time to the nearest school used by villagers was thirty minutes. On average villages were ninety-seven minutes from the nearest road with regular transport, and six hours walking distance from a village with a secondary school. The World Food Program provides school lunches across the nation, and 46 percent of the villages in our survey reported children had access to these meals within the village. None of the villages in the survey had access to publicly provided electricity. Seven percent of households report they have their own generator, and six percent have televisions. Battery-operated radios are common, with three-quarters of households reporting they have one. For transportation, two-thirds of households had bicycles, and only 7 percent have motorbikes. Table 8.1

Population surveyed and analyzed Persons or units Eligible villages Randomly selected villages Households Female caregivers Male caregiver Children in households: of which age 7–17 of which fully completed interviews and tests Schools used by the children of which: were open to be assessed Teachers present and interviewed

Table 8.2

913 202 3,968 4,907 3,875 19,776 10,014 9,947 351 303 781

Characteristics of villages Units

Average population Walking distance to nearest school Is there a school in the village? Is there a meals program in the village? Walking distance to nearest road with regular transport Walking distance to nearest village with secondary school

Average of sample

N Minutes 1 = Yes, 0 = No 1 = Yes, 0 = No Minutes

458

Minutes

363

0.460 97

Standard deviation 230

0.50 80 483

Min.

Max.

200

1,000

0 3

1 600

0

2,880

Note: Observations are excluded where data is missing or not answered unless otherwise reported.

264

Peter Boone et al.

The self-reported total monthly income was CFA 44,292 ($88.11 at the average exchange rate of 502.71 CFA/USD during the survey), equalling roughly $9 per person in the household. In table 8.3 we report questions related to attitudes of the household leader to education. Nearly 100 percent of respondents said they would be willing to pay extra money to add and improve education for their boys and Table 8.3

Characteristics of households Units

How many people in household? How many rooms in the house? Does your house have? Generator Television Radio Mobile Table Motorbike Bicycle Watch/clock Total monthly income of all members Any person in the house who can read or write? Would you be willing to pay extra to have a son get schooling? If so, how much per year? Would you be willing to pay extra to have a daughter get schooling? If so, how much per year? Main reasons that boys stop schooling School isn’t available nearby Needs to earn money Gets married Helps at home Avoid enticement away from family morals Family can’t afford it Main reasons that girls stop schooling School isn’t available nearby Needs to earn money Gets married Helps at home Avoid enticement away from family morals Family can’t afford it She becomes pregnant

N N 1 = Yes, 0 = No

FCF 1 = Yes, 0 = No 1 = Yes, 0 = No FCF 1 = Yes, 0 = No FCF 1 = Yes, 0 = No

Average of sample

Standard deviation

Min.

Max.

9.68 4.71

5.17 2.25

2 1 0

47 14 1

0.07 0.061 0.749 0.04 0.282 0.067 0.677 0.618 44,292

0.255 0.24 0.433 0.197 0.451 0.25 0.468 0.486 34,053

0 0 0 0 0 0 0 0 0

1 1 1 1 1 1 1 1 420,000

0.614

0.487

0

1

0.995 8,308

0.069 7,892

0 0

1 180,000

0.994 8,171

0.079 7,814

0 0

1 180,000

0.371 0.27 0.389 0.382 0.091 0.471

0.483 0.444 0.487 0.486 0.288 0.499

0 0 0 0 0 0

1 1 1 1 1 1

0.314 0.069 0.731 0.217 0.121 0.318 0.527

0.464 0.254 0.444 0.412 0.326 0.466 0.499

0 0 0 0 0 0 0

1 1 1 1 1 1 1

1 = Yes, 0 = No

Note: Observations are excluded where data is missing or not answered unless otherwise reported.

The Dire Situation of Children’s Education in Rural West Africa Table 8.4

265

Characteristics of parents/caregivers Units

Female caregiver (4,907 respondents) Age Ethnic background Balanta Fula Mandinga Other Religion Animist Muslim Christian Other Claims knows how to read and write and was able to pass a literacy test Attained a school level of at least grade 1 Male caregiver (3,875 respondents) Age Claims knows how to read and write and was able to pass a literacy test Attained a school level of at least grade 1

Years 1 = Yes, 0 = No

1 = Yes, 0 = No 1 = Yes, 0 = No

1 = Yes, 0 = No 1 = Yes, 0 = No

Average of sample 38.51

Standard deviation 12.1

Min.

Max.

15

90

0.258 0.414 0.151 0.177

0.438 0.493 0.358 0.381

0 0 0 0

1 1 1 1

0.301 0.619 0.046 0.034 0.0569 0.0273 0.0913

0.459 0.486 0.209 0.181 0.232 0.163 0.288

0 0 0 0 0 0 0

1 1 1 1 1 1 1

14.1 0.517 0.427 0.491

17 0 0 0

100 1 1 1

47.50 0.358 0.239 0.404

Note: Observations are excluded where data is missing or not answered unless otherwise reported.

girls. There was no indication of gender bias. They reported they would be willing to spend 18 percent of their monthly income.7 When asked the reasons for why children stop going to school, the most common reported reason for girls was marriage (73 percent of respondents), pregnancy (53 percent), and lack of a nearby school (37 percent). Since girls in these villages tend to marry soon after puberty, the findings suggest girls would stay at school until their early teens. For boys, the most common reason to stop school was that the family cannot afford it (47 percent), followed by getting married, needing to help at home, and lack of a nearby school (37–39 percent). Very few respondents reported moral concerns regarding the schooling of girls or boys. 8.3.3

Characteristics of Caregivers and Children

Tables 8.4 and 8.5 present findings for caregivers and children. The tribal and religious breakdown of women and men (not shown) was similar and 7. The IMF estimates GDP per capita in 2010 was $42 per month, which would include household income and other factor incomes. These rural villages should be substantially poorer than the average population as our data implies.

266 Table 8.5

Peter Boone et al. Characteristics of children Units

Age Sex Languages at home Kriol Portuguese Other Does she speak/understand Kriol a little or fluently at interview? Ever attended pre-school?

1 = Girl, 0 = Boy

1 = Yes, 0 = No 1 = Yes, 0 = No Ever attended a school? 1 = Yes, 0 = No Attending a school at time of interview? 1 = Yes, 0 = No What type of school does she attend (can be 1 = Yes, none, or more than one for each child)? 0 = No Public Community Private Missionary Koranic What languages do they speak at school? Portuguese Creole

Average Standard of sample deviation Min. Max.

0.526 0.577 0.001 0.421

0.491 0.041 0.497

0.748

0

1

0 0 0

1 1 1

0

1

0.057

0.231

0

1

0.845

0.362

0

1

0.703

0.457

0

1

0.413 0.238 0.026 0.027 0.217

0.492 0.425 0.158 0.163 0.163

0 0 0 0 0 0

1 1 1 1 1 1

0.752 0.899

0.432 0.301

0 0

1 1

Note: Observations are excluded where data is missing or not answered unless otherwise reported.

matches national figures, with approximately three-fifths of the population reported as Muslim, and 30 percent animist. Six percent of female caregivers reported they could read, but less than half of these were able to read a simple sentence when presented with a literacy test. Thirty-six percent of men claimed to be able to read, but we found one-third of these could not read a simple sentence, leaving 24 percent of interviewed men who claimed they could read and then proved able to read a sentence. Among the children we interviewed, 58 percent stated that they spoke the local creole language at home, and interviewers found 75 percent of children were able to speak Kriol somewhat or fluently during their interviews. Less than 1 percent of children reported their family spoke Portuguese at home. When asked about school, only 5 percent of children reported they had attended a preschool, and 85 percent reported they had attended school

The Dire Situation of Children’s Education in Rural West Africa

267

sometime in the past. Seventy percent of children reported they were currently enrolled in school. Forty percent of children attended public schools, while 24 percent were enrolled in community schools. Twenty-two percent of children attended Koranic schools, and 98 percent of these children reported that they also attended another school.8 8.3.4

Characteristics of Schools and Teachers

Tables 8.6 and 8.7 present findings from school and teacher interviews. We compiled the school list from our village surveys and from asking children which school they attended. We then attempted to visit all schools that children attended at times when the school should have been open, and we repeatedly visited over at least three days if teachers were not present on the first occasion. In total we found 351 schools that were reported to serve children in the selected villages; however, when our field workers visited these, only 303 had teachers present on at least one visit. In these schools we found 781 teachers who taught grades 1–4. We did not visit Koranic schools as these do not teach the national curriculum, and are generally not recognized officially as schools. Approximately half the schools we visited received support from the community for cleaning and/or infrastructure, while 28 percent of schools reported that communities provided assistance directly to teachers (financial or in kind). Only 45 percent of schools had a toilet, and 28 percent had drinking water accessible. The average school had three teachers for grades 1–4, and 80 percent of teachers were male. Teacher’s ages ranged from eighteen to sixty-one, with an average of thirty-eight, and on average they completed ninth grade. Despite a young national population, the average teacher had been teaching for nearly thirteen years. Three-quarters of teachers reported that they had a textbook. When field workers visited classrooms, they found slightly less than threequarters of enrolled students were present in the class. If we consider the attendance rate in functioning schools, and assume the attendance rate in nonfunctioning schools is zero, then average attendance would be approximately 62 percent for the overall population (assuming the closed schools had similar enrollment size to schools that were functioning). Nearly all schools had blackboards, and there was chalk visibly available in almost all classes. However, there were very few textbooks for any course or grade. For example, on average there were twenty-nine enrolled children for every grade 2 math textbook visible.

8. Koranic schools do not teach the national curriculum.

Table 8.6

Characteristics of school Units

School (303 respondents) School type Public Community Private Missionary Does the school receive support from the community? Cleaning For teachers Food Water Infrastructure How many teachers for grades 1–4 work in the school? Male Female How many children are enrolled in grades 1–4? Grade 1 Grade 2 Grade 3 Grade 4 Average pupil per teacher ratio for combined grades 1–4a Attendance ratio at time of visit (among schools that were operating when visited) Grade 1 Grade 2 Grade 3 Grade 4 Average number of math textbooks per enrolled child at the school Grade 1 Grade 2 Grade 3 Grade 4 Does the school have a toilet? Does the school have drinking water available?

Average Standard of sample deviation

Min.

Max.

1 = Yes, 0 = No 0.591 0.314 0.059 0.036

0.493 0.465 0.237 0.187

0 0 0 0

1 1 1 1

0.538 0.277 0.185 0.092 0.469

0.499 0.448 0.389 0.29 0.5

0 0 0 0 0

1 1 1 1 1

2.38 0.611

1.55 1.31

0 0

11 8

7 6 6 4

252 221 184 210

1 = Yes, 0 = No

N N

N N N N

60.7 45.7 39.7 42.7

39.1 33.7 294 34.5

N

63.4

24.4

7.2

17.8

Proportion Proportion Proportion Proportion

0.772 0.74 0.724 0.744

0.223 0.236 0.306 0.233

0.139 0.056 0.083 0.067

1.72 2.07 3.29 1.92

Proportion Proportion Proportion Proportion 1 = Yes, 0 = No 1 = Yes, 0 = No

0.033 0.035 0.036 0.041

0.135 0.102 0.123 0.112

0 0 0 0

1 0.96 1 0.88

0.452

0.499

0

1

0.274

0.447

0

1

Note: Observations are excluded where data is missing or not answered unless otherwise reported. a Teachers teach an average 1.67 classes, so pupils per class will be lower.

269

The Dire Situation of Children’s Education in Rural West Africa Table 8.7

Characteristics of teachers Units

Gender Age Highest grade achieved Did she receive training before teaching? How many years has she been teaching? How many years has she been teaching at this school? Does she use a textbook when teaching? Does she test students regularly?

1 = Female, 0 = Male Years Grade level Years Years 1 = Yes, 0 = No 1 = Yes, 0 = No

Average of sample

Standard deviation

Min.

Max.

0.19 38.5 9.41 0.595 12.6

0.392 10.7 1.74 1.23 11.6

0 18 4 0 0

1 61 12 8 46

6.51

6.48

0

35

0.739

0.44

0

1

.986

.118

0

1

Note: Observations are excluded where data is missing or not answered unless otherwise reported.

8.4 8.4.1

Test Score Results Numeracy Test Results

Table 8.8 presents results from the numeracy tests for all children. The total number of children tested declines from 1,527 at age seven, to 1,018 at age twelve, and then falls sharply to 406 by age seventeen. We have roughly 7 percent more males than females, with the bias increasing with age. These patterns are probably due to children dropping out of school and moving outside their village as they get older. As reported from the household survey, women tend to leave school earlier than men due to marriage and pregnancy, while men leave when they need to earn an income. The numeracy test results show educational levels are extremely poor. For example, by age ten, according to the national curriculum of Guinea-Bissau, a child should be in grade 3 or grade 4, and they should be able to pass all aspects of this test. Instead, we find that 36 percent of ten-year-olds cannot even recognize a number. A further 31 percent can recognize single digits but cannot recognize two-digit numbers. This is despite the fact that 74 percent of ten-year-olds reported that they were currently attending school. In this survey, there is one child who, at age ten, was able to complete all questions from the test. 8.4.2

Literacy Test Score Results

Table 8.9 presents results from literacy tests. The materials for these tests were conducted in Portuguese; however, children were able to respond in their local language or Creole if they chose to. These other languages do not

12 Total Number

10 11

8 9

6 7

0 100.0 1,527

0 0

0 0

0 0

78.13 18.66 1.77 0.52 0.59 0.33

7

0 100.0 1,174

0 0

0.17 0

0.17 0

61.24 28.45 5.96 1.45 1.87 0.68

8

0 100.0 976

0 0

0.2 0

0.31 0.2

44.57 34.63 9.73 4 5.43 0.92

9

0.09 100.0 1,169

0 0

0.34 0.17

1.45 0.68

36.36 31.22 13.43 5.47 7.44 3.34

10

0.24 100.0 826

0 0

0.85 0.24

1.57 0.97

26.27 33.05 14.16 5.21 12.59 4.84

11

0.2 100.0 1,018

0.2 0.1

1.18 0.79

2.06 1.08

21.02 25.34 16.4 8.25 16.01 7.37

12

0.59 100.0 848

0.59 0.24

1.18 1.18

3.42 3.42

15.92 21.46 17.33 9.43 19.93 5.31

13

0.75 100.0 802

0.25 0.62

2.62 0.75

5.49 3.49

14.71 19.33 16.46 7.23 20.7 7.61

14

15

1.12 100.0 717

1.12 1.53

2.09 1.26

4.74 3.35

17.57 15.34 13.81 8.37 19.67 10.04

Percentage of children at specific standards (by age)

1.03 100.0 484

1.24 1.45

2.48 2.07

7.64 7.23

13.64 15.29 14.67 5.58 20.45 7.23

16

2.71 100.0 406

1.72 1.97

5.91 2.71

8.13 6.4

14.04 7.88 16.01 4.93 18.97 8.62

17

Note: Tests are administered in a progressive stepwise way. We start by asking children to read a sentence. If they are unable to read a sentence, we ask them to read words. If they are able to read a sentence, we ask them to read a paragraph. Based on 9,947 test scores.

Inexact division, a single-digit number into a two-digit number

Can divide a single-digit number into a twodigit number

Can multiply a two-digit number by a single-digit number

Can subtract one two-digit number from another

0 1 2 3 4 5

Test score:

Age:

Results from numeracy tests by age and standard achieved

Not able to recognize single digits Can recognize single-digit numbers Can recognize two-digit numbers Can add two single-digit numbers Can add a single- and two-digit number

Standard

Table 8.8

6 7 8 Total (%)

5 0 0 0 100.0 1,527

0

0.2 0

92.67 7.14 0

7

0.09 0 0.17 100.0 1,174

0

1.53 0

81.26 16.87 0.09

8

0.61 0 0.1 100.0 976

0.2

4.3 0

71.11 23.67 0

9

1.54 0.34 0.86 100.0 1,169

0.34

5.73 0

62.28 28.83 0.09

10

3.39 0.12 1.94 100.0 826

1.33

9.32 0

52.42 31.36 0.12

11

4.22 0.79 4.13 100.0 1,018

1.57

14.34 0.2

44.2 30.26 0.29

12

4.72 0.94 7.9 100.0 848

1.89

16.51 0.12

35.5 32.31 0.12

13

6.23 0.87 10.97 100.0 802

3.37

15.59 0.12

33.17 29.55 0.12

14

15

8.65 1.67 13.53 100.0 717

1.67

16.18 0

31.94 26.22 0.14

Percentage of children at specific standards (by age)

10.12 1.03 19.42 100.0 484

2.69

15.08 0.21

29.34 21.69 0.41

16

9.36 1.97 26.6 100.0 406

3.69

16.01 0

25.12 17 0.25

17

Note: Tests were conducted sequentially. For example, if a child could not add two single digits, we did not ask that they multiply. Based on 9,947 test scores.

Number of children assessed

Can read and comprehend a story

Can read paragraph but cannot answer simple comprehension questions Can read and comprehend a simple paragraph but not a story

3 4

0 1 2

Cannot recognize letters Can recognize letters but not words

Can recognize and comprehend words but cannot read a simple paragraph

Test score:

Age:

Results from literacy tests by age and standard achieved

Standard

Table 8.9

272

Fig. 8.2

Peter Boone et al.

Literacy scores by age

have a written script, so schools teach reading and writing in Portuguese, and children who attend school should be familiar with the materials. The literacy tests show similar outcomes to numeracy tests. Under the national curriculum, by the end of grade 2 a child should be able to read and comprehend a simple sentence. However, we found 62 percent of ten-yearold children were not even able to recognize letters, while 91 percent were unable to read simple words. Only one ten-year-old in the whole sample was able to successfully complete the numeracy and literacy test in full. Figures 8.2 and 8.3 illustrate average literacy and numeracy test scores by age. Those children, who do eventually learn to read or learn simple maths, tend to do so far later than the school curriculum calls for. We suspect learning occurs according to need—with some children more exposed to money and economic and social activities, such as small-scale trading, where limited literacy and numeracy is required. 8.4.3

Multivariate Analysis

We examined multivariate outcomes at the level of the child and school. The school-based results are shown in table 8.10 where we regress average student scores by school against a number of control variables.9 The general picture from these results—correlations only—is that children at private 9. We report linear ordinary least squares (OLS) regression results in order to simplify the interpretation of coefficients. Logistic regressions provided similar results and are not reported.

The Dire Situation of Children’s Education in Rural West Africa

Fig. 8.3

273

Numeracy scores by age

schools have slightly (0.7) higher scores than children at other schools (results normalized against community schools). The number of teachers at the school is statistically significant; however, the teacher-student ratio is not important. We suspect this reflects the size of the school, with larger schools being closer to urban areas and in regions where written language is probably more commonly seen. The ability of the teacher to speak Portuguese is statistically significant. However, other variables measuring school and teacher quality (such as training, an indicator of equipment including chalk and blackboards, as well as books) do not enter the regressions. Given the generally low level of outcomes and the limited variance across villages and schools (see section 8.4.3), the empirical importance of the right-hand-side variables is small. For example, a private school with fluent Portuguese-speaking teachers is predicted to have an improvement of 1.2 on the average literacy test compared to the same-sized public school with teachers who speak very little Portuguese. An improvement of 1.2 is a minor change compared to the large learning gap revealed here (see figure 8.2). We do not report the outcomes from child-level regressions in this chapter as there is little additional contribution from the findings. The implied “impacts” of realistic changes in the statistically significant right-hand-side variables were not empirically relevant compared to the learning gap. We did find statistically significant correlations between test scores and ability to speak creole (+’ve: positive correlation), a household wealth indicator (+’ve), mother and father’s ability to read (+’ve), and walking distance to nearest road (-’ve).

Table 8.10

School-based regression results Dependent variable

RHS Variable: Average student age Public school Private school Missionary school Student-to-teacher ratio Equipment Number of primary school teachers Average years of teacher training Average years teaching Teacher’s walking distance from school (minutes) Teacher teaches more than one grade (1 = yes) Number of shifts the teacher teaches Teacher speaks local language Teacher speaks Portuguese (rank 0, 1, 2) Average monthly salary (’000 CFA) Constant N R2 RMSE

Literacy score Coef. (se)

Numeracy score Coef. (se)

0.463*** (0.049) –0.040 (0.188) 0.669** (0.285) 0.095 (0.530) –0.002 (–0.003) –0.006 (0.020) 0.727*** (0.16) –0.086 (0.082) –0.002 (0.009) 0.001 (0.001) –0.300* (0.162) –0.231 (0.185) 0.064 (0.115) 0.237* (0.137) 0.008 (.0073) –4.235*** (0.656)

0.526*** (0.062) –0.134 (0.211) 0.741* (0.383) 0.412 (0.717) –0.001 (0.004) –0.018 (0.023) 1.121*** (0.220) –0.121 (0.118) –0.001 (0.012) 0.001 (0.001) –0.221 (0.188) –0.703 (0.260) 0.082 (0.145) 0.425** (0.199) 0.006 (0.00936) –5.571*** (0.880)

280 0.591 1.127

280 0.612 1.322

Note: Regressions based on average variables for schools. Due to missing variables, twentythree schools were excluded. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

The Dire Situation of Children’s Education in Rural West Africa

275

Fig 8.4 Plot of the proportion of children in each school that met national curriculum standards for literacy and numeracy (children age eight to seventeen, schools with at least ten interviewed students) Note: Each point presents the results for one school. For a point 0.1, 0.1 would represent a school where 10 percent of children enrolled in the school achieved the national curriculum standards in both literacy and numeracy. The results cover 185 schools where ten or more children had test scores, and 5,360 students.

8.4.4

Searching for Excellence

Figure 8.4 illustrates the proportion of students at each school who met national curriculum standards for literacy and numeracy. National standards are not demanding. Children enter grade 1 after their sixth or seventh birthday.10 After first grade a child is expected to be able to understand letters and numbers, and read and write simple sentences. At the end of second grade they should be able to read and write short texts, add, and subtract. By grade 3 they should be able to read and write texts, understand basic grammar, and add, subtract, multiply, and divide. We used test scores to calculate whether a child met the age-specific standard implied by the national curriculum (assuming a child did not fail a grade). If a child followed the national curriculum, by the age of ten a child should have been able to easily complete the literacy and numeracy tests.11 10. At the time of the survey when a child passed their seventh birthday they were required to enroll at school, however, parents often enrolled children at age six. Subsequent to this survey, the entry age for grade 1 was changed to six years. 11. Our survey was conducted in the last trimester of the school year, so a ten-year-old child would be nearing completion of either grade 3 or grade 4 if they had begun school at age seven and passed each school year successfully.

276

Peter Boone et al.

We assumed that a child of seven could score zero on a test as they may have just entered school recently, while a child by age eight should have completed only the material of grade 1, and a child age nine should have completed grade 2. We then compared children’s actual scores to what could have been expected if they followed the national standard. We used this data to examine whether any specific schools or villages were “outliers” with very good results compared to the poor national averages. Figure 8.4 shows that there are no good outliers among schools. We found similar results when examining village-level outcomes (not reported). There are no schools where 50 percent of children could meet their standards for numeracy scores, and only six schools where more than half the children met national standards for literacy. We also examined whether schools financed by NGOs showed improved test results. There were thirty-five schools that reported significant NGO financing from two foreign NGOs that have operated for many years in the region. Only 17 percent and 10 percent of children in these schools met national literacy and numeracy standards, respectively. 8.5

Discussion

We conducted a representative survey of educational outcomes in smalland medium-sized villages in Guinea-Bissau. The results illustrate extremely poor literacy and numeracy outcome for children, regardless of whether they attend school. 8.5.1

Study Limitations

There are several limitations to this survey. We have provided a snapshot of the conditions in the spring and early summer of 2010. This was a period of political uncertainty, but during this period the government had made efforts to pay teachers’ salaries on time. Although this would not have impacted test scores quickly, the reported attendance levels and school openings may have been modestly better than other years. Our literacy tests were in Portuguese, and children in rural villages are generally only exposed to Portuguese at school. We managed the interviews in the local languages in order to be able to improve compliance, and to give children their best chance at communicating correct answers. However, the language barrier would naturally reduce literacy scores. We do not see a reason for language barriers to directly impact numeracy scores. Given this is a survey, the correlations reported should not be treated as describing causality. 8.5.2

General Discussion

One goal of this study was to use the unique combination of interviews with parents, children, teachers, and schools in order to better understand

The Dire Situation of Children’s Education in Rural West Africa

277

the key factors determining educational outcomes in the nation. This could provide a background for interventions aimed at improving outcomes in the future. The results provide some evidence that community demand for education is substantial. Nearly 100 percent of heads of households reported they were prepared to pay money in order to improve the education of both boys and girls. On average, they were willing to spend approximately 20 percent of income on children’s education. We found that 24 percent of children attended community schools. Many of these were created and partially financed by local communities. The reasons for dropping out of school also pointed to a healthy desire for education. Girls were reported to mainly drop out when they became pregnant or got married, meaning that they would remain at school until after puberty. Boys were more likely to drop out in order to gain income, but this may also have been associated with fatherhood and marriage. We did not conduct any surveys on the returns to education, but there is no reason to think the pattern in Guinea-Bissau differs from elsewhere in Africa (Kuepie, Nordman, and Roubaud 2009). We did find that their parents reported a willingness to pay for this education, and for the younger children in this study, parents’ views of perceived returns may be most important (Jensen 2010). A literate and numerate child would probably move to the capital, or another major city, where demand for skilled labor is strong. However, if demand for education were truly strong, we should expect communities and households to find means to improve the quality and extent of schooling. There are some good reasons to think that barriers to this are large. The extremely low education levels suggest some communities will not have active members with the information and knowledge needed to intervene successfully in schools. Private schools do exist in urban areas, but they are rare in small villages. This is probably due to the large costs and logistic efforts needed to create, supervise, and maintain schools in these isolated villages. Given the spending power of households is small (with $88 average monthly household income), revenues would be modest. Communities can and do form schools on their own, but to maintain these schools concerted effort and financing is needed. We found students at community schools did not perform better than those at state schools. Anecdotally, we were told in many villages that it is difficult for them to attract good teachers who are willing to stay for long periods. Both seasonal and permanent migration interferes with the sustainability of schools. Further, since families with school-age children change as children grow up, there needs to be a means to ensure regular financing as households who use the schools change. Some ethnic communities, such as the Fula, are known to have stronger community structures than others such as Balanta, however, ethnic background did not correlate strongly with test outcomes (Einarsdóttir 2004).

