Quantitative Research Methods in Consumer Psychology: Contemporary and Data Driven Approaches 1315641577, 9781315641577

Citation preview

QUANTITATIVE RESEARCH METHODS IN CONSUMER PSYCHOLOGY

Quantitative consumer research has long been the backbone of consumer psychology producing insights with peerless validity and reliability.This new book addresses a broad range of approaches to consumer psychology research along with developments in quantitative consumer research. Experts in their respective fields offer a perspective into this rapidly changing discipline of quantitative consumer research.The book focuses on new techniques as well as adaptations of traditional approaches and addresses ethics that relate to contemporary research approaches. The text is appropriate for use with university students at all academic levels. Each chapter provides both a theoretical grounding in its topic area and offers applied examples of the use of the approach in consumer settings. Exercises are provided at the end of each chapter to test student learning. Topics covered are quantitative research techniques, measurement theory and psychological scaling, mapping sentences for planning and managing research, using qualitative research to elucidate quantitative research findings, big data and its visualization, extracting insights from online data, modelling the consumer, social media and digital market analysis, connectionist modelling of consumer choice, market sensing and marketing research, preparing data for analysis, and ethics. The book may be used on its own as a textbook and may also be used as a supplementary text in quantitative research courses. Paul M. W. Hackett’s main area of research is in the theory and application of categorical ontologies. Paul has developed the qualitative or philosophical facet theory approach. He has almost 200 publications, including 10 books. Paul is a visiting professor at the Universities of Suffolk and Gloucestershire, a visiting researcher in psychology at Cambridge University and teaches at Emerson College.

QUANTITATIVE RESEARCH METHODS IN CONSUMER PSYCHOLOGY Contemporary and Data-Driven Approaches

Edited by Paul M. W. Hackett

First published 2019 by Routledge 711 Third Avenue, New York, NY 10017 and by Routledge 2 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN Routledge is an imprint of the Taylor & Francis Group, an informa business © 2019 Taylor & Francis The right of Paul M. W. Hackett to be identified as the author of the editorial material, and of the authors for their individual chapters, has been asserted in accordance with sections 77 and 78 of the Copyright, Designs and Patents Act 1988. All rights reserved. No part of this book may be reprinted or reproduced or utilised in any form or by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying and recording, or in any information storage or retrieval system, without permission in writing from the publishers. Trademark notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation without intent to infringe. Library of Congress Cataloging-in-Publication Data Names: Hackett, Paul, 1960– editor. Title: Quantitative research methods in consumer psychology : contemporary and data driven approaches / edited by Paul M. W. Hackett. Description: 1 Edition. | New York : Routledge, 2019. | Includes bibliographical references and index. | Identifiers: LCCN 2018030734 (print) | LCCN 2018031723 (ebook) | ISBN 9781315641577 (master eb) | ISBN 9781317280415 (epub) | ISBN 9781317280422 (pdf ) | ISBN 9781317280408 (mobi) | ISBN 9781138182691 (hb : alk. paper) | ISBN 9781138182721 (pb : alk. paper) | ISBN 9781315641577 (eb) Subjects: LCSH: Consumers—Psychology. | Quantitative research. Classification: LCC HF5415.32 (ebook) | LCC HF5415.32 .Q36 2019 (print) | DDC 658.8/3420721—dc23 LC record available at https://lccn.loc.gov/2018030734 ISBN: 978-1-138-18269-1 (hbk) ISBN: 978-1-315-64157-7 (pbk) ISBN: 978-1-138-18272-1 (ebk) Typeset in Bembo by Apex CoVantage, LLC

To my wife, Jessica Schwarzenbach, who in so many ways, is significantly responsible for the completion of this book.

CONTENTS

List of Figures,Tables and Boxes ix List of Contributors xiii Acknowledgmentsxvi Prefacexvii   1 Quantitative Research: Its Place in Consumer Psychology Cathrine V. Jansson-Boyd

1

  2 Using Contemporary Quantitative Techniques Or Shkoler

22

  3 Measurement Theory and Psychological Scaling Daniel P. Hinton and Tracey Platt

59

  4 Identify, Interpret, Monitor, and Respond to Quantitative Consumer Data on Social Media Dr. Amy Jauman, SMS   5 Alternative Research Methods: Introducing Market Sensing—A Qualitative and Interpretive Perspective on Research David Longbottom and Alison Lawson

88

124

viii Contents

  6 Big Data: Data Visualization and Quantitative Research Apps Vaidas Lukošius and Michael R. Hyman

166

  7 Exploring Ways of Extracting Insights From Big Data Peter Steidl

194

  8 Contemporary Approaches to Modelling the Consumer Debbie Isobel Keeling

222

  9 Connectionist Modelling of Consumer Choice Max N. Greene, Peter H. Morgan, and Gordon R. Foxall

247

10 Uniting Theory and Empirical Research: Marketing Research and Market Sensing Melvin Prince, Gillie Gabay, Constantinos-Vasilios Priporas, and Howard Moskowitz 11 Ethical Issues in Conducting Psychological Research David B. Resnik

272

298

12 A User-Friendly Practical Guide to Preparing Data for Analysis Kerry Rees

326

13 Integrating and Writing Up Data-Driven Quantitative Research: From Design to Result Presentation Paul M.W. Hackett, Lydia Lu and Paul M. Capobianco

376

Index407

FIGURES, TABLES AND BOXES

Figures

2.1 2.2 2.3 2.4 2.5 2.6 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 4.10 4.11 4.12 4.13 5.1 5.2 5.3 6.1 7.1 8.1

Model for a Direct Relationship Model for a Covariance/Correlation Type Relationship Model for a Spurious Relationship Model for a Mediational Relationship Model for a Conditioned (‘Moderated’) Relationship An After-Only Experimental Design Illustration The Research Process Social Media Poll Sample Online Poll Survey Design Choices Dropdown Boxes Semantic Differential Scale Visual Analog Sliding Scale Multiple Textboxes Reactive Content Objective Analytical Scales Objective Analytical Scales Objective Analytical Scales Responding to Responses Exploratory and Explanatory Research Designs Deep Value Mining Depth Gauge (Hancock & Longbottom, 2017) The Data Reduction Process Process for Extracting Insights From Big Data Identified Dopamine Segments The (Simplified) Theory of Planned Behaviour (Ajzen, 1991)

33 34 34 34 34 47 93 95 97 98 101 102 103 104 111 114 116 116 119 127 148 151 170 199 224

x  Figures, Tables and Boxes

8.2 Delineating the Model Building Blocks of the Simplified Theory of Planned Behaviour (Ajzen, 1991) 8.3 A Simplified Version of the Kano Model (1984) 8.4 A Simplified UTAUT2 Model with Standardised Coefficients 9.1 Schematic Biological and Mathematical Neurons 9.2 NN Architecture That Includes Three Layers: Two Nodes in the Input Layer, Five Nodes in the Hidden Intermediary Layer, and One Input Node 9.3 Receiver Operating Characteristic Curves 12.1 Example Data Labels in Variable View IBM SPSS 12.2 Using the Values Tab to Create Codes for Levels of a Variable 12.3 Example Data in Data View 12.4 Using the Menu to Create a Histogram 12.5 Histogram Window 12.6 Examples of Histograms With a Normal Curve Overlaid on Frequency Bars 12.7 Using the Split File Command 12.8 Organising Output by Groups 12.9 Using the Explore Command 12.10 Making Selection in Options 12.11a Descriptives for Cartoon 12.11b Descriptives Output: Skewness and Kurtosis 12.12 Testing Distribution for an Independent Variable With Three Levels Using Tests of Normality 12.13 Using the Transform Command 12.14 Applying a Log10 Transformation 12.15 Independent T Test Output 12.16 Output for a Paired T Test 12.17 Selecting Homogeneity Tests for Independent Designs 12.18 Levene’s Test Output 12.19 Mauchly’s Test of Sphericity Output 12.20 Using the Explore Command 12.21 Selecting Statistics Options 12.22 Selecting Outliers Option 12.23 Identifying Outliers in Output 12.24 Preparation for Multiple Regression Analyses 12.25 Selecting Options Using the Statistics Menu 12.26 Selecting Distances: Identifying Bias in the Multiple Regression Analysis 12.27 Generating Partial Regression Plots 12.28 Output: Partial Regression Plots 12.29 Output: Correlations and Covariances 12.30 Output: Assessing Multicollinearity Using Collinearity Statistics

229 230 234 249

250 258 327 328 329 336 336 337 338 339 340 341 341 341 341 343 344 347 347 348 348 348 351 351 352 352 354 355 356 356 357 358 358

Figures, Tables and Boxes  xi

12.31 Output: Mahalanobis Distance and Cooks Distance 359 12.32 Output: Assessing Casewise Diagnostics Using the Standard Residual 360 12.33 Output: Assessing the Assumption of Normally Distributed Errors  361 12.34 Output: Assessing Homoscedasticity 362 12.35 Output: Assessing the Durbin Watson Test 362 12.36 Preparing to Conduct an ANCOVA 364 12.37 Generating Regression Slopes for Each Level of an 365 Independent Variable 12.38 Selecting a Simple Plot 365 12.39 Participants Scores in Each Group of the Independent Variable 366 12.40 Chart Editor: Add Fit Lines at Subgroups 367 12.41 Chart Editor: Selecting the Required Relationship to Be 368 Displayed in the Plot 12.42 Output: Regression Slopes for Each Level of the Independent Variable369 12.43 Preparing to Select ‘Model’ 369 12.44 Creating the Interaction Term to Assess Homogeneity of 370 Regression Slopes 12.45 Output: Assessing Homogeneity of Regression Slopes 370 12.46 Preparing to Conduct a MANOVA 372 12.47 Options Menu 372 12.48 Output: Box’s Test 373 12.49 Output: Bartlett’s Test 373

Tables 1.1 Differences Between Quantitative and Qualitative Research 2.1 Quantitative and Qualitative Approaches and Their Research Issues 2.2 A Comparison of Survey Methods’ Utility 2.3 A Comparison of Observational Methods’ Utility 3.1 Summary of Response Formats in Psychometric Scales 3.2 Common Measures of Fit Used in CFA (From Hu & Bentler, 1999) 5.1 Comparing Research Philosophy: Positivist and Interpretive 5.2 Comparing Research Philosophy: Positivist and Interpretive (Epistemology and Ontology) 5.3 Validity, Reliability and Bias 5.4 Approaches to Research: Inductive and Deductive 5.5 The Qualitative Research Process 5.6 Traditional Research Strategies for Qualitative Research 5.7 Market Sensing Methods

6 27 41 45 69 76 130 131 132 133 136 136 139

xii  Figures, Tables and Boxes



5.8 5.9 5.10 5.11 5.12 5.13 5.14 5.15 5.16 5.17 5.18 5.19 6.1 6.2 6.3

8.1 8.2 10.1

10.2 10.3

10.4 10.5

Further Research Strategies for Qualitative Research How Many Depth Interviews Are Enough? Example Interview Plan Interview Difficulty Levels Levels of Questioning: Trust and Time Relationships Levels of Questioning: Style Implications Questioning Strategies Example Interview Transcript Example Summary Derived From Transcript Example Further Summary Analysis Using Cross Tabular Methods Example of Horizontal/Thematic Analysis Example of   Vertical/Case Analysis 5Ps and Big Data Analytics (Fan, Lau, & Zhao, 2015) Marketing Perspective on Big Data (Sheng et al., 2017) Foundational Technologies and Emerging Research in Big Data (Chen et al., 2012) Resultant Three Segments Cases Matched to Model Group 2 by Propensity The Performance of the Elements Based Upon Responses From Total Panel and the Three Emergent Mind-Sets (Interpretable Clusters) Strongest Performing Elements for the Three Segments and the Performance of the Same Element for the Total Panel Linkage Between Elements (Rows) and Feelings (Columns). Linkage Values Around Five to Six Represent Random Associations. Linkages of 10 or Higher Represent Associations That Are Likely to Be Less Random The Mind-Typing Questionnaire to Assign a New Patient to One of the Three Adherence Mid-sets for HF The Communication Guide to Patients Identified as Belonging to a Segment/Mind-Set

141 143 146 147 149 149 150 153 156 158 159 160 174 175 180 236 236

280 283

285 287 288

Boxes 3.1 Types of Reliability and Validity Relevant for Scale Development 3.2 A Case Study Focusing on Positive Affectivity

61 64

CONTRIBUTORS

Paul M. Capobianco has a degree in Marketing Communications. His research interests include religion, philosophy, economics, marketing and as a research assistant he explored the dynamics between creativity, data and strategy. Gordon R. Foxall is the Distinguished Research Professor at Cardiff Business School and Visiting Professor of Economic Psychology at Reykjavik University. He holds PhDs in industrial economics and business studies and in psychology and a higher doctorate (DSocSc). He is the author of some 300 refereed papers and chapters and 35 books. He is a Fellow of the Academy of Social Sciences (FAcSS), a Fellow of the British Psychological Society (FBPsS), and a Fellow of the British Academy of Management (FBAM). His principal research interests include consumer behavior analysis and the theory of the firm (both fields which he inaugurated) and the philosophical implications of consumer neuroscience. Gillie Gabay is a Senior Lecturer and a Co-author of the New Novum Organum:

Policy, Perceptions and Emotions in health. She studies health psychology, the study of psychological and behavioural processes in health, illness and healthcare Max N. Greene comes from an interdisciplinary research background, trained

in marketing and industrial/organizational psychology, and focused on modelling consumer behaviour with artificial neural networks for his PhD research. ­Currently, he is a Data Scientist in industry, involved with global business intelligence and marketing transformation programmes, and holds a Visiting Researcher position in the Marketing and Strategy Section at Cardiff Business School.

xiv Contributors

Daniel P. Hinton is a Senior Lecturer in Psychology at the University of Wolverhampton. His expertise focuses on assessment, psychometrics and scale development within organizational settings. His research has been published in world-leading academic journals. Michael R. Hyman is Distinguished Achievement Professor of Marketing at New Mexico State University in Las Cruces and has numerous journal articles, conference papers, four co-authored/co-edited books, etc. Cathrine V. Jansson-Boyd is a Reader in Consumer Psychology at Anglia Ruskin

University. Cathrine has written two books on consumer psychology, co-edited the International Handbook of Consumer Psychology and many research articles. Dr. Amy Jauman, SMS, is a Speaker, Professor, a Certified Social Media Strategist and Certified National Institute of Social Media Instructor. She has published two books and has a doctorate in organization development. Debbie Isobel Keeling is Professor of Marketing at the University of Sussex. She

has a background in psychology and extensive research and teaching experience. She has authored more than 100 scientific papers, reports and chapters. Alison Lawson is Head of Discipline in Marketing and Operations at Derby ­ usiness School. Alison teaches research methods and the use of interviewing, B case study and appreciative enquiry approaches. David Longbottom is a Reader in Marketing at the University of Derby. He has

published more than 50 academic papers and is co-author of Alternative Marketing Research Methods: Market Sensing. Lydia Lu is a Student at Emerson College and has conducted market research and

academic research into areas, such as quick service restaurant industry and the technology industry. Vaidas Lukošius is Associate Professor of Marketing at Tennessee State Univer-

sity in Nashville, Tennessee. He has published in many notable journals, and his research interests include consumer behaviour. Peter H. Morgan is Reader in quantitative analysis at Cardiff Business School

where he teaches quantitative methods and researches into sampling methods for consumer price indices, consumer behaviour and neural networks and other adaptive data analysis techniques.

Contributors  xv

Howard Moskowitz has a PhD in experimental psychology from Harvard

­ niversity and is a graduate of CUNY. He founded Mind Genomics Associates, U promoting the new science of Mind Genomics. Tracey Platt is Reader of applied psychology at the University of Wolverhamp-

ton, United Kingdom. She has 46 journal articles, book chapters, etc., and won the 2011 Humour Society’s Graduate Student Emerging Scholar award. Melvin Prince is Professor of marketing, Southern Connecticut. His publications are in innovative methodological analyses, focus groups and survey research. He has published over 80 articles on a wide range of topics. Constantinos-Vasilios Priporas is a Senior Lecturer in marketing at Middlesex University Business School, United Kingdom. He has published 45 peer-reviewed research papers, 1 co-edited book, 5 book chapters and 25 papers conference papers. Kerry Rees is interested in nonconscious influences on thoughts, feelings and

behaviours. Specifically, Kerry investigates the role of identity, culture and social values on individual’s decision making and emotional experience. David B. Resnik has an MA and PhD in philosophy and has published over 250

articles and 9 books on philosophy and bioethics. He is a Fellow of the American Association for the Advancement of Science. Or Shkoler is an Independent Researcher whose research interests span statistics

and data analyses, quantitative research methods, individual differences, organizational behaviour and aspects of work behaviour. Peter Steidl has an MBA and PhD from Vienna University and has been a per-

manent staff member in Austria and Australia. His recent publications include, books, book chapters and articles principally on neuromarketing.

ACKNOWLEDGMENTS

An edited collection of essays stands and falls on the consistence of scholarship and writing embodied in its separate chapters. Therefore, the most important acknowledgment I can make is to thank all the authors who invested their time and other resources to ensure the quality of their chapter. I also thank all the staff members at Routledge who assisted me in the production of this book, from the initial proposal through to the finished book. I offer particular thanks to Julie Toich and Christina Chronister. I decided to assemble this book after finding a need for the information it contains. I recognized this need whilst teaching at the Department of Marketing Communication, Emerson College and the Van Loan School, Endicott College. I am extremely grateful to all the students I had the privilege to teach and without whom this book would not have come about. As always, I thank Jessica, my wife, for her unending support.

PREFACE

Why do people select one specific product or brand over another? Why do some consumers take a long time selecting between different products, services and brands, whilst other customers appear to make almost instant decisions? What are the motivators behind an individual’s consumer related choices? What are the roles of thoughts and emotion in the decisions that consumers make? In what ways do social groups (from groups of friends to societies) influence the decisions that individual consumers make? What are the roles of advertisements and other forms of promotional materials that influence consumers’ choices and their other behaviours?

What Is Consumer Psychology? The earlier questions are typical of the concerns of consumer psychologists and the types of questions they may ask. From these questions it is apparent that a large amount of research in consumer psychology addresses the choices that individuals and groups of people make in situations where they are customers or consumers. It is also obvious from these questions that consumer psychology is a branch of theoretical psychology that is concerned with processes such as attitudes, thoughts and emotions. However, consumer psychology is also extremely applied in that its focus is on the selection and use by customers of services and products that companies and individuals make available within a market. Consumer psychology is an applied branch of social psychology; however, it is a discipline that uses methods that are drawn from many areas of psychological research to investigate

xviii Preface

how consumer likes and dislikes, habits and tendencies are related to their behaviours as consumers, clients and users. In its applied role, consumer psychology provides information to companies about their products and services in a way that helps them to tailor their offerings to consumers and to design advertising and other promotional material to have optimal effect. Thus, consumer psychologists study both overt and covert consumer behaviour with the intention of better understanding consumer choice and usage of products and services with the intention of predicting future consumer behaviours. Division 23 of the American Psychological Association: The Society for Consumer Psychology, defines consumer psychology simply as follows: Consumer psychology employs theoretical psychological approaches to understanding consumers. (American Psychological Association, (undated)) The Encyclopaedia Britannica defines consumer psychology in a slightly broader senses as being a branch of social psychology1 that is interested in the behaviour of consumers in a market situation. Consumer psychologists examine the preferences, customs, and habits of various consumer groups; their research on consumer attitudes is often used to help design advertising campaigns and to formulate new products. (The Editors of Encyclopaedia Britannica, 2018) Consumer psychology also incorporates considerations regarding how consumers perceive products and services, the information they get about these and the manner in which this information has an effect on how we think, feel and act towards products and services.To develop both theoretical and applied knowledge in the field of consumer psychology, researchers adopt a large and varied selection of research procedures, and the quantitative methods they may employ are the subject matter of this book.

Research in Consumer Psychology Traditionally, within the field of quantitative consumer research, there has been a long tradition of researchers employing quantitative methods in their research. This may well have been because qualitative approaches have a much more recent pedigree in psychological research. However, quantitative methods have constituted the backbone of consumer psychology research, as these forms of research are able to produce clear results and insights that have unequalled validity and reliability. Some psychologists may also argue that quantitative methods have a much greater predictive power than qualitative approaches. In this new book, I have brought together a series of authors to contribute chapters, each of which

Preface  xix

addresses a specific area or orientation towards quantitative consumer psychology research. When taken together, the book addresses a broad and fairly comprehensive range of approaches to consumer psychology research. Moreover, the book is up to date and presents some of the more recent developments in the area of quantitative consumer research, as well as covering important more established aspects in the area.

Synopsis of Chapters In each of the 13 chapters in this book, an expert, or experts, in consumer psychology present a clear and concise account of a selected quantitative research approach or methodology as it may be used on consumer psychology research. The chapters offer an expert perspective into this rapidly changing discipline of quantitative consumer research. In each chapter, the authors focus on new techniques as well as adaptations of traditional approaches. Ethics are also addressed, especially as they relate to contemporary research approaches. The first chapter, “Quantitative Research: Its Place in Consumer Psychology”, Cathrine V. Jansson-Boyd introduces the ways in which contemporary quantitative research approaches are employed in consumer psychology. This chapter is placed at the start of the book as it introduces the reader to the book’s main themes, which are further explored in later chapters. She presents an overview of the varied features of quantitative research methods that consumer psychologists have to consider in their research. Jansson-Boyd makes the argument that the quality of any consumer psychology research is dependent upon three things: the theories that underlie any specific piece of research, the knowledge that the researcher possesses and brings to the research process and the way a researcher understands how research must be conducted. Or Shkoler is the author of Chapter 2. He presents a thorough review of how contemporary research approaches may be used by consumer psychologists. Shkoler sets the scene for his writing by considering the differences between quantitative and qualitative research. He notes the fact that consumer psychology is a branch of social science research and then goes on to examine the possibilities and properties that the researcher has when deciding on research methodologies, processes and research designs. The author uses the term “empirically sane research” and subsequently sets clear parameters for defining enquiries and hypotheses whilst evaluating relationships between associations, causality and hypotheses. Shkoler stresses the importance of developing good hypotheses in relation to which he stresses the need for using a well-established theoretical background to provide the foundations for good hypotheses and research. Furthermore, he distinguishes between null hypothesis and the alternative/research hypotheses, primary data and secondary data sources and explores measurement, operationalization and sampling. Chapter 3 is by Danny P. Hinton and Tracey Platt on measurement theory, psychological scaling and psychometrics, as well as using tools to measure

xx Preface

psychological constructs. The authors note how a large amount of quantitative psychological research involves the measurement of psychological constructs.This is also the case with academic and applied consumer psychology research, and the use of psychometric measures in this area is at the heart of this chapter. The theoretical underpinnings of psychological measurement are discussed as are the processes for the development of psychometric measures. Social media and digital market analysis are the subjects Dr. Amy Jauman, addresses in Chapter 4. She starts by noting how social media offers marketers unique insights into the psychological processes of consumers. Then in her chapter she outlines four steps that the author proposes should be followed by anyone wishing to evaluate a brand’s digital presence, and she continues by demonstrating how the results of such evaluations can be utilized in a social media strategic plan. Understanding social media has become central to the activities of consumer psychologists and marketers alike, and in her chapter, Jauman emphasizes the importance of this media when she claims, “The savvy marketer knows how to identify relevant information on social media sites, interpret the findings, monitor data over time, and respond to consumers and competitors in a way that’s most likely to be beneficial to their business.” The authors of Chapter 5, David Longbottom and Alison Lawson, pursue a slightly different line of thought to the rest of the book’s content when they consider qualitative research. However, they do not view qualitative research as an approach on its own but when it is used to elucidate quantitative consumer psychology research findings. They present alternative qualitative research methods and initially explain the context and philosophy behind this type of research.They then continue by taking the reader through the sequentially presented process of strategy, planning, data collection, data analysis and data presentation. Digital media are again at the heart of the research approaches discussed in Chapter 6. Authors Vaidas Lukošius and Michael R. Hyman address what is known as big data and the use of apps to visualize this type of information contained in that data. The learning objectives put forward for the chapter provide a clear understanding of its content. The authors start by defining big data and describing the assumptions that are associated with this form of data. They then progress by describing the processes that may be used to extract insights from big data, and they explain and illustrate how big data is currently used, with reference to marketing research. The next section of their chapter considers how emerging research in big data analytics may be appraised, and they conclude by suggesting dimensions for troubleshooting the transition from small data to big data. Chapter 7 is by Peter Steidl who explores how insight may be extracted from online data. He commences his exploration by focusing not on technical components of data extraction but on the prerequisites that facilitate the successful exploration and use of data. Instead of simply providing a list of facts and assertions, the author makes the content of his writing more engaging when he “tells his own story” about his research experiences. Steidl notes how the changes that have happened in software may have made previous understanding obsolete. He