278

Peter Boone et al.

The fact that NGO-financed schools, with better infrastructure and equipment, did not demonstrate good test scores suggests that it is the operations in the classroom, such as teacher activities and supervision of the outcomes of the school, which may be needed to generate better outcomes.12 One common barrier to all schools is language. Children and teachers have little opportunity to use or improve Portuguese outside the classroom, and while teachers can teach using local languages, the communication barrier when books are used, or literacy is taught, can be large. It is hard to underestimate the difficult task a teacher faces. They work in communities where few adults read and write, there are very few books and written language to be seen, children may not have notebooks or pens and paper, and there is no good lighting at night. The results from our survey suggest there is little pressure coming from teachers or their supervisors to improve outcomes at schools. We are not aware of any attempts to incentivize teachers to improve children’s educational outcomes (Glewwe, Ilias, and Kremer 2010). Nearly all teachers reported they were testing children regularly, and classes appeared to be functioning in most schools. Hence, teachers must have been aware that children were faring extremely poorly at schools compared to national standards. If demand for education services is substantial, while collective action problems prevent the formation of functioning schools, then we should observe that a provision of high-quality services would lead to substantial improvement in learning outcomes, especially compared to the low levels observed in these villages in 2010. A cluster randomized controlled trial has recently been completed in tribal regions of Andhra Pradesh, where public school quality is considered to be very low, to learn whether average test scores of all children in villages can be raised if the children are offered additional high-quality afterschool training. Such research will help discern whether supply, rather than issues related to the demand for schooling, drives the poor results in that region (Eble et al. 2010). There are no randomized school allocations or “natural experiments” that could permit us to examine whether improved supply of quality schooling has a causal impact on outcomes in our survey population. We could only determine indicators of quality through variables such as the availability of textbooks, chalkboards, and some basic characteristics of teachers. The 12. In a study comparing children’s test results in twenty-one sub-Saharan African nations, equipment and the number of school shifts of teachers were highly correlated with test score outcomes. We did not find similar results here. It is possible that reported correlations in our study, as well as this other study, effectively capture indicators of the schooling regime rather than causality. Since no regimes appear to work across our populations, there are only weak correlations between indicator variables and test scores (Fehrler, Michaelowa, and Wechtler 2009). A randomized evaluation of providing textbooks to children in Kenya found no impact on test scores for children receiving textbooks (Glewwe, Kremer, and Moulin 2009).

The Dire Situation of Children’s Education in Rural West Africa

279

results from this analysis were mixed, but generally suggested that improvements in these indicators do not correlate with large changes in educational outcomes. 8.6

Conclusion

Guinea-Bissau is an extremely poor nation with frequent political instability and poor schooling, health, and educational infrastructure. As we prepared our survey in 2009, the president was assassinated. In the midst of our survey in April 2010, the prime minister and army chief of staff were imprisoned during an attempted coup. While other parts of Africa are reported to be improving rapidly on socioeconomic indicators, GuineaBissau is an example of a country that appears trapped in poverty despite official data that claims improvements. Our survey suggests nearly the entire current generation of children in rural villages is growing up innumerate and illiterate. Our survey indicated substantial desire by parents and household chiefs for improved education in villages. Children continue to attend schools despite learning very little, and among schools that were open, teachers were usually in the classrooms. Despite this, little learning is occurring. We believe a main factor driving these poor results is poor quality of teaching within the schools. Teachers are isolated, underequipped, receive salaries after long delays, and have little training. A program that substantially improved conditions for teaching, while introducing strong supervision of teachers and monitoring of student progress, may address some of the key reasons that children are learning little despite attending schools. However, further research is needed to learn which factors are at the heart of Guinea-Bissau’s poor outcomes, and to understand which interventions may be able to change these. If left unattended, educational outcomes of young children in rural Guinea-Bissau are likely to remain dire well into the future.

References Boone, P., and S. Johnson. 2009. “Breaking Out of the Pocket: Do Health Interventions Work?” In What Works in Development: Thinking Big and Thinking Small, edited by J. Cohen and W. Easterly, 55–90. Washington, DC: Brookings Institution Press. Eble, A., V. Mann, P. Bhakta, R. Lakshminarayana, C. Frost, D. Elbourne, and P. Boone. 2010. “The STRIPES Trial—Support to Rural India’s Public Education System.” Trials 11:10. Einarsdóttir, J. 2004. Tired of Weeping: Mother Love, Child Death, and Poverty in Guinea-Bissau. Madison: University of Wisconsin Press.

280

Peter Boone et al.

Fehrler, S., K. Michaelowa, and A. Wechtler. 2009. “The Effectiveness of Inputs in Primary Education: Insights from Recent Student Surveys for Sub-Saharan Africa.” Journal of Development Studies 45 (9): 1545–78. Glewwe, P., N. Ilias, and M. Kremer. 2010. “Teacher Incentives.” American Economic Journal: Applied Economics 2 (3): 205–27. Glewwe, P., M. Kremer, and S. Moulin. 2009. “Many Children Left Behind? Textbooks and Test Scores in Kenya.” American Economic Journal: Applied Economics 1 (1): 112–35. Jensen, R. 2010. “The (Perceived) Returns to Education and the Demand for Schooling.” Quarterly Journal of Economics 125 (2): 515–48. Jerven, M. 2013. Poor Numbers: How We Are Misled by African Development Statistics and What to Do about It. Ithaca, NY: Cornell University Press. Kremer, M. 2003. “Randomized Evaluations of Educational Programs in Developing Countries: Some Lessons.” American Economic Review 93 (2): 102–06. Kremer, M., and A. Holla. 2009. “Improving Education in the Developing World: What Have We Learned from Randomized Evaluations?” Annual Review of Economics 1 (1): 513–42. Kuepie, M., C. J. Nordman, and F. O. Roubaud. 2009. “Education and Earnings in Urban West Africa.” Journal of Comparative Economics 37 (3): 491–515. Luman, E. T., A. Worku, Y. Berhane, R. Martin, and L. Cairns. 2007. “Comparison of Two Survey Methodologies to Assess Vaccination Coverage.” International Journal of Epidemiology 36 (3): 633–41. Ramachandran, V. 2003. Getting Children Back to School: Case Studies in Primary Education. Delhi, India: Sage India. Schultz, T. Paul. 2004. “School Subsidies for the Poor: Evaluating the Mexican Progresa Poverty Program.” Journal of Development Economics 74 (1): 199–250. World Bank. 2012. World Development Report 2012: Gender Equality and Development. Washington, DC: World Bank. Young, A. 2012. “The African Growth Miracle.” Journal of Political Economy 120 (4): 696–739.

9

Success in Entrepreneurship Doing the Math Michael Kremer, Jonathan Robinson, and Olga Rostapshova

9.1

Introduction

Outside of agriculture, the family- owned business is the most common form of enterprise in low- income countries. These types of businesses employ hundreds of millions of people (e.g., Banerjee and Duflo 2007; World Bank 2015), yet there is tremendous heterogeneity in how such firms perform. For instance, in the retail sector, some firms hold large inventories and earn significant profits, while others hold minimal stocks and provide little more than subsistence income for their owners. However, it is an open question why there is such heterogeneity in the success of small firms. This chapter examines these issues in the context of one specific industry, retail, in a region of western Kenya. A companion paper (Kremer, Robinson, and Rostapshova 2013b) calculates bounds on the marginal returns to inventories for a sample of retail firms in western Kenya. The results suggest some firms have high returns to Michael Kremer is the Gates Professor of Developing Societies in the Department of Economics at Harvard University and a research associate of the National Bureau of Economic Research. Jonathan Robinson is associate professor of economics at the University of California, Santa Cruz, and a research associate of the National Bureau of Economic Research. Olga Rostapshova is a technical director with Social Impact. We thank Louis Kasekende, David Laibson, Isaac Mbiti, David Weil and participants at various seminars and conferences for helpful comments. Abdullah Al-Sabah, Kenzo Asahi, Pia Basurto, Dan Bjorkegren, Conner Brannen, Elliott Collins, Sefira Fialkoff, Katie Hubner, Eva Kaplan, Anthony Keats, Jamie McCasland, and Russell Weinstein provided excellent research assistance. We gratefully acknowledge funding from the SEVEN Foundation, the Kauffman Foundation, and the NBER Africa program. We thank IPA-Kenya for administrative support. For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see http://www.nber.org/chapters/c13449.ack.

281

282

Michael Kremer, Jonathan Robinson, and Olga Rostapshova

inventories, but that returns vary greatly across firms. The study leverages administrative data from a wholesaler on whether firms purchased enough to take advantage of quantity discounts. While many orders just qualify for a discount, many others are just below the quantity threshold. In our sample, the median shop misses at least some opportunities to earn rates of return in excess of 100 percent (by increasing one purchase slightly to meet the bulk discount threshold and correspondingly reducing the next purchase).1 These results are consistent with a series of papers suggesting that many small firms in developing countries face high returns to capital (e.g., de Mel, McKenzie, and Woodruff 2008; Duflo, Kremer, and Robinson 2011; McKenzie and Woodruff 2008; Fafchamps et al. 2011; Udry and Anagol 2006; Banerjee and Duflo 2012). This study focuses not on the marginal return to inventory but on the correlates of overall inventory and profit levels. Background data was collected on a large sample of retail shops using detailed surveys on a host of firm and owner characteristics. The survey included vocabulary and reading tests in English and Swahili, a math problem- solving test, a digit- recall memory test, Raven progressive matrices, and a maze completion speed test. The survey included modules on demographics; access to savings and credit; ownership of land, durable goods, and other assets; transfers given and received; income; and financial record keeping and other business practices. Incentivized modules to measure time preferences and small- stakes risk aversion were also administered. To measure time preferences, respondents were asked to choose between several schedules of time- dated cash payouts. To measure small- stakes risk aversion, subjects were asked to divide a portfolio of 100 KSh (approximately $1.33) between a safe asset and a risky asset that paid zero with 50 percent probability and 2.5 times the amount invested with 50 percent probability. This chapter examines two main business outcomes: inventory size, and profits. For inventories, several correlates are identified. Firm owners with higher math scores, younger firm owners, and firm owners who are less risk averse over small stakes have larger inventories. Shopkeepers with bank accounts also tend to have somewhat larger inventories. Consistent with our findings that firms with larger inventories have higher profits, these differences translate into profits as well. Firm owners with higher math scores earn higher profits, even conditional on inventories. The rest of this chapter is organized as follows. Section 9.2 reviews the 1. We also measure the return to increased investment by surveying shops on a regular basis about “stock outs”: lost sales due to insufficient inventory. The average bounds on returns are more modest with this approach—the average shop in our sample would achieve a real rate of return of 33 percent to a marginal increase in inventory per year, and 17.6 percent of shops have returns greater than 50 percent. However, if lost customer goodwill or other sales of complementary goods are significant, this will be a very loose lower bound on the rate of return. We were also able to reject the hypothesis that the marginal rates of return are equal across shops.

Success in Entrepreneurship

283

related literature, section 9.3 describes the data, section 9.4 discusses the results, and section 9.5 concludes. 9.2 9.2.1

Literature Review Personality Characteristics and Entrepreneurship

This work fits into a broad multidisciplinary literature that aims to identify factors correlated with entrepreneurship entry and success. Most studies in this field have examined the link between various personality characteristics and the decision to enter entrepreneurship. A number of papers argue that characteristics such as risk and autonomy preferences, innovative orientation, and locus of control predict entrepreneurship.2 Hartog, van Praag, and van der Sluis (2010) use the US National Longitudinal Study of Youth to examine the effects of various personal characteristics among entrepreneurs and employees. They find that verbal abilities appear to be more important for employees, while mathematical, technical, and social abilities are more important for entrepreneurs. They also argue that general ability and balance across the various kinds of ability generate higher incomes for entrepreneurs. Caliendo, Fossen, and Kritikos (2010) find that individuals with intermediate levels of risk tolerance survive longer as entrepreneurs than those with very high or very low levels of risk tolerance. Fairlie and Holleran (2012) find that more risk- tolerant individuals and those with a preference for autonomy benefit more from business training. A smaller literature has focused on entrepreneurs in developing countries. De Mel, McKenzie, and Woodruff (2008, 2009a, 2010) present a large amount of evidence from a sample of small firms in Sri Lanka. In de Mel, McKenzie, and Woodruff (2010), the authors find evidence that microentrepreneurs in Sri Lanka bear more similarities to wage workers in developed countries than to owners of large firms.3 De Mel, McKenzie, and Woodruff (2008) find that capital injections are most beneficial to those that score highest on a Digits Forward memory test and de Mel, McKenzie, and Woodruff (2009a) find that several owner characteristics predict innovation (defined as new or significantly improved product, process, marketing, or organization changes). Predictive characteristics include Raven’s test scores, optimism measures, previous experience, and time preferences. 2. For example, see Zhao and Seibert (2006), Blanchflower and Oswald (1998), Rauch and Frese (2007), Caliendo, Fossen, and Kritikos (2010), Evans and Leighton (1989), and Puri and Robinson (2009). See Rauch and Frese (2007) for a meta- analysis of this literature. 3. A meta- review of empirical work in developing countries (Van der Sluis, Van Praag, and Vijverberg 2005) finds that more educated workers are more likely to be in wage employment than in nonfarm entrepreneurship, and find that the effect is stronger for women, and in least developed countries.

284

Michael Kremer, Jonathan Robinson, and Olga Rostapshova

9.2.2

Returns to Capital, Credit, Business Training, and Microcredit

A host of recent studies have found extremely high marginal rates of return to capital in developing countries (see, for example, de Mel, McKenzie, and Woodruff 2008; McKenzie and Woodruff 2008; Banerjee and Duflo 2012; Udry and Anagol 2006; Fafchamps et al. 2011). However, such high returns are puzzling; it is unclear why firms cannot realize these returns and why owners of small firms do not accumulate more capital until rates of return fall to a more conventional level, as would be suggested by a standard Euler equation. While a possible explanation would be that credit constraints are binding, this seems unlikely to be the whole story, since the Euler equation would hold even under credit constraints. Relatedly, a number of recent randomized experiments on microfinance have found limited take-up for microcredit loans, at least as currently offered by microfinance providers (Bannerjee, Karlan, and Zinman 2015). Another explanation for high unrealized returns is that human capital constraints are binding. Most studies of providing standard business training to entrepreneurs find limited effects (see Karlan and Valdivia [2010] in Peru; Giné and Mansuri [2011] in Pakistan; Bruhn and Zia [2011] in Bosnia and Herzegovina; and the standard business training provided in Drexler, Fischer, and Schoar [2014] in the Dominican Republic). Fairlie, Karlan, and Zinman (2015) find similarly small effects of business training among potential entrepreneurs in the United States. By contrast, Drexler, Fischer, and Schoar (2014) find larger effects from “rule of thumb” training in the Dominican Republic, which emphasized basic cash management strategies (such as keeping separate accounts for the business and for personal consumption). Working with much larger firms, Bloom et al. (2013) find that providing basic management consulting to Indian textile firms increases total factor productivity by 18 percent, even though many of the changes implemented were already available to firms. Bruhn, Karlan, and Schoar (2012) also find large effects of management consulting among Mexican firms. 9.3 9.3.1

Data The Small-Scale Retail Sector in Kenya

The small- scale retail sector comprises a significant share of economic activity in Kenya, particularly in rural areas. For example, Daniels and Mead (1998) estimate that small and medium enterprises with ten or fewer employees (not including agriculture and mineral extraction industries) comprise 12– 14 percent of total Kenyan gross domestic product (GDP), and that a quarter of this contribution comes from the retail trade. This study focuses on a category of retail shops in western Kenya called

Success in Entrepreneurship

285

dukas in Kiswahili. These shops are typically owner operated and are ubiquitous in market centers and small towns in the region. They are often located in clusters, adjacent or in close proximity to several competing shops and retail fast- moving consumer goods (FMCG), or consumer- packaged goods (CPG). The FMCG manufacturing industry is highly concentrated. Manufacturers set retail prices, and a single supplier has a very high market share. Western Kenya is relatively poor, and people tend to buy a fairly small set of goods. Consequently, shops typically sell a relatively homogeneous set of goods, which include basic household products such as perishable and nonperishable foodstuffs, soaps, detergents, cooking fat, sodas, phone cards, and other household items. The more successful shops tend to stock a wider variety of goods, rarely stock out of items, and purchase sufficiently large quantities to qualify for bulk discounts. Several features of the industry may make it possible for less productive firms to survive, rather than be driven out by more efficient competitors. First, efficient shopkeepers are typically not able to manage multiple shops in different locations. Whether this is due to labor market frictions, moral hazard with employees, or other factors, such constraints set a limit on the scope of firm operations. Second, since many customers tend to travel a short distance to purchase items, shops face only limited competition. Finally, manufacturers preclude price competition by fixing retail prices. 9.3.2

Sampling

Some of the data used in this analysis was collected for a companion paper (Kremer et al. 2013b). Firms were sampled from the administrative records of a large distributor of household goods. From the universe of shops in this database, a sample of 854 shops was identified for surveying, spanning a relatively wide geographical area in western Kenya.4 Background surveys were successfully completed with 731 of these shops (85.6 percent) in 2009– 2010. The remaining shops were untraceable, mostly because they had closed. This chapter focuses on the dependent variables of inventory size and profits. The first two measures were collected as part of an end- line survey in our companion paper. Of the 731 firms with baseline data, 486 were sampled for an end- line survey that includes the inventory measures used in this chapter, and which was administered approximately 1.5– 2 years after the baseline surveys (between February and May 2011). Enumerators successfully completed surveys with 380 of these shops (78.1 percent). Approximately halfway through the administration of these surveys, the profit module was added to the questionnaire. Therefore, there is a smaller number of observations for that variable (188). 4. The inclusion criterion was that the shops were close enough for enumerators to be able to travel to the shop to conduct surveys.

286

Michael Kremer, Jonathan Robinson, and Olga Rostapshova

Lastly, survival information on shops was collected when enumerators visited market centers for the end- line survey. The data set contains useable information on 700 of the 731 shops (95.8 percent). Information on the remaining shops, unfortunately, could not be obtained. 9.3.3 Survey Data Background Surveys The background survey gathered information on a number of standard demographic measures such as the owner’s age, sex, ethnicity, educational attainment, literacy, the size of the owner’s family, and how long the shop had been open. The survey also included questions on the shop owner’s access to and use of savings and credit; self- reported credit constraints; land, durable goods, and other asset holdings; and other sources of income. In addition, survey instruments included tests of cognitive ability and various psychological measures, including time and risk preferences, as well as attitudes toward entrepreneurship. Established cognitive tests were adapted to the local context to measure language and math ability, memory, and general reasoning ability. Specifically, the survey included vocabulary and reading tests in English and Swahili, a math problem- solving test, Wechsler Memory Scale Digits Forward test, Raven progressive matrices, and a maze completion speed test. For the analysis, the results of several of these variables were aggregated for the cases where there was substantial correlation (for example, measures of vocabulary and language, as well as competency in English and in Swahili). This chapter focuses specifically on the math test results. This test was adapted from standard psychometric and personnel IQ tests, including the Wonderlic Test and Cognitive Reflection Test and included eleven questions that ranged from simple math calculations (e.g., “A notebook costs 21 KSh for each one. What will four notebooks cost?”), to more challenging analytical reasoning questions (e.g., “In a lake, there is a patch of water hyacinths. Every day, the patch doubles in size. If it takes 48 days for the patch to cover the entire lake, how long would it take for the patch to cover half of the lake?”). The questions used in this analysis are included in the appendix. Time and risk preferences were assessed by asking respondents to make choices over a variety of incentive- compatible choice sets. To incentivize truth telling, one question was randomly chosen at the end of the interview and the corresponding amount was paid to the respondent. To measure small- stakes risk aversion, respondents were asked how much of a given amount of money they would like to put into a risky asset. The money invested in the risky asset was multiplied by 2.5 with 50 percent probability and was completely lost the other half of the time. The remainder of the money was kept with certainty.

Success in Entrepreneurship

287

To estimate time preferences, respondents were asked to choose between a given amount of money (40 Kenyan shillings, or about $0.50) at a particular date and a larger amount one month later. To estimate time consistency, these questions were asked over two time frames: (a) immediately versus one month in the future, and (b) one month in the future versus two months in the future. In the analysis, respondents were coded as being “patient” if they preferred to wait one month for 55 KSh to taking 40 KSh now. In addition, several measures of time consistency were constructed: (a) “timeconsistent” individuals who exhibit the same discount rate in the present as in the future, (b) “present- biased” individuals who are more patient in the future than in the present, (c) “patient now but impatient later” individuals who are more patient in the present than in the future, and (d) people who exhibit the maximal discount rate in both the present and the future (these people prefer 40 KSh earlier to 500 KSh a month later, no matter what the time period is). Finally, respondents were asked questions related to their entrepreneurial disposition and other attitudinal characteristics. Questions developed by the World Values Survey were used to the extent possible, and were adapted to the local context. Inventory and Profit Surveys Self- reported profits and inventory were collected between February and May 2011, approximately 1.5– 2 years after the background surveys. As shops stock a large number of products, it was too burdensome for the respondents to measure inventories product by product. Respondents were therefore asked to estimate the total value of their inventory (at both wholesale and retail prices) with the enumerator’s assistance. In addition, the respondent, together with the enumerator, calculated the value of the thirteen most common items stocked by shops. The methodology to measure profits follows de Mel, McKenzie, and Woodruff (2008).5 Respondents were asked to report their income, less expenses and other employee wages, over the previous thirty days. Unfortunately, these questions were included for only a subset of the shops that were surveyed. Therefore, there are only 188 shops with profit data. The mean value of total inventory was 265,200 KSh (approximately $3,500), more than double of Kenyan per capita GDP, and representing twice the mean value of the durable goods and animals owned by the shopkeepers in our sample. Mean reported monthly profits were 23,900 KSh (~$320). Both the inventory and profit distributions were quite skewed: the median total inventory and median monthly profits were almost half of the means, valued at $2,000 and $160 respectively. The 75th percentile was at 5. See de Mel, McKenzie, and Woodruff (2009b) for a discussion of the difficulty in measuring profits, and a justification for using this method.

288

Michael Kremer, Jonathan Robinson, and Olga Rostapshova

$4,000 and the 90th at $9,300. On the other hand, the 25th percentile was at $1,100. Similar distributions were observed for the inventory calculated using the top thirteen items, as well as for profits. Firm Survival The final outcome measure comes from a census conducted in May 2011 (at the conclusion of the end- line surveys). At that time, enumerators visited every shop from the baseline sample to collect data on whether the firm was still in business at that time. Since collecting this information did not require a survey, it could be obtained for a much larger sample of shops (700). While the enumerators attempted to collect data on the reasons for exits, they were unable to obtain good tracking information for most of the shops that had closed; thus, there is no reliable information on what jobs the shop owners took after exiting. Note that it is difficult to say whether survival of a retail shop is a positive or negative outcome in this setting; it may well be that other, more formal jobs are preferred to small retail employment. Summary Statistics Summary statistics for the 380 shops in the inventory analysis are presented in table 9.1.6 The average shopkeeper has 10.8 years of education, substantially more than the typical rural resident in the area, and 97 percent can read and write in Kiswahili. The mean shop has been around for almost 7.5 years, and the average shopkeeper is about thirty- three years old. Fiftysix percent of shopkeepers are male. Shopkeepers are substantially wealthier than the average rural resident. About 13 percent of owners or their spouses have formal- sector jobs, a figure which is much higher than for the typical rural residents. In addition, 83 percent of shopkeepers in our sample have bank accounts, 42 percent participate in a merry- go- round cooperative (ROSCA), and the average shopkeeper owns 1.95 acres of land. Inventory value and income distributions among shopkeepers are skewed, but even at the 25th percentile shopkeepers have relatively high incomes and wealth in an area where typical agricultural wages are approximately one dollar a day. Thirty- seven percent of shops report that they would like to borrow money but are unable to do so, while 31 percent keep financial records. The table also reports information on small- stakes risk aversion. The average shopkeeper invested a bit more than 50 percent into the risky asset. Interestingly, about one- third of the shopkeepers divided money exactly equally between the risky and safe assets, a result which is consistent with workers in the United States who follow the “1/n” heuristic of investing in retirement assets (i.e., Benartzi and Thaler 2001, 2007). Since individuals 6. Summary statistics for the larger sample of shops with survival data are presented in table 9A.2.

Success in Entrepreneurship Table 9.1

289

Summary statistics for sample of shops with inventory data Quantiles

Inventories, profits, and credit to customers Total inventory Inventory in top 13 items Profits in past month Gives out credit to customers Amount given out in credit in past month Background characteristics Years of education Years shop open Male Married Age Can read and write (Swahili) Asset ownership and formal-sector income Owner or spouse has a formal sector job Acres land owned Value of durable goods and animals owned Financial access Has bank account Participates in ROSCA Would you like to borrow more money but are unable to get it (percentage “yes”) Small-stakes risk aversion Percentage invested in risky asset Time preferences “Patient:” Prefers 55 KSh or less in 1 month to 40 KSh today Time consistent Present biased Patient now but impatient later Maximally impatient in present and future Attitudinal measures Work importance index Financial recordkeeping Always keeps financial records

Mean (1)

Std. dev. (2)

25th (3)

50th (4)

75th (5)

90th (6)

26.52 9.42 2.39 0.92 1.12

29.67 11.34 2.70 — 5.75

8.00 2.80 0.80 — 0.10

15.00 5.16 1.20 — 0.20

30.00 10.81 3.00 — 0.50

70.00 24.93 5.60 — 1.50

10.80 7.47 0.56 0.79 33.36 0.97

3.33 5.60 — — 9.48

8.00 3.16 — — 27.00

11.50 6.41 — — 32.00

12.00 9.90 — — 38.00

16.00 14.56 — — 46.00

0.13 1.95 11.70

— 2.64 15.32

— 0.00 4.20

— 1.00 7.00

— 2.50 11.91

— 4.50 25.50

0.83 0.42

— —

— —

— —

— —

— —

0.37

—

—

—

—

—

0.56

0.20

0.50

0.50

0.70

0.80

0.08 0.20 0.52 0.06 0.22

— — — — —

— — — — —

— — — — —

— — — — —

-— — — — —

0.36

0.19

0.25

0.25

0.50

0.50

0.31

—

—

—

—

—

Notes: There are 380 shops in the sample. Sample size differs for some variables. All monetary values in 10,000 Kenyan shillings. Exchange rate was roughly 75 KSh to $1.

should be close to risk neutral over such small stakes (i.e., Rabin 2000), the lumping at an equal division of assets suggests that investment behavior is not consistent with expected utility maximization. In terms of time preferences in laboratory games, respondents appear very impatient on average. Only 8 percent of people are willing to wait a month

290

Michael Kremer, Jonathan Robinson, and Olga Rostapshova

for 55 KSh instead of receiving 40 KSh immediately. While this may suggest high discount rates, this measure may be confounded by trust concerns as well. Only 20 percent of people are time consistent, whereas 52 percent are present biased. Of the attitudinal variables, a “work importance” index was constructed by averaging responses to the following four questions: (a) a binary variable equal to 1 if the respondent answers “all the time” to the question “How often do you think about your business?”; (b) a binary variable taking value 1 if the shopkeeper reports that she tends to choose work over family; (c) a binary variable taking value 1 for shopkeepers who say that work is “very important”; and (d) a binary variable taking value 1 for those who say their goal in the business is to “make a lot of money.” The average of this index is 0.36 in the sample. Finally, the last row of the table shows that only 31 percent of shops always keep financial records. Mathematical Ability Figure 9.1 shows the distribution of scores on the eleven- question mathematical test.7 There is quite a bit of variation in scores. The interquartile range is 0.36– 0.6, while the mean and median are 0.49 and 0.45, respectively. Table 9.2 shows how scores on the math test are correlated with other measures, and examines the relationship between mathematical ability, education, and other cognitive measures (note that this table includes data from all shopkeepers who completed the math test in the baseline sample, whether or not they were in the end line). In column (1), the standardized math score is regressed on years of education. The correlation is positive, large, and highly significant; an additional year of education is associated with an additional 0.13 standard deviation score on the math test. Column (2) includes other controls, including standardized measures of digit recall, maze completion times, Raven’s matrix score, and a combined English/Kiswahili language score. These covariates somewhat attenuate the effect of education (from 0.13 to 0.07 standard deviations per year of education), but the coefficient is still highly significant. The results do not suggest statistically significant correlations between the math score and either the Raven’s score or the language score. As will be discussed in greater detail in the following sections, these correlations are important for two reasons. First, they provide some reassurance that these measures contain some signal. Second, the correlation with education at least suggests that mathematical ability is not an innate individual characteristic, but is something that could potentially be improved with training. 7. Note that some subjects did not answer all questions, and some questions were filled in incorrectly or were subject to data entry errors. Therefore, data is not available for answers to all questions for all respondents.

Fig. 9.1

Distribution of scores on math test

Table 9.2

Correlates of standardized math score

Years of education

(1)

(2)

0.13 (0.01)***

0.07 (0.01)*** 0.03 (0.04) –0.02 (0.04) 0.26 (0.04)*** 0.09 (0.04)** 620 0.22

Digit recall (standardized) Seconds to finish mazes (standardized) Raven’s matrix (standardized) Combined language score (standardized) Observations R-squared

670 0.19

Notes: The dependent variable is the score on the math exam. Regressions include all firms with either inventory/profit data, or survival data. To avoid dropping observations, we create dummy variables for having missing information for a given variable and code the underlying variable as a 0 when it is missing. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

292

9.4 9.4.1

Michael Kremer, Jonathan Robinson, and Olga Rostapshova

Results Inventory Size

First, the multivariate correlates of inventory size are considered, using the two measures of (log) inventories, both in Kenyan shillings. These results are presented in table 9.3. In columns (1)–(3), the dependent variable is log total inventory on all items; in columns (4)–(6), it is log inventory on the top thirteen items. In the first specification (columns [1] and [4]), only those variables that are most plausibly exogenous are included. In addition to demographic characteristics, these include measures of cognitive ability, small- stakes risk aversion, asset ownership, and income from formalsector jobs. In the second specification (columns [2] and [5]), measures of financial access are added. In the final specification (columns [3] and [6]), time preferences, attitudinal measures, and financial record keeping are also included. The most compelling associations emerge with respect to the cognitive measures. There is strong evidence that shopkeepers’ performance on the math test predicts inventory size. As with small- stakes risk aversion, this association is robust to controlling for measures of credit constraints and other variables. A one standard deviation increase in the math score is robustly associated with 16– 18 percent higher inventories. Raven’s matrix scores are also significant in predicting the inventory of the top thirteen products (columns [4]–[6]), but are not significant with regard to total inventory. The only demographic variable correlated with inventories is the shopkeeper’s age; firms with younger owners tend to have larger inventories. There is some evidence that some measures of credit constraints may be important, though the overall pattern of results does not provide definitive evidence one way or the other. Higher levels of other assets are weakly correlated larger inventories. However, there is no significant correlation between inventories and self- reported credit constraints, land ownership, or formal- sector employment. There is some evidence that those with bank accounts have larger inventories, while members of savings circles called ROSCAs (Rotating Savings and Credit Associations) have smaller inventories (Kremer et al. 2013a). While it is difficult to interpret this causally, one possible explanation is that most shopkeepers have bank accounts (as can be seen from table 9.1, 82 percent have bank accounts) and thus shopkeepers who save in ROSCAs may be worse off than the average shopkeeper. There is strong evidence that small- stakes loss aversion is significantly correlated with lower inventories. This correlation is robust to controlling for measures of credit constraints, as well as for other variables. Shopkeepers who invested 10 percent less of the 100 KSh portfolio (i.e., 10 KSh) in the risky asset had approximately 6.6– 8.1 percent higher inventories. While it

Table 9.3

Correlates of inventories Log inventory on top 13 products

Log total inventory

Background characteristics Years of education (tens of years) Years shop open (tens of years) Age Cognitive measures Math score (standardized) Raven’s matrix (standardized) Digit recall (standardized) Seconds to finish mazes (standardized) Combined language score (standardized) Small-stakes risk aversion Percentage invested in risky asset (out of 100 KSh) Asset ownership and formal-sector income Owner or spouse has a formal-sector job Log (acres land owned + 1)

(1)

(2)

(3)

(4)

(5)

(6)

–0.06 (0.20) 0.16 (0.10) –0.01 (0.01)**

–0.09 (0.20) 0.12 (0.10) –0.02 (0.01)***

–0.14 (0.20) 0.16 (0.10) –0.01 (0.01)**

0.04 (0.21) 0.20 (0.11)* –0.02 (0.01)***

–0.01 (0.21) 0.17 (0.11) –0.02 (0.01)***

–0.06 (0.21) 0.20 (0.11)* –0.02 (0.01)***

0.18 (0.06)*** 0.07 (0.07) –0.02 (0.07) 0.03 (0.07) 0.02 (0.07)

0.18 (0.06)*** 0.05 (0.06) –0.03 (0.07) 0.03 (0.07) 0.04 (0.07)

0.16 (0.06)*** 0.06 (0.06) –0.03 (0.07) 0.03 (0.07) 0.05 (0.07)

0.14 (0.06)** 0.15 (0.07)** –0.04 (0.07) 0.03 (0.07) 0.01 (0.07)

0.14 (0.06)** 0.13 (0.07)** –0.05 (0.07) 0.03 (0.07) 0.03 (0.07)

0.12 (0.06)* 0.14 (0.07)** –0.06 (0.07) 0.03 (0.07) 0.04 (0.07)

0.66 0.51 (0.24)*** (0.24)**

0.54 (0.24)**

0.39 (0.25)

0.00 (0.17) 0.07 (0.08) 0.22 (0.13)*

–0.03 (0.17) 0.08 (0.08) 0.22 (0.13)*

–0.12 (0.18) 0.05 (0.09) 0.17 (0.13)

–0.14 (0.18) 0.06 (0.09) 0.17 (0.13)

0.19 (0.13) –0.48 (0.12)*** 0.00 (0.11)

0.20 (0.13) –0.48 (0.12)*** 0.01 (0.11)

0.27 (0.14)** –0.50 (0.12)*** –0.02 (0.12)

0.29 (0.14)** –0.49 (0.12)*** –0.01 (0.12)

0.79 0.81 (0.23)*** (0.23)***

–0.04 (0.18) 0.03 (0.08) Log (value of durable goods and 0.25 animals owned + 1) (in 10,000 KSh) (0.13)* Financial access Has bank account Participates in ROSCA Would like to borrow more money but is unable to get it Time preferences, attitudinal measures, and financial record keeping “Patient:” Prefers 55 KSh or less in 1 month to 40 KSh today Time consistent Present biased Patient now, impatient later

–0.05 (0.27) –0.04 (0.14) –0.40 (0.17)** –0.31 (0.30)

–0.15 (0.18) 0.02 (0.09) 0.21 (0.13)*

0.20 (0.28) –0.12 (0.14) –0.26 (0.18) –0.17 (0.31) (continued)

294 Table 9.3

Michael Kremer, Jonathan Robinson, and Olga Rostapshova (continued) Log inventory on top 13 products

Log total inventory (1)

(2)

Work importance index Always keeps financial records Mean of dependent variable Std. dev. of dependent variable Observations R-squared

11.97 1.06 380 0.10

11.97 1.06 380 0.14

(3) 0.64 (0.27)** 0.25 (0.12)** 11.97 1.06 380 0.18

(4)

10.89 1.09 380 0.09

(5)

10.89 1.09 380 0.14

(6) 0.56 (0.28)** 0.27 (0.12)** 10.89 1.09 380 0.17

Notes: Dependent variables in (log) Kenyan shillings. To avoid dropping observations, we create dummy variables for having missing information for a given variable and code the underlying variable as a 0 when it is missing. Regressions also include for gender, marital status, and literacy. Standard errors in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

is possible that the decision of how much to allocate to a risky portfolio is endogenous to business performance such that less successful shopkeepers invest less in the asset, this seems unlikely given the small stakes: 10 KSh represents only around 1/700th of the value of the median shopkeeper’s durable goods and animal assets, and only around 1/1400th of the value of the median respondent’s inventory. There is also evidence that some measures of time preferences are associated with inventories. Though these are less clearly exogenous to inventory size, the patterns are intuitive—present- biased owners hold smaller inventories. However, the measure of “patience” in the present is not correlated with inventories. The work importance index and financial record keeping are correlated with inventories in ways that seem sensible. Owners who place more importance on work at the expense of other activities hold bigger inventories, as do owners who always keep financial records. 9.4.2

Profits

Next, the correlates of shops’ reported profits over the last thirty days of operation are examined (table 9.4). The first three specifications in this table are identical to those in table 9.3. Since inventories and profits are positively correlated, two more specifications are added. Column (4) includes a control for log total inventory size, while column (5) includes log inventory of the top thirteen items. This was done to examine whether the relationships between profits and owner and shop characteristics work entirely through inventories

Table 9.4

Correlates of profits Log profits in past 30 days (1)

Background characteristics Years of education (tens of years) Years shop open (tens of years) Age Cognitive measures Math score (standardized) Raven’s matrix (standardized) Digit recall (standardized) Seconds to finish mazes (standardized) Combined language score (standardized) Small-stakes risk aversion Percentage invested in risky asset (out of 100 KSh) Asset ownership and formal-sector income Owner or spouse has a formal-sector job Log (acres land owned + 1) Log (value of durable goods and animals owned + 1) (in 10,000 KSh) Financial access Has bank account Participates in ROSCA Would like to borrow more money but is unable to get it Time preferences, attitudinal measures, and financial record keeping “Patient:” Prefers 55 KSh or less in 1 month to 40 KSh today Time consistent Present biased Patient now, impatient later Work importance index

(2)

(3)

(4)

(5)

0.17 (0.27) 0.06 (0.14) –0.02 (0.01)**

0.12 (0.28) 0.08 (0.14) –0.02 (0.01)*

0.08 (0.27) 0.09 (0.14) –0.01 (0.01)

0.27 (0.21) –0.02 (0.11) –0.01 (0.01)

0.20 (0.22) –0.05 (0.12) –0.01 (0.01)

0.32 (0.09)*** –0.13 (0.09) 0.13 (0.09) 0.15 (0.09) 0.08 (0.09)

0.32 (0.09)*** –0.13 (0.09) 0.11 (0.09) 0.15 (0.09) 0.11 (0.09)

0.32 (0.09)*** –0.06 (0.09) 0.12 (0.09) 0.15 (0.09)* 0.07 (0.09)

0.16 (0.07)** –0.03 (0.07) 0.00 (0.07) 0.10 (0.07) 0.03 (0.07)

0.23 (0.07)*** –0.11 (0.07) –0.01 (0.08) 0.07 (0.07) 0.05 (0.07)

0.79 (0.32)**

0.80 (0.32)**

0.70 (0.32)**

0.17 (0.25)

0.23 (0.26)

0.17 (0.24) 0.07 (0.11) 0.16 (0.16)

0.18 (0.24) 0.10 (0.11) 0.11 (0.17)

0.03 (0.24) 0.11 (0.11) 0.10 (0.17)

0.13 (0.19) 0.08 (0.08) 0.01 (0.13)

0.22 (0.20) 0.06 (0.09) 0.06 (0.13)

0.25 (0.18) –0.20 (0.16) –0.13 (0.15)

0.23 (0.17) –0.19 (0.15) –0.15 (0.14)

–0.02 (0.14) 0.08 (0.12) –0.13 (0.11)

–0.09 (0.14) 0.01 (0.13) –0.09 (0.12)

–0.81 (0.37)** 0.08 (0.19) –0.11 (0.20) –0.84 (0.45)* 0.74 (0.38)*

–0.33 (0.29) –0.13 (0.15) 0.15 (0.16) –0.31 (0.34) 0.40 (0.29)

–0.54 (0.30)* –0.06 (0.16) 0.08 (0.16) –0.42 (0.36) 0.55 (0.31)* (continued)

296

Michael Kremer, Jonathan Robinson, and Olga Rostapshova

Table 9.4

(continued) Log profits in past 30 days (1)

(2)

Always keeps financial records

(3) 0.36 (0.16)**

Inventory Log total inventory

(4)

(5)

0.16 (0.13)

0.16 (0.14)

0.58 (0.06)***

Log inventory on top 13 items Mean of dependent variable Std. dev. of dependent variable Observations R-squared

9.62 1.00 188 0.17

9.62 1.00 188 0.19

9.62 1.00 188 0.27

9.61 0.97 184 0.56

0.52 (0.06)*** 9.61 0.97 184 0.51

Notes: Dependent variables in (log) Kenyan shillings. To avoid dropping observations, we create dummy variables for having missing information for a given variable and code the underlying variable as a 0 when it is missing. Regressions also include for gender, marital status, and literacy. Standard errors in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

(such that certain types of shopkeepers run larger shops), or whether certain types of people are better able to manage a given amount of inventory. The only covariate that is robustly significant both with and without an inventory control is the math score. Unconditional on inventories, a one standard deviation increase in the math score is associated with 32 percent higher profits; conditional on inventories, the figure falls by about half, suggesting that one of the main channels through which quantitative ability factors into profits is inventory management (columns [4]–[5], table 9.5). However, even conditional on inventories, quantitative ability appears to be an important predictor of profits. Though there are fewer other correlates that are statistically significant, the pattern is generally similar. There are several correlates that are statistically significant when inventories are not included as a control: younger shop owners, shop owners who keep financial records, and shop owners who place higher importance on work all tend to earn higher profits. One somewhat surprising result is that shop owners who are rated as “patient” earn less (i.e., those that prefer 55 KSh or less in 1 month to 40 KSh now). Perhaps these shops have higher returns to inventory. In any case, none of these covariates remain significant when inventories are included (columns [4]–[5]), suggesting that they work principally through an inventory channel. Again, small- stakes risk aversion is positively correlated with profits. The data suggests that this association works mostly through the inventory channel (see columns [4]–[5]): the coefficient is attenuated to about one- quarter

Table 9.5

Correlates of survival Shop still open in May 2011 (1)

Background characteristics Years of education (tens of years) Years shop open (tens of years) Age Cognitive measures Math score (standardized) Raven’s matrix (standardized) Digit recall (standardized) Seconds to finish mazes (standardized) Combined language score (standardized) Small-stakes risk aversion Percentage invested in risky asset (out of 100 KSh) Asset ownership and formal-sector income Owner or spouse has a formal sector job Log (acres land owned + 1) Log (value of durable goods and animals owned + 1) (in 10,000 KSh) Financial access Has bank account Participates in ROSCA Would like to borrow more money but is unable to get it Time preferences, attitudinal measures, and financial record keeping “Patient:” Prefers 55 KSh or less in 1 month to 40 KSh today Time consistent Present biased Patient now, impatient later Work importance index

(2)

(3)

0.02 (0.04) 0.04 (0.02)* 0.00 (0.01)

0.03 (0.04) 0.03 (0.02)* 0.00 (0.01)

0.02 (0.04) 0.04 (0.02)* 0.00 (0.01)

0.00 (0.01) 0.02 (0.01)* 0.00 (0.01) 0.02 (0.01) –0.01 (0.01)

0.00 (0.01) 0.02 (0.01)* 0.00 (0.01) 0.02 (0.01) –0.01 (0.01)

0.00 (0.01) 0.03 (0.01)* 0.00 (0.01) 0.02 (0.01) –0.02 (0.01)

–0.13 (0.05)***

–0.13 (0.05)***

–0.16 (0.05)***

–0.03 (0.03) –0.01 (0.02) 0.02 (0.03)

–0.03 (0.03) –0.01 (0.02) 0.02 (0.03)

–0.02 (0.03) –0.01 (0.02) 0.01 (0.03)

–0.01 (0.03) –0.01 (0.02) –0.02 (0.02)

–0.01 (0.03) –0.01 (0.02) –0.02 (0.02)

–0.04 (0.06) 0.00 (0.03) 0.06 (0.04) –0.01 (0.07) 0.07 (0.06) (continued)

298 Table 9.5

Michael Kremer, Jonathan Robinson, and Olga Rostapshova (continued) Shop still open in May 2011 (1)

(2)

(3)

0.91 0.28 700 0.04

0.04 (0.03) 0.91 0.28 700 0.05

Always keeps financial records Mean of dependent variable Std. dev. of dependent variable Observations R-squared

0.91 0.28 700 0.04

Notes: To avoid dropping observations, we create dummy variables for having missing information for a given variable and code the underlying variable as a 0 when it is missing. Regressions also include for gender, marital status, and literacy. Standard errors in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

of its size, and is no longer significant when inventories are included as a control. 9.4.3

Firm Survival

Table 9.5 examines multivariate regressions correlates of firm survival. Overall, 91 percent of businesses were still open in May 2011, implying an annual exit rate of approximately 4– 6 percent. Few background characteristics predict survival. Unsurprisingly, the longer a shop has been open the more likely it is to survive. The Raven’s matrix cognitive measure positively correlates with survival, while the math score is not predictive at all. None of the proxies for credit constraints are correlated with survival. The only strongly predictive characteristic is small- stakes risk aversion, suggesting that shop owners who are less risk averse over small stakes are actually less likely to remain in business. This last result suggests that survival may not be a positive outcome for shopkeepers (unfortunately, data on reasons for exit could not be obtained). In particular, exiting small retail into salaried work would typically make people better off. That there are not many robust predictors of survival would be consistent with this view, as would the correlation between scores on Raven’s matrix pattern recognition task and likelihood of survival. Potentially, this suggests that successful retail shops exit into other, more profitable businesses. 9.5

Conclusion

This chapter examines correlates of various measures of entrepreneurial success with shop and owner characteristics among small retailers in western

Success in Entrepreneurship

299

Kenya. The finding that stands out most strongly is a very high correlation of math scores with our measures of success. Math scores seem to be a robust predictor of inventories. They also strongly predict profits, and while one channel seems to be inventories, math scores appear to important beyond this: the regression of profits on math scores controlling for inventories still shows a significant positive coefficient. Unsurprisingly, math scores are highly correlated with education. One hypothesis worth examining is that programs to improve math skills may be able increase the success of entrepreneurs. Interestingly, other cognitive measures such as language scores, Raven matrix tests, digit recall, and time to complete mazes are not as robustly correlated with entrepreneurial success. There is some evidence that there is a relationship between inventory size and some measures of credit constraints. Other factors seem to be somewhat important as well. Present- biased shopkeepers have smaller shops. Shopkeepers who regularly keep financial records and shopkeepers who report placing a higher value on work over other aspects of their lives have larger shops. However, our analysis does not find correlations of inventories or profits with most shopowner characteristics (i.e., years of education, formal sector income, land ownership, and selfreported credit constraints). A very strong relationship between inventories and small- stakes risk aversion was also observed. As risk aversion over such small stakes is implausible for expected utility maximizers (e.g., Rabin 2000), our companion paper (Kremer et al. 2013a) postulates that the correlation between small- stakes risk aversion and inventory size is due to loss aversion. As discussed in that paper, loss aversion may be an explanation for why millions of firms are unable to exploit potential profit opportunities.