Preface  xxi

stresses that domain-specific experience is required in order not to miss important patterns, meanings and the significance of researchers’ findings. Peter stresses that to identify new and better ways to address the challenges of understanding consumers, the researcher needs to be wary of outdated conventions that could distort analyses to confirm their pre-conceptions. Illustrations of specific projects and a case study are provided. Finally, artificial intelligence is forwarded as an optimization tool and differences between traditional statistics and AI are noted from a user’s perspective. Debbie Isobel Keeling claims in Chapter 8 that modelling plays an important role in consumer psychology and that this may provide understanding about consumers in many diverse contexts. Modelling, Keeling claims, is also fundamental to our understanding of consumer decision making and the willingness of consumers to take risks. In her chapter, she therefore considers contemporary approaches to modelling the consumer, and she provides examples such as the modelling of consumer profiles. She notes how consumer psychologists are able to model consumer attitudes, peer influences and individual differences in personality and health status so as to enable the design of advertising campaigns that encourage healthy living practices. Modelling, she claims, may also be used to effectively segment consumers into groups that then allows marketing communications, products, services and recommendation systems to be tailored to the segments’ needs. Central to the chapter is the assertion that the term modelling when used in consumer psychology must involve a complete process that involves defined objectives, design, data collection and analysis. If conceived in this manner, modelling, she states, allows the development and/or testing of theories relating to why consumers behave as they do. Furthermore, Keeling suggests that modelling yields results that may be analysed using path models, equations, charts, matrices and many other ways to facilitate a holistic indication of the interactions between the important influencers on consumer behaviour. In Chapter 9, Max N. Greene, Peter H. Morgan and Gordon R. Foxall proffer a connectionist modelling of consumer choice.Their principle aim in this chapter is to propose connectionist models as explanations for consumer behaviour. In doing this, they concentrate on feedforward artificial neural-network models, and, by using connectionist constructs, they associate this with the theoretical framework of the Behavioural Perspective Model. The research presented came out of a project which examined a large number of neural-network models of varying complexity which were evaluated in terms of their ability to predict consumer behaviour. The authors also compared their results to logistic regression. Based upon their results, the authors claim that when studying consumer behaviour, neural networks are useful and offer enhanced understanding to the methods that are usually used in consumer behaviour analysis and connectionist models demonstrate potential for predicting and explaining consumer behaviour. In the tenth chapter, Melvin Prince, Costas Priporas, Gillie Gabay and Howard Moskowitz present an approach for uniting theory and empirical research with market sensing and marketing research. Market research, they say, when viewed

xxii Preface

within the context of market sensing, is able to play a role in the organization and extension of knowledge regarding markets and their consumer. Marketing research is able to perform in this way through the production of raw data that may be assembled and analysed in order to provide useful information upon which actions may be based. This data may also allow the estimating of the way in which the market may respond to such actions and allow recommendations regarding marketing actions and the allocation of a company’s resources. Chapter 11 is entitled “ Ethical Issues in Conducting Psychological Research”. In this chapter, David B. Resnik provides information that should be required reading for all quantitative consumer psychology researchers (and as this information is applicable outside of consumer psychology, for researchers in a more general sense). The chapter provides a review of the ethical issues that are associated with psychological research. Resnik emphasizes that many forms of psychological research raise ethical issues. Amongst these issues are generic research concerns, such as how to avert and regulate misconduct, how to manage data, issues with authorship and conflicts of interest. He also considers issues associated specifically with psychological research, such as the deception of human subjects in behavioural experiments. In Chapter 12, Kerry Rees provides a user-friendly practical guide to preparing data for analysis in the form of a hands-on guide to thinking about and screening data. He stresses the importance of considering data preparation, as, he says, this influences the results of statistics used and has a direct impact on the ability to infer from results. Rees commences by noting that statistical tests focus on significance that support or reject the hypothesis(es) at a given level of significance of confidence. In his chapter, Rees concentrates on parametric tests and their assumptions. Analysis, Rees says, is, at a basic level, rooted in the decisions that we make that are based upon our understanding of the data and based upon this understanding the researcher plans data collection and analysis. He states that his chapter is a guide for making these decisions when preparing data for analysis. In the final chapter, Paul M. W. Hackett, Lydia Lu and Paul Capobianco consider the end point in a quantitative consumer psychology research project: integrating the findings from multiple research procedures in a coherent manner. They state that they will attempt to both unite and conclude the writing in the previous chapters and to provide a framework for researchers to use to bring together their quantitative research projects in a coherent manner that does not impose an artificial structure upon their research design, writing and conclusions. The authors commence by claiming that research planning and research design have a profound effect on the research conducted through to the point at which the research will eventually be written up. They address how multiple research procedures within a single research study may be coordinated and integrated.

Preface  xxiii

Hackett and colleagues propose that to achieve an integrated research design through to write-up, the mapping sentence offers a tool that will enable such coherence. They provide illustrations of several mapping sentences, including the research design mapping sentence, the project planning mapping sentence and the traditional and declarative mapping sentences within complex quantitative consumer psychological research. Mapping sentences, they say, are especially well suited for facilitating coherence within quantitative consumer psychological research, as the mapping sentence provides a flexible structure for the research that may be adapted to a specific research context.

Using the Book It has been my intention in compiling this book that it should form a text that is appropriate for use with university and college students at all academic levels. In order to make the book applicable to undergraduate through to doctoral students, I have attempted to ensure that the author for each chapter provides both a theoretical grounding in the contents of the chapter and also that they provide applied examples of the use of the approach and methods they describe in consumer settings. The text is made more appropriate for use as a textbook as exercises are provided at the end of each chapter to test student learning of the chapter’s contents. The book may be used on its own as a textbook and may also be used as a supplementary text to more general textbooks in quantitative research. We have attempted in this book to provide readers with information about the major quantitative research approaches that are used in consumer psychology along with details of some new approaches that they may consider employing in their research.

Note 1. It should be noted that some consider consumer psychology to be a branch of industrial and organizational psychology.

References American Psychologival Association. (undated). Society for Consumer Psychology. Retrieved June 26, 2018, from www.apa.org/about/division/div23.aspx The Editors of Encyclopaedia Britannica. (2018). Consumer psychology. Encyclopædia Britannica website, Encyclopædia Britannica, inc. August 24, 2017. Retrieved June 26, 2018, from www.britannica.com/science/consumer-psychology

1 QUANTITATIVE RESEARCH Its Place in Consumer Psychology Cathrine V. Jansson-Boyd

The use of scientific methodologies to explain consumer behaviours is a relatively recent phenomenon, even though there were some early well-known psychologists, such as William James and John Watson, who indirectly or directly applied their understanding to consumer-related aspects ( Jansson-Boyd & Marlow, 2017). The field of consumer psychology has blossomed in the last few decades out of an amalgamation of different areas such as psychology, marketing, advertising, sociology, and anthropology. It seems that consumer behaviour is a natural extension of many of the theories produced within the aforementioned areas. However, recent years have seen many new theories being developed that are more specifically constructed to explain consumer thought processes and behaviours. The rapid development of models and theories within consumer psychology is a parallel progression to the expansion of the field as a whole. Such expansion is applicable both to the number of people conducting research within the area and the integration of ideas and methods from related disciplines. With the rapid growth, research has become ever more competitive as researchers strive to move our understanding of this discipline forward. Now more than ever it is essential to have a good and clear understanding of research methods as well as an in-depth knowledge of the specific areas that are being investigated. This is particularly important as consumer behaviour is becoming increasingly complex. The complexity stems from the fact that consumer psychology has become a field in which the theoretical underpinnings are drawn from a wide range of sister disciplines (e.g. social psychology, neuroeconomics, cognitive psychology, marketing, economics, advertising, sociology and neuroscience). This means that researchers now need a broader knowledge base in order to ensure that they have a good understanding of their favoured research topic, with the result being that consumer psychology is a truly applied research area. However, it does pose

2  Cathrine V. Jansson-Boyd

some difficulties, such as that researchers need to have a broader understanding of human behaviour in order to understand how their own work fits. This opens up new opportunities for using different types of research methods that may be borrowed from a related discipline.Venturing into using new techniques and equipment may require training.Thus, researchers are faced with different research skills requirements to those needed previously. This chapter will introduce you to the idea that theory has an important role to play within consumer psychological research. It will provide you with an overall perspective of why consumer scientists need to distinguish between quantitative and qualitative research methods. Furthermore, the chapter will explain how quantitative methods set themselves apart from qualitative methods, as well as why consumer psychologists have increasingly relied on quantitative methods as their preferred method of investigation. The chapter will outline how traditional quantitative approaches have changed to include what are seen as less traditional techniques that are often based on techniques used by related disciplines such as neuroscience. ‘Borrowing’ quantitative research methods from related areas has undoubtedly helped deepen the understanding of consumer behaviour by encouraging systematic investigation of consumer psychology. Hence the chapter will outline the significance of having a good understanding of different types of methodologies as well as statistical comprehension.

Chapter Learning Objectives The main take away message from this chapter is that quantitative research is fundamental to consumer psychology. However, as with any scientific discipline, there is also room for qualitative work depending on what is being investigated. Thus, it is not proposed that taking a quantitative approach to consumer research is the be-all and end-all. Once you have read this chapter, you should have an appreciation that consumer psychological research should be driven by a clear and coherent understanding of the topic investigated. You should recognize that it is the concepts that you wish to investigate and the suited theoretical underpinnings that should guide the research processes and chosen methodology. As a scientist, it is imperative to recognize that one cannot take the approach that ‘one model fits all’ kinds of research. Many scientists appear to prefer to repeatedly use certain types of methods. However, science progresses better when ensuring that the methods used are fit for purpose. Hence, it may at times mean that you will have to step out of your comfort zone and that you will have to use a quantitative or qualitative methodology or even a mixed methods approach that you have not previously used. Undoubtedly, it can be difficult to come to grips with different paradigms and perspectives (Kuhn, 1970). Nonetheless, consumer psychologists should embrace the idea that research is a creative process and never be afraid of experimenting with new techniques if this is what is going to give you the most robust results.

Quantitative Research  3

The role of theory A theory is effectively a set of interconnected ideas, definitions, and propositions that explains or predicts events or situations by specifying the relationship among variables (Campbell & Pennebaker, 2017). It is a key ingredient in the research process, and it is essential to know how you should use theories, as it will make you more successful when it comes to generating research questions. How theories are developed differs, as some are based on a conceptual framework whilst others have empirical underpinnings. Increasingly, ‘testability’ of a theory is important, in particular in consumer psychology, as it otherwise may lack reallife application and thus cast a shadow of doubt on the usefulness of the theories. After all, consumer psychology deals with real-life concepts, so it is therefore important that the theories are ecologically valid. For all consumer theories, the idea of generality and/or broad application is important. Hence, a theory is by its very nature abstract and not topic specific. Some theories may emphasize or address the same general ideas. However, each theory is unique by the terminology used to communicate the factors deemed to be important. For most research topics, successful interventions are clearly dependent on the use of appropriate theory. Different theories may be better suited to different types of research such as, for example, groups, individuals or organizations. Also, ask yourself if the theories you intend to draw upon provide your work with the insight required to make your research findings more ecologically valid. Consumer psychology as a discipline has been under fierce critique that it is often working within a narrow framework. Pham (2013) has even labelled the restrictive ways of working as a consumer a psychological sin. Such labelling stems from the fact that the discipline has traditionally focussed too much on three specific theoretical paradigm, which are cognitive psychology, social psychology, and behavioural decision theory (Pham, 2013). Narrow frameworks can restrict the applicability of research findings to anything other than academics and in some cases not even other researchers (Pham, 2013). This is unfortunate, as consumerism is very much a part of our everyday lives, and thus research within this field should genuinely investigate what happens in real life. Consumer psychology should also aim to keep pace with the rest of the scientific community, and there has globally been a shift to emphasize real-life application. This is something that is evident from funding calls made by research councils as well as research-based governmental funding (e.g. the Research Excellence Framework (REF) in the UK). Ideally, as an applied researcher, your work should aim to reach far and wide, and it is therefore essential to pick the ‘right’ theories that can help you to achieve such an aim. This may mean that you at times need to be ‘bold’ in your decisions and look for new and novel theories that you may wish to borrow from other (related) disciplines. Recently, within consumer psychology, there has been an increasing amount of work that is drawing

4  Cathrine V. Jansson-Boyd

on emotion theory and culturally based theories (e.g. see the recent International Handbook of Consumer Psychology edited by Jansson-Boyd & Zawisza, 2017). Additionally, as consumer psychologists have a relatively newfound interest in consumer neuroscience, we have also seen new ideas emerging on the back of neuroscientific studies. Hence, there has been some expansion in regard to the underpinnings used, even though this can be advanced further. Pham (2013) proposed that researchers should be open to use multiple theories and move away from pressures to differentiate their work by the use of a unique theoretical explanation, which he refers to as ‘theoretical tyranny’. Others have proposed that the obsession with theory and a narrow understanding may be hampering the development of new models and theories (Schwarz & Stensaker, 2014). This casts some doubts on what is the best method for moving science forward. Incorporating different types of theories (whether one or several) may at least help to overcome one common problem within consumer psychology, which is the tendency that researchers have to focus on the acquisition stage and largely ignore other consumption related activities (Pham, 2013; Sheth, 1982; Wells, 1993). Perhaps scientists should be braver and, at least occasionally, deviate from the ‘constraints’ of theories. However, there needs to be some caution so that the “baby is not thrown out with the bathwater”. Researchers should, at least for now, be encouraged to make use of sound theoretical knowledge, though they should be bold in their choices and branch out to look for underpinnings that can aid the exploration of new and exciting areas.

Theories, Be-All and End-All? There is no doubt that the use of theories has advanced the field of consumer psychology. Equally, they are a critical tool in providing researchers with a good framework for further testing. However, by continuing to use the same theoretical structure as before, it is possible that consumer psychologists may hamper progress or at least decrease the pace at which we further the understanding of the discipline.This is because there can be an overreliance on theories, ultimately constraining researchers from thinking outside the box and exploring different perspectives that could lead to a new theoretical framework that may take us in an entirely new and different direction. That is not to suggest that theories are redundant but rather that researchers may at times need to be brave enough to deviate from what is currently considered to be the norm. Alba (2012) has noted that some of the more exciting findings within the area of consumer behaviour have come from work that can be described as descriptive. For example, Dickson and Sawyer (1990) observed supermarket shoppers and found that they have poor awareness of the prices of products they have just purchased. It is findings such as the aforementioned that have led to the proposal that consumer psychologists should more commonly consider broadening the epistemology used (Pham, 2013).

Quantitative Research  5

What Method? There is a long-standing debate within the scientific research community in regard to the method that should be employed when conducting research. Usually, this debate is centred on whether to use a quantitative or qualitative approach to research, which is a somewhat crude and broad definition, as it undermines some important differences of each methodological approach. Consumer psychology as a discipline has moved towards the use of quantitative methods, something that has been paralleled with an advanced understanding of the discipline and increased sophistication in statistical analysis. A high number of published studies use experimentally based research structures and thus use statistical means to analyse the data. Effectively, an experiment tries to measure the effects of X on Y by controlling X and measuring Y, while at the same time keeping everything else constant. Hence, X (usually called the independent variable or IV) and Y (usually called the dependent variable or DV) will need to be predetermined. A quick search in the Journal of Consumer Psychology (which is devoted to psychological perspectives on the study of the consumer) reveals that the majority of the papers published in the last ten years are in fact using quantitatively based methodologies. Equally, a high number of consumer-based studies, published in other related journals, also predominantly employ a quantitative approach for investigative purposes. Many areas lend themselves particularly well to be investigated by using experimental designs, such as the role of touch in consumption (e.g. Jansson-Boyd & Marlow, 2007) or how emotions influence consumers (e.g. Gorn, Pham, & Sin, 2001). A significant proportion of quantitatively based studies make use of questionnaires to gather information about the consumers they wish to target. Questionnaires often allow researchers to gather a great deal of specific information. Just as with experimental studies, these will also be analysed by using statistical means. There are, however, those who question if the use of questionnaires is genuinely an effective tool when it comes to measuring consumer responses (e.g. Reid, 2013). This critique is largely based on the fact that consumers provide socially desirable answers to questions. Equally, questionnaires often fail to measure more subtle aspects that consumers are unaware of, such as implicit thoughts and emotions.

Distinguishing Between Quantitative and Qualitative Research In order to have a clear appreciation for quantitative methodologies and see how they are best suited for your research, you also need to understand the fundamental key differences between qualitative and quantitative methods. Broadly, both quantitative and qualitative concepts refer to the typology of the research in meanings of technique, instruments, data gathering, and analysis of the data collected. For key differences between quantitative and qualitative methods, see Table 1.1.

6  Cathrine V. Jansson-Boyd TABLE 1.1  Differences Between Quantitative and Qualitative Research

Quantitative method

Qualitative method

Features

• Data is collected through measuring different aspects • Quantifiable data • Statistical measurement • Positivist theory

Examples of methods used

• Experiments • Questionnaires

Methods of analysis used

• Data is analysed through numerical comparisons • Statistical analysis (e.g. regression, ANOVA, t test)

• Concerned with understanding human behaviour from an individual perspective • Describes the nature of a phenomenon • Interpretivist theory • Interviews • Grounded theory • Case studies • Focus groups • Data is analysed by extracting information from information gathered • Thematic analysis • Content analysis

Adapted from Minichiello, Aroni, Timewell, and Alexander (1990)

Quantitative data generates numerical data or data that can be put into categories, such as ‘yes’ or ‘no’ answers. For example, a scientist may use a Likert scale (a rating scale) to measure participants’ responses. The fact that quantitative data can be interpreted with the use of statistics means that it is viewed to be scientifically objective (Carr, 1994; Denscombe, 2010). Experiments typically yield quantitative data, as they are structured to measured specific outcomes. It is worth bearing in mind that experimental methods limit the ways in which participants can react to and express appropriate socially linked behaviour, which is something that often influences consumers in real life. Thus, the outcome of an experiment is context bound and simply a reflection of the assumptions originally made by the researcher when structuring the study. Consequently, the ecological validity of the study may be called into question. However, as previously mentioned, experiments are not the only option, and scientists can also make use of questionnaires and even controlled observations, as both produce quantitative data. Quantitative methodologies usually require larger sample groups. A larger sample size is important in that smaller sample sizes can make the study less reliable because of the low quantity of data (Denscombe, 2010). A smaller sample size would also hamper the ability to extrapolate the findings to a wider audience. However, this can also be the case when the testing population is a very specific sample group which may behave or respond differently to the population at large. For quantitative methods, the collected data is mostly quantifiable. Statistics helps us to summarize our data, describing patterns, relationships, and connections.

Quantitative Research  7

Statistics can be descriptive (when the data summarizes a given data set so that patterns may emerge from the data) or inferential (allows the researcher to identify statistically significant differences between groups and make conclusions about the data set). Researchers are usually interested in causal relationships between different concepts and by doing so can accept or reject hypotheses. A great strength of using quantitative research is that it can be—for example when using questionnaires—an efficient way to get great coverage in terms of participation, and it is also relatively economic. It is important to understand that all research methods have the capacity to be flawed. In addition to the already mentioned experimentally based limitations, they also don’t allow participants to explain their choices or their understanding of the questions (Carr, 1994). Black (1999) has pointed out that the researcher him or herself can also be an obstacle to conducting quantitative studies successfully, in that poor knowledge of the area researched can lead to a poor structure of the actual study. Equally, a poor understanding of the application of statistical analysis may negatively affect the analysis and subsequent interpretation. Quantitative methodologies have become the research paradigm commonly employed by those studying consumer behaviour. A paradigm, in this context, is a group of researchers sharing common assumptions about a specific topic, using similar methods, and dealing with comparable problems (Kuhn, 1970). Supporters of a paradigm share a common view of ‘the consumer world’ and methods employed to explore this world view are guided from the common perspective (Thompson, Locander, & Pollio, 1989). Many consumer psychologists currently believe that quantitative methods are essential as they are designed to ensure objectivity and reliability. This development is a reflection of the fact that many researchers have a need to view human beings in a reductionist manner, as it simplifies predictions of consumer behaviours. Presumably, this is also largely driven by the fact that specific recommendations can be made in regard to factors that can be altered to change behaviours and communicate effectively with consumers. This is something that is difficult to do on the back of qualitative research, as there may be ‘multiple’ solutions to one question. Quantitative techniques can be sharply contrasted with qualitative perspectives, such as the existential-­phenomenological view. Using this approach, consumer researchers are challenged not to interpret human behaviours through quantitative means, as it is deemed to be a narrow framework (Thompson, Locander, & Pollio, 1989). Existential-phenomenological psychology is humanistic in its approach and argues that humans should be looked at through the meanings that people make in their lives that are subsequently reflected in the choices they make and behaviours they engage in.This particular qualitative approach is closely associated with the philosophical teachings of Sartre (1962/orig. 1943), Heidegger (1962/orig. 1927), and Merleau-Ponty (1962/orig. 1945). However, philosophical underpinnings are also the basis for other qualitative research methods, such as phenomenology and ethnomethodology. There is a lot to be said in favour of using an

8  Cathrine V. Jansson-Boyd

existential-phenomenological approach or indeed any qualitative approach in consumer sciences. However, the impracticalities of doing so is one reason why psychology has been driven in the direction of positivist approaches whereby rigorous testing is key to achieving a clear understanding of behaviour. One reason is that qualitative methodologies are usually not as cost-effective as quantitative methodologies. They also require a fair amount of time to conduct. Even though qualitative designs do not make use of large sample sizes, researchers usually spend a considerable amount of time with each research participant. Furthermore, it is difficult to replicate any contexts, situations, events, conditions, and interactions that are studied. Neither can you make generalizations to a wider context, something that is particularly important when studying consumer behaviour, as consumer psychologists are usually interested in knowing what the vast majority of people do.

What About Business-Led Research? Increasingly, businesses have shown an interest in conducting consumer psychology– based research in order to further the understanding of human behaviour, ultimately in the hope of increasing their business profits. Unilever, Chrysler, Procter & Gamble, PepsiCo, and Johnson & Johnson are all big organizations that have both funded and published research based on the outcomes of their research. Large companies are often interested in being innovative in their work to give them an advantage over competitors. Not all work conducted within business organizations is suitable for publishing. Some may only wish to conduct research that is quick and that provides a snapshot of different aspects of consumer behaviours, whilst others may not make use of research tools properly, and there can be some ambiguity in regard to results. Such work also requires (just as more detailed work) a clear (and often broader) understanding for other related work previously conducted. Nevertheless, business-led research can help us get a better in-depth perspective of consumer behaviours as well as provide scientists with potential testing grounds for academic theories. Bearing in mind that consumer psychology is a discipline that focuses on ‘real-life’ behaviours, it is important to have a good relationship with the business world in order to ensure that we work within the parameters of ecological validity. Ideally, there should be a continuous flow of information between industry and academia to avoid researchers becoming detached from how the business world operates.Thus, we should be interested in considering how the use of specific methods might have an influence on whether businesses find our findings useful. In many cases, organizations find quantitative research more valuable, as it allows them to get an overview of how larger consumer groups may act, feel, and think. It is logical that quantitative methodologies will dominate business-led, consumerbased research, as being able to change even a small percentage of a large number of consumers can be hugely profitable. Furthermore, some quantitative methods, such

Quantitative Research  9

as surveys, are easy and quick to conduct, which can also be beneficial, as time may be limited if a fast turnaround is required.