300

Michael Kremer, Jonathan Robinson, and Olga Rostapshova

Appendix Table 9A.1

Math questionnaire

F_1

Look at the row of numbers below. What number should come next? 8 4 2 1 1/2 1/4 ?

|________|

F_2

A notebook costs 21 KSh for each one. What will four notebooks cost?

|________| KSh

F_3

How many of the five pairs of items listed below are exactly the same? Circle the same ones.

F_4

|________|

1

Nieman, K. M.

Neiman, K. M.

2

Thomas, G. K.

Thomas, C. K.

3

Hoff, J. P.

Hoff, J. P.

4

Pino, L. R.

Pina, L. R.

|________|

5

Warner, T. S.

Wanner, T. S.

|________|

|________| |________|

Which one of the numbered figures below is most different from the others? Circle the number.

|________| 1

2

3

4

5

F_5

A bus travels 20 meters in 1/5 second (one-fifth of a second). At the same speed, how far will it travel in three seconds?

|________| meters

F_6

Rope costs 10 KSh a meter. How many meters can you buy for 60 KSh?

|________| meters

F_7

Which number in the following group of numbers represents the smallest amount? Circle the smallest. (a)7 (b) 2/5 (c) 31 (d) 1/3 (e) 2

|________|

F_8

A boy is seventeen years old and his sister is twice as old. When the boy is twenty-three years old, what will be the age of his sister?

|________| years old

F_9

Trousers and a shirt cost 1,100 KSh in total. The trousers cost 1,000 KSh more than the shirt. How much does the shirt cost?

|________| KSh

F_10

If it takes five machines five minutes to make five cars, how long would it take 100 machines to make 100 cars?

|________| minutes

F_11

In a lake, there is a patch of water hyacinths. Every day, the patch doubles in size (becomes twice as big). If it takes forty-eight days for the patch to cover the entire lake, how long would it take for the patch to cover half of the lake?

|________| days

Success in Entrepreneurship Table 9A.2

301

Summary statistics for sample of shops with survival data Quantiles

Inventories, profits, and credit to customers Total survival Survival in top 13 items Profits in past month Gives out credit to customers Amount given out in credit in past month Background characteristics Years of education Years shop open Male Married Age Can read and write (Swahili) Asset ownership and formal-sector income Owner or spouse has a formal sector job Acres land owned Value of durable goods and animals owned Financial access Has bank account Participates in ROSCA Would you like to borrow more money but are unable to get it (percentage “yes”) Small-stakes risk aversion Percentage invested in risky asset Time preferences “Patient:” Prefers 55 KSh or less in 1 month to 40 KSh today Time consistent Present biased Patient now but impatient later Maximally impatient in present and future Attitudinal measures Work importance index Financial record keeping Always keeps financial records

Mean (1)

Std. dev. (2)

25th (3)

50th (4)

75th (5)

90th (6)

26.64 9.47 2.47 0.91 1.11

29.75 11.38 2.87 — 5.72

8.00 2.83 0.80 — 0.10

15.00 5.17 1.20 — 0.20

30.00 10.82 3.43 — 0.45

70.00 25.01 6.00 — 1.50

10.51 6.56 0.52 0.80 33.42 0.97

3.35 5.89 — — 9.57

8.00 2.16 — — 27.00

11.00 4.88 — — 32.00

12.00 9.30 — — 38.00

16.00 14.24 — — 46.00

0.15 1.97 10.83

— 2.92 14.29

— 0.00 3.97

— 1.00 6.56

— 2.50 11.35

— 4.00 19.88

0.82 0.43 0.37

— — —

— — —

— — —

— — —

— — —

0.57

0.20

0.50

0.50

0.70

0.80

0.08

—

—

—

—

—

0.18 0.51 0.05 0.25

— — — —

— — — —

— — — —

— — — —

— — — —

0.35

0.19

0.25

0.25

0.50

0.50

0.28

—

—

—

—

—

Notes: There are 687 shops in the sample. Sample size differs for some variables. All monetary values in 10,000 Kenyan shillings. Exchange rate was roughly 75 KSh to $1.

302

Michael Kremer, Jonathan Robinson, and Olga Rostapshova

References Banerjee, A., and E. Duflo. 2007. “The Economic Lives of the Poor.” Journal of Economic Perspectives 21 (1): 141– 67. ———. 2012. “Do Firms Want to Borrow More? Testing Credit Constraints Using a Directed Lending Program.” Unpublished Manuscript, MIT. Banerjee, A., D. Karlan, and J. Zinman. 2015. “Six Randomized Evaluations of Microcredit: Introduction and Further Steps.” American Economic Journal: Applied Economics 7 (1): 1– 21. Benartzi, S., and R. H. Thaler. 2001. “Naïve Diversification Strategies in Defined Contribution Saving Plans.” American Economic Review 91 (1): 79– 98. ———. 2007. “Heuristics and Biases in Retirement Savings Behavior.” Journal of Economic Perspectives 21 (3): 81– 104. Blanchflower, D., and A. Oswald. 1998. “What Makes an Entrepreneur?” Journal of Labor Economics 16 (1): 26– 60. Bloom, N., B. Eifert, A. Mahajan, and D. McKenzie. 2013. “Does Management Matter? Evidence from India.” Quarterly Journal of Economics 128 (1): 1– 51. Bruhn, M., D. Karlan, and A. Schoar. 2012. “The Impact of Consulting Services on Small and Medium Enterprises: Evidence from a Randomized Trial in Mexico.” Yale Economics Department Working Paper no. 100, Yale University. Bruhn, M., and B. Zia. 2011. “Stimulating Managerial Capital in Emerging Markets: The Impact of Business and Financial Literacy for Young Entrepreneurs.” World Bank Policy Research Working Paper no. 5642, World Bank. Caliendo, M., F. Fossen, and A. Kritikos. 2010. “The Impact of Risk Attitudes on Entrepreneurial Survival.” Journal of Economic Behavior and Organization 76 (1): 45– 63. Daniels, L., and D. Mead. 1998. “The Contribution of Small Enterprises to Household and National Income in Kenya.” Economic Development and Cultural Change 47 (1): 45. de Mel, S., D. McKenzie, and C. Woodruff. 2008. “Returns to Capital in Microenterprises: Evidence from a Field Experiment.” Quarterly Journal of Economics 123 (4): 1329– 72. ———. 2009a. “Innovative Firms or Innovative Owners? Determinants of Innovation in Micro, Small, and Medium Enterprises.” World Bank Policy Research Paper no. 4934, World Bank. ———. 2009b. “Measuring Microenterprise Profits: Must We Ask How the Sausage is Made?” Journal of Development Economics 88 (1): 19– 31. ———. 2010. “Who are the Microenterprise Owners? Evidence from Sri Lanka on Tokman v. de Soto.” In International Differences in Entrepreneurship, edited by J. Lerner and A. Schoar, 63– 87. Chicago: University of Chicago Press. Drexler, A., G. Fischer, and A. Schoar. 2014. “Keeping it Simple: Financial Literacy Training and Rule of Thumbs: Evidence from a Field Experiment.” American Economic Journal: Applied Economics 6 (2): 1–31. Duflo, E., M. Kremer, and J. Robinson. 2011. “Nudging Farmers to Use Fertilizer: Theory and Experimental Evidence from Kenya.” American Economic Review 101 (6): 2350– 90. Evans, D., and L. Leighton. 1989. “Some Empirical Aspects of Entrepreneurship.” American Economic Review 79:519– 35. Fafchamps, M., D. McKenzie, S. Quinn, and C. Woodruff. 2014. “Microenterprise Growth and the Flypaper Effect: Evidence from a Randomized Experiment in Ghana.” Journal of Development Economics 106:211–26.

Success in Entrepreneurship

303

Fairlie, R., and W. Holleran. 2012. “Entrepreneurship Training, Risk Aversion and Other Personality Traits: Evidence from a Random Experiment.” Journal of Economic Psychology 33 (2): 366– 78. Fairlie, R., D. Karlan, and J. Zinman. 2015. “The Effects of Entrepreneurship Training: Evidence from a Randomized Control Experiment.” American Economic Journal: Economic Policy 7 (2): 125– 161. Giné, X., and G. Mansuri. 2011. “Money or Ideas? A Field Experiment on Constraints to Entrepreneurship in Rural Pakistan.” Working Paper, World Bank. http://siteresources.worldbank.org/DEC/Resources/Money_or_Ideas.pdf. Hartog, J., M. van Praag, and J. van der Sluis. 2010. “If You Are so Smart, Why Aren’t You an Entrepreneur? Returns to Cognitive and Social Ability: Entrepreneurs versus Employees.” Journal of Economics and Management Strategy 19:947– 89. Karlan, D., and M. Valdivia. 2010. “Teaching Entrepreneurship: Impact of Business Training On Microfinance Clients and Institutions.” Review of Economics and Statistics 93 (2): 510– 27. Kremer, M., J. Lee, J. Robinson, and O. Rostapshova. 2013a. “Behavioral Biases and Firm Behavior: Evidence from Kenyan Retail Shops.” American Economic Review, Papers and Proceedings 103 (3): 362– 68. ———. 2013b. “The Return to Capital for Small Retailers in Kenya: Evidence from Inventories.” Unpublished Manuscript, Harvard University. McKenzie, D., and C. Woodruff. 2008. “Experimental Evidence on Returns to Capital and Access to Finance in Mexico.” World Bank Economic Review 22 (3): 457– 82. Puri, M., and D. Robinson. 2009. “The Economic Psychology of Entrepreneurship and Family Business.” Unpublished Manuscript, Duke University. Rabin, M. 2000. “Risk Aversion and Expected-Utility Theory: A Calibration Theorem.” Econometrics 68 (5): 1281– 92. Rauch, A., and M. Frese. 2007. “Let’s Put the Person Back into Entrepreneurship Research: A Meta-Analysis on the Relationship between Business Owners’ Personality Traits, Business Creation, and Success.” European Journal of Work and Organizational Psychology 16:353– 85. Udry, C., and S. Anagol. 2006. “The Return to Capital in Ghana.” American Economic Review 96 (2): 388– 93. Van der Sluis, J., M. Van Praag, and W. Vijverberg. (2005). “Entrepreneurship Selection and Performance: A Meta-Analysis of the Impact of Education in Less Developed Countries.” World Bank Economic Review 19 (2): 225– 61. World Bank. 2015. World Development Indicators. http://data.worldbank.org/data - catalog/world- development- indicators. Zhao, H., and S. Seibert. 2006. “The Big Five Personality Dimensions and Entrepreneurial Status: A Meta-Analytical Review.” Journal of Applied Psychology 91:259– 71.

10

The Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa Some Computations Using Data from Ghana Yaw Nyarko

10.1

Introduction

Over the past several decades, African nations have been spending large amounts of their limited government revenues on education, particularly higher education. Many African leaders and many in the press in many African countries often express the view that higher education is critical for African economic development. There are those, however, who criticize spending on higher education because of statistics showing that a high percentage of those who are educated leave the country (the brain drain)—they point to statistics showing that for some countries, around 50 percent of the tertiary educated leave, and that many of those who leave were educated at government expense. The question we pose is fairly simple. Could it be that the huge investments in education, particularly at the tertiary level, were actually the right thing to do during the period we study—roughly the period from postindependence to around the middle of the first decade of the twenty- first century? Specifically, could it be that spending on higher education, knowing full well the extent of the brain drain, could have been the right thing to do for many sub-Saharan African nations, at least in terms of yielding positive and high net returns on investment? Yaw Nyarko is professor of economics at New York University. I thank Bill Easterly with whom I collaborated on a precursor of this chapter. I also thank David N. Weil, Oliver Babson, Murray Liebbrandt, and participants in NBER Africa Project conferences for very useful comments. I also thank the NBER Africa Project for funding related to the research presented here. I thank my research assistants over the past few years for research work related to this chapter: Victor Archavski, Moussa Blimpo, Silvana Melitsko, and Nicole Hildebrant. For acknowledgments, sources of research support, and disclosure of the author’s material financial relationships, if any, please see http://www.nber.org/chapters/c13450.ack.

305

306

Yaw Nyarko

We show that taking into account remittances of brain drainers provides a metric under which the large expenses in tertiary education have been a success via the metric we use. As we will discuss in our concluding remarks, imaginative thinking about these computations could therefore in principle result in ways of increasing the exceptionally low tertiary enrollment rates in many African countries. The chapter begins by setting out a simple model of the role of education in improving incomes of individuals. We focus on tertiary education, as this is the most pertinent for the brain drain from Africa to many countries. It is the loss of the skilled that attracts the greatest amount of attention in the media and in policy debates. At the heart of the exercise is a net present value (NPV) computation, similar to that used in the economics of education literature. In particular, we will study the question of the spending by governments on higher education, focusing on sub-Saharan Africa. The principal pecuniary costs and benefits of spending on higher education are collected. These include costs of the education itself and the benefits of the education among those who stay in the local economy. The analysis explicitly takes into account the fact that many leave—that is, there is a brain drain. Further, and in particular, those who are outside the country also bring in remittances. In addition, many of those who leave return with higher skill levels. We discuss the costs and benefits of the brain drain from different perspectives. We begin by discussing this in the context of a nation or a “village.” Under this perspective we think of the village as paying for the tuition, but also receiving the benefits of the increased remittances. There is also another perspective that is often forgotten in the analysis. That is the perspective of the individual himself or herself. People migrate to seek better lives. If successful, then this should be included in the calculus of the pluses and minuses of the brain drain. Under this perspective we again see that there are positive net benefits to the brain drain. Indeed, this resolves a paradox in the literature on the economics of education that has found very low internal rates of return to tertiary education in many African countries. This is a paradox because it is contrary to what would initially be expected in countries with very low human capital levels seeking rapid economic transformation. In our computations, allowing for the probability of draining and therefore earning large incomes abroad, we obtain relatively large rates of return. This suggests that it is the probability of being a part of the brain drain that results in evident high interest of many to invest their time in tertiary education. This chapter provides the detailed data analyses and empirical implementations of ideas in Easterly and Nyarko (2009) and earlier. As will be described later, some computations have been made in Bollard, McKenzie, and Morten (2010) that are related but different from those of the current chapter, using survey data for Ghana. We are not aware of any other papers that explicitly model the costs and benefits of spending on education taking into account the brain drain in Africa. As we shall point out in each

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa

307

of the subsections, there has been a voluminous amount of research work done on the various elements that go into our computations—remittances, brain drain migration statistics, the value of diaspora both when they are abroad (trading networks) and when they return. Clemens and Pettersson (2008) and Clemens (2007) have also documented positive aspects of the brain drain. 10.2

The Simplest Village Economy

To fix ideas, we now describe a stylized small African village. We imagine the local leaders or the village chief or voters deciding on how much to spend on tertiary education at the university in their village, which has recently been created. We imagine a village economy with small numbers of educated beyond the primary level, modest secondary schooling, and with an extremely limited tertiary- educated stock. The economy also has extremely limited industrial capacity or a tertiary sector. The decision makers need to decide how much to spend on higher education. Those who have finished secondary school level may be able to enter the university system. Since there are so few spots at the tertiary or university level relative to the possible entrants with secondary schooling, the numbers that enter the university level is constrained only by the village government spending. Hence, the total number of seats at the university level is determined by the total spending of the village governments on tertiary education. In particular, at this stage we ignore private schools and tuition paid at government tertiary institutions, each being negligible for many subSaharan nations for the period of interest. We will let c denote the cost per year for educating an individual. It takes Tc years to complete the university education; typically, Tc = 4 but can run from three to five years. It is presumed that the costs are raised from general taxes of the villagers. (Our robustness section will deal further with this assumption and myriad other issues.) Of the tertiary educated, a fraction will be drained off to foreign villages or countries, with the residual fraction remaining in the home village. Those who remain in the local economy earn wages and contribute to the economy. Those who are abroad are assumed to send back home remittances to family members each year they are abroad. The villagers obtain “utility” from having educated people locally around them. In particular, their valuation of educated people is precisely equal ∞to the wage rate they earn, conditional on their being in the village. Let wt(i ) t=0 denote the expected wages of individuals of education level i in the local economy. We shall use i = 0,1,2,3 to denote the education levels of categories “uneducated,” “primary,” “secondary,” and “tertiary,” so that the two designations that will be important here will be i = 2 (secondary) and i = 3 (tertiary). As our emphasis is on the tertiary educated, we shall suppose that it is only the tertiary educated who drain. The village chiefs also value the remittances of those who leave the village.

{ }

308

Yaw Nyarko

These remittances are of the form of transfers to other members of village, ∞ construction of houses in the village, and so forth. Let {Rt }t=0 denote the sequence of expected remittances of an individual, which only happen when outside of the village; in particular, we may think of Rt = 0 when the individual is within the economy. In particular, the village chiefs do not assign a value to the wages received in foreign countries by those who leave for those foreign countries—they care only about their remittances. The village chiefs do not care who gets the remittances just as they do not care who gets to talk to or be serviced by the educated within the local village economy. In particular, distributional issues do not worry them. Incomes (and costs) in the future are discounted by a discount factor. Let r0 denote the rate of interest for those computations, with an implied discount factor of d0 ≡

1 . 1 + r0

Let C and W (i ) denote the discounted costs and wages: ∞

∞

t=0

t=0

C ≡ ∑dt ct and W (i ) ≡ ∑dt wt(i ) for all i, and ∞

R ≡ ∑dt Rt . t=0

The village elders seek to maximize the expected discounted present value of the streams of incomes, W + R of the different types of agents less costs of educating them at the tertiary level, C. At the optimum, the village chiefs will compare the costs of educating the marginal student, from secondary level to tertiary level, to the expected discounted benefit accrued from that marginal student. In this village, drainage occurs right after schooling. There is a probability d that our representative individual will leave the village. If the individual does not leave but instead stays in the village, the individual will stay in the village forever. If the individual leaves (or drains), there is a probability  that the individual will return to the village economy, and a probability 1 –  that the individual never returns. As mentioned earlier, the chiefs of the village receive satisfaction from knowing that their young ones are either employed locally at high wages or are sending remittances to other members of the village. The village chiefs think of there being two possibilities or types of the tertiary- educated representative individual. With probability (1 – d) the individual will not drain and will stay in the local economy. We refer to these types as the locally resident educated (or LRE). The net additional return of the chiefs from such individuals, over and above having them be secondary educated is:

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa

309

NPV LRE = W (3) −W (2 ) − C. Next, with probability d the individual will drain. Conditional on draining, there is a probability of (1 – ) that the individual will drain and never return to work in the village. The chiefs get no wage satisfaction in that state in any period, but will be receive satisfaction from the remittances in that state. If we let RNR denote the net present value of the expected remittances in this state, then the net satisfaction of the chiefs, NPVNR equals the expectation of these remittances less costs of education and relative to their expected contributions (wages) if they did not get tertiary educated:

NPV NR = R NR −W (2 ) − C. Finally, conditional on draining, with probability  the individual will leave the village but eventually come back. Let ∞

W ret ≡ E ∑dt wtret t=0

denote the expected net present value of this sequence of wages in the local economy, with an analogous definition for Rret, the remittances they send when they are outside of the country and which we set to zero in any period t when they are within the village. The net contribution to the chiefs of these eventual returnees is therefore given by: NPV ret = R ret +W ret −W (2 ) − C. Hence, the expected return of the “Drainers” to the chiefs, including both those who never return as well as those who return, will be NPV D = (1 − x ) NPV NR + xNPV ret . The return to the chiefs, taking into account the net contribution of the two types of tertiary educated, those who are local and those who drain, is therefore: (1)

NPV = (1 − d ) NPV LRE + dNPV D .

To see this even more clearly, write the expected utility of the educated in equation (1) above as follows. First, recall we defined Wret to be the expected discounted wages of the eventual returnee individual in the periods when returned to the village. Then W not _ ret ≡ W (3) −W ret is the expected sum of discounted wages that were not received because the individual was abroad. These are the wages that could have been received if the individual was in the home country as opposed to being abroad. A simple rearrangement of equation (1) shows that

310

Yaw Nyarko

NPV = NPVvillage + DNPVabroad where

{

}

NPVvillage = W (3) −W (2 ) − C , and

{

}

{

}

DNPVabroad = d (1 − x ) R NR −W (3) + dx R ret −W not_ret . In particular, the NPV can be broken into two parts. The first part, NPVvillage, is the expected net present value of the increment if there was absolutely no brain drain, and the second, NPVabroad, representing the expected discounted increment of the remittances over local village wages in each of the periods that the individual is abroad. The first term, NPVvillage, is the NPV that would be obtained if there was absolutely no brain drain. We expect this to be positive, although the internal rates of return obtained both here and in the literature are low. The expression NPVabroad represents the impact of the brain drain. To the extent that remittances exceed incomes locally, this expression will be positive. In the exposition below we shall frequently talk about the returns to education without the brain drain and mean the expression NPVvillage and refer to the incremental effect of the brain drain as the term NPVabroad. 10.3

Results for the Village Economy

In the subsequent sections we will be discussing our simple model in great detail, and we will further discuss our data in some detail. In our robustness section we will stress test our model with different parameter value assumptions. In this section we quickly state our main conclusions under some stylized parameter values for Ghana. We will follow some standard procedures in the literature on the economics of education and compute some internal rates of return. We will proceed by providing a quick list of some of the data and parameter values we use. 1. We use the cost data for tertiary education from the United Nations (UN) data sets. 2. We obtain data on the brain drain probabilities, d, from the Docquier and Marfouk (2005) data sets. 3. We obtain the value of  from survey data that suggests a value  = 0.5 as the probability of return and with this taking place at year seven being reasonable assumptions to use. 4. We use the Ghana Livings Survey data (GLSS V) to get the wages,1 W (3) and W (2). 5. We use data on remittances given by the UN International Fund for Agricultural Development (IFAD) surveys, which imply a per migrant re1. See http://www.statsghana.gov.gh/docfiles/glss5_report.pdf.

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa Table 10.1

Internal rates of return and net present values IRR

Remittances data Locally res. Never ret. Returnees Drainers Educated

311

NPV at r = 0.08

UN IFAD (%)

R = US$3,600 (%)

UN IFAD ($)

R = US$3,600 ($)

14 42 41 42 29

14 33 32 32 23

17,229 57,714 37,271 47,492 32,361

17,229 33,026 29,368 31,197 24,213

mittance value of US$5,260 for Ghana, as well as the lower value of $3,600 per person, per year. The first pass of our results can be summarized in table 10.1. We report both the internal rate of return2 (IRR) as well as the net present value at an interest rate of 5 percent. We do this for each of the different types of tertiary- educated types mentioned earlier: locally resident (LRE), never returns (NR), returnees (ret.), the drainers (D), and the tertiary educated as a class (E). We obtain positive and large internal rates of return and net present values. Using the IFAD remittances data we obtain for the tertiary educated as a whole, an internal rate of return of 29 percent and a net present value at r = 5 percent of $32,361. The values are lower, but still large, when we use the lower value of $3,600. What is driving the results should be clear: the remittances of those who leave compensate for the loss from being at home locally. In our robustness section we stress test our model with even lower values of the remittances than the two illustrated above. 10.4

Do W and R Capture True Costs and Benefits?

One big question in this entire analysis is whether the wages and the remittances truly capture the benefits of tertiary education and the value (or losses) associated with those who drain. What are the arguments for and against using W and R as measures of the value of the tertiary educated? We begin by noting that we have included the costs of tertiary education in our computations, noting in our robustness section the obvious issues 2. The internal rate of return is defined as the interest rate at which the net present value is equal to zero. This is often used as a measure of the profitability of an activity or enterprise yielding costs and incomes over time. Although it has many faults, its use is standard in the economics of education literature.

312

Yaw Nyarko

surrounding its use and how sensitive our conclusions are possible errors in the estimates. So here we take out costs in our discussion. So, our first and somewhat feeble defense in the use of wages as the value of the tertiary educated is that it is what is used in most of the returns to education literature. The question we seek to ask is not the moral or philosophical one of what is the value of a human being. Instead, it is the narrow economic question of whether resources used in educating an individual from secondary to tertiary education is worth the resources used for that tertiary education. The question then becomes one of evaluating the benefits of that additional education against its costs. At a first pass, the additional wages that the tertiary- educated person will earn is a proxy for that additional benefit from the schooling. For the village elders in the village economy who are expending resources, the additional resources of their village sons and daughters and the remittances they bring back to the parents in the village will feature heavily in their calculus of the pluses and minuses of the spending. Remittances often end up with members of the village, often the poorest members. This is a plus to the village elders. In our individual calculations, one would presume that wages would be a huge part of one particular person’s cost and benefit analyses of education. Yes, there may be other motivating factors like quality of life, prestige, and so forth. But these are often correlated with wages, and, raw cash itself has to be important too. Let us pursue this question further though. What are other measures that should be used? How else could we objectively measure the value of the tertiary education? Or, to say this in a different way, what are the other possible aspects of the tertiary educated that could be contributory factors in assessing the pluses and minuses? Well, there are a number of possible ways. There are valid factors to include in our analysis, but it is not clear whether these factors help our hurt our general conclusions. In particular, the relative importance of tertiary education and also the brain drain could actually be increased by adding these other factors. We leave these questions for later sections (and later papers). We will, however, mention one of them now. One issue that comes to us continually in the discussion of the brain drain is the issue of having enough critical mass in the home countries. The argument is usually made that if more of the talented Ghanaians, say, would stay in Ghana, they would exert pressure locally to result in change at home. There are, of course, many responses to this argument. First, there is quite a large amount of unemployment among the tertiary educated. It is not at all clear that being in their home countries and dependent on the government for jobs that the tertiary educated would form the effective pressure group often dreamed about in the media. Indeed, having a large diaspora community exerting a fearless independent voice may be much better. Those who have been part of the brain drain and have returned to their home country may have outside options and hence be less afraid of criticizing the local

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa

313

leaders. These are potential pluses of the brain drain, and, when appropriately measured could strengthen our basic argument. 10.5

The Individual-Decision Problem

Above we have treated the costs and benefits from the point of view of the village collectively. The villagers tax themselves to pay for the education of their children, and perceive rewards according to the local wages the children receive when they are locally resident, or the remittances they bring when they are abroad. We can now think of the individual or family perspective. As a first pass at the individual perspective, we can simply follow the existing literature on internal rates of return as follows: In this first pass, suppose the tertiaryeducated benefits are, as before, the incremental wages over and above secondary educated, and that those tertiary- educated wages equal either the local wages if the individual is locally resident or foreign wages when the person is abroad. In particular, notice that this is exactly the same as the exposition for the village economy, but where the remittances are replaced by foreign wages. Let us continue by assuming that the costs are precisely the costs of education as above, which is described in detail, as will be the data on foreign wages. We are computing the social returns to tertiary education so we use the same costs as in the village economy. By replacing remittances by foreign wages, we are getting to the individual- decision problem and individual costs and benefits. This provides a useful comparison with the existing literature: what happens to the rates of return when we add the benefits of the brain drain—we know that for some countries 50 percent or more of the tertiary educated are outside their home countries, so omitting the brain drain is potentially omitting an important factor in the returns to education calculus. The returns we compute would be genuine private returns if individuals pay for their education, which they do not in practice. However, we perform our computations as if they did, and we include as costs the costs of the education. So, what do the net present values (NPVs) and the internal rates of return (IRR) look like? Well, to continue with our computations we require the wages of the typical tertiary- trained individual when abroad. We have data on incomes of Ghanaians in the United States from the US census data. We also have data on Africans in Europe from various Organisation for Economic Co- operation and Development (OECD) databases. These are all described in detail later. In terms of our computations, note that all we have to do is to replace the figure for remittances in the village economy computations with that of wages for the individual computations. The village chiefs have an optimization problem similar to the individual private returns computation except

314

Yaw Nyarko

that when the tertiary educated migrate, the chiefs value them for their remittances while the individual values himself or herself via the foreign wages. Clearly, since we obtained positive net present values (NPV) and internal rates of return (IRR) with the lower remittances values, we expect much higher NPVs and IRR when we perform the individual computations. We will perform our computations under two scenarios. The first is with the parameters we used in the baseline case. The second multiplies the local Ghanaian wages of the tertiary educated and the local costs of tertiary education by a factor of 2. We intend for this to capture any possible miscalculations in those variables. Further, since we are here comparing wages earned while a resident abroad to wages earned while a resident locally, there is an argument for scaling the wages of the tertiary educated up to account for purchasing power parity (PPP) as is standard in, for example, gross national income (GNI) computations.3 We seek to correct for biases that may cause us to underestimate wages and costs of the tertiary educated. In particular, we are not also increasing the base wages of the secondary educated. All types of the tertiary educated, which we focus on here, use the secondary educated as the base in computing net present values, so changes to the secondary wages would affect each group equivalently. For the Ghana data, our results can be summarized in table 10.2, where we have listed the internal rates of return (IRR) as well as the net present values at an interest rate of r = 5 percent (NPV5). We see that both are high for the individual, both under the standard parameters (multiplier = 1 in the table) as well as when we double both the local wages and local costs (multiplier = 2 in the table). For the tertiary educated as a whole, the internal rate of return is 67 percent for the standard parameters and 49 percent when we double the wage and cost parameters. Each of these numbers is very high relative to what is normally presented in the literature. The net present values NPV5 are $126,244 and $149,522, respectively. When we double the costs and wages there is an increase in the IRR and the NPV5 for the locally resident, the IRR goes from 14 percent to 18 percent while the NPV5 goes from $17,229 to $51,192. In this case the benefits to lifetime increases in wages outweigh the effect of the increase in costs (relative to the wages of secondary educated).4 There is a decrease in the values 3. We can obtain the purchasing power parity (PPP) factor from the World Bank World Development Indicators (WDI) by looking at the gross national income and dividing the value in PPP terms by that in current US $. For a variety of reasons, the PPP factor moves quite a bit from year to year. From the World Bank WDI data sets at http://data.worldbank.org/indicator as of Jan 12, 2011, for the years 2005– 2009, the GNI those years in current USD billion was 10.0, 13.3, 18.4, 26.8, and 28.4; while in PPP international USD billion was 25.8, 28.5, 31.4, 34.7, and 36.6, resulting in what we call PPP ratios of 2.5, 2.1, 1.7, 1.2, and 1.3. Ratios for prior years also exhibit similar variability around the value of 2. 4. We remind the reader that we are not also doubling the wages of the secondary school educated—if we did then there would be no change in the IRR and NPV5 of the locally educated as all relevant variables would then have been doubled.

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa Table 10.2

315

Internal rates of return (IRR) and net present value (NPV) comparisons IRR comparisons

NPV comparisons (r = 0.05)

Multiplier on cost and wage

1 (%)

2 (%)

1 ($)

2 ($)

Locally res. Never ret. Returnees Drainers Educated

14 90 90 90 67

18 68 67 67 49

17,229 342,188 128,329 235,258 126,244

51,192 338,409 157,294 247,851 149,522

for those who never return, as they are affected by the increase in costs of education but do not get any of the benefits of the increase in wages—the IRR goes from 90 percent to 68 percent, while the NPV5 goes from $342,188 to $338,409. We stress here that what we are computing are actually “social” rates of return, as we have included the costs of education in the computation of the pluses and minuses. If we excluded the costs of education we would obtain even higher values of the IRR and NPV5. Given the large percentage who travel abroad, and the higher wages abroad relative to within the local economy, the higher returns of the drainers pushes up dramatically the ex ante expected returns to tertiary education as a whole—that is, among the collective group of tertiary- educated drainers and nondrainers. We obtain figures far higher than those in the literature. See, for example, the survey paper Psacharopoulos and Patrinos (2004). The figures in the literature for rates of return to tertiary education in sub-Saharan Africa are of the magnitude of what we obtain in our above tables for the locally resident tertiary educated, as they should be. They are the same measures, except that different authors use different data sets and slightly different methods of estimation. We believe that since there is such a high incidence of brain drain in the countries we are interested in, it is important to include the brain drain in the compuations of the returns to education. When we do include these measures, we obtain extremely high rates of return. This may resolve a certain paradox. Despite the low returns obtained in the literature, many Africans continually clamor for governments to invest in higher education. The media and popular presses all insist on the importance of investments in tertiary education. Students also clamor for the limited slots in the tertiary institutions. This would be paradoxical given the literature’s stated low rates of return to higher education. A possible resolution to the paradox involves people taking into consideration the fact that they may get a change to drain abroad to obtain higher salaries. Our figures show that when these are taken into account, the rates of return to tertiary education as a whole become relatively large numbers.

316

Yaw Nyarko

We now turn to a detailed description of the data in the next few sections. Following that we will deal with some robustness issues—stress testing our model conclusions with different parameter values. 10.6

Cost of Tertiary Education, C

The cost to the government per year for the typical tertiary- educated person is obtained from the United Nations Educational, Scientific and Cultural Organization (UNESCO) data sets. The data on annual costs is often presented in a form that is (after normalizing by or dividing by) that year’s nominal gross domestic product (GDP) per capita. This enables some comparability across nations, albeit imperfect. For all African nations for which we have data, the per student annual tertiary costs have been decreasing over time. For Ghana the costs have gone from 14.8 in 1970 to 2.09 in 2005. Given the recent extremely rapid increases in student numbers in Ghana, our expectation is that per unit costs today for Ghana have dropped significantly below this number. In Burkina Faso, it has gone from a large 29.39 in 1980 to 1.93 in 2005. For example, for Ghana the cost today is about 2 times GDP per capita. For the countries for which there is reasonably current data, the per unit annual costs hover around this number or less. The questions about the costs and how they move over time will be important in our robustness section. For this reason we re- produce below the most current data we have for African nations as well as the year of the data, ordered from the smallest to the highest costs. (See tables 10.3 and 10.4). 10.6.1

Cost Recovery via Loans, Tuition, and Others

There are a number of issues that may distort the calculations above and give us potentially an incorrect view of the costs of the provision of tertiary education. On the one hand, there are various cost recovery and tuition policies by universities. To the extent that there is cost recovery, this would imply an overestimation of costs and an underestimation of the benefits of education in our own computations in the village economy above. We list Table 10.3 Country Botswana Burkina Faso Ethiopia Ghana Mauritius Rwanda Senegal Zambia

Per unit annual costs of tertiary education Early cost

Year

Later cost

Year

6.98 29.39 9.51 14.82 3.56 14.46 4.32 13.27

1975 1980 1995 1970 1980 1970 1980 1970

3.72 1.93 5.74 2.09 0.30 4.04 2.35 1.68

1985 2005 1993 2005 2006 2005 2005 2000

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa Table 10.4

317

Most current cost data Countries

Cost

Years

Libyan Arab Jamahiriya Mauritius Mauritania South Africa Somalia Egypt Tunisia Cameroon Zimbabwe Angola Morocco Cape Verde Togo Namibia Benin Swaziland Zambia Madagascar Guinea Uganda Burkina Faso Ivory Coast Kenya Ghana Sierra Leone Chad Senegal Gambia Congo Liberia Mali Central African Republic Burundi Botswana Niger Rwanda Eritrea Lesotho Mozambique Ethiopia Nigeria Malawi

0.24 0.30 0.41 0.45 0.50 0.54 0.56 0.60 0.64 0.65 0.67 0.74 0.87 0.93 1.14 1.40 1.68 1.75 1.89 1.89 1.93 1.94 2.05 2.09 2.31 2.35 2.35 2.38 2.46 2.58 2.65 2.91 3.49 3.72 3.96 4.04 4.27 5.05 5.32 5.74 10.15 11.77

1999 2006 2005 2004 1970 1980 2005 1999 1985 2005 1996 2005 1970 2002 2002 1980 2000 2005 2005 2004 2005 1994 2000 2005 1985 1996 2005 2004 2002 1975 1999 2006 2005 1985 2006 2005 2002 1994 2004 1993 1970 1992

these below, but argue that they are very small and would not appreciably change our main conclusions. Further, it is unclear whether the costs of education effectively take into account (a) the school capital construction costs and (b) the possibly distortionary effects of the method in which the government raises money to pay for the costs of higher education. We will

318

Yaw Nyarko

revisit these issues in our robustness section, but first we list some of the possible cost- recovery schemes. Again, we will argue that under plausible assumptions our main assumptions still hold. Before we provide the list of these schemes, it may be important to discuss in the context of our model whether and how cost recovery should be treated in the basic optimization problem. There are two ways of viewing the interpretation of costs and, indeed, the entire optimization exercise in entering our village economy above. One interpretation is that the village elders think of “taxing” themselves to pay for the costs of the tertiary education for which they receive as benefits both remittances and the benefit of the services of the educated youth, and that those services are measured by the wages the educated receive. For example, the value of the services of a doctor is higher than the services of a clerical worker, the value of each is measured by their respective wages. Under this interpretation of cost recovery is a net benefit to the village chiefs; they receive the same services at lower net cost. (In this hypothetical exercise we suppose remittances would not be affected by cost recovery.) However, there is another interpretation of the village economy. Suppose that the chiefs have primarily altruistic motivations and caring mostly about the youth, and in particular the net wages the youth receive upon graduation (as well as the remittances). If the students have to pay back to the chiefs a part of the cost, then this should be equivalent to a negative wage for the youth while going to school, and therefore a negative in the altruistic chiefs net present value computation. In particular, for these altruistic chiefs the cost recovery reduces chiefs’ costs and also reduces one for one their perceived benefits. Cost recovery, therefore, would not affect the net present value computation of these altruistic chiefs.5 As mentioned earlier, cost recovery is very small for many sub-Saharan African nations, so the above mentioned arguments are somewhat moot. We now list some possible cost- recovery mechanisms practiced. Tuition and Private Universities If universities charge students tuition, then the government costs are much lower than those computed above. Indeed, one could imagine situations where the government cost of providing education is zero. Indeed, in some private for- profit universities the tuition may even exceed costs, so that there is a surplus for the investors or shareholders of the private university. During the period we have most of our data (up to 2000), the numbers of private universities in Ghana and many other African countries has been small. It is only in the past decade and less that private universities have taken off.

5. I thank David Weil for bringing this point up.

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa

319

Student Loans As with student payment of tuition and enrollment in private universities, student loans could be used as a form of cost recovery. This, however, is not widely used in Africa and even in situations where it is used, it is not clear how much of costs are really recovered (given the low or negative real interest rates, administrative costs, and loan defaults). (See Albrecht and Ziderman 1993). National Service An alternate form of cost recovery is national service. This scheme requires those in universities to work before, during, or after their schooling. The salary is often lower than government civil service salaries, so in principle this could be a form of cost recovery. As noted by Albrecht and Ziderman (1995), however, the cost- recovery benefits of national service are very low. Further, in areas where there is an excess supply of the tertiary educated and therefore unemployment in those ranks, the national service could be considered a benefit to the student and potentially a cost to the government, as the jobs are effectively being guaranteed. Partnerships with Industry Partnerships with industry are often touted as being possible ways of generating income for universities. At this time, however, even for Europe the estimates of the contribution to total costs of universities is low (6.5 percent was the estimate for Germany by Albrecht and Ziderman [1995]), and given their lower industrial bases, this would be expected to have almost negligible contribution to costs for African countries. 10.7

Young Locally Educated Who Stay

Estimates of incremental wages of those with tertiary education will be obtained from living standards surveys. As described earlier, the time series of wages of secondary educated, tertiary educated, and returnees are all needed in making our net expected returns computations. We have explicit data from the Ghanaian data sets, which we proceed to describe. 10.7.1

The GLSS Data Sets

We describe the Ghanaian data set below. The Ghana Living Surveys were taken at three different years: versions 3, 4, and 5 taken, respectively, in years 1991, 1998, and 2005. The surveys asked respondents a series of questions including education levels, age, and wages, among very many others. In the computations presented here we focus on the GLSS 5 data sets, the most recent. The GLSS5 was conducted in 2005– 2006, and covered the entire country with a sample size of 8,687 households. In figure 10.1, we summarize

320

Yaw Nyarko

Fig. 10.1

Number of data points in GLSS survey data

the number of data points—broken down by education level and age level. We will, of course, be most concerned about the secondary- and tertiaryeducation levels in our computations. Next, for our computations we need to compute the wage rate for each education level at each age. The first and direct method is, of course, to take the average wage for each education level and age combination. These aver∞ ages will give us the wage sequences wt(i ) used earlier and shown in figt=0 ure 10.2. We note here that the average wages of the tertiary educated in figure 10.2 may seem low to the casual observer. We did some back- of-the- envelope checks of this data. After university education Ghanaians are required to do national service, considered by some as guaranteed employment for such students, many who may otherwise be unemployed. Those wages were, in the relevant years, around $600 per year. In contrast, public service workers were around $300–$500 per year, with higher amounts in the private sector. There was also anecdotal evidence of a decent amount of unemployment

{ }

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa

Fig. 10.2

321

Earnings of each educational group as a function of age

among the tertiary educated. These facts seem to corroborate the data from the GLSS survey for the tertiary educated.6 One other issue we have to address is whether to include or exclude the wages of those coded in the data as having income of zero. Most, although probably not all, of these are presumably unemployed. In our computations we average wages including those of all the zeros (although we exclude all those with income “N/A” or not available). In table 10.5, we note the ratios of incomes including and excluding the zero incomes. We do this for the three different waves of the GLSS data sets. We compute for each the income ratio between tertiary and secondary educated (as these are the cohorts of interest to us). In particular, columns (2) and (3) of table 10.5 provide the ratio w 3 / w 2 of the average wage of the tertiary educated, w 3 , and the secondary educated, w 2 , with and without the zero income earners. The second column shows the ratio of the average of the tertiary educated to the average of the secondary educated in the GLSS samples, including those with zero income. These range between 2.30 and 2.66, indicating a slight increase in 6. The GLSS 5 Report (September 2008) summarizes the findings of the GLSS 5 survey. Section 9.8 covers household income. Table 9.18 shows mean annual per capita income for all Ghana at GHC397 or about US$433 in prevailing exchange rates. The stated mean annual per capita income for the highest quintile is GHC688 (or about US$750).

322 Table 10.5

GLSS version 3 4 5

Yaw Nyarko Income ratios (tertiary over secondary) and Mincer regression coefficients Income ratio with all incomes included

Income ratio with only positive incomes

Ratio of logincomes with all incomes

Ratio of logincomes with only positive incomes

Mincer regression coefficient

2.30 2.33 2.66

2.29 2.18 2.37

1.07 1.06 1.06

1.06 1.05 1.05

0.69 0.81 0.82

the ratio over the different waves—the tertiary educated are out earning the secondary educated by larger fractions over time. The third column shows the ratio of average wages when we exclude those with zero incomes (presumably due to unemployment). We see then that there is a small decrease in ratio from Wave 3 to 4, then an increase from 4 to 5. We look at respondents who are of age eighteen and older in computing our average wages. Columns (4) and (5) report the ratios of the logs of average incomes, log ( w 3 ) / log ( w 2 ) . Given our purposes and the results presented here, we therefore do not believe that the issues of the zeros and nonzeros will significantly change our results. 10.7.2

Smoothing Data via the Mincer Regressions

As an alternate to using the raw average wages, one could consider smoothing the wages using the Mincer regressions, as is popular in the economics of education literature. In particular, let w denote income, let AGE be the age of the individual, and let SCHOOL be the dummy that is equal to 1 if tertiary educated, and zero otherwise. As is common in this literature, we use age as a proxy for experience. The Mincer regression we run is then given by ln w = a + b0 .AGE + b1.AGE2 + g.SCHOOL + ´. The Mincer regression we report is, , the coefficient on the tertiary schooling dummy variable. Some standard theory, or the interpretations of the theory, in the economics of education literature considers the parameter  to be the returns to schooling and the internal rate of return of that schooling (the interest rate at which the net present value of the incremental return to schooling is zero). We can apply the Mincer regressions to get an estimate of the wage rate as a function of age or experience. We then assume that an individual currently beginning the schooling process will follow that trajectory of wage rates into the future. We then set the wage of an individual of age t and education level i to be equal to the value predicted by the Mincer regression, at those given values of t and i. Figure 10.3 shows the Mincer equation smoothing of the raw income data. First, we note that we obtain very little difference in our results upon using

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa

Fig. 10.3

323

Mincer regression smoothing of incomes

the method with raw averages versus those with the Mincer regressions. We have therefore chosen to report the former. We also note in passing that the issues of inclusion or exclusion of the unemployed arises, just as with the raw nonsmoothed averages. Again, since it does not seriously affect our results, we report only the values with the zero incomes included. Our results are, of course, consistent with standard results in the economics of education literature, when remittances and the brain drain are excluded. Without the latter, the returns to tertiary education in Africa are surprisingly low, as is well documented in the literature. Bloom, Canning, and Chan (2006) and references cited there are a good source for a review of the literature.

324

10.7.3

Yaw Nyarko

Panels and Pseudopanels

Note that we have used a fixed year (2005) cross- section of individuals and used this as a proxy for the evolution of incomes across time. This use of a fixed- time data set to measure life- cycle earnings, although common in the economics of education literature, is still not one would wish for. There are, however, surprisingly few panel data sets that track individuals over time to enable us to seriously answer this defect. One option, which we have not chosen here, is to create a pseudopanel with some individuals from each of the three waves of the GLSS data sets at different ages to mimic the evolution of individuals across time. In particular, earlier data sets are used for younger cohorts while later data sets are used for the older cohorts. We would be piecing together different people at different times to construct a fictitious panel. Given the somewhat stark nature of our net present value results, we doubt that this would have made a significant difference to our conclusions. Again, we concede there are many issues with the use of the data sets in computing the wage profiles. The possible weak arguments in our defense are, first, that it is the best data we have, and second, that it is currently the norm in the literature we seek to contribute to. 10.8

The Rate of Drainage, d, and the Rate of Return 

In the description of the village economy above, we made a number of simplifying assumptions on when there is the first exit out of the local economy, the drain, and when there is a return to the local economy of those who do indeed return. In the more general version of the simple village economy, there is potentially a complicated stochastic process explaining the emigration and return decisions. In this general model, what we need to do is to set a sequence of probabilities {dt}t of draining in each period t. Then, conditional on draining in period T, we need to specify in addition ∞ the probability of return in each subsequent period, {xt }t=T . Since there is a chance the representative individual never returns, we let x ∞ . represent this ∞ probability. In particular, we need to set probabilities {xt }t=T and x ∞ such ∞ that x ∞ + ∑ t=T xt = 1. Further, in the general case, there could also be remigrations after the first return, and later returns after later migrations. Instead, we shall impose very severe assumptions in our initial computations. The motivation for these come from surveys and casual observations. Recall that our primary focus is with emigration to OECD countries. Many who leave make one important emigration decision and then stay abroad for a while. When they decide to return, it is usually for good. (We are, of course, excluding short tourist visits). The survey by Black, King, and Litchfield (2003) finds that the overwhelming majority of respondents who have returned (83 percent) state that their return is permanent, with only 11 percent stating that they intend to reemigrate. We therefore model our repre-

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa

325

sentative agent as making a decision to migrate one time, then after going abroad staying there until a one- time and irreversible decision to return. We further calibrate our model as follows. First, we suppose that migration takes place almost immediately after completion of tertiary education. In particular, in our first cut we assume that the drainage occurs right after schooling. In particular, we suppose that tertiary education ends at age t = 22, then using the notation mentioned above, d22 = d > 0 and dt = 0 for all other periods t. We then use as a flow probability, d, the average rate of migration of tertiary- educated migrants obtained from the work of Docquier and Marfouk (2005). The data are obtained from censuses in OECD countries, and is available for the year 2000, as well as from national enrollment data. We will use the year 2000 rates of migration of skilled or tertiary educated in our computations. If one believes that emigration has been increasing over time, then these average migration rates will underestimate the true migration rates. As will be shown later, this would strengthen our basic conclusions. When does our representative agent return? In particular, using the nota∞ tion above, what are the return probabilities x ∞ and {xt }t=T ? Of those who decide to return, we shall approximate this by assigning all of the probability  on a return date at year seven. There are two reasons for this. First, this is suggested by the survey of Black, King, and Litchfield (2003) and Pires, Kassimir, and Brhane (1999). Second, it is our hunch that a lot of the tertiary educated go to the OECD for further education (graduate degrees), which take about five or six years, and then spend a year or two doing practical training (if they are on F1 visas) or if they want to get a quick job to ready themselves for return. Again, in our robustness section we shall discuss alternative formulations of the return probabilities and analyze the impacts on our results, spreading this probability over several years. It is fairly easy to see the impact of these changes in the date of return. We mention a few more surveys from the return migration literature in the subsection below. 10.8.1

Review of Some of the Return Migration Literature

• Gundel and Peters (2008) use data from the German Socio-Economic

Panel (GSOP) to examine return migration among immigrants to Germany. They find that highly skilled individuals are more likely to leave Germany than low- skilled migrants. However, return migration is found to be lower for migrants from non-EU countries. • Borjas and Bratsberg (1996) and Docquier and Rapoport (2007) also study the return of migrants. The latter refers to work showing that the return rate rose from less than one- fifth to about two- thirds for the return of Taiwanese PhDs who graduated from US universities in the fields of science and engineering from the 1970s to the 1990s. Very low rates of return are quoted from some studies of China and India, while

326

Yaw Nyarko

some other Indian software industry surveys “showed strong evidence of brain circulation, with 30– 40 percent of the higher- level employees having relevant work experience in a developed country (Commander et al. 2004, 3).” • The survey by Lowell and Findlay (2002) shows that some 50 percent of skilled workers return to their countries of origin, usually after about five years. 10.9

The Premium of the Returnees

When those who have been abroad return to their home countries, how much do they earn? Well, there are several parts to this, only one of which we will be able to meaningfully capture at this time. First, the returned come back with better skills. Second, they may earn a premium relative to their skill level because of the fact that they have had experiences abroad. On the other hand, because of lost social networks the returnees may face diminished wages. For our initial cut on the net present value computations, we shall assume for now no premium on returned migrants. In later work, and with better data, we will provide estimates of what we think are the net positives from the returned superior skills of migrants. 10.9.1

Literature and Data on Returnee Premia

Regarding data on the premium of the returned, we have identified a number of surveys that we list below. 1. Gibson and McKenzie (2010) find that migrants who return home do not tend to earn higher incomes than nonmigrants. However, they do tend to return with higher levels of human capital. Levels of repatriated savings appear to be similar in level to annual remittances, and there is some evidence that return migrants are more likely to be investing in business start-ups and sharing knowledge than nonmigrants. 2. A recent household survey on urban population (De Vreyer, Gubert, and Robilliard 2008) studied the impact of return international migration in seven major countries in Africa (Benin, Burkina Faso, Cote d’Ivoire, Mali, Niger, Senegal, and Togo). The surveys took place from 2001 to 2002. The sample consists of 58,459 individuals ages fifteeen years and older; 52,267 individuals in the sample never left the country where they were born and interviewed. The return migrants from OECD countries are 390 in number, and they constitute 0.6 percent of the sample. Average individual earnings of return migrants are 227.1 and nonmigrants are 55.9 (in 1,000 FCFA PPP; only active individuals). The average years of education of OECD return migrants is 11.1, compared to 5.6 for the nonmigrants. In Benin and Togo, conditioning on education individuals who have been abroad earn 28 percent and 21 percent more than those who have not, and

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa

327

the effect is statistically significant. In the other five countries the outcomes are mixed (negative in Cote d’Ivoire, Mali, and Senegal and positive in Burkina Faso and Nigher) but the coefficients are not significantly different from zero. 3. Wahba (2007) examines the labor market performance of return migrants to Egypt, and finds that on average, return migrants earn about 38 percent more than similar nonmigrants. The wage premium is lower for highly educated migrants: university graduates earn on average 19 percent more than their nonmigrant counterparts. 4. Barrett and Goggin (2010) estimate the wage premium for Irish migrants using a 2006 survey of Irish firms. After controlling for other factors likely to affect earnings, they find a 7 percent wage premium associated with return migrants. Estimated wage premiums differ by education level and migration destination. The premium for migrants with a postgraduate qualification was estimated to be 10 percent. Moreover, migrants that moved to far away countries (United States, Australia) were found to benefit from a higher premium than migrants that stayed in the United Kingdom or Europe. Finally, they find that the premium diminishes at a rate of about 1 percent per year. 10.10

Remittances

10.10.1

How Big Are the Remittances?