Consumer Neuroscience: An Extended Family Member The fact that businesses have shown an interest in neuroscience (e.g. Nielsen and Unilever) may have accelerated the incorporation of it into the discipline of consumer psychology. The methodologies used within consumer neuroscience now has a role to play in shaping the methods used in consumer psychology. Over the last couple of decades consumer researchers have been forced to ask themselves whether they are approaching different lines of investigation from the correct angle. Especially as the area of neuroscience has progressed rapidly and highlighted that there are still unexplored areas to tap into. As a result, in the last decade it appears that consumer psychologists are keen to incorporate elements of neuroscience into the field and hence consumer neuroscience is now emerging as a new field. Consumer neuroscience investigates problems of consumption and marketing ­ hamberlain, through methods and findings from neuroscience (Lee, Broderick, & C 2007; Fugate, 2008). However, this pose a challenge for consumer psychologists more broadly as they need to get to grips with the field of neuroscience, an area that was previously largely ignored. The discipline of consumer neuroscience has mainly emerged from the subdiscipline of neuroeconomics and thus methodologies are also borrowed from the way in which they conduct research (Camerer, Loewenstein, & Prelec, 2005; Braeutigam, 2005; Fehr, Fischbacher, & Kosfeld, 2005; Kenning & Plassmann, 2005; Singer & Fehr, 2005). The incorporation of neuroimaging tools in consumer psychological research is relatively new and deemed to be an exciting development (Ariely & Berns, 2010). This includes methods such as electroencephalography (EEG), positron emission tomography (PET) and functional magnetic resonance imaging (fMRI). EEG measures voltage fluctuations on the scalp. This is done by placing electrodes on the skin which can detect electrical potentials that are produced by neurons. As with a lot of research tools, there are limitations, and this has sparked a debate about the usefulness of EEG (Knight, 2004). Nevertheless, there are some interesting consumer-based studies that have made use of EEG that shows that it can be effectively employed to further the understanding of consumer behaviour. For example,Young (2002) used it as an investigative tool to look at ‘branding moments’ in television commercials. He found that there was a correlation between moments identified by brain waves and moments identified by use of a behavioural, attention-sensitive method of picture sorting. Thus, showing that brand moments are important in television commercials. PET and fMRI both provide a different perspective, as they allow a direct observation of brain processes (Plassmann, Ambler, Braeutigam, & Kenning, 2007)

10  Cathrine V. Jansson-Boyd

and such brain imaging techniques are currently the most commonly used neuroscientific tools (Camerer et al., 2005; Logothetis, 2008; Logothetis, Pauls, Augath, Trinath, & Oeltermann, 2001) across the neuroscientific field. fMRI is perhaps the most promising technology used in neuroscience (Shamoo, 2010) as typically it compares a person’s brain activation under two conditions, a control condition and an experimental task condition. This allows for a comparison in activation between the two conditions. Usually, there are multiple trials that are averaged when conducting statistical analysis (Kenning et al., 2007; Huettel et al., 2009). There are different strengths and weaknesses to each neuroscience-based tool. In order to generate more robust findings, it may be an idea to combine two or more types of methods. Similarly to a game of Sudoku, different clues can assist in filling in the other gaps, thus learning from other tools (Kenning & Linzmajer, 2011). Undeniably, the use of different neuroscience tools provides consumer psychologists with opportunities to investigate different aspects to consumer behaviour that could otherwise not be explored. At times, such measures may provide a more objective perspective, at least comparatively to self-assessment methods (Huettel, Song, & McCarthy, 2004). It opens doors for the integration and testing of different theories that are based on neural mechanisms. Drawing more broadly on biology-based areas may help to address areas of consumer research that have a high level of unexplained variance (Riedl et al., 2010a, b). In particular, it is worth remembering that many consumer behaviours happens with minimal or no awareness (Kahneman, 2003), and thus neurosciencebased methods may provide insight into matters that we simply can’t just ask consumers about. One aspect that can be difficult to investigate is pricing, as consumers often fail to remember a price (Vanhuele & Drèze, 2002; Evanschitzky, Kenning, & Vogel, 2004; Ofir, Raghubir, Brosh, Monroe, & Heiman, 2008). There are also other difficulties with asking about pricing, as similar price levels can be perceived differently depending on the type of product. Higher prices can be perceived as a sign of good quality; however, it can also deter consumers from purchasing something if an increase is perceived as something they must give up and consequently experience a sense of loss (Lichtenstein et al., 1993; Völckner & Hofmann, 2007). One example of how such difficulties can be overcome can be noted from a study conducted by Knutson et al. (2007).They used an fMRI scanner to examine “whether distinct neural circuits respond to product preference versus excessive prices as well as if anticipatory activation extracted from these regions could independently predict subsequent decisions to purchase” (p. 148). Firstly, they showed participants an image of a product, this was followed by the same image but this time with a price from which they had to determine if they wished to purchase the product. The findings revealed that product preferences correlated with the activation of the nucleus accumbens (triggered by anticipation of gains) whilst the insula (associated with anticipation of losses) was correlated with excessive pricing. These findings are consistent with other neuroimaging studies in terms of which areas are activated by the anticipation of gains and losses.

Quantitative Research  11

Furthermore, Knutson et al., also noted that the medial prefrontal cortex was correlated with reduced prices. It was found that the activation of these regions predicted subsequent purchase much better than the self-report measures used. Thus it seems that there are specific neural representations, i.e. the insula, that can reflect negative price effects. Other studies, just like Knutson et al., have also demonstrated that pricing activates the medial prefrontal cortex. This was found by Plassmann et al. (2008) when their participants evaluated more expensive wine as being more pleasant. Thus suggesting that the change of price can influence neural representations of the pleasantness experienced by consumers. What Plassmann et al.’s (2008) and Knutson et al.’s (2007) studies essentially show us in terms of research methods is that using a neuroscience-based method can help consumer psychologists to overcome some of the difficulties we face when wishing to tap into subconscious influences. It also shows the importance of having a good understanding of quantitative methods as they correlate the neuroimaging with other quantitative measures. Other positive aspects of looking to neuroscience for solutions, even though a relatively minor point, is that these sort of studies overcome the problem of social desirability, as participants can’t control their brain activity (Camerer et al., 2005). It may seem that neuroscience-based studies are the solutions to many of the problems we face as researchers. However, it is worth remembering that they are very expensive to conduct and can be very time consuming (e.g. Ariely & Berns, 2010; Riedl et al., 2010b). As there are a limited number of studies conducted to date, there is a need to conduct more in order to validate and further the understanding of consumer neuroscience (e.g. Kenning & Linzmajer, 2011; Vul et al., 2009). Equally, the small number of participants used, usually due to the costs involved in using a high number, raise some concern in regard to whether the findings can be extrapolated onto the wider population (e.g. Kenning & Linzmajer, 2011). However, this may be offset by the fact that different researchers seem to have come to the same conclusions regarding brain activation, regardless of where the studies have been conducted, and with the use of different types of experimental approaches (e.g. Ambler et al., 2000; McClure et al., 2004; Koenigs & Tranel, 2007; Kenning et al., 2007; Plassmann et al., 2007). Even so, the argument has been made that the consistency in findings may be a problem in that the brain seems to process semantic variations between different brands in a similar fashion and hence indicate that the methods employed are too crude to accurately measure more subtle brain activations that may constitute significant brand differences (Savoy, 2005). Undoubtedly, fMRI is a very complicated research method to employ (Savoy, 2005) and thus even more reason as to why a researcher needs to be entirely at ease with the research literature prior to embarking on designing a new research programming that is reliant on using neuroscientific data. It is not within the realms of this chapter to go into detail about EEG, PET, and fMRI. Should you wish to find out more about these tools, there are several overviews that can provide a good starting point for further exploration (e.g. Lee,

12  Cathrine V. Jansson-Boyd

Broderick, & Chamberlain, 2007; Camerer et al., 2005; Kenning & Plassmann, 2005; Kenning et al., 2007; Lee et al., 2007; Perrachione & Perrachione, 2008; Riedl et al., 2010a).

Difficulties Using Quantitative Methods in Consumer Sciences Statistical methods have become increasingly sophisticated but, essentially, they are only as good as the method employed. If the method is not robust, then it does not matter what analysis is used, as the outcome will still be flawed. Using the correct quantitative method is imperative to a successful research outcome. It starts with the research design, and it is important to carefully consider that the method employed will closely fit with the analysis that the researcher intends to use. If you choose to use an experimental design, you should aim to be as objective as possible. Researchers should be emotionally indifferent about the respondents, and you should test variables in fully controlled conditions (Mesly, 2015). To do so effectively, two ‘identical’ groups are needed. One group to test one particular variable and another to act as a control group. In addition, all the other variables would need to be exactly the same. This is simply not possible when working with human beings, as they are not the same, and thus one can never have a control group that is identical to the experimental group. Naturally, one way of overcoming such a problem could be to employ a repeated measure design whereby you have one group of participants that takes part in two or more measures and allows you to compare whether individuals produce a different outcome depending on the variables introduced. Other problems with quantitative research methods resides within potential sources for method bias (see Podsakoff, MacKenzie, Lee, & Podsakoff, 2003, for an overview and additional biases), and in particular when using questionnaires, a tool commonly used within social sciences. In particular, there are two types of bias that researchers need to overcome: social desirability bias and consistency motif. Self-report methods can be difficult in creating ‘genuine’ responses, as people often respond in a socially desirable way. Social desirability is when people have a “need for social approval and acceptance and the belief that it can be attained by means of culturally acceptable and appropriate behaviors” (Crowne & Marlowe, 1964, p. 109). It is not uncommon that people wish to present themselves in a favourable light. This may be something that younger generations are particularly prone to, as they are often avid users of social media and thus virtually brought up thinking about their self-images and how to control them. When people’s responses to questions reflect how they wish to be perceived, rather than answer truthfully, it can distort the relationship between one or more variables (Ganster, Hennessey, & Luthans, 1983). Thus it can skew or even hide the real relationships that a researcher is hoping to tap into. People also tend to maintain consistency between their attitudes and cognitions (e.g. Heider, 1958; McGuire, 1966) when answering questions, something that has

Quantitative Research  13

been referred to as ‘consistency motif’ ( Johns, 1994; Podsakoff & Organ, 1986; Schmitt, 1994; Podsakoff et al., 2003). People do so because they wish to appear rational and therefore search for similar features in the questions so that they can respond in a similar fashion on those questions. The problem with this (similar to the social desirability bias) is that it produces associations between questions that may not otherwise exist. The consistency motif becomes particularly apparent in situations whereby they are asked to provide information about feelings and behaviour a while after they have experienced them (Podsakoff et al., 2003). Presumably, people then fail to remember accurately what they thought at the time or why they did something and thus wish to present a rational account.

Creative Methods Not only has there been a move towards the use of quantitative methods but also a move towards creatively overcoming previous problems in research to extract genuine responses from participants. This is often done by the use of technology, and one great example of this is the Implicit Association Test (IAT). The relationship between attitudes and behaviour is an important one to consumer psychologists in that we would like to have a clear understanding how people’s attitudes can be influenced (in terms of marketing). Hence, we have seen a stream of investigations that have attempted to explore this (e.g. Ajzen, 1991; Ajzen & Fishbein, 1980; Ajzen & Madden, 1986; Bargh, Chen, & Burrows, 1996; Fazio, 1989, 1990; Fazio, Chen, McDonel, & Sherman, 1982; Fazio & Zanna, 1978a). Such work has often been correlational in nature, and the work itself has involved asking participants about their attitudes and specific behaviours (e.g., Bagozzi, 1981; Bentler & Speckart, 1981; Davidson & Jaccard, 1979; Fazio & ­Williams, 1986; Fazio & Zanna, 1978b; Lavine, Thomsen, Zanna, & Borgida, 1998). The problem is that such work does not necessarily demonstrate whether attitudes predicts behaviour. A problem is that we can’t know for certain that we are indeed tapping into peoples’ genuine attitudes by asking them to respond to questions about different topics. People are not always conscious of the attitudes they hold and may therefore not be captured by self-report measures (e.g. Bargh & Chartrand, 1999; Dovidio, Kawakami, Johnson, Johnson, & Howard, 1997; Fazio, Jackson, Dunton, & Williams, 1995; Greenwald & Banaji, 1995; Hetts, Sakuma, & Pelham, 1999; Jones, Pelham, Mirenberg, & Hetts, 2002; von Hippel, Sekaquaptewa, & Vargas, 1997; Wittenbrink, Judd, & Park, 1997). This is why the IAT has been a welcomed addition to the research scene, as it has the capacity to measure implicit aspects of attitudes. Basically, the test measures strengths of associations between concepts by detecting response latencies in computer-based categorization tasks. This enables researchers to test attitudes to topics that people may not otherwise be able to or want to admit to. For example, the IAT has been used to measure racially based attitudes. The IAT is now relatively commonly used

14  Cathrine V. Jansson-Boyd

amongst social science researchers and has been found to be a reliable measure across different domains, including consumer psychology (Hao & Wang, 2013).

Don’t Be Afraid to Replicate As quantitative data is based on measured values, it can be fairly easily checked by others, as it is less open to ambiguous interpretation (Antonius, 2003). This is (at least partially) why some concerning discoveries have been made within the scientific discipline of consumer sciences. There has recently been many public debates about whether some research studies are ‘genuine’ or whether they contained manufactured results (e.g. Singal, 2017; Jha, 2012). This means that you should consider replicating studies for two reasons; firstly, because you want to ensure that you can generate the same results as previously conducted studies. Secondly, because you want to help create a suitably standardized knowledge base in which your own work is fittingly integrated. The fact that researchers have found themselves in a scientific landscape whereby their work is being scrutinized has led to encouragement of pre-registration of work that is yet to be conducted. This system has been developed as a means to combat questionable research practices. Basically, the pre-registration system is an opportunity to submit the research rationale, hypotheses, and design, and how the data is going to be analysed prior to conducting the work. This can then be submitted to a journal that is engaged with a pre-registration system, and the researcher can receive feedback on their intended work. If accepted for publication, researchers can continue their work knowing that it is more likely to be published when completed. There are some concerns that this process may act as a hindrance for exploratory work as well as lead to bias in terms of methods used (Gonzales & Cunningham, 2015) and as such may stifle research that is taking a different and creative approach.

Sophisticated Statistics Consumer scientists (the same as most other scientists) are expected to be able to communicate their results through a carefully thought through analysis, something that pre-registration may help with (Gonzales & Cunningham, 2015). As the methods have progressed, so also have the ways in which we analyse data. One example of data analysis progression can be noted from the use of structural equations modelling (SEM.) Rather than just using correlations or regressions, we may consider using SEM, something that Iacobucci (2009) has described as a “natural extension of factor analysis and regression” (p. 673). Even though SEM is available as a software package called Amos through SPSS (probably still the most commonly used statistical package by psychologists), Lisrel is a package that is more comprehensive (Iacobucci, 2009) and thus a slightly better tool to use. This emphasizes the need

Quantitative Research  15

for consumer psychologists to think about learning additional or different statistical packages to those they already know or are more familiar with.

Software for Conducting Analysis It is important for a consumer psychologist to have a good understanding of at least one type of statistical software, if not two. These days, sophisticated statistical software programmes removes much of the need for prolonged data analysis, especially when it involves large data sets (Antonius, 2003). Most psychologists make use of SPSS, but there are undoubtedly other equally, and in some instances better (as already mentioned in regard to SEM), statistically based programmes that can be used for analytic purposes. Sometimes, especially for researchers who are not affiliated with a university and thus need to pay for the statistical packages they use, it may be better to make use of a free statistical tool. Currently, there are some that can be downloaded and used, and JASP is one that can be used pretty much as SPSS but may lack some of its flexibility. There is also another free package called PSPP (note similarity in name to other statistical software), which is faster but has restricted capability compared to SPSS or JASP. Another free software programme worth exploring is R, which is often deemed by many as a more powerful tool than SPSS. However, when using R, it is essential to be familiar with writing code. It is worth noting that there are academics who have spoken out in favour of the use of R (Datahowler, 2016). Regardless of which statistical package you decide to use as a researcher, the tool will only be as good as you are, and it can only be used appropriately if you have a clear understanding of the statistical concepts required for the research you are conducting.

Conclusion There are many factors that have come together to make quantitative methods the dominant choice in consumer psychology.What is apparent is that scientists today face an increasingly challenging research climate. Not only are they expected to come up with new and exciting ideas but also be willing to learn and embrace complicated methods and tools. Because of the increasing complexity of the field, it is essential for researchers to carefully consider what underpinnings to use to drive their research forward. Currently, when many questions are asked in regard to the authenticity of published work, quantitative methods may be more important than ever. The results are expected to present good scientific findings and be replicable, irrespective of who the researcher is, so that retesting can take place, if so needed. However, it is worth bearing in mind that what may currently seem like the best approach to science may not hold up in 20, 30, or even 40 years’ time. Investigative science

16  Cathrine V. Jansson-Boyd

should be viewed as an ongoing journey whereby we are prepared to embrace new and exciting adventures in order to have the best overall outcome.

Chapter Exercise Read one qualitative study and quantitative study. There are two suggested next, but you can look at others. Once you have read both, answer the three following questions: 1. Could there be an advantage for Stoner and colleagues to also consider a qualitative line of investigation? If so, what would it be? 2. Could Joy et al. have employed a quantitative method to conduct their investigation? If yes, what do you think the benefit would be? 3. Do you think that one approach (quantitative or qualitative) is better than the other for progressing the understanding of the topics investigated? QUANTITATIVE STUDY Stoner, J. L., Loken, B., & Stadler Blank, A. (2018). The name game: How naming products increases psychological ownership and subsequent consumer evaluations. Journal of Consumer Psychology, 28(1), 130–137. QUALITATIVE STUDY Joy, A., Belk, R. W., Wang, J. J., & Sherry Jr, J. F. (2018). Emotion and consumption: Toward a new understanding of cultural collisions between Hong Kong and PRC luxury consumers. Journal of Consumer Culture, 1–20. https://doi.org/10.1177/1469540518764247.

References Ajzen, I. (1991). The theory of planned behavior. Organizational Behavior and Human Decision Processes, 50(2), 179–211. Ajzen, I., & Fishbein, M. (1980). Understanding attitudes and predicting social behaviour. Englewood Cliffs, NJ: Prentice Hall. Ajzen, I., & Madden, T. J. (1986). Prediction of goal-directed behavior: Attitudes, intentions, and perceived behavioral control. Journal of Experimental Social Psychology, 22(5), 453–474.

Quantitative Research  17

Alba, J. W. (2012). In defense of bumbling. Journal of Consumer Research, 38(6), 981–987. Ambler,T., Ioannides, A., & Rose, S. (2000). Brands on the brain: Neuro-images of advertising. Business Strategy Review, 11(3), 17–30. Antonius, R. (2003). Interpreting quantitative data with SPSS. London: Sage Publications Ltd. Ariely, D., & Berns, G. S. (2010). Neuromarketing: The hope and hype of neuroimaging in business. Nature Reviews Neuroscience, 11(4), 284–292. Bagozzi, R. P. (1981). Attitudes, intentions, and behavior: A test of some key hypotheses. Journal of Personality and Social Psychology, 41, 607–627. Bargh, J. A., & Chartrand, T. L. (1999). The unbearable automaticity of being. American Psychologist, 54, 462–479. Bargh, J. A., Chen, M., & Burrows, L. (1996). Automaticity of social behavior: Direct effects of trait construct and stereotype activation on action. Journal of Personality and Social Psychology, 71(2), 230–244. Bentler, P. M., & Speckart, G. (1981). Attitudes “cause” behaviors: A structural equation analysis. Journal of Personality and Social Psychology, 40(2), 226–238. Black, T. R. (1999). Doing quantitative research in the social sciences: An integrated approach to research design, measurement and statistics. London: Sage Publications Ltd. Braeutigam, S. (2005). Neuroeconomics—from neural systems to economic behaviour. Brain Research Bulletin, 67(5), 355–360. Camerer, C., Loewenstein, G., & Prelec, D. (2005). Neuroeconomics: How neuroscience can inform economics. Journal of Economic Literature, 43(1), 9–64. Campbell, S., & Pennebaker, J. (2017). Theory in psychology. Retrieved December 9, 2017, from https://saylordotorg.github.io/text_research-methods-in-psychology/s08theory-in-psychology.html Carr, L. T. (1994). The strengths and weaknesses of quantitative and qualitative research: What method for nursing? Journal of Advanced Nursing, 20(4), 716–721. Crowne, D., & Marlowe, D. (1964). The approval motive: Studies in evaluative dependence. New York, NY: Wiley. Datahowler. (2016). Retrieved November 29, 2017, from https://datahowler.wordpress. com/2016/09/10/no-more-excuses-r-is-better-than-spss-for-psychology-­u nder grads-and-students-agree/ Davidson, A. R., & Jaccard, J. J. (1979).Variables that moderate the attitude—behavior relation: Results of a longitudinal survey. Journal of Personality and Social Psychology, 37(8), 1364–1376. Denscombe, M. (2010). The good research guide: For small-scale social research. Maidenhead: McGraw-Hill. Dickson, P. R., & Sawyer, A. G. (1990). The price knowledge and search of supermarket shoppers. Journal of Marketing, 42–53. Dovidio, J. F., Kawakami, K., Johnson, C., Johnson, B., & Howard, A. (1997). On the nature of prejudice: Automatic and controlled processes. Journal of Experimental Social Psychology, 33, 510–540. Evanschitzky, H., Kenning, P., & Vogel, V. (2004). Consumer price knowledge in the ­German retail market. Journal of Product & Brand Management, 13(6), 390–405. Fazio, R. H. (1989). On the power and functionality of attitudes:The role of attitude accessibility. In A. R. Pratkanis, S. J. Breckler, & A. G. Greenwald (Eds.), Attitude structure and function (pp. 153–179). Hillsdale, NJ: Erlbaum. Fazio, R. H. (1990). Multiple processes by which attitudes guide behavior: The mode model as an integrative framework. In M. P. Zanna (Ed.), Advances in experimental social psychology (Vol. 23, pp. 75–109). New York, NY: Academic Press.

18  Cathrine V. Jansson-Boyd

Fazio, R. H., Chen, J. M., McDonel, E. C., & Sherman, S. J. (1982). Attitude accessibility, attitude-behavior consistency, and the strength of the object-evaluation association. Journal of Experimental Social Psychology, 18(4), 339–357. Fazio, R. H., Jackson, J., Dunton, B., & Williams, C. (1995). Variability in automatic activation as an unobtrusive measure of racial attitudes: A bona fide pipeline? Journal of Personality and Social Psychology, 69, 1013–1027. Fazio, R. H., & Williams, C. J. (1986). Attitude accessibility as a moderator of the attitude— perception and attitude—behavior relations: An investigation of the 1984 presidential election. Journal of Personality and Social Psychology, 51(3), 505–514. Fazio, R. H., & Zanna, M. P. (1978a). On the predictive validity of attitudes: The roles of direct experience and confidence. Journal of Personality, 46(2), 228–243. Fazio, R. H., & Zanna, M. P. (1978b). Attitudinal qualities relating to the strength of the attitude-behavior relationship. Journal of Experimental Social Psychology, 14(4), 398–408. Fehr, E., Fischbacher, U., & Kosfeld, M. (2005). Neuroeconomic foundations of trust and social preferences: Initial evidence. American Economic Review, 95(2), 346–351. Fugate, D. L. (2008). Marketing services more effectively with neuromarketing research: A look into the future. Journal of Services Marketing, 22(2), 170–173. Ganster, D. C., Hennessey, H. W., & Luthans, F. (1983). Social desirability response effects: Three alternative models. Academy of Management Journal, 26, 321–331. Gonzales, J. E., & Cunningham, C. A. (2015). The promise of pre-registration in psychological research. Downloaded on the 3rd of September, 2018 from: http://www.apa. org/science/about/psa/2015/08/pre-registration.aspx Gorn, G., Pham, M. T., & Sin, L. Y. (2001). When arousal influences ad evaluation and valence does not (and vice versa). Journal of Consumer Psychology, 11(1), 43–55. Greenwald, A. G., & Banaji, M. R. (1995). Implicit social cognition: Attitudes, self-esteem, and stereotypes. Psychological Review, 102, 4–27. Hao, H., & Wang, Y. (2013). Implicit Association Test and consumer psychology. Advances in Psychological Science, 21, 1865–1873. Heidegger, M. (1962). Being and time. New York, NY: Harper & Row. Heider, F. (1958). The psychology of interpersonal relations. New York, NY: Wiley. Hetts, J. J., Sakuma, M., & Pelham, B. W. (1999). Two roads to positive regard: Implicit and explicit self-evaluation and culture. Journal of Experimental Social Psychology, 35, 512–559. Huettel, S. A., Payne, J. W., Yoon, C., Gonzalez, R., Bettman, J. R., Hedgcock, W., & Rao, A. R. (2009). Commentaries and rejoinder to “trade-off aversion as an explanation for the attraction effect: A functional magnetic resonance imaging study”. Journal of Marketing Research, 46(1), 14–24. Huettel, S. A., Song, A. W., & McCarthy, G. (2004). Functional magnetic resonance imaging (Vol. 1). Sunderland: Sinauer Associates. Iacobucci, D. (2009). Everything you always wanted to know about SEM (structural equations modeling) but were afraid to ask. Journal of Consumer Psychology, 19(4), 673–680. Jansson-Boyd, C. V., & Marlow, N. (2007). Not only in the eye of the beholder: Tactile information can affect aesthetic evaluation. Psychology of Aesthetics, Creativity, and the Arts, 1(3), 170–173. Jansson-Boyd, C.V., & Marlow, N. (2017). In Jansson-Boyd, C. V., & Zawisza, M. J. (Eds.). Routledge international handbook of consumer psychology. Abingdon: Taylor & Francis. pp. 3–17. Jansson-Boyd, C. V., & Zawisza, M. J. (Eds.). (2017). Routledge international handbook of consumer psychology. Abingdon: Taylor & Francis.