By one estimate, African workers send home around US$40 billion to the region (see table below). The value of remittances in sub-Saharan Africa (excluding North) are small by world standards, but high relative to GDP in Africa. The highest value of the remittances to GDP ratio is Nigeria, at 10.9 percent, with an additional four countries at ratios of 9 percent or higher (see table 10.6, as well as Barajas et al. [2010]). The data on remittances come from a number of sources. None of them is really completely satisfactory. We shall discuss the different sources of data and note the limitations of each. One difficulty with the official statistics is that so much of the flows of remittances take place through informal Table 10.6

Estimates of remittances/GDP ratios, top five African countries Country

Year

Remittances/GDP (%)

1 2 3 4 5

2007 2007 2007 2004 2007

10.9 9.7 9.6 9.4 9.4

Nigeria Sierra Leone Togo Guinea-Bissau Senegal

Source: From IMF Balance of Payments data as reported in Barajas et al. (2010).

328

Yaw Nyarko

channels—friends and family cash transfers, or the hawala system. In our empirical exercise, we shall specify the possible biases that could result from the use of different data sets on remittances. 1. Balance of Payments Data—Broad Definition. The official statistics for remittances are obtained from the Balance of Payments (BOP) data collected by the International Monetary Fund (IMF). The broad category used is that listed under “Workers Remittances, Compensation of Employees, and Migrant Transfers,” made up of three constituent parts: (a) Workers’ Remittances, defined as “current private transfers from migrant workers resident in the host country for more than a year, irrespective of their immigration status, to recipients in their country of origin”; (b) Compensation of Employees, defined as “wages, salaries, and other benefits paid to individuals who work in a country other than where they legally reside, for example, seasonal workers”; and (c) Migrants’ Transfers, defined as “the net worth of migrants who are expected to remain in the host country for more than one year that is transferred from one country to another at the time of migration.” Migrants’ transfers are reported as “capital transfers” in the capital account of the balance of payments accounts. The data are published by the World Bank Development Indicators (WDI), which relies on the IMF’s Balance of Payments Yearbook (item codes 2391, 2310, and 2431, respectively). Data are available from 1970 onward. This source is very often used in the literature on remittances. As has been pointed out by Chami, Fullenkamp, and Gapen (2008), this is on the one hand too broad of a definition, as it adds “wages, salaries, and other benefits paid to individuals who work in a country other than where they legally reside, for example, seasonal workers.” On the other hand and as mentioned earlier, it is also widely believed that a huge part of the remittance flows of individuals from host to home countries does not pass through the official channels at all, so would not be picked up in the BOP data. Freund and Spatafora (2008), based on market survey reports, indicated that the informal transfers may lie in the of 50 to 250 percent of recorded flows, depending on the country. Authors may use the more expansive BOP definition to compensate for the fact that informal transfers are excluded, but are important. This definition is used by Kapur (2004), who explains further the pluses and minuses of its use. 2. Balance of Payments Data—Narrow Definition. This uses only the entry “Workers’ Remittances Receipts” in the Balance of Payments. In particular, it applies the correction to (1) advocated by Chami, Fullenkamp, and Gapen (2008). The problem with using this narrower definition is that there are fewer observations as, presumably, for many countries the aggregate of the three portions of (1) are listed, without a disaggregation into component parts. For example, countries like South Africa, Kenya, and Ivory Coast appear not to have entries for the narrower definition.

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa

329

3. The United Nations IFAD Data.7 In an attempt to capture informal flows of remittances, the IFAD (International Fund for Agricultural Development) has developed and reported data based on sources like population censuses in destination countries, household surveys, central banks and other official government sources, money transfer companies, international organizations, and academic institutions. Sample estimates are obtained from which extrapolations are made. 4. National Central Banks. Individual countries, especially the central banks, also gather local data on remittances. For (2003) in Ghana, the Bank of Ghana estimates that the remittances equaled US$1,017.2 (Addison 2004), which is an order of magnitude higher than the US$65 million estimate of the World Bank in the same year. Even this amount may be an underestimate. Mazzucato, van den Boom, and Nsowah-Nuamah (2004) suggest that unregistered remittances flowing into Ghana is around 65 percent of the total, meaning that the true remittances are around three times the value of the Bank of Ghana numbers. Based on this, the (2003) remittances of Ghana equal $3 billion. Informal quotes by the then-Ghanaian president John Kuffour put the 2006 number at US$4 billion, while that of the minister of tourism and diasporan affairs put the number at US$4.3 billion in 2007 (see Voices of the South on Globalization, 2007). Note that this would make remittances a sizable percentage of GDP. Irving, Mohapatra, and Ratha (2010) reports on the findings of a 2008– 2009 World Bank survey of 114 central banks worldwide (thirty- three in Africa). Approximately 43 percent of respondents in remittance- receiving countries collected information on remittances transferred through informal channels. Of these respondents, 42 percent base these estimates on information and data gathered in household and/or overseas migrant surveys. The report notes that there can be very large discrepancies between what central banks report to the IMF and what was reported to the World Bank in the survey: For example, for Ghana, remittances reported to the IMF totaled $105 million in 2007, while remittances reported in the survey were $1.8 billion. 5. Other Studies. Bollard et al. (2009) describe and analyze a new data set on remittances. The database is a compilation of microlevel immigration data from fourteen surveys in eleven OECD destination countries. According to the authors, these countries were the destination for 79 percent of 7. From their web page, http://www.ifad.org: “The International Fund for Agricultural Development (IFAD), a specialized agency of the United Nations, was established as an international financial institution in 1977 as one of the major outcomes of the 1974 World Food Conference. The Conference was organized in response to the food crises of the early 1970s that primarily affected the Sahelian countries of Africa. The conference resolved that ‘an International Fund for Agricultural Development should be established immediately to finance agricultural development projects primarily for food production in the developing countries.’ ”

330

Yaw Nyarko

all global migrants to OECD countries in 2000. The surveys cover 33,000 immigrations, including 12,000 African migrants to nine OECD countries (Bollard et al. 2009, 9; Bollard, McKenzie, and Morten 2010, 4). Table 10.7 provides the data on remittances for African countries from various sources, with WDI- broad and WDI- narrow representing the broad and narrow definitions of remittances mentioned above, and in current US $ millions, 2006.8 As argued above, the numbers from the balance of payments data probably do not capture what we need when we speak of remittances. In our baseline “simplest model” scenario, we use the higher UN IFAD numbers in computing remittances. The UN IFAD numbers are much more in accordance with central bank figures we have obtained from local African nations. This results in a per migrant remittance value of US$5,260 for Ghana (total of US$851 million from 16,1800 migrants). In our baseline figures above we also indicated the internal rates of return and net present values at a per migrant remittance of US$3,600 (or $300 per month). Our robustness section discusses even lower values of the remittances. The numbers that we use will be average remittances over all classes, and we believe that this captures more fully the remittances of the tertiary educated. We now proceed to the question of remittances from different educational classes. 10.10.2

Decomposition of Remittances from Different Educational Classes

In our computations we will be using average remittances of nationals abroad when determining returns to tertiary education. One potential problem that needs to be addressed is the possibility that different educational groups send different levels of remittances. In particular, if it turns out that the tertiary educated remit much less than the average, then our use of the average remittances would bias upward the positive effects of the brain drain of the skilled. Although one would a priori think that the higher skilled, being better educated, are more likely to remit more, some (e.g., Faini 2007) think that the skilled are more likely to bring their families with them to their host country, and therefore remit less. The basic finding is confirmed by Niimi, Ozden, and Schiff (2008), who suggest that a 1 percent increase in the proportion of university- educated migrants will lead to a 2.8 percent decline in total remittances. Bollard et al. (2009) on the other hand, show the opposite. The authors focus on the relationship between remittances and educational attainment of migrants (all source countries). They look at both the likelihood of remitting and the level of remittances. They find that migrants with a university 8. World Bank World Development Indicators (WDI) downloaded June 2010.

Table 10.7

Remittances per year, in current US $ millions, 2006

Country Algeria Angola Benin Botswana Burkina Faso Burundi Cameroon Cape Verde Central African Republic Chad Comoros Congo Congo, Democratic Republic of Côte d’Ivoire Djibouti Egypt Equatorial Guinea Eritrea Ethiopia Gabon Gambia Ghana Guinea Guinea-Bissau Kenya Lesotho Liberia Libyan Arab Jamahiriya Madagascar Malawi Mali Mauritania Mauritius Morocco Mozambique Namibia Niger Nigeria Rwanda São Tomé and Principe Senegal Seychelles Sierra Leone Somalia South Africa Sudan Swaziland Togo Tunisia Uganda United Republic of Tan Zambia Zimbabwe

IFAD estimates 5,399 969 263 n/a 507 184 267 391 73 137 85 423 636 282 n/a 3,637 77 411 591 60 87 851 286 n/a 796 355 163 134 316 102 739 103 356 6,116 565 n/a 205 5,397 149 n/a 667 n/a 168 790 1,489 769 89 142 1,559 642 313 201 361

WDI-broad 2,527 n/a 173 117 50 0 103 137 n/a n/a 12 11 n/a 167 28 5,330 n/a n/a 172 7 64 105 42 28 1,128 361 685 16 11 1 212 2 215 5,454 80 17 66 3,329 21 2 633 5 33 n/a 424 1,156 99 193 1,510 665 15 58 n/a

WDI-narrow n/a n/a n/a 79 n/a 0 n/a 136 n/a n/a n/a n/a n/a 2 4 5,330 n/a n/a 169 n/a 63 105 42 n/a 570 4 685 6 n/a n/a 193 n/a n/a 5,454 16 7 n/a n/a 17 2 n/a 14 30 n/a n/a 1,155 1 n/a 1,510 665 8 58 n/a

332

Yaw Nyarko

degree are less likely to remit than migrants without a degree (27 percent versus 32 percent). However, the average level of remittances is higher for migrants with a university degree. The authors find that remittance behavior is primarily accounted for by income effects—that is, more educated migrants earn more money abroad and are thus able to send more home. Next, we note that there are many studies that find no impact of education on remittances per migrant (e.g., Naufal [2007] for Nicaragua, and Rodriguez and Horton [1994] for the Philippines). 10.10.3

When Do Remitters Remit?

One could also ask: When do remitters remit? We model in our baseline scenario remittances to be independent of time of return. Bollard, McKenzie, and Morten (2010) show that future returnees remit more. Gundel and Peters (2008) find that individuals that send remittances home are more likely to remigrate. 10.10.4

Data on Wages Abroad Used for the Individual Problem

We proceed by providing a few snapshots on the data, each producing slightly different estimates of the average wage rate of tertiary- educated Ghanaians moving abroad. We will use these snapshots in explaining how we arrive at a figure we will use in our computations. We have, we believe, used very conservative numbers (i.e., low foreign- wage rates). 1. From the Docquier and Marfouk (2005) data sets, we know that approximately 44 percent of the Ghanaian migrants to the United States have tertiary education. The US census data states that 31 percent of their sample of people born in Ghana and resident in the United States has tertiary education. This is not a perfect measure of incomes, but it should come close. It is imperfect because, of course, many of those who migrated to the United States with less than tertiary education could later become tertiary educated in the United States. Since we are interested in wage data, we use the US census data that also asks individuals for wages and we compute the average income of the top 31 percent of Ghanaians in the United States. This gives us a mean household US income of between $74,000 and $104,000 (the high and lows of the US 2000 census income buckets), or $89,000 with the midpoint value. Assuming a two- income household gives us an income level of $45,000 per person. 2. The average individual income of all Ghanaians in the United States among full- time, year- round workers is, according to the US 2000 census, US$32,262 for men and $26,235 for women. We know from the Docquier et al. data sets that a majority of the migrants are men. A simple average of the two would give us $29,242.50. 3. Although a large percentage of migrants from Ghana move to the United States, a significant percentage also go to other Western European countries. We, however, use the same US figure for them. We do not cur-

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa

333

rently have precise data for the United Kingdom, but we doubt that this will significantly throw off our IRR computations. 4. We of course need to exclude taxes from income statements. Or do we? Taxes after all are, for the most part, returned as benefits to the individual in terms of services, unemployment benefits, and so forth. Further, we did not take out taxes from the Ghana data. At the income levels we are using in the United States, the average federal tax rate was 16.6 percent in the year 2000. Even if we assume Ghanaian taxes are zero (they are not), addition of the taxes did not measurably change the very large IRR values. In our first pass at the individual problem, we will use the after- tax average annual US income values. That is, the wage figure in (2) above less US federal income taxes of 16.6 percent. We obtain an annual after- tax wage of $24,388. As argued above, we believe this to be an underestimate of the wages. 10.11

Robustness

We will begin by discussing the effect of changes in various parameters used in the model. We hope that this will enable us to test the basic assumptions of our model. Our baseline parameters are those used in the implementation of the village economy presented earlier. The values of our parameters involve estimates from different sources. In this robustness section, we will vary some of the parameters across a range obtained from the literature or that seem reasonable as ranges. We will look at the impact of changes in these parameters on two types of results we could be interested in: 1. The first is on how the return to education as a whole is affected—this is the ex ante definition of education taking into account those who stay and those who leave. This is measured by the NPV of the educated (NPVE )— again, this includes both the locally resident and the drainers. 2. The other question is the effect on the comparison between the return of the locally resident educated versus the drainers. This is measured by NPVLRE ) and NPVD), and their internal rates of return. As described in the introduction, the second question gets most of the attention in the media and in the press, but it is the first that should be the most relevant for policymakers in many developing African nations. As we argue here, the returns to education are large, even allowing for the brain drain. An increase in education would therefore help to raise incomes. In this robustness section, however, we will discuss the implications on both questions as we change our baseline parameters. We proceed in the next section by discussing the effect on our two questions in (1) and (2) above, of changes in the costs of education, C, of the wage profile of the tertiary educated, of the level of remittances, and of the drainage probabilities. We will measure the effect of these changes by looking at

334

Yaw Nyarko

the changes in the internal rates of return (IRR) and the net present values at baseline rates of interest of r = 5 percent (which we denote by NPV5) for the relevant quantities required in answering questions (1) and (2). 10.11.1

Cost of Education

Suppose we have underestimated the average cost of tertiary education. Suppose that the costs are higher than what we have from official statistics. Suppose further, that following Feldstein (1995), that the true cost of each $1 of spending is actually say $2 because of the distortionary effects of the taxation required to raise that $1. How will this affect our conclusions on the brain drain? In our baseline village economy, all the tertiary educated have the same cost C of education, whether they stay in the local economy or eventually leave. Changes in C therefore affect the locally educated exactly the same as the drainers. In situations where the NPV of drainers is higher than that of the locally resident, as was in our baseline model, if the drainers become “unprofitable” in an NPV sense (i.e., NPV less than zero) because of increases in the cost C, then so too would the local nondrainers, since the latter have lower NPVs. Indeed, the locals will become unprofitable before the drainers in this case—that is, at a lower level of cost. So, how much of a difference in the cost of education do we need to overturn our results? Well, keep the wage profile at our baseline, as obtained from the livings survey data, and maintain the remittance level at our (lower) baseline of $3,600 for Ghana. Define the “cost multiplier” to be the corrective multiplicative factor to costs—so that a cost multiplier equal to 1 is the baseline cost data as reported by official statistics and, for example, a cost multiplier equal to 2 denotes doubling the costs of tertiary education—as perhaps recommended by Feldstein (1995). Table 10.8 shows the decrease in the internal rate of return (IRR) and in the value of the net present value (NPV) at the baseline interest rate of r = 5 percent, NPV5, caused by the increase in costs from our baseline values to twice the baseline value. The internal rates of return still drop, but remain Table 10.8

Effect of changes in costs C

Cost multiplier Locally res. Never ret. Returnees Drainers Educated

IRR comparisons

NPV comparisons (r = 0.05)

1 (%)

2 (%)

1 ($)

2 ($)

14 33 32 32 23

10 21 19 20 14

17,229 33,026 29,368 31,197 24,213

13,450 29,247 25,589 27,418 20,434

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa

335

positive and large for education as a whole (from 23 percent to 14 percent) and particularly for the drainers (32 percent to 20 percent). Similarly, the NPV5 computations all remain positive. One could ask how high costs have to be to begin to overturn the positive NPV5 numbers. It turns out that we would need costs 5.6 times the baseline cost values for the returns to begin to be negative. The negative returns begin with the lowest values—the NPV5 of the locally resident educated. At that level the other NPVs still remain positive, and it takes a cost factor of 7.4 for the NPV5 of the educated (which includes both resident and drainers) to become zero. Again, we note that the changes in the robustness exercise here are changes in costs only, keeping all other relevant parameters at their baseline village economy levels mentioned in the earlier section. 10.11.2

Are There Quality Issues and Capacity Constraints on Production of Tertiary Educated?

We presented in tables 10.3 and 10.4 the cost data across Africa, as these values are critical for our computations and the entire NPV exercise. We will not pursue here, in any detail, the very interesting question of the changing production function of tertiary education in Africa as capacity rises. We do explore the effects of increasing output in tertiary education in Africa. If there are capacity constraints and the costs figures rise significantly as enrollments rise, contrary to the impression given by tables 10.3 and 10.4 above with everything else remaining the same, then our policy recommendations may no longer be valid. We have a number of responses to this concern: (a) It is our own view that economies of scale will work as a check on rapid rises in the costs of tertiary education. We have seen declines over time in the per person costs of tertiary education since the independence of many African countries, and although it is unlikely to fall much further in the future, it is also unlikely, in our opinion, to rise that steeply; (b) the capacity constraints of real significance are related to the shortage of professors for the universities—other costs, like housing and infrastructure, one would expect to have major economies of scale. The shortage of professors is probably related to the existence of better opportunities in the local economy for both the professors and the graduates themselves. However, if wages of graduates are rising, then the entire net present value exercise needs to be redone, as this could increase the baseline returns to tertiary education without the brain drain (what we called NPVvillage earlier). That is, the factors that cause an increase in costs (shortage of professors) could also increase the wage of locals. The negative effect of increased costs of professors is mitigated by the increased wages of graduates in our NPV computations. In summary, if the production of tertiary education is supply constrained and costs go up, there is the potential for our policy recommendations to be made invalid. At this time we do not believe that the changes in costs will

336

Yaw Nyarko

change that rapidly, due to economies of scale, and further as it changes there are other parts of the calculus that will also move around, which will result in small net effects. Quality Issues. There is a second and related issue concerning the possibly declining quality of the tertiary- educated graduates in local African universities, as there is a massive push in enrollments. Here it is important to distinguish two parts of this question—that related to past graduates and our computations above, and those related to future graduates and our policy recommendations. Regarding the computations in the village economy earlier and those using current data, we have already incorporated quality issues into the computations. Presumably, the wage rates locally and remittances (which should be related to wages abroad) are all a function of the quality of the tertiary educated. They have therefore already been accounted for. The bigger issue is with the policy prescriptions for the future. If tertiary enrollments are expanded and quality falls, how would this affect our basic argument? Well, first there is the question of the returns to tertiary education itself, even without taking into account the brain drain; that is, the term NPVvillage. The reduced quality will presumably reduce the local wages and perhaps lead to increased unemployment of the tertiary educated. How do these reduced wages compare to the new reduced costs of education? It is the comparison of these two that will determine the net effect. How about the effect of the reduced quality of graduates on the incremental returns to brain drain, NPVabroad? Well, the main channel will be via remittances. Since remittances are a small proportion of total wages abroad, it is possible that these remittances will stay relatively robust, even as there are reductions in wages abroad due to reduced quality of those graduates. Our position on the quality issues is related to that of costs and supply constraints mentioned earlier. Since the NPVs of education are generally positive, and since there are returns to scale in the provision of tertiary education as evidenced by past cost data, we believe that there are opportunities for increasing the quantity of education without major impacts on the quality. 10.11.3

Wages of Nondraining Locals

Suppose we have underestimated the level of wages of the locally resident educated. Suppose this is either because of poor data or nonrepresentative samples. Alternatively, this could be because we are incorrectly measuring the value of the tertiary educated by their wages. How would an upward revision in the wages of the locally resident affect our results? It should be repeated here, however, that given the high unemployment of the tertiary educated, it is not obvious that the existing wages are an underestimate in our village economy model. However, we perform this robustness check anyway. An increase in the imputed wage rate of the tertiary educated with no

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa

337

change in that of the secondary educated will of course increase the NPVs and internal rates of return to educated of the locally resident. Further, to the extent that some of the drainers return, the higher local tertiary wages will also increase the NPVs and IRR of the drainers. In particular, a revision upward in the wage sequence of the tertiary educated will increase all NPVs and IRRs. How about the comparison between the locals and the drainers? Well, an increase in the level of the wage sequence will obviously have a bigger effect on the locally resident than the drainers. So, how much of a difference in the wages of locals do we need to overturn our result that the expected NPVs of the drainers exceed those of the locally resident? Again, note that we keep all other parameters equal to our baseline levels: costs are those from the UNESCO data sets and we maintain the remittance level at our baseline of $3,600 for Ghana. Let the wage multiplier denote the corrective multiplicative factor to the sequence of local wages—so that a wage multiplier equal to 1 is the baseline wage sequence and, for example, a wage multiplier equal to 2 denotes doubling the local wages at each and every date. Table 10.9 shows the increase in the internal rate of return (IRR) and in the value of the net present value at the baseline interest rate of r = 5 percent (NPV5) caused by the increase in local wages. The internal rates of return rise from 14 percent to 24 percent for locally resident tertiary educated and from 32 percent to 33 percent for drainers, and from 23 percent to 29 percent for the tertiary educated as a whole. The NPVs at 5 percent interest rates rise from $17,229 to $54,971 for the locally resident tertiary educated, from $31,197 to $47,596 for the drainers, and from $24,213 to $51,270 for the tertiary educated as a whole. We also note that the wage multiplier of 1.65 is needed for the expected NPV at r = 0.05 of locals to exceed that of the drainers (where again we should stress that we believe the comparison of NPV5 of locals to that of drainers is actually not the appropriate question to be asking). It should be mentioned that part of the reason for the exercise in this section is that the data we have indicates what some may consider to be low domesticwage rates. Part of the issue is that we have correctly included wages of Table 10.9

Effect of changes in local wage sequence NPV comparisons IRR comparisons (r = 0.05)

Wage multiplier Locally res. Never ret. Returnees Drainers Educated

1 (%) 14 33 32 32 23

2 (%) 24 33 34 33 29

1 ($) 17,229 33,026 29,368 31,197 24,213

2 ($) 54,971 33,026 62,112 47,569 51,270

338

Yaw Nyarko

unemployed as zero. It should be stressed that the main conclusions do not change much when we omit the unemployed. The average wages will rise, but definitely not as much as the twofold rise modeled in this robustness section. 10.11.4

The Effect of Errors in Measurement of Remittances

The earlier section explained the effect of changes in local wages. The effect of changes in remittances is almost the exact opposite. An increase in remittances affects the relative importance of drainers in comparison to locally resident in only those periods an individual is away, just like the effect of wages. A $1 increase in the remittances has the same positive relative effect (i.e., on NPV of drainers minus NPV of locally resident) as a $1 decrease in the local wage rate. We now ask what happens if we suppose that our estimates of the remittances are too high relative to our baseline (of $3,600). We should mention here we actually believe that our remittance levels are too low, and do not include all the informal remittances and investments of people who are abroad. Nonetheless, we provide the robustness checks here. In particular, we look at remittance multipliers: a remittance multiplier of 1 is the baseline level, and a multiplier of 0.5, say, means that we use remittances equal to one- half of our baseline level. A decrease in the imputed remittances of the tertiary- educated drainers with no change of other parameters will, of course, decrease the NPVs and internal rates of return to the drainers, and will not affect the locally resident. In particular, a revision downward in the remittances of the drainers will decrease all NPVs and IRRs except those of the locally resident. Table 10.10 below shows the decrease in the internal rate of return (IRR) and in the value of the net present value at the baseline interest rate of r = 5 percent, NPV5, caused by the decrease in remittances from our baseline to half its value. The internal rates of return all remain positive. The biggest drop, as would be expected, is among the drainers who never return—the NPV5 figure goes down from $33,026 to $6,256. Table 10.10

Effect of changes in the assumed level of remittances

Remittances multiplier Locally res. Never ret. Returnees Drainers Educated

IRR comparisons

NPV comparisons (r = 0.05)

1 (%)

0.5 (%)

1 ($)

0.5 ($)

14 33 32 32 23

14 15 19 18 16

17,229 33,026 29,368 31,197 24,213

17,229 6,256 20,799 13,528 15,378

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa

339

We also note that the remittance multiplier of 0.6 is needed for the expected NPV5 of locals to exceed that of the drainers. (Again, we stress that the comparison between locals and drainers is not the right question; it is the NPV of the educated as a whole taking into account the drainers—in the table it remains positive even with a halving of our baseline remittance numbers.) 10.11.5

The Drainage Probabilities

The net present value of the educated, NPV E, is a weighted average of the NPVs of the locally resident educated NPV LRE, and the drainers, NPV D. The larger is the probability of drainage, d, the more the weights move the NPV E toward NPV D and away from NPV LRE. So, suppose we have miscalculated the drainage probability and that instead it is a larger number. Since for most of our computations NPV D exceeds NPV LRE, the increase in the drainage probability d will increase the NPV of education. It will actually strengthen the main conclusions of this chapter. It is indeed possible that our drainage probabilities, being the average drainage rates (nationals abroad divided by total nationals for the tertiary educated) may underestimate the marginal drainage probabilities (those in the most recent years) if the drainage probabilities have been rising over time. 10.11.6

Timing of Return

We have adopted a very stylized model of the timing of return of those who drain and come back. We have assumed all who drain leave immediately after school and those who return do so in seven years.9 The more general case involves a complex model of the tertiary educated leaving at all different dates and returning at different dates, with a complicated model of return probabilities and random durations of stay. Our feeling at this time, based on the various surveys we have seen, is that we have probably underestimated the duration of stay abroad. Given the relative values of the remittance term and the local wage rates, this would imply that in a more general model we should have larger returns to tertiary education more generally, and to those who drain more specifically. A more general model, given the other parameters in the model, will most probably strengthen our general conclusions.

9. One may be concerned that a return date of seven years means that the returnees are not important. This is not correct. First, a return date of seven years means that at a 5 percent interest rate, since 1/(1.05)7 = 0.71,we see that after seven years approximately 71 percent of the value is retained with 29 percent discounted relative to the present value. Further, we are comparing income streams, so both returnees and nonreturnees incomes are both discounted and at the same rate.

340

10.11.7

Yaw Nyarko

Post- 2000– 2005 Data Issues

We have worked with a lot of data from around the years 2000– 2005. This has been constrained by the data sources we have—the data on stocks of migrants are usually obtained from census figures, many of which were last taken around the year 2000. There are a number of post- 2000 developments that should be discussed. In Ghana there has been a tremendous increase in enrollments at the tertiary level over the past five to ten years. As the enrollment levels have increased, so too, presumably, has been the per person costs of education. Many have remarked that this has been associated with the reduced quality of education. We discussed issues of costs and quality in our robustness section above. The big open question is the extent of the brain drain currently, as the total stock of the tertiary educated has increased so rapidly. Rather than speculate, we await the census figures that should be in within the next couple of years in Ghana and in many other countries. 10.12 Items Omitted from Discussion 10.12.1 Skills of the Returnees Due to a lack of data, we have modeled the returnees as having no extra education after their time spent abroad. We know, however, that many come back with superior skills, which could be extra formal education or skills in more advanced economies working in sectors for which there would be few opportunities for advancement in their home countries. There has been quite a bit of attention put on the importance of returnees to India in the information communication technologies (ICTs) industries there. If these benefits of the returnees are added, they will of course increase the already high returns to those who drain out of the country. In much earlier work (see Easterly and Nyarko 2009), we have commented on the skills of the returnees. We reserve for subsequent work the study of the improved skills of those who return. We merely remark here that if these skills were added into the computations our results would be stronger, and we believe considerably so. In that sense, the fact that we have obtained strong returns to education and the brain drain without including these factors may be a reflection of the power of our results. 10.12.2

Internal African Migration

The focus of much of our work has been on the brain drain outside of Africa. There has been quite a bit of brain circulation within Africa. Adepoju (2002, 2006) have observed that highly skilled African professionals have increasingly found South Africa and Botswana to be “attractive alternatives” to Europe, the United States, and the Gulf States. At this time we do not have data indicating large transfers of the tertiary educated from one

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa

341

sub-Saharan African country to the next. We leave the question of withinAfrica brain drain or circulation to future work. 10.12.3

Labor Hoarding?

As the economy of Ghana and many other African countries improve, we are beginning to get anecdotal evidence of an increase in the return of the tertiary educated. This of course begs an obvious question. Could it be that it was a great idea to educate people and have them sent abroad when the economy was doing poorly, so that they could form a reserve pool of skilled labor ready to come back to the home country when the economy improved? Was there an invisible hand leading the central planner to educate people and to “hoard” them in foreign countries so that when the local economy could absorb them they are available to return? Our data can not directly test this hypothesis, of course, but the model we present could easily be tweaked at to get a handle at this. Again, we leave an in-depth discussion of this for future work. 10.12.4

Incentives

In other work (see Easterly and Nyarko 2009) we have discussed the very important literature on the question of the role of incentives to invest in education in the presence of the brain drain. The basic idea is that in the presence of the brain drain, and the opportunity to receive very high wages in the future with some probability, individuals make bigger investments in their education (either in terms of money spent or effort in studying and attending university). This incentive effect could increase the supply of the tertiary educated so much that it more than compensates for those who leave. In particular, the final number of tertiary educated left in the home or source country after the brain drain exceeds the number who would be in the country if the incentive effect of the brain drain was not there (say by banning the drain by law or by making it extremely difficult for people to leave). 10.13

Conclusion

In this chapter we have discussed the brain drain in Ghana. We have performed some rates of return to education computations using various data sets. Our main conclusions are that when using wages and remittances in standard cost- benefit returns to education computations, we have found that there have been high rates of return to tertiary education in general, taking into account the brain drain. Both from a social or “village” point of view, as well as from the individual point of view, the rates of return are large. Our results on the individual returns to education resolve a paradox in the returns to tertiary education literature, which often finds low or sluggish returns. This is paradoxical given the clamoring for tertiary education by leaders and the general public in many sub-Saharan African nations. It is

342

Yaw Nyarko

also paradoxical as one may expect a high rate of return to tertiary education in countries that have such low stocks of tertiary educated and where development is a priority. In our robustness checks we have stress tested the model, and the main conclusions seem to withstand these tests. Various variables we have omitted from our analyses may strengthen the conclusions we have. There are several issues we wish to highlight in our concluding remarks. First, we point out that in most of the conversations on the brain drain in Africa, it is almost universally considered something that is bad and to be avoided. The arguments in the media and in policy circles often use a costbenefit argument. It is often stated that “the government has wasted money” if people trained at the tertiary level then drain out. Our numbers show that these statements must be made carefully, and indeed that the opposite may be true. There are also arguments of the form “if only the highly skilled would stay” the local economies would do much better. Our results at least cast a little doubt on such assertions. More importantly though, our results indicate that there is room for creative thinking around the question of tertiary education provision. We have found high internal rates of return to tertiary education. This suggests that creative thinking around the provision of higher education could possibly be both self- financing (or even return a profit) and lead the education of large numbers of people. At currently levels of local incomes, however, this may involve some leaving the home country, at least for a while. In particular, rather than thinking of the brain drain as a curse upon the economies of sub-Saharan African countries it could instead be a part of the instrument to use for expanding the number of tertiary educated who are in the local economies. If it is known that one out of every two tertiary- educated people leave the country, then the logical implication is that there is the need to train twice as many to get the desired number locally. Our numbers show that this may indeed be feasible financially for the sponsoring entity (the government or non- governmental organization [NGO]). Our numbers also suggest that the individuals would also be able and willing to pay for loans incurred in this process. Our computations suggest interesting possibilities with financing schemes for tertiary education that (a) explicitly take into account the possibility that some will drain out of the country, and (b) that asks those who are out of the country, and presumably earning more money, to pay higher amounts to reimburse the government for their education. The analysis also suggests that the payments by those who leave could in principle form the bulk of the income, which in later years will finance the massive expansion of tertiary education in the local economies. For emphasis, we should note all the beneficiaries of schemes as described above. First, since this is potentially self- financing, the local economies will benefit from expanded numbers of educated. Second, by introducing a new financing system for higher education, those who are initially credit con-

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa

343

strained may be able to attain an education that otherwise would have been denied them. Finally, it should be remarked that those who end up being part of the brain drain should be counted in the welfare computations. In the discussions and the rhetoric on the brain drain, it is often presumed that once Ghanaians leave their country they cease being Ghanaians and so their welfare no longer matters. Should the goal of development not be the development of Ghanaians as opposed to those who happen to reside in Ghana? If a large number of people are educated who otherwise would not be, and a large fraction of those get improved incomes and livelihoods abroad who otherwise would not or would be unemployed in Ghana, is that not a positive to be included in evaluating policy? In this chapter we have evaluated the costs and benefits of the tertiaryeducation system including the calculus of all Ghanaians, those abroad and those in the home country. Our data show that continued investments in tertiary education may yield significantly large net present values and internal rates of return, and further, that higher- education financing schemes could therefore be ultimately self- financing.

References Addison, E. K. Y. 2004. “The Macroeconomic Impact of Remittances.” Unpublished Manuscript, Bank of Ghana, September. Adepoju, A. 2002. “Issues and Recent Trends in International Migration in SubSaharan Africa.” International Social Science Journal 52 (165): 383– 94. ———. 2006. “Leading Issues in International Migration in Sub-Saharan Africa.” In Views on Migration in Sub-Saharan Africa: Proceedings of an African Migration Alliance Workshop, edited by C. Cross, D. Gelderblom, N. Roux, and J. Mafukidze. Cape Town, South Africa: HSRC Press. Albrecht, D., and A. Ziderman. 1993. “Student Loans: An Effective Instrument for Cost Recovery in Higher Education?” World Bank Research Observer 8 (1): 71– 90. ———. 1995. “National Service: A Form of Societal Cost Recovery for Higher Education?” Higher Education 29:111– 28. Barajas, A., R. Chami, C. Fullenkamp, and A. Garg. 2010. “The Global Financial Crisis and Workers’ Remittances to Africa: What’s the Damage?” IMF Working Paper no. 10/24, International Monetary Fund, January. Barrett, A., and J. Goggin. 2010. “Returning to the Question of a Wage Premium for Returning Migrants.” Discussion Paper no. 4736, Institute for the Study of Labor, Bonn, Germany. Black, R., R. King, and J. Litchfield. 2003. “Transnational Migration, Return and Development in West Africa.” Final Research Report, Sussex Center for Migration Research, University of Sussex. Bloom, David, David Canning, and Kevin Chan. 2006. Higher Education and Economic Development in Africa. Human Development Sector, Africa Region, World Bank. Bollard, A., D. McKenzie, and M. Morten. 2010. “The Remitting Patterns of African Migrants in the OECD.” World Bank Working Paper no. 5260, World Bank.

344

Yaw Nyarko

Bollard, A., D. McKenzie, M. Morten, and H. Rapoport. 2009. “Remittances and the Brain Drain Revisited: The Microdata Show that More Educated Migrants Remit More.” World Bank Working Paper no. 5113, World Bank. Borjas, G., and B. Bratsberg. 1996. “Who Leaves? The Outmigration of the ForeignBorn.” Review of Economics and Statistics 78 (1): 165– 76. Chami, R., C. Fullenkamp, and M. Gapen. 2008. “Measuring Workers’ Remittances: What Should be Kept in and What Should be Left Out?” IMF Working Paper, International Monetary Fund, July. Clemens, M. 2007. “Do Visas Kill? Health Effects of African Health Professional Emigration.” CGD Working Paper no. 114, Center for Global Development. Clemens, M., and G. Pettersson. 2008. “New Data on African Health Professionals Abroad.” Human Resources for Health 6:1. Commander, S., R. Chanda, M. Kangasniemi, and L. A. Winters 2004. “Must Skilled Migration Be a Brain Drain? Evidence from the Indian Software Industry.” Discussion Paper no. 1422, Institute for the Study of Labor, Bonn, Germany. http://ftp.iza.org/dp1422.pdf. De Vreyer, P., F. Gubert, and A.-S. Robilliard. 2008. “Return Migrants in Western Africa: Characteristics and Labour Market Performance.” Unpublished Manuscript. Docquier, F., and A. Marfouk. 2005. International Migration by Educational Attainment (1990– 2000). Release 1.1., Update of World Bank Policy Research Working Paper no. 3381, Washington, DC, World Bank. Docquier, F., and H. Rapoport. 2007. “Skilled Migration: The Perspective of Developing Countries.” IZA Discussion Paper no. 2873, Institute for the Study of Labor, June. Easterly, W., and Y. Nyarko. 2009. “Is the Brain Drain Good for Africa?” In Skilled Migration Today: Prospect, Problems and Policies, edited by Jagdish Bhagwati and Gordon Hanson. Oxford: Oxford University Press. Faini, R. 2007. “Remittances and the Brain Drain: Do More Skilled Migrants Remit More?” World Bank Economic Review 21 (2): 177– 91. Feldstein, M. 1995. “Tax Avoidance and the Deadweight Loss of the Income Tax.” NBER Working Paper no. 5055, Cambridge, MA. Freund, C., and N. Spatafora. 2008. “Remittances, Transaction Costs, and Informality.” Journal of Development Economics 86 (2): 356– 66. Ghana Livings Survey. Various Years. Versions 3, 4, and 5. http://www.statsghana .gov.gh. Gibson, J., and D. McKenzie. 2010. “The Economic Consequences of BrEain Drain of the Best and Brightest: Microeconomic Evidence from Five Countries.” Working Paper, World Bank, February. Gundel, S., and H. Peters. 2008. “What Determines the Duration of Stay of Immigrants in Germany? Evidence from a Longitudinal Duration Analysis.” The International Journal of Social Economics 35 (11): 769– 82. IFAD. 2007. “Sending Money Home: Worldwide Remittance Flows to Developing Countries.” International Fund for Agricultural Development, Rome, Italy. Irving, J., S. Mohapatra, and D. Ratha. 2010. “Migrant Remittance Flows: Findings from a Global Survey of Central Banks.” World Bank Working Paper no. 194, April. Kapur, D. 2004. “Remittances: The New Development Mantra?” United Nations Conference on Trade and Development, G- 24 Discussion Paper Series, no. 29, April. Lowell, L. B., and A. M. Findlay. 2002. “Migration of Highly Skilled Persons from Developing Countries: Impact and Policy Responses-Synthesis Report.” Inter-

Returns to the Brain Drain and Brain Circulation in Sub-Saharan Africa

345

national Migration Papers no. 44, International Labour Office, Geneva, Switzerland. Mazzucato, V., B. van den Boom, and N. N. N. Nsowah-Nuamah. 2004. “The Impact of International Remittances on Local Living Standards: Evidence for Households and Rural Communities in Ghana.” Paper Presented at the Conference on Migration and Development in Ghana, Accra, September 14– 16. Naufal, G. 2007. “Who Remits? The Case of Nicaragua.” IZA Discussion Paper no. 3081, Institute for the Study of Labor, Bonn, Germany. Niimi, Y., C. Ozden, and M. Schiff. 2008. “Remittances and the Brain Drain: Skilled Migrants Do Remit Less.” IZA Discussion Paper no. 3393, Institute for the Study of Labor, Bonn, Germany. Pires, M., R. Kassimir, and M. Brhane. 1999. Investing in Return. Rates of Return of African Ph.D.s Trained in North America. New York: Social Science Research Council. Psacharopoulos, G., and A. H. Patrinos. 2004. “Returns to Investment in Education: A Further Update.” Education Economics 12 (2): 111– 34. Rodriguez, E., and S. Horton. 1994. “International Return Migration and Remittances in the Philippines.” In Development Strategy, Employment and Migration. Country Experiences, edited by D. O’Connor and L. Farsakh. Paris: OECD Development Centre. Voices of the South on Globalization. 2007. “Migration–Benefits and the Bane.” No. 8, IPS Europe. http://www.fes- globalization.org/dog_publications/voices_south /voices_of_the_south_on_globalization_no8.pdf. Wahba, J. 2007. “Returns to Overseas Work Experience: The Case of Egypt.” In International Migration, Economic Development & Policy, edited by C. Ozdon and M. Schiff. Washington, DC: World Bank. World Bank. 2010. World Development Indicators. Data retrieved June 2010, from WDI online database.

Contributors

Richard Akresh Department of Economics University of Illinois at UrbanaChampaign 1407 West Gregory Drive, 214 David Kinley Hall Urbana, IL 61801 Nava Ashraf Harvard Business School Baker Library 443 Soldiers Field Boston, MA 02163 Sarah Baird Department of Global Health George Washington University 950 New Hampshire Ave, NW 4th Floor Washington, DC 20052 Peter Boone Centre for Economic Performance London School of Economics Houghton Street London WC2A 2AE England Ephraim Chirwa Department of Economics Chancellor College University of Malawi Zomba, Malawi

Jacobus de Hoop UNICEF Office of Research, Innocenti Piazza SS. Annunziata, 12 50122 Florence, Italy Damien de Walque Senior Economist—DECRG The World Bank 1818 H Street, NW Washington, DC 20433 William H. Dow University of California, Berkeley School of Public Health 239 University Hall, no. 7360 Berkeley, CA 94720–360 Sebastian Edwards UCLA Anderson Graduate School of Business 110 Westwood Plaza, Suite C508 Box 951481 Los Angeles, CA 90095–1481 Ila Fazzio Effective Intervention CEP-London School of Economics 10 Houghton Street WC2A 2AE London, United Kingdom

347

348

Contributors

Giulia Ferrari London School of Hygiene and Tropical Medicine Department of Global Health and Development Social and Mathematical Epidemiology Group 15–17 Tavistock Place WC1H 9SH London, United Kingdom Günther Fink Harvard School of Public Health Department of Global Health and Population 665 Huntington Avenue Boston, MA 02115

Michael Kremer Harvard University Department of Economics Littauer Center M20 Cambridge, MA 02138 Edward Kutsoati Department of Economics 320 Braker Hall Tufts University Medford, MA 02155 Carol Medlin Children’s Investment Fund Foundation 7 Clifford Street London W1S 2FT England

Radha Iyengar RAND Corporation 1200 South Hayes Street Arlington, VA 22202

Randall Morck Faculty of Business University of Alberta Edmonton, AB T6G 2R6 Canada

Kameshwari Jandhyala ERU Consultants C208 Pasha Court, Punjagutta Hyderabad 500082, Telengana State, India

Rose Nathan Ifakara Health Institute PO Box 78 373 Dar es Salaam, Tanzania

Chitra Jayanty Effective Intervention, CEP-London School of Economics 10 Houghton Street London WC2A 2AE England Gangadhar Jayanty Effective Intervention CEP-London School of Economics 10 Houghton Street London WC2A 2AE England Simon Johnson MIT Sloan School of Management\ 100 Main Street, E52–562 Cambridge, MA 02142 Harounan Kazianga Oklahoma State University Department of Economics 324 Business Building Stillwater, OK 74078

Yaw Nyarko Department of Economics New York University 19 W. 4th Street, 6th Floor New York, NY 10012 Berk Özler The World Bank 1818 H Street, NW Washington, DC 20433 Vimala Ramachandran ERU Consultants Private Limited K 21 Hauz Khas Enclave Third Floor New Delhi 110016 India Jonathan Robinson Department of Economics University of California, Santa Cruz 457 Engineering 2 Santa Cruz, CA 95064

Contributors

349

Olga Rostapshova 2300 Clarendon Boulevard Suite 1000 Arlington, VA 22201

Nicholas Wilson Department of Economics Williams College Williamstown, MA 01267

Filipa Silva Effective Intervention CEP-London School of Economics 10 Houghton Street London WC2A 2AE England

Zhaoguo Zhan School of Economics and Management Tsinghua University Beijing, 100084 P.R. China

David N. Weil Department of Economics Box B Brown University Providence, RI 02912

Author Index

Abramsky, T., 200 Acemoglu, D., 53 Adepoju, A., 340 Aizer, A., 167, 168 Akresh, R., 118n4 Albrecht, D., 319 Alger, I., 216n1 Anagol, S., 282, 284 Angeles, L., 63 Arends-Kuenning, M., 141 Ashraf, N., 141, 142n5, 168 Askew, I., 87 Attanasio, O., 114, 115n2, 284 Avenell, A., 94 Awusabo-Asare, K., 218, 231 Baird, S., 86, 91, 114n1, 115n3, 142, 143n7, 144n8, 146n14, 151, 151n16, 152n17, 153, 153n18, 157, 159, 160 Bandiera, O., 142 Banerjee, A., 281, 282, 284 Banko, K. M., 97 Barber, S. L., 141 Barrett, A., 327 Bartlett, E. E., 94 Baylis, K., 141 Becker, G., 60, 75, 80, 167 Benartzi, S., 288 Ben-Porath, Y., 63 Berer, M., 87 Bertozzi, S., 88, 95 Bertrand, J. R., 88