Quantitative Research  19

Jha, A. (2012). Retrieved November 28, 2017, from www.theguardian.com/science/2012/ sep/13/scientific-research-fraud-bad-practice Johns, G. (1994). How often were you absent—a review of the use of self-reported absence data. Journal of Applied Psychology, 79, 574–591. Jones, J. T., Pelham, B. W., Mirenberg, M. C., & Hetts, J. J. (2002). Name letter preferences are not merely mere exposure: Implicit egotism as self-regulation. Journal of Experimental Social Psychology, 38, 170–177. Kahneman, D. (2003). Maps of bounded rationality: Psychology for behavioral economics. The American Economic Review, 93(5), 1449–1475. Kenning, P., & Linzmajer, M. (2011). Consumer neuroscience: An overview of an emerging discipline with implications for consumer policy. Journal für Verbraucherschutz und Lebensmittelsicherheit, 6(1), 111–125. Kenning, P., & Plassmann, H. (2005). Neuroeconomics: An overview from an economic perspective. Brain Research Bulletin, 67(5), 343–354. Kenning, P., Plassmann, H., & Ahlert, D. (2007). Applications of functional magnetic resonance imaging for market research. Qualitative Market Research: An International Journal, 10(2), 135–152. Knight, J. (2004). The truth about lying. Nature, 428(6984), 692–694. Knutson, B., Rick, S., Wimmer, G. E., Prelec, D., & Loewenstein, G. (2007). Neural predictors of purchases. Neuron, 53(1), 147–156. Koenigs, M., & Tranel, D. (2007). Prefrontal cortex damage abolishes brand-cued changes in cola preference. Social Cognitive and Affective Neuroscience, 3(1), 1–6. Kuhn, T. S. (1970). The structure of scientific revolutions. Chicago, IL: University of Chicago Press. Lavine, H., Thomsen, C. J., Zanna, M. P., & Borgida, E. (1998). On the primacy of affect in the determination of attitudes and behavior:The moderating role of affective-cognitive ambivalence. Journal of Experimental Social Psychology, 34, 398–421. Lee, N., Broderick, A. J., & Chamberlain, L. (2007).What is “neuromarketing”? A discussion and agenda for future research. International Journal of Psychophysiology, 63(2), 199–204. Lichtenstein, D. R., Ridgway, N. M., & Netemeyer, R. G. (1993). Price perceptions and consumer shopping behavior: A field study. Journal of Marketing Research, 234–245. Logothetis, N. K. (2008). What we can do and what we cannot do with fMRI. Nature, 453(7197), 869–878. Logothetis, N. K., Pauls, J., Augath, M., Trinath, T., & Oeltermann, A. (2001). Neurophysiological investigation of the basis of the fMRI signal. Nature, 412(6843), 150–157. McClure, S. M., Li, J.,Tomlin, D., Cypert, K. S., Montague, L. M., & Montague, P. R. (2004). Neural correlates of behavioral preference for culturally familiar drinks. Neuron, 44(2), 379–387. McGuire, W. J. (1966). The current status of cognitive consistency theories. In S. Feldman (Ed.), Cognitive consistency (pp. 1–46). New York, NY: Academic Press. Merleau-Ponty, M. (1962). The phenomenology of perception. London: Routledge & Kegan Paul. Mesly, O. (2015). The hypothetico-deductive method. In Creating models in psychological research: Springer briefs in psychology. Cham: Springer. Minichiello, V., Aroni, R., Timewell, E., & Alexander, L. (1990). In-depth interviewing: Researching people. Melbourne: Longman Cheshire. Ofir, C., Raghubir, P., Brosh, G., Monroe, K. B., & Heiman, A. (2008). Memory-based store price judgments: The role of knowledge and shopping experience. Journal of Retailing, 84(4), 414–423.

20  Cathrine V. Jansson-Boyd

Perrachione, T. K., & Perrachione, J. R. (2008). Brains and brands: Developing mutually informative research in neuroscience and marketing. Journal of Consumer Behaviour: An International Research Review, 7(4–5), 303–318. Pham, M. T. (2013). The seven sins of consumer psychology. Journal of Consumer Psychology, 23, 411–423. Plassmann, H., Ambler, T., Braeutigam, S., & Kenning, P. (2007). What can advertisers learn from neuroscience? International Journal of Advertising, 26(2), 151–175. Plassmann, H., O’Doherty, J., Shiv, B., & Rangel, A. (2008). Marketing actions can modulate neural representations of experienced pleasantness. Proceedings of the National Academy of Sciences, 105(3), 1050–1054. Podsakoff, P. M., MacKenzie, S. B., Lee, J. Y., & Podsakoff, N. P. (2003). Common method biases in behavioral research: A critical review of the literature and recommended remedies. Journal of Applied Psychology, 88(5), 879. Podsakoff, P. M., & Organ, D. W. (1986). Self-reports in organizational research: Problems and prospects. Journal of Management, 12, 69–82. Reid, A. (2013). Is it implicit? Implicit vs explicit techniques in consumer insights research. Retrieved November 27, 2017, from www.sentientdecisionscience.com/implicit-vsexplicit-techniques-in-consumer-insights-research/ Riedl, R., Banker, R. D., Benbasat, I., Davis, F. D., Dennis, A. R., Dimoka, A., & MüllerPutz, G. R. (2010a). On the foundations of neuroIS: Reflections on the Gmunden retreat 2009. CAIS, 27, 15. Riedl, R., Hubert, M., & Kenning, P. (2010b). Are there neural gender differences in online trust? An fMRI study on the perceived trustworthiness of eBay offers. MIS Quarterly, 34(2), 397–428. Sartre, J. P. (1962). Being and nothingness. New York, NY: Washington Square Press. Savoy, R. L. (2005). Experimental design in brain activation MRI: Cautionary tales. Brain Research Bulletin, 67(5), 361–367. Schmitt, N. (1994). Method bias: The importance of theory and measurement. Journal of Organizational Behavior, 15, 393–398. Schwarz, G., & Stensaker, I. (2014). Time to take off the theoretical straightjacket and (re-) introduce phenomenon-driven research. The Journal of Applied Behavioral Science, 50(4), 478–501. Shamoo, A. E. (2010). Ethical and regulatory challenges in psychophysiology and neuroscience-based technology for determining behavior. Accountability in Research, 17(1), 8–29. Sheth, J. N. (1982). Consumer behavior: Surpluses and shortages. In: Advances in consumer research (Vol. 9, pp. 13–16). Provo, UT: Association for Consumer Research. Singal, J. (2017). Retrieved November 28, 2017, from www.thecut.com/2017/03/cornellfood-brand-lab-problems.html Singer, T., & Fehr, E. (2005). The neuroeconomics of mind reading and empathy. American Economic Review, 95(2), 340–345. Thompson, C. J., Locander, W. B., & Pollio, H. R. (1989). Putting consumer experience back into consumer research: The philosophy and method of existential-phenomenol ogy. Journal of Consumer Research, 16(2), 133–146. Vanhuele, M., & Drèze, X. (2002). Measuring the price knowledge shoppers bring to the store. Journal of Marketing, 66(4), 72–85. Völckner, F., & Hofmann, J. (2007). The price-perceived quality relationship: A meta-­ analytic review and assessment of its determinants. Marketing Letters, 18(3), 181–196. von Hippel, W., Sekaquaptewa, D., & Vargas, P. (1997). The linguistic intergroup bias as an implicit indicator of prejudice. Journal of Experimental Social Psychology, 33, 490–509.

Quantitative Research  21

Vul, E., Harris, C., Winkielman, P., & Pashler, H. (2009). Puzzlingly high correlations in fMRI studies of emotion, personality, and social cognition. Perspectives on Psychological Science, 4(3), 274–290. Wells, W. D. (1993). Discovery-oriented consumer research. Journal of Consumer Research, 19(4), 489–504. Wittenbrink, B., Judd, C. M., & Park, B. (1997). Evidence for racial prejudice at the implicit level and its relationship with questionnaire measures. Journal of Personality and Social Psychology, 72, 262–274. Young, C. (2002). Brain waves, picture sorts, and branding moments. Journal of Advertising Research, 42(4), 42–53.

2 USING CONTEMPORARY QUANTITATIVE TECHNIQUES Or Shkoler

The goal of this chapter is to acquaint the reader with fundamentals of quantitative research and scientific orientation. Consumer psychology is a field of study investigating how our thoughts/ cognitions, beliefs, feelings, and perceptions influence how people buy and/or relate to goods and services. It is devoted to the individual, group, and organizational levels of research alike. Though a psychology sub-field, it is often multidisciplinary, with contributions from (1) psychology (individuals’ behavior and mental processes), (2) sociology (collective behavior and group decisions), (3) social psychology (how individuals influences and maybe influenced by others or groups), (4) anthropology (individual-cultural relationships), (5) economics (production, exchange and consumption of goods and services) (see Loudon & Della Bitta, 1993). Further, consumer behavior may be defined as the decision process and physical activity individuals engage in when evaluating, acquiring, using, or disposing of goods and services. To that end, it is important to differentiate between (a) a customer and (b) a consumer. The term ‘customer’ refers to “someone who purchases from a particular store/company/firm”, while the term ‘consumer’ is a more generalized ‘futuristic’ definition of the former. Meaning, someone who buys a bottle of water in Walmart, for example, is their customer. But someone who could potentially buy a bottle of water, or anything purchasable for that matter (does not matter where), is a consumer. Thus, a consumer may or may not become a customer de facto. Consumers’ definition is not limited to monetary exchanges alone, but also other services and intangibles (e.g., education and philosophy, ideas). It becomes obvious why the understanding of these distinctions is important in the marketing world and competitive markets (see Loudon & Della Bitta, 1993).

Using Contemporary Quantitative Techniques  23

Studying and understanding consumer behavior is significant due to the simple fact that we spend many hours in the marketplace, either thinking about buying goods and services or purchasing them in the end. The economic potential of this behavior is infinite, because there will always be a need to buy and the supply to match it, and therefore the market system could benefit tremendously from understanding these behaviors, their underlying antecedents, and their outcomes. As our attitudes, emotions, and behaviors may be influenced by both external and internal factors (i.e., personality, environment, job context, peers), it becomes more important to understand their interactions and mutual effects in relation to consumer behavior. To this end, we will tap into scientific research.

The Research Framework and Its Process The study of a subject or a discipline is made easier and clearer by examining it in an organized and well-established fashion—the research. It is a framework of thinking inquisitively, critically, and analytically; questioning what we observe; studying further; and attempting to explain and understand these observations, while drawing conclusions and inferences to deepen and enhance our knowledge, skills, and understandings. The prominent goal of the research is, thus, to reach valid, objective, and accurate answers to our research questions, as much as possible, through investigating the effects of one or more factors/phenomena on other one or more factors/phenomena (see also: Marczyk, DeMatteo, & Festinger, 2005, p. 46). And as implied by Reeve (2016), conducting research is “necessary for a coherent science, and only a coherent science is capable of constructing a general theory” (p. 1). The research usually follows a routine-like and rigorous chain process: (1) Deciding on a research direction and formulating a research ‘problem’  (2) conceptualizing a research design/model (as per step 1)  (3) writing the research proposal (i.e., research hypotheses) with adequate theoretical background/ literature review  (4) constructing/using an instrument for data collection  (5) selecting a sample to be collected  (6) collecting data  (7) processing and displaying data  (8) discussing and concluding the research, and writing a research report. The research needs to be unbiased, meaning, not concealing (or emphasizing) something because of the researcher’s interests (for further reading, see Kumar, 2014).

Research Types It is clear that research may very well help us in better understanding complex, yet broad, phenomenon, such as consumer behavior. However, there are many characteristics to the research. The main kind employed in psychology is the empirical research in which any conclusions drawn must be based on hard evidence by gathering information/data from real-life experiences, subjects, or observations.

24  Or Shkoler

In terms of application, there are also the pure research and the applied research approaches. Simply put, pure research is mainly to intellectually challenge the researcher; this may or may not have practical applications (and, thus, is usually more abstract). The more common approach in social sciences is applied research, for it has more practical applications, which are of paramount importance, especially in consumer behavior research. In addition, in terms of objectives/goals, the research is divided into the following: Descriptive research—aims to only describe the problem/phenomenon the research is directed at or provide ‘dry’ information in that regard (e.g., frequencies, percentages and proportions [such as how many women are in the research, marital statuses, other characteristics collected in the sample, and more], means, sampling errors). Correlational—attempts to ascertain or establish relationships/associations between at least two aspects of a problem/phenomenon in the research and, at times, to replicate those relationships to strengthen their reliability and validity. It is highly important to note that correlational studies assume no causal relationships (!), but rather, as its name implies, correlational associations (or, how at least two aspects mutually affect each other). Explanatory—whereas the correlational type aims to discover relationships, the explanatory type attempts to explain why those relationships form/exist and their underlying mechanisms and rationales. The main idea behind this type is to ask the question ‘why’. This type, as opposed to the correlational one, aims to investigate the causal links between phenomena/occurrences. Usually, researchers would want to find a one-directional effect/causation and not the common mutual effect (in correlational studies). For example, if there is a relationship between stress and depression, from an explanatory standpoint, the goal is to point that only one aspect in this equation affects the other; one aspect causes the other, and not the other way around. This type of research is usually grounded in strong theory basis, as casualty is a very sensitive and strict issue in social sciences. Researchers tend to perform experimental studies (will be discussed in the “Primary Data Collection Methods” [4.2] section) in order to achieve the closest proximity to causal associations. Exploratory—usually done to explore something (phenomenon, problem, area of research, etc.) where little to nothing is known about it. This is a general research type in which one may employ a descriptive approach, correlational or even explanatory—all for the super-goal of exploring the unknown, giving more data on it, and trying to develop or deepen the theory behind it. Usually, a pilot study is undertaken to start exploring, which consists of a significantly lower sample size. Such small-scale studies may help with deciding whether the area explored is worthy and may benefit from further investigation. Based on recommendations and assessments from the

Using Contemporary Quantitative Techniques  25

exploratory research, others may follow-up in its light (i.e., descriptive, correlational, explanatory). However uncalled for, exploratory research is often dismissed and/or ruthlessly criticized for its lenient and non-confirmatory approach. Nevertheless, it is the exploratory research alone that gave birth to all ­theories—it is the genesis of research, because nothing is known until investigated. When facing the unknown, the exploratory approach is the best method to use in order to formulate a basis for a theory, which makes a great contribution to science and other research types. However, usually research is not built on a single type. Most is actually a combination of at least two types. For example, the descriptive and the correlational types go well hand in hand.When comparing groups (e.g., youngsters and adults), they are usually given their descriptive data (means and standard deviations, for examples) but also the inferential data as well (whether one group “is higher” than the other, in a statistically significant manner1).

A General Example Let us assume our research goal is the study of self-esteem and tendency to buy expensive/luxurious goods and/or services, and possible gender differences.Thus, different research types will aim at different aspects of the research: Descriptive goals—how many men and women are there in the sample? What type(s) of industry is/are in question? What is the respondent’s age range and mean? Correlational goals—is there a relationship between self-esteem and the tendency to buy expensive/luxurious goods and/or services? What is the strength of that relationship? Is this relationship more prominent for men than women? Explanatory goals—why the relationship between self-esteem and propensity to buy expensive/luxurious goods and/or services exist? What causes the propensity to buy expensive/luxurious goods and/or services? Why are there gender differences? Exploratory goals—is the area of self-esteem in marketing viable and worth investigating? What, or how much, do we know about it and its supposed correlates? Can different people behave differently in this regard?

Quantitative vs. Qualitative Research (and Mixed Methods) In addition to the types of studies mentioned earlier, there are also two core paradigms of research, each aims at the same goal but from different perspectives and hence different, questions, methodologies, designs, samples, flexibility/­r igorousness, and objectives.

26  Or Shkoler

Quantitative approach—used to quantify a problem/phenomenon via numerical data or data that can be transformed into usable statistics, meaning, to measure/quantify attitudes, opinions, behaviors, and other defined variables,2 and then attempt to draw conclusions and inferences in order to generalize the results. This research assumes that the reality is fixed and measurable. The research aims to establish general laws of behavior and phenomenon across different settings/­ contexts, and it is used in testing a theory to ultimately support or reject/disprove it. In this method we use measurable data to formulate facts and uncover patterns in question, meaning determining the extent of a phenomenon.The data collection methods are much more structured and less flexible in this manner, and hence this approach requires a much larger sample size. Quantitative data collection methods include a variety of survey forms (online surveys, paper surveys, mobile surveys, group surveys), face-to-face interviews, telephone interviews, longitudinal studies, website interceptors, online polls, data aggregation (e.g., meta-analysis), and systematic observations. Data analyses are heavily bent on statistics and ultimately done in computer environments and statistical software. In addition, the data reported is in and through the statistical and inferential languages. Qualitative approach—is primarily exploratory in nature and thus much more flexible. Used to gain a deeper understanding of underlying reasons, opinions, and motivations/drives of individuals, with emphasis on personal narratives and experiences, rather than measuring them. It provides insights into the problem in question or helps in developing ideas or hypotheses for potential future quantitative research (or consequent qualitative investigation). This research is concerned with understanding human behavior from the informants’ perspective, while assuming that reality is negotiable and dynamic. Moreover, it is usually used to uncover trends in thoughts, attitudes, and opinions of individuals, aiming to deepen the understanding of an issue in question and determining the nature of a phenomenon. Qualitative data collection methods vary using unstructured or semi-structured techniques: Focus groups (group discussions, very common in marketing research), individual interviews, and observations.To that end, there is a need for a far smaller sample size (as opposed to the quantitative counterpart), and respondents are usually selected to fulfill a preset quota (even though the ‘quota’ can be managed in quantitative participants selection as well). Data analyses in this regard usually begin with the researcher and his/her interpretation of the qualitative data collected. At times, the interpretation is followed by some statistical analyses as well, but not as intense and rigid as in the quantitative research, and it is of less importance than the work done by the researcher. In addition, the data reported is in the informants’ language. Both approaches (and most research), however, deal with the effects of one or more independent variables on one or more dependent variables.The  former is also called predictors or antecedents, while the latter is called outcomes or criterion(a). The predictor is independent of the outcome and thus the name, which means that it affects the outcome and not the other way around. For example, a consumption of a dissociative drug may result in hallucinations. In this case, the hallucinations are

Using Contemporary Quantitative Techniques  27

the outcome of the drug (the predictor/cause).The independent variable is the factor that may be manipulated or controlled by the researcher (in experiments), but it usually is quantified and measured just as the rest of the variables. Table 2.1 displays a comparison between quantitative and qualitative research approaches in a number of important research-wise parameters. There are vast discussions on which paradigm is ‘better’, as each has advantages over the other and lacks in places the other might fill in the gaps. The choice between them should be relevant to the study at hand, research questions and hypotheses, and the research goal. However, some may find it would be ‘easier’ not TABLE 2.1  Quantitative and Qualitative Approaches and Their Research Issues

Quantitative Data is usually statistically interpreted. “More objective” since it is based on the principles of mathematics Theorizing Useful for testing and validating already existing theories or adding small pieces of new information to them. Correlational and/or descriptive in nature Analyses Modern statistical software speed removes the long analysis process (even for big data) Replicability Based on measured values and can be validated and re-validated by others, when replicated, because numerical data is less open to ambiguities or interpretations. In addition, hypotheses can also be tested and re-tested due to the use of well-established and well-defined measurements and statistical analyses Objectivity

Context

Experiments, for example, do not take place in natural settings and do not allow participants to explain their choices, answers, or meaning. Surveys, for example, are a “snapshot” and crosssectional

Qualitative Data is usually interpreted via the researcher’s understanding and expertise, which is considered “more subjective” Useful for generating new theories or adding more complex and intricate pieces of information to them. Explanatory and/or exploratory in nature More lengthy, delicate and requires a great amount of expertise in the field of study Usually based on unstructured or semi-structured surveys, and the involvement of the researcher is very high with the participants. As the main idea is finding subtleties, complexities, contradictions, and narratives, which represent the social reality better than a “status/situation report” (as our behavior is not linear), while using small-scale samples, replicability is usually not viable/reliable The researcher is closely involved with the participants (i.e., an insider’s view of the field), which allows for finding issues that are often missed (e.g., subtleties, complexities and contradictions) by quantitative approaches (Continued)

28  Or Shkoler TABLE 2.1 (Continued)

Quantitative Researcher’s expertise

Sample size

Focus bias

Reporting standards

Qualitative

Requires expert knowledge of the field under study, and great care must be taken when doing so (e.g., when looking for symptoms of mental illness, or when investigating sensitive issues) in order to discover possible relationships, causes, effects, and dynamic processes Usually requires a small-scale Usually requires large sample sample, mainly because of the sizes. The smaller the sample, time and costs (from reaching to the less reliable the results or participants, to the data analyses) their generalization beyond the involved in conducting such specific sample used for the research research Focusing on the generation of new Focusing on specific phenomena ideas and/or theories, which or hypotheses might lead to might lead to inconclusive missing out on other phenomena. results, but more complex and This research focuses on theory intimate testing, not theory generation Requires a minimum of adequate knowledge in the application of statistical analyses; otherwise, there would be a considerable issue with the analyses, hypotheses testing, and subsequent analyses

Each research paradigm has its own unique reporting standards and should be adhered and followed through to coherently convey the research’s goals and findings to the reader. For quantitative research, see Appelbaum, Cooper, Kline, Mayo-Wilson, Nezu, & Rao, 2018. For qualitative research, see Levitt, Bamberg, Creswell, Frost, Josselson, & Suárez-Orozco, 2018.

to choose between them, thus combining them instead. This is called the mixed methods approach. Mixed methods approach—aims to utilize the best method, but in reality, uses the strength of both—quantitative and qualitative approaches—in order to answer the issues in question. Ultimately, it combines at least two different data collections and analyses methods for the research problem. The idea behind combining at least two methods lies in the rationale that some methods are more appropriate to be used in some situations and some in other situations. An important underlying notion is the ability to gain as much information about the phenomena under research, trying to reach as much generalizability as possible along the process. All of the different approaches aim to answer different (or even same) research questions/issues, meaning viewing the phenomena from different angles, addressing them with different methods and goals, and deducting different results for further discussion about them (for further reading, see Kumar, 2014).

Using Contemporary Quantitative Techniques  29

Although there are plenty of research variations and approaches to choose from, the current chapter focuses on the quantitative paradigm.

Research Questions and Hypotheses As stated, research aims to study a phenomenon, and, to this end, one must formulate a research question that will be translated into corresponding hypothesis(es). A research question is, literally, a question (usually ending with a question mark) that tries to encompass the research goal or topic. A hypothesis, however, has many definitions (e.g., Bailey, 1978, p. 35; Black & Champion, 1976, p. 126; Grinnell, 1988, p. 200; Kerlinger, 1986, p. 17; Warner, 2013, p. 1070; Webster’s Third New International Dictionary, 1976). To synthesize these definitions, we may conclude that a hypothesis is actually “an educated guess” (Weinstein, 2010, p. 17): (1) It must be relatively simple to understand; (2) it is specific, focused and clear; (3) it is a tentative declaration of a lawlike relationship of some kind between, at least two, phenomena/variables; (4) aimed to prove a research claim (a replication or otherwise) (see also Janicak, 2017, p. 140); (5) it may be supported or dismissed, based on empirical evidence; and (6) must be formulated based on prior knowledge and sufficient theoretical reasoning (justification for replicating past findings and/or claiming otherwise). The hypotheses are the researcher’s attempt to explain the investigated phenomena in a simple short sentence (Marczyk et al., 2005). Furthermore, both questions and hypotheses must answer several criteria to be considered viable and adequately established for investigation and research: Empiricism (verifiability vs. falsifiability)—need to be based on empirical data and there must be sufficient theoretical background to support them. An important notion of the empirical capacity of a hypothesis is that it is concerned with the observable reality, meaning a good hypothesis will be based on what we can observe in one way or another, directly or indirectly, by our five senses and human organs. Unobservable phenomena, such as the human soul, Freud’s subconscious, and so on, are not good objects for hypotheses generation. Furthermore, we are dealing with human behavior, not mathematical axioms, and as such hypotheses also need to be falsifiable and not dogmatic; we could either support/confirm them or reject/­ dismiss them. A hypothesis which cannot disproven is not considered a good hypothesis (e.g., Popper, 1963) in social science. Compatibility/determinism—we assume that there is a natural order to reality. Centuries of research and science have accumulated a large body of knowledge, with many reliable laws and hypotheses. Thus, new hypotheses are expected to be in line with what we already believe about the natural order of things. However, this criterion is by no means absolute and usually very relevant to a replicability of the results. The best idiom in this case is “nothing ventured, nothing gained”, so that in order to produce new

30  Or Shkoler

knowledge and laws, we must not only replicate past results but also think how to add to them in various ways, limited or extensive as they may be. This criterion is also very delicate, because we hypothesize according to known beliefs and knowledge, but the alignment of a new hypothesis with past ones should not be by 100%, unless, as stated, it is made and tested as a replication of past research. Relevance/usefulness—the ability of the hypothesis to produce results/ information that will aid us in our lives in some way.The phenomena under investigation must be relevant to the field of study and science in general, with some emphasis on practical applications. Simplicity/parsimony—a hypothesis should always strive to be as simple, or ‘economical’ (and short), as possible, even when the phenomenon investigated is rather complex. Predictive power/scope—research’s goal is also to be able to predict future events (i.e., wide range of predictive capacity). Hence, a good hypothesis will also be applicable to as many other circumstances as possible, leading to more predictions and consequences we would be able to infer from it. Laws of physics are a good example for a very powerful and wide-ranged predictive capacity (e.g., anywhere on earth, an apple will always drop if thrown to the air).

Hypotheses’ Direction of Association In addition to those criteria, a hypothesis may be phrased (and statistically tested) in two fashions: One-tailed (directional)—a hypothesis with a clear direction of association between its parameters. The correlation can go both ways—positive or negative—but it is still clear as to what the relationship between the variables is expected to be. Positive correlation/association—congruent relationship, meaning same directional correlation; an increase in one variable is followed by an increase in another variable, and vice versa. For an example, see the “general example” section that follows. Negative correlation/association—inverse relationship, meaning opposite directional correlation; an increase in one variable is followed by a decrease in another variable, and vice versa. For an example, see the “general example” section that follows. Two-tailed (nondirectional)—a hypothesis with no clear direction of association between its parameters. The most common is a plain correlation/ relationship between at least two variables. The outcome will be sufficient for further discussion, whether the association is positive or negative.

Using Contemporary Quantitative Techniques  31

Main Types of Hypotheses Regardless of the direction of hypotheses, they also haves two essential types: Relational—a hypothesis that posits a relationship of some kind between at least two variables. This association is statistical in nature and by no means hints at a causal inference. Linguistically, such hypotheses will use terms such as correlation, association, relationship, difference, or “more/less than”. Correlative (or associative)—depicts a correlative relationship between variables. For example, the happier a customer gets, the greater the likelihood he/she will return to the same store to buy (this is a positive relationship and can be rephrased as follows: There is a positive correlation/­relationship/ association between customers’ happiness and likelihood for returning to buy); the more annoyed a customer is with the ‘customer service’, the less likely he/she will buy again from the company (this is a negative relationship and can be rephrased as follows: There is a negative correlation/­ relationship/association between customers’ annoyance level and likelihood for returning to buy). Differentiative (or differential)—depicts a difference between groups (between-subjects design) or states/time stamps (within-subjects design). For example, men have a higher tendency to visit car sales websites than women (different groups comparison = between-subjects’ hypothesis); women, after intense commercials, would buy more perfume than before the commercials (same group comparison, different states/time stamps = within-­ subjects’ hypothesis). These are the most common hypotheses in social science, as they require no causal inference and should be regarded as such even upon their approval/support. Causal—sometimes called ‘experimental’ (but it is not correct in all cases and very dependent on the goal of the experiment, and as such will be regarded as causal throughout the chapter). A hypothesis posits a causal relationship of some kind between at least two variables. Furthermore, causal inference must meet three important criteria: • • •

Chronological/time order of precedence (the predictor must come before the outcome). Significant theoretical and statistical relationship between the variables. Absence of alternative explanations for the relationship (which is controlled in experiments, for example).

The third point is problematic, because there are so many factors we cannot account and/or control for in a research, but its meaning is to control as much as

32  Or Shkoler

possible and ensure a min-max approach. In other words, to minimize the ‘noise’ (possible interference by random factors) in order to maximize the accuracy and validity of the results. Experimental research designs are known for making causal inferences, but sufficient and solid theoretical logic may also help in formulating a causal link between variables in a correlative study as well (such as precedence of a personality trait over a certain behavior or experience; see, for example, Shkoler, Rabenu, & Tziner, 2017). A simple example is a PTSD (post-traumatic stress disorder) after an acute trauma, in which the trauma is the cause of the PTSD, which is the outcome/effect. Another example may be that waiting-in-que negatively influences/causes customer satisfaction (which implies a negative relationship between the variables, but with the emphasis of causality). Linguistically, such hypotheses will use terms such as influence and cause.

Main Types of Associations/Relationships The assessment of a certain association or a relationship between one or two variables can be differentiated into two main types—direct and indirect—and a sub-type—covariance. Direct—the most basic of associations. The assertion that a variable X is directly linked to a variable Y (correlational variation) or directly causes Y (causal variation), with no interference from random factors or indirect effects. Indirect—the assertion that X is linked to Y through indirect mechanisms. There are three main indirect relationships: Spurious—the assertion that the link between X and Y is ‘fake’, meaning it exists because another variable (C, for example) is associated with both X and Y. So, the direct association between X and Y occurs (i.e., is created) when C is related to them both. In this sense, C is considered a confounding variable. It is spurious because it is statistically derived, not theoretically. However, it might be difficult to distinguish between this type of relationship, the direct one (see the earlier discussion), through statistical testing alone. This is a situation where we have to postulate spuriousness as a possibility and, thus, question our results. The rationale behind such an association type is that variable C, for example, creates an increase in both X and Y (i.e., positive relationship), thus rendering the seemingly positive association between X and Y, a statistical by-product. Mediated—the assertion that X is linked to Y through another mechanism/ variable (i.e., the mediator).This indicates a specific causal pathway.The relationship between X and Y may be fully or partially mediated, meaning it goes fully/completely through the mediator variable (thus nulling the previously thought-as-direct association between X and Y), or only partially. Commonly depicted as ‘M’ in statistical models.

Using Contemporary Quantitative Techniques  33

An interesting notion in mediational hypotheses is that they assume causal inference because the mediator has to “come after” the predictor but “come before” the outcome. It is a chain of relationships in which the predictor leads to the mediator, which then leads to the outcome, as stated earlier. However, how can a researcher in a correlative study aim to formulate causal links when it is only a ‘snapshot’ of the observable reality in regard to the research? This emphasizes the paramount and indisputable importance of the theoretical background in hypotheses formulation (see also “Theoretical Background as Basis for Hypotheses Formulation” section [3.5]). The Theory of Reasoned Action (Ajzen & Fishbein, 1980) suggested a mediational model in which the basic assumption was that people usually act under rational reasoning.Yoh (2003), for example, used this theory to explain customers behaviors (specifically online clothes purchasing). They indicated there is a causal chain from thoughts and attitudes (i.e., thinking about the purchase) to eventual behavior (i.e., the purchase). This may help us understand that mediational hypotheses may be based on the causal, logical, and theoretically supported chain of events/variables, even in correlative studies, and not experiments or other explanatory designs. Moderated (conditioned)—the assertion that the link between X and Y is conditioned on another variable. A ‘moderator variable’ is a very misleading name for this type of relationship and should be considered as ‘conditional’ (see Appendix A in Shkoler et al., 2017). This type of variable revolves around the conditioning and changing of a link between two other variables such that the original association between them may be either enhanced or diminished/moderated. In the most classical definition, the conditioning (‘moderator’) variable has no relationship with the outcome or the independent variables. Commonly depicted as ‘Z’ in statistical models. Covariance—the basic understanding of a correlation. It consists of two basic words: ‘co’ (= together) and “variance” (= variability), which means if X covariates with Y; the correlation is mutual and not directionally one-sided. The main differences between this type and the rest is that the researcher assumes no theoretical causal link between the variables and thus only indicates that they correlate with each other, not necessarily affecting/causing each other. Figures 2.1–2.5 depict the relationship types in a modular manner. A square/rectangle is an observed variable (a variable to which data was collected in any method), and arrows are association links (so if X  Y, then X associates with Y, and not vice versa).

X FIGURE 2.1 

Model for a Direct Relationship

Y

34  Or Shkoler

X FIGURE 2.2 

Y

or

X

Y

Model for a Covariance/Correlation Type Relationship

X C Y FIGURE 2.3 

Model for a Spurious Relationship

M

X FIGURE 2.4 

Y

Model for a Mediational Relationship

Z X FIGURE 2.5 

Y

Model for a Conditioned (‘Moderated’) Relationship

A General Example Let us assume our research goal is, again, the study of self-esteem and tendency to buy expensive/luxurious goods and/or services: Direct relationship: Self-esteem associates with the tendency to buy expensive/luxurious goods and/or services. Spurious relationship: Self-esteem associates with the tendency to buy expensive/luxurious goods and/or services, but further investigation reveals that the income level positively relates to both self-esteem and the tendency to buy, thus making the original relationship a spurious one.

Using Contemporary Quantitative Techniques  35

Mediational relationship: Self-esteem positively associates with the tendency to buy expensive/luxurious goods and/or services through higher self-­ indulgence (this implies that the self-esteem leads to higher self-­indulgence [as mediational mechanism], which then leads to higher tendency expensive/ luxurious goods and/or services). Conditioned (‘moderated’) relationship (via groups): Self-esteem positively associates with the tendency to buy expensive/luxurious goods and/ or services only for females and not for males (another version is “this association is stronger for females and weaker for males”). Conditioned (‘moderated’) relationship (via another variable): Selfesteem positively associates with the tendency to buy expensive/luxurious goods and/or services, and this association will be increased depending on the degree of social intelligence one has.

Null Hypothesis vs. Alternative/Research Hypothesis When one hypothesizes that a variable X correlates with a variable Y, this is a research hypothesis (statistically signed as H1, also known as an alternative hypothesis), which is what the researcher wants to test. However, because research hypotheses are not made in a vacuum, until proven otherwise, the opposite (or even ‘neutral’) hypothesis is that X does not correlate with Y. This is named the null hypothesis (statistically signed as H0), which usually one wishes to disprove in a research (see Janicak, 2017, p. 140). If H1 is a correlation or a difference or a causal link between two variables, then H0 is the absence of those correlations, differences, or causal links until shown otherwise. In other words, until there is sufficient evidence that we can reject the null hypothesis (i.e., disproving H0), it remains the current hypothesis/result (which defaults to “there is no change in the current relationships between the investigated variables/phenomena”). Research hypotheses are formulated within uncertainly and as such, again, must be backed up by enough theoretical and statistical evidence/data to be tested and supported (or disproved) (Raykov & Marcoulides, 2013, p. 113). In addition, there are some cases in which the researcher deliberately hypothesizes for the H0 and not the other way around (again, it must be backed up by theoretical rationale).

A General Example Let us assume our research goal is, again, the study of self-esteem and the tendency to buy expensive/luxurious goods and/or services: Research question—how can self-esteem be related/influence tendency to buy expensive/luxurious goods and/or services? Another research ­question— what is the relationship between self-esteem and the tendency to buy expensive/luxurious goods and/or services?

36  Or Shkoler

H1: Two-tailed relational hypothesis—self-esteem associates/relates/correlates with the tendency to buy expensive/luxurious goods and/or services. H0: Null hypothesis—self-esteem does not associate/relate/correlate with the tendency to buy expensive/luxurious goods and/or services. H1: Two-tailed causal hypothesis—self-esteem influences/causes the tendency to buy expensive/luxurious goods and/or services. H0: Null hypothesis—self-esteem does not influence/cause the tendency to buy expensive/luxurious goods and/or services. H1: One-tailed relational hypothesis—self-esteem positively associates/relates/ correlates with the tendency to buy expensive/luxurious goods and/or services. H1: Another variation—There is a positive relationship/correlation between self-esteem and the tendency to buy expensive/luxurious goods and/or services. H0: Null hypothesis—There is a no positive relationship/correlation between self-esteem (i.e., does not positively associate/relate/correlate with) and the tendency to buy expensive/luxurious goods and/or services. H1: One-tailed causal hypothesis—self-esteem causes an increase in the tendency to buy expensive/luxurious goods and/or services. H1: Another variation—Self-esteem positively influences the tendency to buy expensive/luxurious goods and/or services. H0: Null hypothesis—Self-esteem does not positively influence/cause the tendency to buy expensive/luxurious goods and/or services.

Theoretical Background as Basis for Hypotheses Formulation Theoretically speaking, most hypotheses are one-tailed, as they are more ‘educated’ because they require sufficient theoretical rationale to support the direction of association’s claim. Nevertheless, it is a very common and unfortunate phenomenon that hypotheses, as one of the most important parts of a research, do not get sufficient and/ or adequate theoretical reasoning to support them. That is to say, good wellestablished hypotheses are such that the reader of the research understands them perfectly and the mechanisms through which the supposed associations depicted in the hypotheses work. It is not so often, but frequent to a degree, to read a published paper/article and notice that the hypotheses were based on similar past hypotheses and not solid theoretical rationale. For example, one would hypothesize X is positively/negatively related to Y because other paper(s) have found similar relationships, not due to educated and careful literature review, and without explaining and justifying the mechanisms of the hypotheses thoroughly. This is very problematic because it demonstrates scientific and/or academic stagnation and dogmatism, as Descartes once very rightfully stated: “In order to seek truth, it

Using Contemporary Quantitative Techniques  37

is necessary once in the course of our life, to doubt, as far as possible, of all things” (Descartes, 1644/2017, p. 13). The role of a researcher is to make a significant donation to the body of knowledge of science (whatever the discipline) through his/her own initiative and capability, and he/she cannot rely on other people’s (sometimes vague) reasonings to hypotheses formulation—even for replications! Therefore, the utmost crucial and vital part of a research is its theoretical background, which is the heart of the research itself, and as such needs to be taken good and delicate care of.

Overview of Methods for Data Collection in Quantitative Research In research, in order to test a research hypothesis, H1 (i.e., the rejection of H0), the researcher must (1) acquire an adequate amount of data, (2) use different data analyses methods (different for quantitative vs. qualitative), and (3) draw conclusions from the data collected. In quantitative research, inferential statistics methods are primarily used in order to test the H1. There are two main sources of data/information: Primary data—firsthand information gathered by the researcher himself/­ herself. The data for collection does not appear anywhere but ‘within’ the sample itself, such as attitudes, experiences, age, and gender. This kind of data is usually raw and needs to be processed and refined. Methods for gathering this kind of data usually revolve around the following: • Questionnaires/surveys (mailed, online, group/collective, individual, panel). • Interviews (structured, semi-structured, unstructured). • Observations (participant, non-participant). • Experiments (experimental, semi-/quasi-experimental). Secondary data—known information, collected by other(s), usually past data, which exists somewhere that is not the firsthand sample, such as historical records, client histories, government publications, websites, books, journal articles, internal record, service records, administrative data, or ethnostatistics. This data is usually already processed and refined. In social sciences, data gathering is used primarily via a sample of participants (primary data). The researcher will need to find an adequately sized sample to draw inferences and conclusions from in order to generalize the findings from it unto the general population.The process of selecting a smaller group (the sample) from a bigger one (the population) as the basis of estimating research phenomena is called sampling. An adequate sample is one that is a good representative of the population it was taken from.This is a very amorphic and ambiguous terminology,

38  Or Shkoler

however, because it is difficult to discern whether a sample is good enough or not. Larger sample sizes are considered more representative of the population. In addition to that, the degree of variability in the population affects the adequacy of the sample, meaning a homogenous population will require a small sample size to represent it (and vice versa), because the variability is low in such a case that a small sample will be enough to encompass it. In any case, there will always be a difference between the population parameters and the sample statistics, or, in other words, between their estimates, because there is no 100% representation and/or accuracy of the sample unless one will pick the entire population itself as the sample. The reason is that it is virtually impossible to sample the whole population, because it is assumed to be infinite, whilst the sample itself must have a finite number. There are several methods for gathering data, which will be further reviewed next: survey methods, observational methods, experimental designs, and metaanalyses.

Cross-Sectional Research The (quantitative) research is often criticized as “too one-timely” or “too unidimensional”, as it usually focuses on a specific time period of a specific sample. It is like a status report or even a ‘snapshot’ of reality in a very focused and specified manner.This type of research is called a cross-sectional design.The words crosssectional derive from the fact we are taking a cross-section of the population in one time only, in order to investigate a phenomenon of some sorts (e.g., experiences, attitudes, behaviors, emotions). The cross-sectional factor is, again, regarded for both the population (or, rather, the sample) and the time the research was based on (i.e., one sample in one time) (Babbie, 1989, p. 89; Kumar, 2014, p. 134), and does not involve manipulating any variable/phenomenon. This design is very easy to formulate and follow through, as it does not require much from the researcher, usually. As such, this type of research is both (1) the predominant kind in social research, as it delivers the fastest results in a fastest and cheapest fashion, and (2) is facing a multitude of criticism in the academic research community. The criticism about this type of research design is understandable. How can we draw objective conclusions from a specific sample in a specific time stamp and generalize them beyond these specific sample and time? This is debatable whether cross-sectional research has more cons than pros, as we, as human beings, more predominantly look at negative aspects than at positive ones (e.g.,Tversky & Kahneman, 1991), but it is its unique advantages over other research designs that rewarded it as the most commonly used method.To illustrate, it is very difficult to create sterile lab conditions for experimental designs and manipulating certain variables, or sample the exact same participants in, at least, two different time periods in order to see if the phenomena under investigation have changed in time. Another

Using Contemporary Quantitative Techniques  39

advantage of this kind of design nests in its great capacity for replicability across studies. In addition, it is worth mentioning that when a cross-sectional study is being conducted in two different or follow-up times with the same respondent sample, it is called a panel study.

Primary Data Collection Methods Each primary data gathering method considers the following: Economical cost—the extent to which the data collection is financially/ economically burdensome (e.g., costs of printing questionnaires, driving to participant’s addresses, ensure lab conditions). Time cost—the extent to which the data collection takes time to follow through. Response rate—the extent to which the respondents, de facto, participate and deliver their raw data to the researcher (e.g., filling a survey). Researcher influence—the extent to which the researcher has an impact/ influence on the respondents. For example, the presence of the researcher is known to increase the response rate due to participants’ sense of obligation to him/her (in personal survey delivery or interview, for example, as opposed to online questionnaires). Privacy and/or anonymity—the extent to which the respondent is assured and feels he/she is being supplied with the utmost privacy and anonymity possible. Respondents’ perceived control—the extent to which the participants feel they have control over how and when to deliver the raw data. Possibility for asking sensitive questions—the extent to which the researcher may ask the respondent sensitive questions (e.g., about one’s sex life, domestic violence, drug abuse, bullying). Possibility for asking complex questions—the extent to which the researcher may ask the respondent complex questions and/or the degree the researcher may explain those questions to the participants when needed. Possibility for re-administering/re-sampling—the extent to which the researcher may ask the same respondents for more cooperation in his/her research (or another one) for the purpose of a replication or other research types.

The Survey One of the most commonly used method in social sciences are the cross-sectional survey types. This is a primary data collection method. A survey (= a questionnaire) is a written list of questions (from general demographics, such as gender, age, and marital status, to specific measures of attitudes and personality trait, such as

40  Or Shkoler

openness to experience, anxiety). In this method, the participants read and follow through the research with minimal intervening by the researcher, as opposed to an interview, for example.The data for the research is granted and recorded by the respondents themselves (i.e., their answers for the questions in the survey). Each participant will answer the same set of questions as the other participants. Because the respondent goes through the questionnaire alone, the survey must be clear and easy enough to understand. The researcher who puts the survey together must always think about the cognitive well-being of the participants— the more cognitive effort they will have to invest into filling out the survey, the more agitated they become and the less reliable are their answers. The length of the survey is one such example; the longer the questionnaire, the less likely we will receive reliable results, if at all.This relationship is nonlinear and very sampledependent (there are people, or even cultures, which are easier agitated than others, for example). In addition, there are several ways for administering a research survey; each depends on the ease, speed, and economical considerations. Individual/personal survey—the researcher personally delivers the physical (pencil-paper) questionnaire to the participant(s), expecting it to be filled in a few minutes to a few days. It is fairly easy to manage, but the cost might be high in terms of time and/or money. Administering a survey may occur in a classroom, in a mall, in a hospital, in a bus station, etc. Collective/group survey—instead of delivering a questionnaire to singular participants, a more efficient way is to collect a group of potential respondents in a room (whatever it may be: A classroom, a waiting room, a lab, etc.) and deliver the survey to all of them at the same time. Students in class, workers in stations, and the like are very good candidates for group survey delivery. Online survey—instead of printing the questionnaire, it is uploaded to an online medium (e.g., Google Forms, Qualtrics). A URL link is produced and then can be sent to multiple potential participants simultaneously by a simple click on the mouse or keyboard, via email, social media/groups, and more. This is very easy to manage and very efficient and ‘green’, because it does not consume paper and ink. However, one must keep in mind that the absence of the researcher in the room, when handing the survey personally, has an impact on the respondents. Again, they would feel more obliged to fill out the questionnaire when the researcher is there with them personally. This does not mean that online surveys are unreliable, but this point should be taken into account when trying to deliver a survey with sensitive questions. Mailed survey—in this case, the survey is, again, printed, but delivered to the respondents via mail (not email, see the earlier discussion), assuming the researcher has access to the relevant addresses. The researcher would not expect the survey to be filled in a manner of days, or even a week or so. This necessitates that the receiver of the questionnaire fill it out and then return it via mail as well. This is usually done if the geographical accessibility to participants is problematic.

Using Contemporary Quantitative Techniques  41 TABLE 2.2  A Comparison of Survey Methods’ Utility

Cost (economical) Cost (time) Response rate Researcher influence Privacy and/or anonymity Respondents’ perceived control Possibility for sensitive questions Possibility for complex questions Possibility for re-administering

Mailed

Personal

Group

Online

Low Slow Very low Very low Very high

Moderate low Moderate Moderate Moderate Moderate

Moderate low Fast Moderate high Moderate Moderate

Low Fast Moderate low Very low Very high

Very high

High

Moderate

High

Very high

Moderate

Moderate

High

Very low

Moderate

Moderate

Very low

Moderate low

Very unlikely

Very unlikely Moderate

Each survey method has an advantage and disadvantage in several terms. Table 2.2 displays a comparison of the survey methods across those parameters. The choice between these methods is heavily based on the researcher needs and the characteristics of the research itself.

How to Quantify—Measurement by Measures/Scales/Constructs A survey or a research questionnaire is the aggregation of different measures/ scales (also called ‘inventories’) for each variable instigated in the research itself. The purpose of these instruments is to gauge and quantify intangible/ abstract phenomena, such as attitudes, emotions, cognitive processes, experiences, and so on. A measure is a written research instrument for data collection, which may consist of a few to many questions: Demographics, research ­constructs/variables, etc. Each question is regarded as an item, and a set of items is regarded as a measure or a research construct (e.g., 20 items for measuring self-esteem, 33 items for measuring emotional intelligence, 8 items for measuring anxiety). Before each set of items, there is a given ‘protocol’; instructions for the reader on how to answer and complete these items. This is given, sometimes differently to each measure, through the entire research questionnaire, at least one protocol for each different measure. This is called operationalization (i.e., an operational definition), which sums up to “how is the researcher going to measure/ quantify a certain phenomenon” or “what instrument is used to measure a certain phenomenon”. For example, weight is operationalized by grams or kilograms via a weight, height by meters, centimeters or inches via ruler, blood sugar level by the amount

42  Or Shkoler

of glucose present in the blood via a blood test device, etc. The challenges of the research in social sciences are—the measurements of attitudes, emotions cognitions, and behaviors, because these phenomena are intangible. The best solution the quantitative approach has come up with is the operationalization of said factors through well-established, reliable, and valid scales/questionnaires, as explained before. Most commonly, the respondents are asked to rate (as opposed to rank) their answers on a Likert-type scale (see Carifio & Perla, 2007; Derrick & White, 2017; Likert, 1932), in a specific instrument, which is usually, for comfort’s use, given in a grid/matrix. In this case, the rows represent different items of a certain measure, while the columns represent the range of the responses (e.g., from 1 to 5, from −2 to 3, from 1 to 6, from 1 to 10). Each numeric response is accompanied by a short “explanation” using very few words. For an example, see the following: Please circle the answer that best describes your feelings about your local grocery store:

Item No. 1 Item No. 2 . . . . . .