Bertrand, M., 216n1 Bigelow, G. E., 93n2, 94, 101 Black, R., 324, 325 Blanchflower, D., 283n2 Bleakley, H., 53 Bleeker, S., 219 Bloom, D., 323 Bloom, N., 284 Bollard, A., 306, 329, 330, 332 Boone, P., 255 Borjas, G., 325 Bowlus, A., 168 Bozzoli, B., 176 Brahmbhatt, H., 60 Bratsberg, B., 325 Brhane, M., 325 Brooker, S., 114 Brown, R. A., 94 Browning, M., 167 Bruhn, M., 149, 284 Bundy, D. A. P., 114 Caceres, C., 85 Caliendo, M., 283, 283n2 Cameron, J., 97 Canning, D., 60, 63, 80, 323 Carpenter, L., 62 Cartwright, J., 141 Case, A., 114 Chami, R., 328 Chan, K., 323 Chanda, P., 33

351

352

Author Index

Chao, L.-W., 61 Chen, W. J., 62 Chiappori, P., 142, 167 Chiteji, N. S., 216n1 Clemens, M., 307 Coates, T. J., 85 Cooper, D., 61 Corvo, K., 167 Crépon, B., 284 Cummings, K., 92 Dabis, F., 59, 60, 63, 64, 78, 80 Dana, P., 61 Daniels, L., 284 Dasgupta, P., 114 Deaton, A., 114 De Brauw, A., 115n3 De Hoop, J., 157, 159, 160 De Janvry, A., 108 De Mel, S., 282, 283, 284, 287, 287n5 Deren, S., 93n2 De Soto, H., 223 De Vreyer, P., 326 De Walque, D., 86, 87, 97 Docquier, F., 310, 325, 332 Doepke, M., 63 Drexler, A., 284 Duffield, A., 91 Duflo, E., 114, 139, 140, 141, 141n5, 142n5, 167, 281, 282, 284 Dugan, L., 167 Dutton, D. G., 167 Easterly, W., 306, 340, 341 Eble, A., 278 Edmonds, E. V., 142n5 Einarsdóttir, J., 277 Eiss, R., 87 Ekpini, E. R., 59, 60, 63, 64, 78, 80 Elbel, B., 101 Ellul, A., 217 Emont, S., 93 Eswaran, M., 153 Evans, D., 283n2 Fabiani, M., 62 Fafchamps, M., 282 Faini, R., 330 Fairlie, R., 284 Farmer, A., 168 Fehrler, S., 278n12 Feldstein, M., 334

Fenrich, J., 216 Fernald, L. C. H., 91 Ferraro, K. J., 167n1 Field, E., 142n5 Findlay, A. M., 326 Fink, G., 62 Finocche, C., 92 Fischer, G., 284 Fiszbein, A., 86, 90, 114n1 Flores, R., 114, 114n2 Follick, M. J., 94 Folwer, J. I., 94 Fortson, J., 62 Fossen, F., 283, 283n2 Frese, M., 283n2 Freund, C., 328 Frost, L. J., 34 Fullenkamp, C., 328 Galasso, E., 114n2 Gallup, J. L., 53 Gapen, M., 328 García-Moreno, C., 178, 180 Gertler, P. J., 91, 95, 141 Gibson, J., 326 Giné, X., 284 Glass, R., 87 Glewwe, P., 114, 142, 278, 278n13 Glynn, J., 62 Goggin, J., 327 Gomez, M., 216 Gonzalez-Robledo, L. M., 90 Grantham-McGregor, S., 114 Gregson, S., 62 Gubert, F., 326 Gundel, S., 325, 332 Hacker, D., 217 Haines, A., 86, 91, 107 Hall, A., 114 Hamer, D. H., 34 Hamilton, D., 216n1 Hargreaves, J. R., 171, 174, 175n12 Hartog, J., 283 Haug, N., 106 Heintz, J., 230 Higgins, S. T., 93, 101 Higgins, T. E., 216 Hoddinott, J., 115n3 Hoff, K., 216n1 Hoffman, I., 63 Holla, A., 258

Author Index Horton, S., 332 Hosegood, V., 114 Hou, X., 91 Ilias, N., 278 Irving, J., 329 Iyigun, M., 142, 153 Jackson, R., 139, 139n1 Jacoby, H., 114 Jeffery, R. W., 94, 103 Jensen, R., 277 Jerven, M., 256 Jewkes, R., 104 Johnson, M. P., 167n1 Johnson, S., 53, 255 Juhn, C., 62 Kabeer, N., 141, 157 Kahneman, D., 96 Kakwani, N., 90 Kalemi-Ozcan, S., 62 Kamb, M. L., 93n2 Kandpal, E., 141 Kane, R. L., 91, 94 Kapur, D., 328 Karlan, D., 141, 168, 284 Kassimir, R., 325 Khandker, S. R., 141 Kim, J., 166, 169n4, 173, 174, 205, 210, 211 King, E., 114 King, R., 324, 325 Kludze, A. K. P., 218 Kohler, H. P., 86n1 Kongnyuy, E., 62 Kremer, M., 257, 258, 278, 278n13, 281, 282, 285, 299 Krishnan-Sarin, S., 104 Kritikos, A., 283, 283n2 Kuepie, M., 277 La Ferrar, E., 223 Lagarde, M., 86, 91, 107 Lastarria-Cornhiel, S., 217 Lee, J., 142n5, 281 Leighton, L., 283n2 Leslie, J., 114 Levine, R., 139 Levy, D., 115n2 Lewis, H. G., 60, 75, 80 Linnemayr, S., 62 Litchfield, J., 324, 325

353

Lloyd, C. B., 139 Long, S. K., 168 Lowe, J. B., 94 Lowell, L. B., 326 Lubell, Y., 34 Lucas, A., 53, 54 Luman, E. T., 259 Lund, F., 114 Lundberg, S., 141 Macours, K., 114 Maluccio, J., 114, 114n2 Mansuri, G., 284 Marfouk, A., 310, 325, 332 Martin, B., 92 Mauldon, J. G., 106 Mazzucato, V., 329 McIntosh, C., 91, 114n1, 115n3, 142, 143n7, 144n8, 151, 152n17, 153, 153n18, 157, 159 McKenzie, D., 149, 282, 283, 284, 287, 287n5, 306, 326, 330, 332 Mead, D., 284 Medlin, C., 86 Michaelowa, K., 278n12 Miller, D., 216n1 Mohapatra, S., 329 Molyneux, E., 139, 139n1 Moock, P. R., 114 Morris, S., 115n2 Morten, M., 306, 330, 332 Moulin, S., 278n13 Mullainathan, S., 216n1 Muralidharan, K., 140 Nagin, D. S., 167 Naufal, G., 332 Neufeld, L. M., 91 Nigenda, G., 90 Niimi, Y., 330 Nkuna, B., 175n12 Nordman, C. J., 277 Nsowah-Nuamah, N. N. N., 329 Nyarko, Y., 306, 340, 341 O’Donoghue, T., 96 Ohls, J., 115n2 Ollennu, N. A., 221 Oreffice, S., 142 Ostuka, K., 217 Oswald, A., 283n2 Ozden, C., 330

354

Author Index

Özler, B., 91, 114n1, 115n3, 142, 143n7, 144n8, 151, 151n16, 153, 153n18, 157, 159, 160 Packard, R. M., 31 Padian, N. S., 85, 88 Pagano, M., 217 Palmer, N., 86, 91, 107 Panunzi, F., 217 Parker, S. W., 114 Patrinos, A. H., 315 Paul-Ebhohimhen, V., 94 Paxson, C., 114, 114n2 Peltzer, K., 61 Peters, H., 325, 332 Petry, N. M., 91, 92, 94, 101, 102 Pettersson, G., 307 Philipson, T., 95 Pierce, W. D., 97 Pires, M., 325 Pitt, M. M., 141 Plateau, J.-P., 215 Pollak, R. A., 141, 168 Posner, R. A., 95 Prakash, N., 140 Psacharopoulos, G., 315 Puri, M., 283n2 Quisumbing, A. R., 217 Rabin, M., 96, 299 Ramachandran, V., 258 Rangel, M. A., 141, 167 Rapoport, H., 325 Ratha, D., 329 Rauch, A., 283n2 Rawlings, L., 90, 103 Reich, M. R., 34 Richter, L., 85 Robilliard, A.-S., 326 Robinson, D., 283n2 Robinson, J., 95, 281, 282 Rodriguez, E., 332 Rolland, E., 34 Rosen, M. I., 92 Rosenfeld, R., 167 Rosero, J., 103 Ross, P. H., 142 Rostapshova, O., 281 Roubaud, F. O., 277 Rubio, G., 90, 103 Rutledge, L., 142

Sachs, J. D., 53 Sadoulet, E., 108 Schady, N. R., 86, 90, 103, 114, 114n1, 114n2 Schiff, M., 330 Schoar, A., 284 Scholz, B., 216 Schubert, B., 91 Schultz, T. P., 114, 257 Seibert, S., 283n2 Seitz, S., 168 Sen, A., 141n4, 216n1 Shah, M., 95 Sherin, K., 174n8, 177, 180 Silverman, K., 92 Simanowitz, A., 175n12 Sindelar, J., 101 Slater, R., 91 Soares, F., 90 Soares, R., 63 Son, H., 90 Sorenson, J., 106 Spatafora, N., 328 Spira, R., 60 Sridhar, D., 91 Stekelenberg, J., 68n5 Steketee, R. W., 36n17 Stitzer, M., 86, 92, 93n2, 94, 101, 102 Strauss, J., 114, 167 Stringer, E., 63 Sunstein, C., 96 Taha, T., 60 Tauchen, H. V., 168 Thaler, R., 96, 288 Thomas, D., 114, 167 Thompson, P. D., 94 Thornton, R., 86, 86n1 Tiefenthaler, J., 168 Turan, B., 62 Tversky, A., 96 Udry, C., 282, 284 Vakis, R., 114 Valdivia, M., 284 Van den Boom, B., 329 Van der Sluis, J., 283, 283n3 Vandrey R., 93n2 Van Praag, M., 283, 283n3 Vijverberg, W., 283n3 Volpp, K., 94

Author Index Wales, T. J., 141 Walker, N., 62 Walsh, R. P., 142, 153 Warhurst, A., 139, 139n1 Wechtler, A., 278n12 Weeden, D., 106 Weibull, J., 216n1 Weiss, Y., 142 Windsor, R. A., 94 Wing, R. R., 94 Witte, A. D., 168 Wiysonge, C., 62 Wolpin, K., 63 Wong, R., 114

Woodman, G. R., 223, 224, 226 Woodruff, C., 282, 283, 284, 287, 287n5 Wydick, B., 142 Yeh, E., 95 Yin, W., 141, 168 Young, A., 62, 62n1, 255 Zaba, B., 62 Zhao, H., 283n2 Zia, B., 284 Ziderman, Z., 319 Zinman, J., 284 Zurovac, D., 33

355

Subject Index

Note: Page numbers followed by “f ” or “t” refer to figures or tables, respectively. Act 560 (Children’s Act, Ghana), 238–44; methodology and economic results of, 238–44; survivors’ benefits under, 227–30 Adolescent girls, 139–40 Africa. See Sub-Saharan Africa AIDS. See HIV/AIDS AIDS prevention. See HIV/AIDS prevention Artemisinin-based combination therapy (ACT), 33–34 Bed nets. See Insecticide treated mosquito nets (ITNs) Brain drain. See Drainage Breastfeeding behavior, PMTCT and, 77–80 Burkina Faso, 115; health care system of rural, 120. See also Nahouri Cash Transfers Pilot Project (NCTPP) Burundi: comparison of results with South Africa, 180–81; microfinancing in, 165– 66. See also Village savings and loan associations (VSLA) program CCT programs. See Conditional cash transfer (CCT) programs Child mortality: PMTCT and, 73–75; risk of, and fertility, literature on, 63 Children’s Act (Act 560, Ghana): methodology and econometric results of, 238–40; survivors’ benefits under, 227–30

Comprehensive Review of Education Outcomes (CREO) survey, 256–59; data collection and management for, 261–62; discussion of results, 276–79; literacy test score results of, 269–72; methodology and eligible population for, 259–60; multivariate analysis in, 272–75; numeracy test results of CREO survey, 269; results of, 262–68, 262–69; sample size for, 262; searching for excellence and, 275–76; test score results of, 260–61, 269–76. See also Guinea-Bissau Conditional cash transfer (CCT) programs, 86, 114; impacts of, in Zomba Cash Transfer Program (ZCTP), 149–59; literature review of, 90–91. See also Nahouri Cash Transfers Pilot Project (NCTPP); Unconditional cash transfer (UCT) programs; Zomba Cash Transfer Program (ZCTP) Contingency management (CM) interventions, literature review of, 92–94 Correspondence analysis, 179 CREO survey. See Comprehensive Review of Education Outcomes (CREO) survey DHS. See Zambian Demographic and Health Survey (ZDHS) Domestic violence. See Gender-based violence, theories for

357

358

Subject Index

Drainage, rate of, village economy model and, 324–26. See also Remittances Education, development and, 8–9. See also Tertiary education Empowerment: conditional cash transfer (CCT) programs and, 142–43; defined, 141; effectiveness of, 168; female, 142; literature on, 141–42; literature on female, 142. See also Zomba Cash Transfer Program (ZCTP) Entrepreneurship: personality characteristics and, literature review of, 283; returns to capital, credit, business training, and microcredit and, literature review of, 284. See also Kenya Exposure theory, for gender-based violence, 167 Extended families, and sub-Saharan Africa, 215–16, 216n1. See also Ghana Family-owned businesses, 281. See also Entrepreneurship Fertility: child mortality risk and, literature on, 63; HIV/AIDS and, literature on, 62–63; prevention of mother-to-child transmission of HIV in Zambia and, literature on, 61–62 Gender-based violence, theories for, 166–68 Gender issues, in Africa, 7–8 Ghana: criticisms of traditional inheritance norms of, 222–23; inheritance rules in, 216; introduction to study of, 215– 18; items omitted from discussion of returnees to, 340–41; legal reforms for intestate inheritances in, 223–30; matrilineal customary inheritance norms of, 219–20; methodology and econometric results of Children’s Act (Act 560), 238–44; methodology and econometric results of Law 111, 244–47; methodology and econometric results of reforms for economic status of widows, 247–49; 1985 Intestate Succession (PNDC) Law 111, 216–17, 224–27; patrilineal customary inheritance norms of, 220– 22; pension bequests data descriptive analysis for study of, 235–38; premium of returnees, 326–27; rate of brain drain and, 324–26; rate of return and, 324–26; regional map of, 219f; robust-

ness of results, 333–40; Social Security Act (PNDC Law 247), 227–30; Social Security and National Insurance Trust (SSNIT), 227–30, 235–38; Social Security and National Insurance Trust (SSNIT) data for, 217; survey data descriptive analysis for study of, 230– 35; survivors’ benefits under Children’s Act (Act 560), 227–30; traditional inheritance rules in, 218; types of marriages in, 218; young locally educated who stay in, 319–24 Ghana Living Surveys (GLSS), 319–22 Guinea-Bissau: background, 256; literacy rates in, 256, 256n1; map of, 257f. See also Comprehensive Review of Education Outcomes (CREO) survey HDI (Human Development Index), 14, 14n3 Health, of sub-Saharan Africans, 6–7 Health Management Information System (HMIS, Zambia), 14, 16–28; change in seasonality of all-cause mortality, 31f; change in seasonality of malaria mortality, 30f; data issues of, 20; death by province data of, vs. ZDHS, 23, 23f; extent of coverage of, 21–24; improvement of, 18–19; indoor residual spraying (IRS) in, 50–52; inpatient and mortality data from, 25, 26t; ITNs in, 46–47; malaria data in, 16–18; malaria deaths in, 15f; mortality changes in data of, vs. ZDHS, 24, 24f; nonreporting facilities and, 24–27; seasonality and, 28, 29t, 30f, 31f; seasonality and changes over time of, 29t; seasonality of mortality in, 30f. See also Zambia; Zambian Demographic and Health Survey (ZDHS) HITS (Hurt, Insult, Threaten, and Scream) screening tool, 173–74, 174n8, 177 HIV/AIDS, 6–7; diagnosis concerns of HMIS data and, 20; fertility and, literature on, 62–63 HIV/AIDS prevention: background, 85–87; information, education, and communications (IEC) investments in, 87–88; neoclassical income effect pathway to, 95–96; neoclassical price effect pathway to, 95; traditional approaches to, 87–90. See also RESPECT study

Subject Index HMIS. See Health Management Information System (HMIS, Zambia) Household chattels, 225 Human Development Index (HDI), 14, 14n3 Hurt, Insult, Threaten, and Scream (HITS) screening tool, 173–74, 174n8, 177 IMAGE program. See Microfinance for Gender Equity (IMAGE, South Africa) program Income effects, neoclassical, 95–96 Indoor residual spraying (IRS), 33, 36, 39–40, 47–49; in Demographic and Health Survey (DHS), 49–50; in Health Management Information System (HMIS), 50–52 Information, education, and communications (IEC) investments, HIV/AIDS prevention and, 87–88 Inheritance rules, 216; legal reforms of, 223– 30; traditional, 218–23. See also Ghana Insecticide treated mosquito nets (ITNs), 32–33, 37–39; in Demographic and Health Survey (DHS), 41–46; in Health Management Information System (HMIS), 46–47 Internal rates of return (IRR), 311, 311n2, 313–15 Intimate partner violence. See Gender-based violence, theories for Kenya: background survey data for, 286–87; data for small-scale retail sector in, 284–85; firm survival data for, 288; firm survival results for, 297f, 298; inventory and profit survey data for, 287–88; inventory size results for study, 292–94; mathematical ability scores used for study of, 290, 291f, 291t; profits results for study, 294–98; samples used for study of retail sector in, 285–86; summary statistics for, 288–90, 289t. See also Entrepreneurship Law 111 (Ghana), 216–17, 224–27; methodology and econometric results for study of, 244–47 Lineage property, 224, 224n8 Lineages, 215–16. See also Ghana Malaria Indicators Survey (MIS, Zambia), 16n4

359

Malawi, 7–8, 143. See also Zomba Cash Transfer Program (ZCTP) Male backlash theory, for gender-based violence, 167 Microfinance for Gender Equity (IMAGE, South Africa) program, 8, 166, 168–71; distinguishing between theories for, 175–76; quantitative data for, 177–78; results of, 202–12; sampling and design of, 174–75; summary of comparative measures, 177t; theoretical framework of, 170f. See also South Africa Microfinancing, 165 Mincer regressions, 322–23, 323f MIS (Malaria Indicators Survey, Zambia), 16n4 Mortality, 63; all-cause, change in seasonality of, 31f; child, 73–75; malaria, change in seasonality of, 30f; seasonality of, 30f Mother-to-child transmission (MTCT), 60; change in knowledge of, 68–71. See also Prevention of mother-to-child transmission of HIV (PMTCT, Zambia) NAAT (nucleic acid amplification tests), 100 Nahouri Cash Transfers Pilot Project (NCTPP), 115; descriptive statistics, 119–23; design of, 116–19, 117f; empirical results of, 126–33; identification strategy for, 123–26. See also Conditional cash transfer (CCT) programs National Malaria Control Program (NMCP) of Zambia, 13–14; artemisinin-based combination therapies (ACTs) and, 33–34; assessing link from rollout to incidence of, 40–52; elements of, 32–35; indoor residual spraying (IRS) component of, 33, 39– 40, 39t, 40f; Information, Education Communication/Behavior Change Communication (IEC/BCC) component of, 35; insecticide-treated mosquito nets (ITNs) component of, 32–33, 37, 38t; origins and scope of, 28–32; prompt and effective case management and (PECM) and, 33; rapid diagnostic tests (RDTs) and, 34–35; rollout of, 36–40; treatment component of, 33 NCTPP. See Nahouri Cash Transfers Pilot Project (NCTPP) Neoclassical income effects, 95–96

360

Subject Index

Neoclassical price effects, 95 1985 Intestate Succession (PNDC) Law (Ghana), 216–17 NMCP. See National Malaria Control Program (NMCP) of Zambia Nucleic acid amplification tests (NAAT), 100 Nudges, 96–97 Participatory wealth ranking (PWR), 170–71 PECM (prompt and effective case management), 33 PMTCT. See Prevention of mother-tochild transmission of HIV (PMTCT, Zambia) PNDC. See 1985 Intestate Succession (PNDC) Law (Ghana) PNDC Law 247 (Social Security Act, Ghana), 227–30 Poor health, 113–16 Pregnancy, PMTCT expansion and, 75–77 Prevention of mother-to-child transmission of HIV (PMTCT, Zambia): about, 59–61; background, 63–64; breastfeeding behavior and, 77–80; child mortality and, 73–75; data for, 65–66; expansion of access to, 66–68; fertility and, literature on, 61–62; knowledge of, 71–73; pregnancy and, 75–77; SubSaharan Africa and, 64–65. See also HIV/AIDS; Mother-to-child transmission (MTCT) Price effects, neoclassical, 95 Prompt and effective case management (PECM), 33 PWR (participatory wealth ranking), 170–71 Rapid diagnostic tests (RDTs), 20, 33, 34 Remittances: decomposition of, from different educational classes, 330–32; estimates of, of tertiary-educated Ghanaians, 332–33; per year, across different countries, 331t; size of, 327–30, 327t; timing of, 332; using, as measures of tertiary education, 311–13 Residue assets, 225 RESPECT study, 7, 86–90, 87f, 89t; baseline summary statistics, 98t; conditionality design of, 99–101; discussion of results of, 105–8; ethical considerations, 97; incentive size and frequency of, 101–

4; incentives used in, for risky sexual behaviors, 95–97; randomized study design, 87, 87f; self-reporting of behavioral change from incentives of, 104–5; skill building to aid behavior change for, 104; in Tanzania, 88–90, 89t; target populations of, 98–99 Risky sexual behaviors: theoretical pathways for incentive effects on, 95–97 Sexually transmitted infections (STIs), 99– 101; theoretical pathways for incentive effects on, 95–97 Social Security Act (PNDC Law 247, Ghana), 227–30 Social Security and National Insurance Trust (SSNIT, Ghana), 227–30, 235–38; data for, 217 South Africa: comparison of results with Burundi, 180–81; microfinancing in, 165. See also Microfinance for Gender Equity (IMAGE, South Africa) program SSNIT. See Social Security and National Insurance Trust (SSNIT, Ghana) Sub-Saharan Africa: extended families in, 215–16, 216n1; population of, 6; prevention of mother-to-child transmission of HIV in, 64–65 Tanzania, 88–90, 89t. See also RESPECT study Tertiary education: costs of, 316–19, 334–35; quality issues and capacity constraints on production and, 335–36; using wages and remittances as measures of, 311–13 Test score results as outcome measure for CREO survey, 260–61 Tribal property, 224n8 Unconditional cash transfer (UCT) programs, 114; impact of, in Zomba Cash Transfer Program (ZCTP), 159–61. See also Conditional cash transfer (CCT) programs Village economy, model for: drainage and, 324–26; individual-decision problem and, 313–16; results for, 310–11; simplest model for, 307–10; using wages and remittances for capturing true costs and benefits in, 311–13

Subject Index Village savings and loan associations (VSLA) program, 166, 171–73, 172n6; distinguishing between theories for, 175–76; qualitative data for, 178–80; quantitative data for, 176–77; results of, 181–202; sampling and design of, 173– 74; summary of comparative measures, 177t. See also Burundi Violence, gender-based, theories for, 166–68 Wages, using, as measures of tertiary education, 311–13 Zambia: Demographic and Health Survey (DHS) data in, 14, 16; future directions for research on malaria initiative of, 53–55; Health Management Information System (HMIS) data in, 16–28; indoor residual spraying (IRS) in, 47–49; malaria deaths in, 14, 15f; National Malaria Control Centre (NMCC) of, 32; nonmalaria changes in health environment of, 35; nonmalaria health initiatives/programs in, 35. See also Health Management Information System (HMIS, Zambia); National Malaria Control Program (NMCP) of

361

Zambia; Prevention of mother-tochild transmission of HIV (PMTCT, Zambia) Zambia Malaria Control Program. See National Malaria Control Program (NMCP) of Zambia Zambian Demographic and Health Survey (ZDHS): deaths by province data of, vs. HMIS data, 23, 23f; indoor residual spraying (IRS) in, 49–50; ITNs in, 41– 46; malaria data of, 16 (see also Health Management Information System (HMIS, Zambia)); mortality changes in data of, vs. HMIS, 24, 24f Zomba Cash Transfer Program (ZCTP), 140; conditional cash transfer (CCT) intervention for, 144–45; data used for, 146; estimation strategy for, 146–49; impact of conditional cash transfers (CCTs), 149–59; impacts of unconditional cash transfers (UCTs), 159–61; research design of, 144; selection of enumeration areas (EAs) for, 143–44; unconditional cash transfer (UCT) intervention for, 145. See also Conditional cash transfer (CCT) programs