1 Highly disagree

2 Somewhat disagree

3 Neutral

4 Somewhat agree

5 Highly agree

1 1 1 1

2 2 2 2

3 3 3 3

4 4 4 4

5 5 5 5

There are many more scale types, but the most commonly used in measuring attitudes, behaviors, and feelings is the Likert-type (DiIorio, 2005, p. 28). Of course, answers to questions such as—gender, age, tenure, marital status, and so on, are not Likert-type, but case specific, meaning the answer for “age”—is the respondent’s age (unless given age groups as possible answers), gender: 1 = Female, 2 = Male, etc. Moreover, the ends on the (Likert) scale are opposites and may vary from “disagreement” to “agreement”, from “does not describe me at all” to “describes me very well”, from “unsatisfied at all” to “very satisfied”, from “rarely” to “very often”, and so on, depending on the construct it gauges and the need of the research. In addition, these questions can be phrased in two fashions: Open ended—possible responses/answers are not provided after the question. The respondent writes down an answer he/she formulates in his/her mind. In this case, answers usually vary from one participant to the other, unless the answer is obvious and unambiguous. For example, “What is your age (in years)? ______”, or “What are the advantages of the new recruit

Using Contemporary Quantitative Techniques  43

program? ______”.This type of questions is usually problematic in terms of data coding and processing, because answers may be difficult to quantify or code. For example, if we ask, “Please write the advantages of purchasing a gift-card: ______”. The answers to that question may vary to a great extent between each respondent, making coding and processing of the information significantly more burdensome and lengthy, though not impossible. Closed—the answers for the questions are given and are plainly written after the question. Most surveys operate on closed questions as they are usually easier to understand and answer and facilitate data coding and processing. For example, “What is your education level? (a) high-school, (b) B.A., (c) M.A.”, or “Are you working in a managerial position? (1) = No, (2) Yes”. Likert-type questions (see the earlier discussion) are also closed types, because the answer is written and given to the respondents and they need but circle/choose it. If so, how is a set of items a good indication/measurement for one’s personal feelings or cognitive processes? The quantitative paradigm stresses this is the only way to quantify intangible/abstract phenomena, while the qualitative paradigm will stress it is also a mathematical representation of said phenomena and must be understood more in-depth via narrative interviews and other qualitative methods. In short, the research survey/questionnaire is a composition of questions (or sets of items). These items are indicators for the research’s variables/constructs, each set of items represents a distinct variable. The responses for each item/­question are the raw primary data the researcher aims to collect. This method is the most commonly used in social sciences.

The Observation In the survey method, the questionnaire is the instrument with which a researcher collects the raw data for the research. However, there are methods in which the researcher himself/herself is the instrument for the collection of the data. The observational methods make use of our five senses, and with our eyes and ears, specifically in social research, when we observe a phenomenon, we watch and listen to it in real time. This method is, like the other methods, based on the needs of the research—for example, group interactions, customer behaviors, employees’ working patters, etc. Furthermore, when potential participants are not cooperative or non-responsive to filling out a survey, or even when they are so preoccupied in some interaction (e.g., working, making a purchase), they cannot detach and pay attention to the researcher, the observational method seems better in gauging the phenomenon under examination. However, this method is heavily bent toward the measurement of behaviors more than anything. The recording of an observation may vary from researcher’s notes, to video footage, audio log files, and more. These recordings are the raw data derived from

44  Or Shkoler

observational methods, as opposed to the survey methods in which the participants themselves write the data.There are two methods of data collection through observation: Participant observation—when the researcher participates in the activities/ interactions of the sample he/she wishes to observe (with or without their knowledge they are actually being observed). The researcher gets involved (active observation) to a certain or great extent, depending on the research needs. Non-participant observation—when the researcher does not participate in the activities/interactions of the sample he/she wishes to observe and becomes a passive observer. In addition to the observational methods, the situations in observing a phenomenon may be as follows: Natural—without interfering with the natural normative activities of the observed participants/sample. Controlled (artificial)—introducing a stimulus of some kind to make a reaction in the observed group in order to observe their actions.

A General Example If we want to research the behavior of customers, we may choose: To ‘disguise’ ourselves as customers and mingle with the many customers where we choose to observe. We will interact with them as well and take notes of their behaviors and reactions. That is the active/participative observation. To keep watch over customers’ behavior from a distance, or even via security cameras, and write notes of their behavior and reactions in this manner. That is the passive/non-participative observation. In addition, we may choose to observe without any manipulation on our part (= natural situation), or deliberately, for example, entice some customers as to the high prices of the products they buy (= controlled/artificial situation). Table 2.3 displays a comparison of the observational methods across several parameters. The choice between these methods is heavily based on the researcher needs and the characteristics of the research itself.

The Experiment There is a plethora of experimental research designs, which will not be discussed in this chapter (for further reading, see Kumar, 2014, pp. 135–150). In an

Using Contemporary Quantitative Techniques  45 TABLE 2.3  A Comparison of Observational Methods’ Utility

Cost (economical) Cost (time) Response rate1 Researcher influence Privacy and/or anonymity Respondents’ perceived control1 Possibility for sensitive questions Possibility for complex questions Possibility for re-administering

Participant natural

Non-participant natural

Participant controlled

Non-participant controlled

Moderate Slow 100% Very high

Moderately low Slow 100% Very low

Moderate Slow 100% Very high

Moderately low Slow 100% Moderately low

Very low

Moderately low Very low

Moderately low

Very high

Very high

Very high

Very high

Very high

Very low

Very high

Very low

Very high

Very low

Very high

Very low

Very unlikely Very unlikely

Very unlikely Very unlikely

Notes: (1) the respondents are not aware that they are being observed and the data from them collected.

experiment, the researcher wishes to draw a causal conclusion and to that end will make an intervention or manipulation in the study.The manipulation is done on the predictor(s) (independent variable[s]) only (!), which is believed to be the ‘cause’ of the dependent variable (the outcome). A manipulation cannot be done on variables, such as gender and age, as they cannot be changed or manipulated through conventional orthodox means. A manipulated variable is different from a measured variable, because the researcher is the one who establishes the variability or dictates the degree of the variable and hence the name manipulation. In any event, the outcome variable is always measured across all the participants and groups in the experiments, as it is usually the goal of the research and must not be tampered with. We want to see the outcome change, but not by our intervention. Because the manipulation is artificial, there is usually a necessity for dividing the research into groups: (1) Experiment/research groups and (2) control groups. The need is because we are artificially making changes in the variables, and thus we need to observe a group in which the manipulation is null/non-existent, for comparison’s and validity’s sake. Participants are randomly selected to be included in either group. Experiment/research groups—the participants in these groups must have a certain degree of the manipulated variable. Only the participants in the experiment group will have to go through a manipulation of some kind of the independent factor in the research.

46  Or Shkoler

Control groups—the participants in these groups usually have a degree of zero of the manipulated variable. For example, if we wish to induce frustration, we could make participants go through a very agonizing, long, and difficult exercise. Those who will go through this exercise are considered as the experiment group. And those who do not are the control group. There is also a very common case of experiment groups is that each of them is a control group of the others. In this sense, the manipulation is done to all of the groups, but in varying magnitudes. For example, there are three different degrees of frustration—a group which will go through a very frustrating test, a group which will have to fill a moderately frustrating test, and a group which will enjoy a very mild frustrating test. A common case is biological and chemical studies, in which bacteria might, for example’s sake, be exposed to a degenerative compound in varying doses (e.g., low exposure, medium, and high). It is very important to note, however, that the participants in either experiment or control group, must have similar characteristics as possible. Otherwise, they become biased and incomparable, and that might affect the results of the research. For example, if we wish to investigate the effect of solving mathematical exercises on the degree of success in a math test (i.e., solving exercises is the independent variable and success in the test is the outcome), we would take one group that will have to solve a lot of exercises (this is the manipulation), as opposed to the other one which will not solve anything at all (the control group). However, if, for some reason, there will be a vast majority of “math geniuses” in the control group, then the experiment group would not display a significant advantage over the control group, nulling the research results. If so, how an experiment allows for causal inference? It is almost solely based on the timing. In a survey, the cross-sectional method of it implies that all the variables of the research are measured at the same time (the ‘snapshot’), regardless of time precedence between them. In an experiment, the researcher goes to great lengths in order to ensure the manipulation of the predictor will be the cause of the outcome (usually, the manipulation precedes the outcome in the timing of the research). Figure 2.6 illustrates a simple experimental design (the after-only experimental design, see Kumar, 2014, pp. 143–144). As was aforementioned, this chronological precedence is of paramount importance when trying to draw causal inferences. Furthermore, let us assess and conclude the causal adequacy of experiments (see also “Main Types of Hypotheses” section [3.3]): Chronological/time order of precedence—ensured when the manipulation precedes the outcome. Significant statistical and theoretical relationship between the variables—is the basis of all research, correlational, or causal alike, and is gauged via educated and careful theoretical reasoning and statistical analyses.

Using Contemporary Quantitative Techniques  47

Randomized

Group A (Experiment)

Manipulation

sampling

observation) of

(Randomized selection of participants)

Time 1: pre-manipulation FIGURE 2.6 

Measurement (or

Group B (Control)

NO manipulation

Time 2: the manipulation

the (same) outcome variable

Time 3: post-manipulation

An After-Only Experimental Design Illustration

Absence of alternative explanations for the relationship—the presence of the control group is key. The reason is because it allows for a comparison between the outcomes. If the experiment will be based on the experiment group alone, how would we know that the manipulation ultimately worked as intended and not some random effects occurred instead? Utilizing a control group allows us to compare a group with treatment (i.e., manipulation) and without, thus increasing the likelihood that if the outcome variable of the experiment group differs significantly from one in the control group, we may conclude (to a certain degree) that the manipulation was the cause of said outcome and not a random, unforeseen effect. The experiments too receive criticism, with special emphasis on the small sample size used in experiments and, more importantly, the lab conditions imposed on the subjects that do not usually reflect the natural situation the participants find themselves in each day. In addition, as opposed to the survey and observation sections that were mentioned before, the experiment is a vast and wide subject, and thus a table that displays a comparison of the experimental methods across several parameters will not be concluded in this chapter.

Secondary Data Collection Methods As mentioned, secondary analyses are the reworking of known or collected data. The information is not primarily gathered by the researcher, but by others, and the secondary manipulation and processing of the data is done by the researcher. One of the most prominent quantitative secondary analyses is the meta-analysis that will be elaborated in the next section.

48  Or Shkoler

Furthermore, it is important to light a few issues with secondary analyses: Availability of and/or accessibility to the data—not all data are made available for the researcher (e.g., if the information is confidential or governmental). Validity and reliability of the data—because the researcher did not collect the data firsthand, the reliability and validity of secondary data may be questionable (or not). Personal bias of the data—'persona’ information (e.g., diaries, newspapers) may be biased due to the author’s personal influence. Personal bias of the researcher—the choice of what secondary data and which criteria to use to include in the analysis is the researcher’s and thus might be biased. Format of the data—the extent to which the data matches the needs of the researcher and/or the research (e.g., if the researcher needs tenure groups between 0–5, 6–8 years, and the data only exists in 0–4, 5–10 tenure groups).

The Meta-Analysis A meta-analysis is the synthesis of results from numerous past researches and experiments, from a few months so several years old. This combination is based on few-to-many quantitative studies that used similar methodologies (to a certain degree). This notion is based on the ‘common truth’ behind each research that connects them together logically. The meta-analysis aims to connect them—­ statistically. This approach aims to make pooled (aggregated) estimates that will represent this ‘common truth’ in the most accurate manner possible, which is also based on errors in the derivation of these estimates (see: Gomm, 2008, pp. 349–350; Greenland, & O’Rourke, 2008). This aggregation results in more robust statistical estimates and higher statistical power than may, otherwise, be concluded by a single study. However, the conducting of a meta-analysis is easily influenced by the researcher. One must make choices that might affect the end results, such as (a) deciding how to search for studies, (b) inclusion criteria of the studies—selecting studies based on a set of objective criteria, (c) facing and dealing with incomplete data, (d) the data analyses themselves, and (e) accounting (or not) for publication bias (Gomm, 2008, p. 356; Walker, Hernandez, & Kattan, 2008). A meta-analysis is a whole theory in its own right and will not be discussed in this chapter. However, to illustrate the advantages and main goals of this analysis, Walker et al. (2008) have summed that the main objectives of a meta-analysis are to: Summarize and integrate results from a number of individual studies. Analyze differences in the results among studies. Overcome small sample sizes of individual studies to detect

Using Contemporary Quantitative Techniques  49

effects of interest, and analyze end points that require larger sample sizes. Increase precision in estimating effects. Evaluate effects in subsets . . . Determine if new studies are needed to further investigate an issue. Generate new hypotheses for future studies. These lofty objectives can only be achieved when the meta-analysis satisfactorily addresses certain critical issues. Drawing upon Walker et al. (2008), we may conclude (1) the meta-analysis’ results can be generalized to a larger population (aggregation of samples will represent the population more accurately); (2) the precision and accuracy of estimates can be, thus, improved (more data may increase the statistical power); (3) inconsistency of results across studies can be quantified and analyzed; (4) hypotheses testing may be applied on pooled estimates; and (5) publication bias may be investigated.

Comprehension Questions 1) Consumer research is based on socio-psychological research only: A. True. B. False, it gained its knowledge and applications from various disciplines. C. False, it formed from the socio-economics discipline. D. Consumer behavior is not a science and therefore cannot be researched. 2) Is there a meaningful difference between a ‘customer’ and a ‘consumer’? A. Yes, a consumer is someone who buys from a specific store/ company; a customer is someone who is a potential future consumer. B. Yes, a costumer is someone who buys from a specific store/ company; a consumer is customer after he/she uses bought goods and/or services. C. Yes, a costumer is someone who buys from a specific store/ company; a consumer is someone who is a potential future customer. D. No, there is no meaningful difference between a ‘customer’ and a ‘consumer’.

50  Or Shkoler

3) Research does NOT involve: A. Making assumptions. B. Data collection. C. Sampling. D. Literature review. 4) Consumer research is usually based on: A. B. C. D.

Empirical research. Applied research. Pure research. Answers A+B.

5) A researcher aims to find out what are the relationships between several variables and as such will use the _____________ research type to this end: A. Explanatory. B. Correlational. C. Exploratory. D. Descriptive. 6) A researcher aims to find out what causes the relationships between several variables and as such will use the _____________ research type to this end: A. Explanatory. B. Correlational. C. Exploratory. D. Descriptive. 7) A researcher aims to find out what are the relationships between several variables in a niche field of study which was not investigated thoroughly and as such will use the _____________ research type to this end: A. Explanatory. B. Correlational. C. Exploratory. D. Descriptive.

Using Contemporary Quantitative Techniques  51

  8) A researcher aims to find out what are the underlying experiences costumers have in online transactions and as such will employ the _____________ research paradigm: A. Qualitative. B. Quantitative.   9) The sentence “social intelligence affects the tendency to make an online purchase” is a: A. Research question. B. Research hypothesis. 10) The sentence “how does social intelligence affect the ­tendency to make an online purchase?” is a: A. Research question. B. Research hypothesis. 11) What, among the following, is NOT a criterion for a well-­ established research hypothesis: A. B. C. D.

Written in parsimonious manner. Data driven. Cannot be dismissed/disproved. Based on natural orders.

12) The hypothesis “higher social intelligence associates with higher tendency to make an online purchase” is: A. B. C. D. E. F.

One-tailed, correlative. One-tailed, differentiative. One-tailed, causal. Two-tailed, correlative. Two-tailed, differentiative. Two-tailed, causal.

13) The hypothesis “social intelligence influences the tendency to make an online purchase” is: A. One-tailed, correlative. B. One-tailed, differentiative.

52  Or Shkoler

C. D. E. F.

One-tailed, causal. Two-tailed, correlative. Two-tailed, differentiative. Two-tailed, causal.

14) The hypothesis “elders have higher social intelligence than youngsters” is: A. B. C. D. E. F.

One-tailed, correlative. One-tailed, differentiative. One-tailed, causal. Two-tailed, correlative. Two-tailed, differentiative. Two-tailed, causal.

15) In the hypothesis “higher social intelligence associates with higher tendency to make an online purchase”, social intelligence is: A. B. C. D. E. F. G. H.

A predictor (independent). An outcome (dependent). A constant. A variable. Answers A+C. Answers A+D. Answers B+C. Answers B+D.

16) In the hypothesis “women before the age of 35 have a higher tendency for online purchasing of goods than women after the age of 35”, gender is: A. B. C. D. E. F. G. H.

A predictor (independent). An outcome (dependent). A constant. A variable. Answers A+C. Answers A+D. Answers B+C. Answers B+D.

Using Contemporary Quantitative Techniques  53

17) In the hypothesis “women before the age of 35 have a higher tendency for online purchasing of goods than women after the age of 35”, purchasing tendency is: A. B. C. D. E. F. G. H.

A predictor (independent). An outcome (dependent). A constant. A variable. Answers A+C. Answers A+D. Answers B+C. Answers B+D.

18) Review the sentence “as the older a person gets, the higher one’s social intelligence, and hence a greater tendency to purchase online goods”. This is an indication of a: A. B. C. D.

Direct relationship. Spurious relationship. Mediational relationship. Conditioned (“moderated”) relationship.

19) Review the sentence “as the older a person gets, the greater tendency to purchase online goods, and this is stronger for married people”. This is an indication of a: A. B. C. D.

Direct relationship. Spurious relationship. Mediational relationship. Conditioned (“moderated”) relationship.

20) Review the sentence “as the older a person gets, the greater tendency to purchase online”. This is an indication of a: A. B. C. D.

Direct relationship. Spurious relationship. Mediational relationship. Conditioned (‘moderated’) relationship.

54  Or Shkoler

21) The hypothesis “women make more online purchases than men” is: A. A null hypothesis. B. An alternative hypothesis. 22) What is the H0 if we wish to correlate life satisfaction with increased expenses on food: A. Life satisfaction increases the expenses one would make on food. B. Life satisfaction does not associate with the expenses one would make on food. C. Life satisfaction positively correlates with the expenses one would make on food. D. Answers A+C. 23) What is the two-tailed H1 if we wish to correlate gender with types food delivery: A. Women tends to order more pasta and a sushi from delivery food purchases than men. B. Women do not differ from men in food delivery types they order. C. Women differ from men in food delivery types they order. D. Answers A+C. 24) What is the one-tailed H1 if we wish to correlate gender with types food delivery: A. Women tends to order more pasta and a sushi from delivery food purchases than men. B. Women do not differ from men in food delivery types they order. C. Women differ from men in food delivery types they order. D. Answers A+C. 25) When a researcher goes to an organization and gathers data on employees from their managers, this is an example of ___________: A. Primary data collection. B. Secondary data collection.

Using Contemporary Quantitative Techniques  55

26) When a researcher goes to an organization and gathers data on employees by observing their work routines, this is an example of ___________: A. Primary data collection. B. Secondary data collection. 27) When a researcher goes to an organization and gathers data on employees by asking them questions, this is an example of ___________: A. Primary data collection. B. Secondary data collection. 28) When a researcher goes to an organization and gathers data from 30 employees from the marketing department, this is an example of ___________: A. Sampling. B. Data analysis. C. Data collection. D. Generalization of results. 29) The sentence “in order for the sample to be a perfect representative of the population, it must be as large as possible” is: A. True. B. False. 30) The sentence “sampling is the notion of taking the sample and try to generalize it to represent the entire population as much as possible” is: A. True. B. False. 31) When a researcher goes to an organization and gathers data on employees by asking them questions, this is an example of ___________: A. Survey method for data collection. B. Observation method for data collection. C. Experiment method for data collection.

56  Or Shkoler

D. Cross-sectional study design. E. Meta-analysis F. Answers D+E. G. Answers A+D. H. Answers A+E. 32) When a researcher goes to an organization and gathers data on employees by reviewing multiple past research in his/her field of study, this is a probable example of ___________: A. Survey method for data collection. B. Observation method for data collection. C. Experiment method for data collection. D. Cross-sectional study design. E. Meta-analysis F. Answers D+E. G. Answers A+D. H. Answers A+E.

Answers to Comprehension Questions Q1

Q2

Q3

Q4

Q5

Q6

Q7

Q8

Q9

Q10

B

C

A

D

B

A

C

A

B

A

Q11

Q12

Q13

Q14

Q15

Q16

Q17

Q18

Q19

Q20

C

A

F

B

F

C

H

C

D

A

Q21

Q22

Q23

Q24

Q25

Q26

Q27

Q28

Q29

Q30

B

B

C

A

B

A

A

B

B

B

Q31

Q32

G

E

Using Contemporary Quantitative Techniques  57

Notes 1. Statistical significance (α level) will not be discussed in this chapter. 2. Variable = anything that can vary, meaning it can take on different values—for example, age, gender, IQ, degree of satisfaction, etc. If something cannot vary, it is considered a constant. Constants, for the most part, do not concern the researcher because they have no relevant information to add to the research itself.

References Ajzen, I., & Fishbein, M. (1980). Understanding attitudes and predicting social behavior. ­Englewood Cliffs, NJ: Prentice Hall. Appelbaum, M., Cooper, H., Kline, R. B., Mayo-Wilson, E., Nezu, A. M., & Rao, S. M. (2018). Journal article reporting standards for quantitative research in psychology: The APA publications and communications board task force report. American Psychologist, 73, 3–25. Babbie, E. (1989). Survey design methods (2nd ed.). Belmont, CA: Wadsworth. Bailey, K. D. (1978). Methods of social research (3rd ed.). New York, NY: Free Press. Black, J. A., & Champion, D. J. (1976). Methods and issues in social research. New York, NY: Wiley. Carifio, J., & Perla, R. J. (2007). Ten common misunderstandings, misconceptions, persistent myths and urban legends about Likert scales and Likert response formats and their antidotes. Journal of Social Sciences, 3, 106–116. Derrick, B., & White, P. (2017). Comparing two samples from an individual Likert question. International Journal of Mathematics and Statistics, 18, 1–13. Descartes, R. (2017). Principles of philosophy. (A. MacNoravaich, trans.).Whithorn, Dumfries and Galloway, Scotland: Anodos Books. (Original work published 1644). DiIorio, C. K., (2005). Measurement in health behavior: Methods for research and evaluation. New York, NY: Jossey-Bass Publishers. Gomm, R. (2008). Social research methodology: A critical introduction (2nd ed.). New York, NY: Palgrave Macmillan. Greenland, S., & O’Rourke, K. (2008). Meta-analysis. In K. J. Rothman, S. Greenland, & T. Lash (Eds.), Modern epidemiology (3rd ed., pp. 653–682). Philadelphia, PA: Lippincott Williams  & Wilkins. Grinnell, R. Jr. (Ed.). (1988). Social work research and evaluation (4th ed.). Itasca, IL: F. E. Peacock. Janicak, C. A. (2017). Applied statistics in occupational safety and health. London: Rowman & Littlefield Publishers, Inc. Kerlinger, F. N. (1986). Foundations of behavioral research. New York, NY: Holt, Rinehart and Winston. Kumar, R. (2014). Research methodology: A step-by-step guide for beginners (4th ed.). London: Sage Publications Ltd. Levitt, H. M., Bamberg, M., Creswell, J. W., Frost, D. M., Josselson, R., & Suárez-Orozco, C. (2018). Journal article reporting standards for qualitative primary, qualitative meta-­ analytic, and mixed methods research in psychology: The APA publications and communications board task force report. American Psychologist, 73, 26–46. Likert, R. (1932). A technique for the measurement of attitudes. Archives of Psychology, 140, 1–55.

58  Or Shkoler

Loudon, D. L., & Della Bitta, A. J. (1993). Consumer behavior: Concepts and applications (4th ed.). New York, NY: McGraw-Hill. Marczyk, G., DeMatteo, D., & Festinger, D. (2005). Essentials of research design and ­methodology. In A. S. Kaufman & N. L. Kaufman (Eds.), Essentials of behavioral science. Hoboken, NJ: John Wiley & Sons, Inc. Popper, K. (1963). Conjectures and refutations. London: Routledge & Kegan Paul. Raykov, T., & Marcoulides, G. A. (2013). Basic statistics: An introduction with R. Plymouth: Rowman & Littlefield Publishers, Inc. Reeve, J. (2016). A grand theory of motivation:Why not? Motivation and Emotion, 40, 31–35. Shkoler, O., Rabenu, E., & Tziner, A. (2017). The dimensionality of workaholism and its relations with internal and external factors. Journal of Work and Organizational Psychology (Revista de Psicología del Trabajo y de las Organizaciones), 33, 193–203. Tversky,A., & Kahneman, D. (1991). Loss aversion in riskless choice:A reference-­dependent model. The Quarterly Journal of Economics, 106, 1039–1061. Walker, E., Hernandez, A.V., & Kattan, M.W. (2008). Meta-analysis: Its strengths and limitations. Cleveland Clinic Journal of Medicine, 75, 431–439. Warner, R. M. (2013). Applied statistics: From bivariate through multivariate techniques (2nd ed.). Los Angeles, CA: Sage Publications Ltd. Webster’s third new international dictionary. (1976). Springfield, MA: G. & C. Company. Weinstein, J. A. (2010). Applying social statistics: An introduction to quantitative reasoning in ­Sociology. Plymouth: Rowman & Littlefield Publishers, Inc. Yoh, E., Damhorst, M. L., Sapp, S., & Laczniak, R. (2003). Consumer adoption of the Internet: The case of apparel shopping. Psychology & Marketing, 20, 1095–1118.

Further Reading Braun,V., & Clarke,V. (2006). Using thematic analysis in psychology. Qualitative Research in Psychology, 3, 77–101. de Leeuw, E. D., & Collins, M. (1997). Data collection methods and survey quality: An overview. In L. E. Lyberg, P. P. Biemer, M. Collins, & E. D. de Leeuw (Eds.), Survey measurement and process quality (pp. 199–219). New York, NY: Wiley. Denscombe, M. (2010). The good research guide: For small-scale social research (4th ed.). ­Berkshire: McGraw-Hill. Engler, B. (2009). Personality theories: An introduction. Boston, MA: Houghton Mifflin ­Harcourt Publishing Company. Fowler, F. J. Jr. (2014). Survey research methods (5th ed.). Thousand Oaks, CA: Sage Publications Ltd. Glaser, B. G., & Strauss, A. L. (1967). The discovery of grounded theory: Strategies for qualitative research. London: AldineTransaction, A Division of Transaction Publishers. Punch, K. (1998). Introduction to social research: Quantitative and qualitative approaches. London: Sage Publications Ltd.

3  EASUREMENT THEORY AND M PSYCHOLOGICAL SCALING Daniel P. Hinton and Tracey Platt

Quantitative research in every branch of psychology involves the measurement of psychological constructs, and consumer psychology is no exception. The use of tools to measure psychological constructs is known as psychometrics. This chapter will outline the use of psychometric measures within consumer psychology and related fields—both in academic and practice settings—and discuss the theory underlying psychological measurement, before exploring the process by which psychologists develop these measures.

Why Use Psychometrics in Consumer Psychology? Over the past 50 years and beyond, the field of consumer psychology has seen a widespread adoption of psychometric tools for the measurement of psychological constructs of interest. In practices such as marketing, in the past, these variables tended to be measured poorly (Churchill, 1979). However, companies are becoming increasingly aware that the need to properly understand consumer behaviour necessitates the use of robust measurement tools in order to understand the psychological processes that drive this behaviour. Psychometrics have been developed and deployed to understand consumer preferences and attitudes (e.g. Kidwell, Hardesty, & Childers, 2007; Gattol, Sääksjärvi, & Carbon, 2011), how brand loyalty is created and maintained through the development of affection for a brand (Albert & Valette-Florence, 2010), how consumers react to advertising through new forms of media (Bauer Reichardt, Barnes, & Neumann, 2005), the process by which consumers make purchasing decisions (Schwartz et al., 2002), and even more negative consumer behaviours such as compulsive purchasing (Maraz et al., 2015). Overall, psychometric measures equip the consumer psychologist with a set of tools with which they can better understand customers and their behaviour.

60  Daniel P. Hinton and Tracey Platt

Learning Objectives By the end of this chapter, you should be able to • • • • •

understand the key differences between psychometric and non-psychometric measures understand the theoretical underpinnings of psychological measurement and compare and contrast Classical Test Theory with Item Response Theory understand how psychometrics are typically used within the field of consumer psychology understand the process of scale design and scale development, from initial conception through to finalising a scale’s item set critically evaluate modern alternatives to the psychometric approach to scale development, such as Rossiter’s C-OAR-SE method

Psychometrics and Their Key Characteristics At the broadest level of abstraction, a psychometric tool is one that is designed to measure some kind of psychological construct. The constructs measured by psychometric instruments are many and varied, encompassing personality traits, mental ability, and a host of other individual differences such as attitudes, values, and beliefs. Psychometrics typically consist of a number of questions or statements (known as items) to which the test taker must respond from a finite number of response options. Psychometric tools may consist only of a single scale that has been designed to measure a single construct (in which case they are referred to as unifactorial measures). Equally, they may be made up of a number of conceptually related scales, all of which are designed to measure distinct constructs, such as personality questionnaires. To the casual observer, it may seem as though there is no real difference between a psychometric tool and any other form of questionnaire or survey. However, there are a number of key differences between psychometric and non-psychometric tools. Firstly, published psychometrics should have all been through a rigorous process of design and development, similar to the one described later on in this chapter. Secondly, and related to the previous point, psychometric measures should all demonstrate good psychometric properties. When we talk about psychometric properties, we are referring to a tool having an evidence base to support its reliability and validity. A measure’s reliability refers to the consistency or accuracy with which it measures the construct of interest. A measure shows validity if it measures what it is supposed to. That is to say, it measures the construct that it was designed to. In psychological research, there are a great many forms of reliability and validity, only some of which are relevant for the scale development process. A brief summary of the most relevant types of reliability and validity for the design and development of psychometric measures is shown in Box 3.1.

Measurement Theory and Psychological Scaling  61

BOX 3.1  TYPES OF RELIABILITY AND VALIDITY RELEVANT FOR SCALE DEVELOPMENT Internal Consistency: A measure of the degree to which all items within a scale are consistent in their measurement. For example, in a scale that demonstrates good internal consistency, a test taker who scores highly on one item would be expected to score high on the other items, as they are all measuring the same construct. The method by which internal consistency is assessed depends upon the response format of the scale. Scales with two or more response options (e.g. Likert-type scales) are most frequently assessed using Cronbach’s α (Alpha). Scales with dichotomous response formats (e.g. ‘true/false’, ‘yes/no’, correct/incorrect) tend to be assessed using Kuder-Richardson Formula-20 (KR-20). Content Validity: A measure of the degree to which the items within a scale appear to measure the construct of interest, as judged by subject matter experts. Traditionally a qualitative assessment, there has been a relatively recent move towards more robust, quantitative assessments of content validity. Construct Validity: An overall assessment of whether the items within a scale measure the construct of interest. Construct validity takes a number of forms, all of which are assessed quantitatively using some form of factor analysis procedure such as PAF or CFA. Structural validity is an examination of whether the underlying latent factor structure of the tool is consistent with that which is theoretically expected. In single-construct measures, this tends to equate to the degree to which the items within the scale are unifactorial. Convergent validity represents the degree to which the items within a scale measure the same construct as items from another scale that has been designed to measure the same (or a very similar) construct. Conversely, discriminant (or divergent) validity is the degree to which a scale’s items are unrelated to items from other scales that have been designed to measure different constructs. Criterion-Related Validity: A measure of the degree to which scores on a scale are related to important outcomes. There are three kinds of criterion-related validity, all of which are assessed by computing correlation coefficients between scale scores and outcome variables. Predictive validity represents the degree to which a scale predicts an outcome in the future. Concurrent validity represents the relationship between scale scores and an outcome in the present. Postdictive validity refers to the relationship between scale scores and an outcome in the past. Some outcome variables in consumer psychology that are frequently used as criteria for these purposes are discussed later in this chapter in Step 6.

62  Daniel P. Hinton and Tracey Platt

Finally, psychometric measures are standardised in a number of different ways. Firstly, in the case of normative measures, a test taker’s score on a scale is converted to a standardised score, such as a percentile or a sten.Whereas the test taker’s scale score (known as their raw score) does not tell the test user much about his or her level of a trait or resultant likely behaviour, standardised scores allow for comparison with a norm group of test takers who have previously completed the tool. This allows a test taker’s score to be placed within the distribution of test takers and inferences to be made about his or her level of a trait in comparison to others. The other key aspect of standardisation of psychometric measures aims to make them more objective and less prone to random and systematic error variance. The instructions provided to test takers is always the same, and responses to items are scored in a strict and highly procedural way. Indeed, this principle of minimising error variance is a central one in the practice of psychometrics. To understand this further, however, it is necessary to understand something about their theoretical foundations.

Measurement Theory Psychometric practice is heavily influenced by theory. The earliest attempt to formalise measurement theory was in the form of Classical Test Theory. First proposed by Novick (1966), Classical Test Theory revolves around the principle that psychological constructs are latent (that is to say that they are ‘hidden’ and are not directly observable). As a result, they cannot be measured in the same way as one might a person’s height or weight. Rather, the way that we measure these hidden constructs is by measuring a person against indicators of that construct. Classical Test Theory makes the assumption that a test taker has a ‘true’ level of a trait (the ‘true score’), but that measurement of this true score is necessarily obscured by random error in the process of measurement against these indicators. Therefore, the equation used to describe the relationship between a test taker’s true score and his or her observed score according to Classical Test Theory is described by the following equation: X =T + E where X represents a test taker’s observed scale score, T represents their true (i.e. ‘hidden’) score, and E represents measurement error. Therefore, the key principle of Classical Test Theory is that, in order to be able to measure psychological constructs with any degree of accuracy, we must use a range of indicators that adequately samples behaviour relevant to the construct of interest, while at the same time minimising measurement error. A relatively modern alternative to classical test theory is Item Response Theory (IRT). Whereas the psychometric focus of Classical Test Theory is based upon the properties of the test as a whole, the principles of IRT rely on the use of individual items, or a subset of items from within the test. One of the principle

Measurement Theory and Psychological Scaling  63

tenets of IRT is that the probability of a particular response to an item depends upon a number of parameters, most notably the ‘person parameter’, represented by Greek letter theta (θ). θ represents the true level of a trait (i.e. the construct of interest) of a particular test taker. A central concept within IRT is that of item and test information. Information is analogous to reliability in Classical Test Theory, with one notable difference: It is recognised that the reliability of an item or scale may vary according to the level of θ of each individual test taker. For example, a scale that measures extraversion may be extremely reliable when administered to a test taker with a moderate level of extraversion but may be much less reliable when measuring test takers with extremely high or extremely low levels of extraversion. IRT techniques take this contingent reliability into account when describing an item or test’s information, representing it as an item information curve (IIC) or test information curve (TIC). Examining item and test information in this way can be a helpful additional way of identifying problematic items as a part of the scale development process alongside more traditional methods (see Steps 5 and 6 later in this chapter).

The Application of Psychometrics in Consumer Psychology Psychometric tools have been used extensively in consumer psychology. Utilising psychometric methods and models allows for reliably measuring an individual’s abilities, attitudes, and personality traits, and, thus, more confidently predicting patterns of behaviour.Therefore, the successful prediction of consumer behaviour has many applications within consumer psychology, where psychometric testing adds to the validity of research. The technique of testing a sample of a population in order to predict behaviour generalisable to a more global level forms the basis of much research in consumer behaviour. However, conducting research does not necessarily require that a new instrument be constructed. A huge range of psychometric tools are available that a researcher can draw upon to measure variables relevant to consumer psychology. For example, Klein, Ettenson, and Krishnan (2005) studied ethnocentrism from a viewpoint of consumers’ preferences for foreign products. To investigate this phenomenon, they tested participants with a previously developed instrument, the CETSCALE (Netemeyer, Durvasula, & Lichtenstein, 1991). The authors of the CETSCALE had rigorously tested its factor structure and internal consistency estimates across four countries. As well as this, they examined discriminant validity by investigating the sample’s general attitudes towards their home country, exploring their perception of their home country’s people, practices, and values. Furthermore, as there should be a link between countries for this consumer behaviour, the scale’s nomological validity was also examined to find out whether the constructs measured by the CETSCALE could be applied across different countries. For Klein and colleagues, the purpose of using this soundly

64  Daniel P. Hinton and Tracey Platt

psychometrically tested scale gave them an instrument to extend the knowledge around this construct further by using it to test in different, previously untested countries from the original four, thus strengthening the validity and predictive value of the tool. Psychometrically testing the characteristics and behaviours that form the constructs of human personality is well established. However, Aaker (1997) discovered that by applying the same psychometric principles to brands, one could identify a brand’s ‘personality’. Aaker defined this construct as “the set of human characteristics associated with a brand” (p. 347). This may seem like an odd thing to want to examine, but once one understands these perceived symbolic characteristics, they can be used to better understand the personality-based drivers behind purchasing decisions and brand loyalty. The steps taken to uncover the traits that exist within a coherent model of brand personality were done sequentially and rigorously to ensure the properties remained reliable, valid, and, thus, generalisable. Factor analysis conducted on the item set indicated that brand personality consisted of five factors: Sincerity, excitement, competence, sophistication, and ruggedness.

BOX 3.2  A CASE STUDY FOCUSING ON POSITIVE AFFECTIVITY The role that affect plays on consumer judgement and decision making is well documented in the consumer behaviour literature (Cohen, Tuan Pham, & Andrade, 2008). Positive affect and the facial expression of joy— namely, smiling—have been shown to increase brand loyalty and repeat purchasing behaviour (Jacoby & Kyner, 1973), as well as customer satisfaction (Söderlund & Rosengren, 2008), and extends to product advertisement (Schmitt, 1999). However, recent investigations into gelotophobes, individuals with the fear of being laughed at, showed how these individuals misperceive displays of positive affect (Ruch, Hofmann, & Platt, 2015; Hofmann, Platt, Ruch, & Proyer, 2015). These individuals do not experience positive emotional contagion from hearing laughter or seeing smiling, but irrationally feel that they are being ridiculed. In these situations, most people will mirror the enjoyment and begin to smile also, but gelotophobes react to them by displaying facial expressions of contempt. With up to 13% of the population experiencing some form of gelotophobia, any marketing strategy utilising positive affect should also consider its potential aversive impact upon their target market, in that the same stimuli could have both a positive or aversive impact on consumer behaviour such as purchasing decisions. A deeper understanding can be obtained using

Measurement Theory and Psychological Scaling  65

psychometric testing, which can circumvent any problems. Gelotophobes are less aversive to low arousal displays of positive emotion. Psychometrically testing the perceived level of arousal of the positive affect embedded within the stimuli could benefit the marketing campaign, gaining all the well-documented benefits of using the positive affective states, while having none of the disadvantages of it becoming distressing to those sensitive to the laughter of others. Conversely, as eliciting negative emotion can also positively influence consumer behaviour, having psychometric tools that can measure when emotion elicitation becomes aversive can guide marketing.

Research projects in consumer psychology—as is the case in other branches of psychology—routinely include questions relating to demographic information. However, a recent big data trend is towards the use of psychographics to complement demographic information about consumer populations (e.g. Lin, 2002). Whereas demographic questions gather information on the characteristics of the population being measured, such as gender, age, education, race/ethnicity, occupation, income level and marital status, psychographics start with the very sensible proposition that it would be naïve to assume that these demographic groups are homogeneous in other respects. If one is able to understand particular consumer groups in terms of their differences on key psychological constructs, a greater depth of understanding of the consumer can be achieved, and the efficacy of things like marketing practices can be enhanced. Psychographics seeks to do this by enhancing market segmentation processes. Whereas traditional approaches to market segmentation (Smith, 1956) seek to understand groups within a market in terms of shared behaviour and needs, psychographic segmentation attempts to classify consumer groups according to similarities in their psychometric profiles (Kotler, 1997). This approach provides marketers with greater insight into the nature of sub-markets within demographic market segments, allowing for more effective, targeted sub-marketing strategies (Lin, 2002).

The Scale Design and Development Process While there is a huge amount of psychometric measures—both freely available academic scales and commercially focused tools marketed by test publishers— available for the purposes of research in consumer psychology, it may occasionally be the case that a researcher is unable to obtain a suitable scale to measure the psychological construct in which they are interested.This may be for a number of reasons. It may be the case that the researcher wishes to explore a new construct that has not been well researched at that point in time. Equally, measures may

66  Daniel P. Hinton and Tracey Platt

exist that measure the construct, but they may be of poor quality, prohibitively expensive, or may be dated and may not reflect our current understanding of the construct’s nature. In any of these cases, a researcher might have to consider the possibility that he or she develops a new scale to measure the construct from scratch. The process of scale design and development is a long and relatively labourintensive one, and certainly not to be taken lightly. This reflects the nature of psychometric scales in that they have been carefully designed to ensure reliability and validity of measurement. Over the years, a great many extensive academic papers, book chapters, and entire textbooks have been written on the specifics of scale design and development (e.g. DeVellis, 2016; Schweizer & DiStefano, 2016; Hinkin, 1995, 2005; Churchill, 1979; Nunnally, 1978), and it is not the intention of the authors to present a comprehensive account of the single best practice approach within this chapter. Rather, the authors aim to synthesise and condense the wider scale design and development literature to provide a broad overview of the process that a researcher might use as a starting point in the development of a scale development project. In the following steps, the authors will describe common approaches to the design of a scale. For the sake of simplicity, in the main, the process describes the design of a single, unifactorial scale. If one were seeking to design a multifactorial tool such as a personality measure, this process would be replicated for each construct to be measured by the final tool.

Step One: Defining the Construct The first—and in many respects the most critical—step in the process of scale design is to establish a clear definition of the construct of interest. This may seem like a fairly obvious point to make, but in the authors’ experience, students frequently express an interest in conducting scale development projects with no clear idea of what it is they want to measure before they begin the process. In all branches of psychology, research is continually moving forward, and as a result, our understanding of the psychological constructs in which we are interested will slowly (or, occasionally, rapidly) evolve. The foundation of a good, robust psychometric measure is a coherent link between the construct being measured and the content of the items that comprise it. There are two broad approaches to scale design. The approach taken is directly informed by current understanding of the construct of interest. If a construct is well understood and clearly defined within the literature, the deductive approach to scale design should be taken. The deductive approach involves a thorough search of the published literature relevant to the construct of interest in order to develop a definition of the construct that covers its full breadth and depth. It may be possible that an adequate working definition of this construct

Measurement Theory and Psychological Scaling  67

already exists, though, in many cases, existing definitions may need to be modified somewhat to accurately capture current understanding of it. However, it is sometimes the case that no clear definition of the construct exists in the literature (perhaps because the construct is relatively new), or that the understanding of the construct has evolved over time to the point where definitions in the literature no longer accurately reflect its breadth and detail. In these instances, the inductive approach to scale design should be taken. The inductive approach relies on the use of field-gathered, observational information, most frequently in the form of interviews with subject matter experts (SMEs). SMEs can be anyone who has a good working knowledge of the area in which the construct is based, so they can be either academics or practitioners specialising in that area, or a mixture of both. The inductive approach seeks to identify patterns of behaviour that can then be used to guide the development of new theoretical models. However, as Gioia, Corley, and Hamilton (2012) highlight, this approach must “apply systematic conceptual and analytical discipline that leads to credible interpretations of data” (p. 15). Gibbert and Ruigrok (2010) argue that even with the lack of control of the context of this research method, it is possible to ensure qualitative rigour that also has both reliability and validity in the way the research is conducted. This can be achieved by addressing considerations such as triangulating sources of evidence, used here to mean looking at the same phenomenon but using different data collection strategies and source materials. One such systematised information collection process that is particularly useful for the inductive approach is the Grounded Theory method, first proposed by Glaser and Strauss (1967). This method involves stages in the grounded theory process and, although there has since been divergence reported in the literature in terms of how the processes should be conducted (e.g. Heath & Cowley, 2004), the principles remain the same.The first is data gathering from an initial research question, which is then posed to SMEs via interviews, observations, diary keeping, or focus group discussion. The second stage is note taking on the specifics of the collected data. The third stage is the coding of the information into categories, which then develop into the new construct or constructs. Gathering more data will do one of two things at this point. It will either yield new information, which can be grouped into an additional, new concept, or it will be repeated information, which relates to an older concept that has already been obtained and categorised. This process of collecting, analysing, and interpreting is continued until the point of information saturation is reached, at which point no new information can be gleaned from the data analysis.This is where the fourth and fifth stages of the process begin: Sorting, followed by writing up.The advantage of the Grounded Theory approach is that it is designed to develop clear definitions of the construct or constructs of interest from first principles, untainted by any existing pre-conceptions that the researcher may have going into the process.

68  Daniel P. Hinton and Tracey Platt

Step Two: Generating the Initial Item Set and Deciding Upon Response Format The next step in the process is to generate an initial set of items.This item set will undergo several stages of iterative refinement before the final item set that makes up the scale is identified, so it stands to reason that substantially more items than are intended to go into the final scale will need to be generated in this step. The broad approach to scale design taken in the previous step—whether it be the deductive or inductive approach—will have an impact upon the process of item generation in this step. This is where the advantages of the deductive approach become apparent, as it is generally the case that the initial item set will be made up of fewer items than that for the inductive approach (Burisch, 1984). A rule of thumb for the deductive approach is to generate around twice as many items as are intended to make up the final item set (Hinkin, 2005), so, if a researcher was intending to design a relatively short scale of 10 items, at least 20 items should be generated. Other authors have recommended a more cautious approach to item generation, recommending that as many as three or four times the number of items required for the final scale are generated at this stage (DeVellis, 2016). Regardless of the approach taken, the actual generation of items in this step is a relatively straightforward process. In many respects, the quality of items does not matter as long as sufficient items are generated, as the process of content validity and subsequent development will separate the weak items from the stronger ones. A good place to begin writing items is to attempt to paraphrase the construct as it has been defined (DeVellis, 2016). From there, item wording can be changed to express similar ideas in different ways. However, Hinkin (2005) identifies a number of guiding principles to the writing of items. Firstly, the perspective— whether behaviourally based or affective—of the items should be consistent across the whole item set. Secondly, each item should only reflect a single issue. Thirdly, items should not be phrased in the form of leading questions, neither should they be on topics that all test takers would be likely to endorse. Finally, every item should be written so that it can be easily understood by all test takers, so the researcher needs to carefully consider the target population for the tool and their likely level of literacy. As a rule, shorter items tend to be easier to read (DeVellis, 2016). A key decision to be made at this stage is that of the response format of the items. There are many potential response format options that may be chosen for the tool, each of which have both strengths and weaknesses (see Table 3.1 for a summary). The nature of the response format chosen necessarily impacts upon later stages of the scale development process in terms of the kind of analyses that can—and should—be conducted upon the data collected using them. It is important to note that, whichever response format is decided upon, the chosen format should be the same across all items within a scale. While it is possible to include

Weaknesses

Quick to complete

Loss of data fidelity (i.e. accuracy of measurement) Cannot compare test takers

Allows for precise measurement; Takes longer to complete allows for comparison between test takers

Strengths

PAPI (Lewis & Anderson, 1998) Allows comparison of relative strength and importance of traits and behaviours within an individual; less threat of socially desirable responding (SDR) behaviour

Forced choice; rank order

Ipsative

Most like me Least like me

NEO PI-3 (McCrae, Costa & 1 = Strongly disagree Martin, 2005) 2 = Disagree 3 = Neither Agree nor Disagree 4 = Agree 5 = Strongly Agree Likert-type scale 1 = Does not apply at all . . . Albert and Valette-Florence’s 10 = Totally applies (2010) Brand Love Scales Dichotomous/ True Hogan Personality Inventory binary False (HPI; Hogan, 1995)

Example Tools

Likert scale

Example Response Options

Normative

Measure Type Response Format

TABLE 3.1  Summary of Response Formats in Psychometric Scales

70  Daniel P. Hinton and Tracey Platt

multiple response options within the same tool (e.g. Saville’s Wave personality inventory; Kurz & MacIver, 2008), for the purposes of development, multiple different response formats should be treated as different scales. One particular distinction to draw is between normative and ipsative measures. Whereas normative measures seek to compare test takers to others (e.g. that a particular test taker is more outgoing than most people), ipsative measures draw comparisons between traits within an individual. Ipsative items force test takers to choose between mutually exclusive response options, allowing scales to differentiate between traits or behaviours in terms of their strength or importance relative to one another. For example, an ipsative item might ask test takers to rank a set of statements in order of the degree to which they feel each statement applies to them, or it might ask them to select one statement as being ‘most like me’ and another as ‘least like me’. While ipsative measures do not allow comparison between test takers, they have a distinct advantage over normative measures in that they can be designed in such a way as to minimise socially desirable responding (SDR; see the following discussion) behaviour (Bäckström, Björklund, & Larsson, 2009). However, the process for scale development of ipsative measures is much more complex than development of normative measures (Hicks, 1970). For this reason, and for the purposes of clarity, the remaining steps of the scale development process described here will make the assumption that a normative scale is being developed. One particular issue of contention within the literature is whether any items should be generated that are negatively keyed. Negatively keyed items are worded in such a way that not endorsing them is associated with a higher score on the construct of interest than endorsing them. For example, in a scale to measure extraversion, the item “I don’t like being in large crowds of people” would be endorsed more by test takers with lower levels of extraversion. Inclusion of negatively keyed items presents a number of benefits to a scale, as they combat response set bias (Price & Mueller, 1986), and encourage test takers to pay more attention to the content of the items. However, their inclusion has the potential to affect the psychometric properties of the final scale (Harrison & McLaughlin, 1991, cited in Hinkin, 2005).

Step Three: Establishing Content Validity Once the initial item set is generated, the process of iteratively reducing this item set down to make the resultant scale as robust as it can be can begin. The first stage in this process—and one that is frequently overlooked—is to establish content validity. It may appear, at least to the designer of the item set—that the items that have been generated are all perfectly fit for purpose, but this is a dangerous assumption to make. Once the process of item trialling (see next step) has begun, there is no going back—or, at least, not without substantial additional effort—and the discovery that one’s item set does not behave in the way expected necessitates

Measurement Theory and Psychological Scaling  71

returning to the item generation stage and discarding any data collected to that point. Clearly, this a scenario that should be avoided, and the establishment of content validity is a critical step in minimising the risk of this happening. To explore the content validity of the item set, the researcher will need to draw upon the expertise of some subject matter experts (SMEs), as is the case for the process of construct development in the inductive approach, albeit using different methods to do so. Generally speaking, the number of SMEs needed for this step is much smaller than the number of participants required in the later item trialling steps, as the analyses conducted on their responses tends to be less sophisticated, statistically speaking. For the most basic of content validity processes, feedback from between five and ten SMEs should be sufficient. Exploration of content validity tends to be more qualitative in nature, though there has been a relatively recent move towards imposing more quantitative frameworks upon the process (Hinkin, 2005). Typically, this process is made up of two components. In the first, SMEs are asked to make a judgement (by using, for example, ratings of one to five) on the degree to which each item taps into the construct of interest, based upon the definition with which the researcher has supplied them. In the case of tools that are made up of multiple scales, it is helpful to have SMEs rate the degree to which each item taps into every construct measured, not just the construct measured by the scale to which the item belongs. In the second, SMEs are asked to make comments upon individual items and upon the scale in general. They are asked to comment on any items that they feel could be better worded (or should be removed altogether) and asked to highlight whether there is anything missing from the item set in terms of the breadth and depth of coverage of the construct. Guidance on how the researcher then interprets and utilises this data varies within the literature. The researcher may look over the items and remove any item which is poorly worded (for which a simple rewording of the item isn’t an option), and/or which does not appear to tap strongly into the construct of interest. Alternatively, a more quantitative approach may be taken, in which the degree of agreement between SMEs’ ratings of each item is calculated (using, for example, Fleiss’ Kappa, a statistical measure of the consistency of rating across different raters), discarding any item for which consensus has not been reached. However, care should be taken when making decisions about whether to discard an item on the basis of a poor Fleiss’ Kappa value (see “Troubleshooting Tips”). Content validity on multifactorial tools, in which SMEs are asked to rate the degree to which each item relates to each construct of interest, allows for a degree of quantitative comparison between ratings of relevance of an item across constructs. Using an ANOVA, a researcher can examine which of the items are rated as significantly more relevant to their intended construct than to other constructs in the tool, discarding those that are not (Hinkin, 2005). This is a much more robust and objective approach to the establishment of content validity, though it

72  Daniel P. Hinton and Tracey Platt

requires the collection of data from substantially more SMEs than the approaches described earlier in order to achieve adequate statistical power.

Step Four: Initial Item Trialling Once the item set has been refined and the researcher is happy that the scale to this point demonstrates acceptable content validity, the next step involves the further development of the scale by examining its psychometric properties and identifying any items that may be negatively affecting its reliability or validity.This process involves examining the scale’s internal consistency and structural validity, and removing items from it iteratively, one at a time, until an item set can be identified that demonstrates robust psychometric properties. To do so requires some data. The scale items are administered to a relatively large number of participants, chosen to represent the group for whom the final scale is intended. These participants complete all items that remain in the scale. The only data required for this process are the item responses themselves, so any other data collected at this stage, such as demographic data, will not be used for any analyses. The key determinant of the sample size required for this phase of item trialling depends upon the number of items within the item set. As some form of factor analysis (see the following discussion) is going to be the most complex analysis used in this step, the normal rules should be followed for establishing minimum sample size for these procedures. Though guidance in the literature varies, the most liberal estimate offered of the ratio of number of participants : number of items is 10:1 (Nunnally, 1978). That is to say, if the item pool entering this stage of analysis was made up of 20 items, approximately 200 participants would be needed. First, the item set is checked for internal consistency. This is achieved by computing Cronbach’s α for the scale as a whole, and for each individual scale if the tool is made up of a number of separate scales.The common rule of thumb is that a scale with a value of Cronbach’s α above 0.7 shows acceptable internal consistency, though the closer to 1.0 this value is, the better, generally speaking. Certainly, a value of Cronbach’s α below 0.7 is cause for concern, indicating that at least one of the items is adversely affecting the scale’s reliability. It is worth pointing out that, if the response format of a scale’s items is dichotomous in nature, the value of KR-20 for the scale should instead be calculated. The way in which KR-20 functions is very similar to Cronbach’s α, however, so the rest of this section will refer exclusively to the calculation of α values. Next, the tool’s structural validity is examined. To achieve this, a dimension reduction technique such as Principal Axis Factoring (PAF), or another form of Exploratory Factor Analysis (EFA) is used.There is a degree of debate in the literature as to the ‘correct’ or ‘best’ method to use to explore the underlying structure of the tool, and, at least to some degree, the choice of analytic approach is influenced by the nature of the data obtained. Hinkin (2005) suggests that the rotation

Measurement Theory and Psychological Scaling  73

method chosen for EFA depends upon whether the latent factors underlying the data are expected to be correlated (oblique rotation) or uncorrelated (orthogonal rotation), but that, to be on the safe side, it is prudent to run factor analysis with each and compare the solutions obtained. A point of much greater contention is the extraction method that should be used. It has been argued that, generally speaking, a maximum likelihood estimator (ML) is the best choice when data are normally distributed, whereas PAF is the best choice for non-normally distributed data (Costello & Osborne, 2005). However, purists might argue that, since Likert-type scales are ordinal in nature rather than being scale-level data, a diagonally weighted least-squares (WLSMV) estimator is more appropriate for factor analysis (Bandalos, 2014).To further muddy the water, in the case of dichotomous data, it appears that an unweighted least-squares (ULSMV) estimator appears to produce the best results (Parry & McArdle, 1991). In practice, however, most forms of factor analysis techniques yield similar results, particularly with larger sample sizes. However, one approach that should never be used for the purposes of scale development is Principal Components Analysis (PCA). PCA is not, in actual fact, a form of factor analysis at all, and, as such, cannot be used to identify the latent structure underlying a dataset. Lee and Hooley (2005) lament that a great deal of scale development documented in the marketing literature wrongly uses PCA and speculate that this may be because it is the default extraction option within SPSS. Whichever method is chosen, and whether the tool is made up of one or a number of scales, the broad approach is the same: The structure of the data is examined to determine the number of factors that underlie it. There can be something of an art to determining the correct number of factors to extract, and traditional methods, such as the Kaiser Criterion (examining the number of factors with eigenvalues above 1.0), or examining the point of inflexion on the scree plot (Cattell, 1966), can provide wildly different solutions. The most robust way of determining the most likely number of factors underlying your data is to conduct parallel analysis (e.g. O’Connor, 2000). Parallel analysis works by simulating a random dataset that contains the same number of variables (i.e. items) and cases (i.e. participants) as the dataset under examination. The eigenvalues for the random dataset are computed and are then compared to those of the factors extracted from the real data. The point at which the eigenvalues of the simulated data become larger than those of the real data indicates that no further meaningful factors underlie the data. Clearly, any indication that the number of factors underlying the data differs from the number of constructs the tool is intended to measure suggests that there is a need to examine the items more deeply to identify those that are problematic. Assuming that the previous two analyses have uncovered some room for improvement in either the internal consistency or structural validity of your tool, the next step is to identify which items are causing the problem. Indeed, it is very unlikely that the tool cannot be improved in some way with the removal of

74  Daniel P. Hinton and Tracey Platt

specific items, so the researcher should expect to have to eliminate some items at this stage, highlighting the necessity for generating more items than are required for the final scale. The process by which items are identified as problematic is through use of the previous two statistical techniques, though their focus is somewhat different to that described earlier.The process revolves around four statistical checks for each item, upon failure of any of which the item is flagged as problematic. Each scale within the tool is examined separately for these analyses. In the first check, Cronbach’s α is computed for the scale, along with the value of α should each individual item be deleted. Any item that shows a substantial increase in the value of Cronbach’s α means that that item should be flagged for deletion. The most problematic item is removed, and α is computed for the remaining scale items. As before, if any item or items appear problematic, the worst offender is removed. This process repeats until no further substantial increase in α can be achieved. The remaining checks are conducted using exploratory factor analysis techniques. In contrast to the guidance on factor analysis provided earlier, in these analyses, orthogonal rotation such as varimax rotation should be used in order to achieve maximum differentiation between the construct of interest represented by the primary extracted factor and any nuisance factor (Hinkin, 2005). In the second check, a two-factor solution is forced. As the aim is for a unifactorial scale, any item that loads substantially (i.e. with a factor loading greater than 0.30; Hair et al., 1998) upon the second, nuisance factor is flagged. As before, the analysis is then rerun with this item removed, and the loadings of the remaining items examined. This process continues until no further items are shown to load substantially upon the second extracted factor. For the third and fourth checks, factor analysis is again conducted, but this time forcing a one-factor solution. Here, any item that shows either a low loading on the single extracted factor (less than 0.3; Hair et al., 1998), or shows low communality with all the other items in the scale (less than 0.3) is flagged for deletion. Having removed the worst performing item, this analysis is repeated as before until no further items remain. Once these checks have been completed, the researcher is then faced with the decision of which items to retain and which items to delete from the final scale. The number of flags obtained by each item will largely inform this decision but may also be influenced by the nature of the tool. If the number of items required for the final scale is small relative to the number of items in the analyses, the researcher can afford to be choosier and impose stricter criteria for inclusion in the final scale. However, Lee and Hooley (2005) urge caution when removing items on the basis of low communality, so any decision made on the basis of this criterion should be made in conjunction with evidence provided as part of the other checks. Once the decision has been made of which items to retain, Cronbach’s α is computed, and factor analysis is conducted (preferably supported by parallel analysis) to check that the final item set demonstrates acceptable internal consistency and is acceptably unifactorial.

Measurement Theory and Psychological Scaling  75

Step Six: Second Item Trial At this point, the researcher has designed and developed a scale that demonstrates (assuming that the previous steps have been at least partially successful) good psychometric properties, in that it reliably measures a single construct. However, the process of scale development is not complete at this stage. A key question remains over whether the construct that it measured by the scale is, in actual fact, the construct it is intended to measure. To address this, it is necessary to examine the scale’s convergent and discriminant validity, which requires the identification of suitable other published psychometric measures. In order to explore convergent validity, a scale needs to be identified that measures the same construct. This shouldn’t be a problem if the deductive approach has been taken, as there should be an abundance of scales that exist that measure the construct of interest. However, this can be challenging if the inductive approach has been taken. For discriminant validity, an existing scale needs to be identified that measures a construct different to the one that is the focus of the scale. If a personality measure is being designed, a useful resource to help in this process is the International Personality Item Pool (Goldberg et al., 2006). IPIP is a huge repository of robust, published, free-to-use personality scales that covers a very wide range of constructs. At this stage, it is also a sensible idea to assess the scale’s criterion-related validity to establish whether the scale is predictive of important outcomes, such as consumer behaviour. As such, some sort of data should be collected to aid in this. The precise nature of the criterion data gathered will depend upon the nature of the scale and its intended use. Some examples of the kind of criterion data that is typically collected for the validation of scales within consumer psychology are things such as the prediction of brand associations (e.g. Berry, 2000), consumer decision-making behaviour (e.g. Kidwell, Hardesty, & Childers, 2007), and purchasing decisions (e.g. Nenkov et al., 2009). Once the measures that will be used for this round of analyses have been identified, they are administered alongside the remaining items within the scale to a new set of participants. The number of participants required for this step again varies according to the size of the scale but is determined largely by the number of cases needed for model identification in the analyses based on confirmatory factor analysis (CFA). Unlike for EFA, there is little clear guidance on this in the literature, as model identification is dependent on the specific nature of the model, the number of indicators (i.e. items), and the number of latent factors specified. Attempts to provide approximations of number of cases needed for model identification have been found to vary wildly (e.g.Wolf et al., 2013). Suffice to say, if the specified model has any degree of complexity, it is likely to require a sample size in the hundreds, or even the thousands for very complex models. The scale’s internal consistency and structural validity is first checked using Cronbach’s α and factor analysis. The results of these are likely to be slightly different to those obtained in the previous step, but this is not a cause for concern

76  Daniel P. Hinton and Tracey Platt

as long as the scale still demonstrates acceptable reliability and validity. Once this has been checked, CFA is conducted on the scale items, specifying a single latent factor onto which each item is made to load. As for EFA in the previous step, the estimator used for CFA will vary according to the nature of the data. In the case of ordinal data such as Likert-type items, a robust diagonally weighted least-squares estimator is preferable (Schweizer & DiStefano, 2016), assuming that the statistical package you are using to conduct CFA supports computation of polychoric correlations (see “Software” section). The CFA model’s fit indices are then examined to determine whether or not the data are acceptably unifactorial. The best fit indices to use for this purpose are, again, a source of some debate in the literature. The most commonly used measures of CFA model fit are shown in Table 3.2, along with recommended values by which to judge the adequacy of model fit based on guidance from Hu and Bentler (1999). If model fit is poor, special attention should be paid to the model’s modification indices. These provide an indication of the improvement to the model if its structure is changed in specific ways. This is very helpful when conducting CFA on the items within a multifactorial measure, as it allows the researcher to examine model fit when item cross-loadings (i.e. when an item loads upon more than one construct) are taken into account. Once the factor structure of the tool has been confirmed, convergent and discriminant validity can be examined. In the past, the common procedure for examining convergent and discriminant validity was to generate large matrices of Pearson correlations between the scale items and those within the additional existing scales that were identified prior to this round of data collection. This correlation matrix would then be examined to establish patterns of intercorrelations that suggested convergent and discriminant validity (i.e. strong correlations between items designed to measure the same construct and weak or no correlations between items designed to measure different constructs). However, modern approaches to investigating convergent and discriminant validity are somewhat TABLE 3.2  Common Measures of Fit Used in CFA (From Hu & Bentler, 1999)

Measure

Values for quality of model fit

Chi Squared (χ2)

Extremely variable. Lower values indicate better model fit. 0.9 = acceptable 0.95 = good 0.9 = acceptable 0.95 = good 0.05/0.06 = good 0.08 = acceptable > 0.12 = poor 0.08 = good

Comparative Fit Index (CFI) Tucker-Lewis Index (TLI) Root Mean Squared Error of Approximation (RMSEA) Standardised Root Mean Squared Residual (SRMR)

Measurement Theory and Psychological Scaling  77

more robust that this, drawing upon CFA procedures. In this approach, competing CFA models are examined and their fit indices compared. For convergent validity, a model is specified in which the scale’s items and those from the existing scale designed to measure the same construct are all made to load onto a single latent factor. In the competing model, the scale’s items are made to load onto a single latent factor, and the other scale’s items onto a second latent factor, which is correlated with the first. To demonstrate acceptable convergent validity, it should be the case that the first model should be a better fit to the data than the second.The procedure for the examination of discriminant validity using CFA is very similar. In this case, the model in which the scale’s items and those within the existing scale designed to measure a separate construct are made to load onto separate, correlated latent factors should be a better fit to the data than that in which both scale’s items are made to load onto a single latent factor. The examination of criterion-related validity is, statistically speaking, much more straightforward than the previous set of analyses. To establish criterionrelated validity, scale scores are first computed for each participant by adding their scores on each item. Pearson correlations are then computed between these scale scores and the criterion data that has been collected. These correlation coefficients can then be judged according to their strength. Though the strength of these relationships can vary substantially according to the nature of the specific criterion used, general rules exist within the literature for the judgement of the adequacy of evidence to support criterion-related validity. Coefficients of 0.50 or above are considered to be excellent evidence for a measure’s criterion-related validity, correlations above 0.35 to be good evidence, above 0.2 to be adequate, and 0.2 and below to be inadequate (Hemphill, 2003).

Step Seven: Norming the Final Tool and Developing Scoring Procedures If all has gone to plan, the researcher will have, at this point, a robust tool that measures the construct for which it was intended. The final step of the process it to norm the tool. The process of norming the tool adds meaning to specific scale scores, allowing users of the tool to understand the behavioural implications for consumers based on how they respond to each scale’s items. When developing a norm for each scale, it is a relatively straightforward process. From the two-item trialling phases in the previous two steps, there should be plenty of data that the researcher can use to construct a norm group, against which future test takers’ scores can then be compared. Most statistical packages feature an option that allows the researcher to generate the percentile score that corresponds with each participant’s scale score. Alternatively, sten scores for each participant can be calculated by first computing their z-score and then by applying the formula that follows. Sten = 2z + 5.5

78  Daniel P. Hinton and Tracey Platt

This will give each participant in the norm group a sten score between 1 and 10. Sten ranges can then be banded as being representative of the below average range of the trait (stens 1–3), the average range (4–7) or the above average range (stens 8–10).The researcher can then attach narratives to each of these ranges that can be used to make inferences about a consumer’s likely preferences for behaviour on the basis of their score.

Further Issues in Scale Development Measurement Bias and Its Detection A frequently neglected issue within scale development is that of measurement bias. Measurement bias, in its simplest form, occurs when a scale functions (i.e. measures the construct of interest) differently for different groups of test takers. The overwhelming majority of psychological research has been conducted on participants from Western, educated, industrialised, rich, and democratic societies (so-called WEIRD participants; Henrich, Heine, & Norenzayan, 2010), and scale development is no exception. It would be naïve to assume that all psychometric scales functioned in exactly the same way for all groups of participants, particularly those from different countries with different cultural artefacts and values. Therefore, it is prudent to consider the impact that any possible measurement bias might have on research findings when using scales in different cultural contexts. Measurement bias is most frequently assessed using techniques based on IRT methodology (see the earlier discussion). The most common set of techniques aim to detect differential item functioning (DIF), which contributes to differential test functioning (DTF), an antecedent of measurement bias. Though these techniques can be extremely useful in establishing the validity of a piece of research that has been conducted in a context dissimilar to the one in which a psychometric it uses was developed, they are all either very time consuming, mathematically complex, or, most frequently, both. Most DIF/DTF procedures require specialist statistical software to run (see section on software).

Socially Desirable Responding Behaviour As with any self-report measure in psychology, an ever-present issue is that of socially desirable responding (SDR) behaviour. Previously known as faking, SDR occurs in instances in which it is tempting for test takers to present a more favourable portrayal of themselves than might actually be the case. While it is seen as a more serious problem for the use of psychometrics when they are used in highstakes situations, such as job selection (Hogan, Barrett, & Hogan, 2007), SDR is, nevertheless, an issue for the use of psychometrics in lower-stakes situations, such as those that are the focus of consumer psychology research.

Measurement Theory and Psychological Scaling  79

There are a number of solutions that have been formulated to address SDR behaviour. One somewhat contentious approach is the inclusion within the tool of an in-built SDR scale. SDR scale items are those that it would be very unlikely that a participant would endorse were they not trying to appear more socially desirable than was the case. For example, an item within an SDR scale might read “I have never been late for an appointment”. Punctuality is a socially desirable trait to have. However, almost without exception, everyone who has reached adulthood has been late for at least one appointment in his or her lifetime, so it is extremely unlikely that a test taker would endorse this if they were being completely honest in their responses. It should be clear, then, that as the number of items endorsed within an SDR scale increases, the probability that the test taker is responding genuinely becomes vanishingly small. The concept of SDR scales sits uneasily with many researchers and practitioners. Many see such scales—quite rightly—as a form of deception, and, therefore, as being unethical to use. A somewhat more palatable approach to addressing SDR is for psychometrics to establish an honesty contract within their instructions (e.g. Bartram, 2009). An honesty contract describes the necessity for honest responding from the test taker and informs them that the authenticity of their responses can be checked, the hope being that they will be compelled to answer honestly. Whichever of these approaches seems more appropriate, the issue of SDR behaviour is one that should not be ignored. SDR necessarily impacts upon the validity of measurements provided by psychometric scales (Hogan et al., 2007), thus eroding the validity of research findings that depend upon them.

Translating Scales into Other Languages The final issue to be discussed on the subject of scale development is something of a special case and one with which many researchers will never be faced. The translation of scales may seem like a relatively straightforward process, but it is fraught with difficulty. One cannot simply translate existing scale items and expect the whole scale’s meaning to be preserved, as a sentence is more than the sum of its parts. One common approach to ensuring that meaning is preserved when a scale is translated is the process of translation-backtranslation (van de Vijver & Leung, 1997). In this approach, a scale’s items are first translated into the language intended for the final tool. The items are then translated again by a second researcher back into their original language. The initial item set and these translated-backtranslated items are then compared to ensure that the meaning of each item is preserved. Harkness (2003) takes this process a step further, describing a five-step process for the translation of scales according to best practice. This process hinges upon three key roles, each of which should be fulfilled by separate researchers. In the first step, the items are translated by a translator. Next, a reviewer

80  Daniel P. Hinton and Tracey Platt

produces several different versions of the translated items. Following this, an adjudicator selects which of these versions to use for each item in the translated version of the tool. Once this has been completed, the scale undergoes pretesting. Finally, any problems with specific items and decisions about solutions are documented. As should be clear, the translation of scales is an undertaking that should not be entered into lightly. That said, translating scales allows consumer psychologists to gain a much deeper understanding of consumers within different cultures in the global marketplace.

Troubleshooting Tips Scale design and development processes rarely go entirely according to plan. As such, there are myriad roadblocks that a researcher may face. It is important to understand what the issue is and how best to address it when these problems arise. To that end, the authors will attempt to address some of the more common problems encountered.

My CFA Model Is Not Identified There are a couple of reasons why this might be the case. It may be that your model is misspecified, or it may lack a sufficient number of cases for identification. Try constraining additional model parameters first, as this may rectify the problem. If not, it may be that two or more of your latent factors which you thought to be orthogonal are actually correlated, in which case you should adjust the model accordingly. If neither of these approaches work, it may simply be the case that you need to collect more data.

When Assessing Content Validity, My Value of Fleiss’ Kappa Is Very Low This generally indicates poor agreement between your raters, but it is not necessarily the case. Fleiss’ Kappa assesses the degree to which your raters agree on specific scale numbers. As such, if two raters rated an item as being related to your construct of interest, one providing a rating of 4 and the other of 5, Fleiss’ Kappa would treat this as disagreement. Clearly, this is nonsensical. In this situation, try collapsing the points on your rating scale into a smaller number, incorporating, for example, points 4 and 5 into an overall rating. This should address much—if not all—of the issue.

In EFA, I Cannot Get to the Point at Which My Scale’s Items Are Unifactorial Unfortunately, if you’ve followed the process described in Step 5, and you’re still having problems, it may be an issue with either the wording of your items, the

Measurement Theory and Psychological Scaling  81

construct definition, or both. Scrutinise your items and your construct closely. Does it look like your construct might reflect two (or more) related-yet-different constructs? Do your items all adequately sample the construct? If the answer to either of these is yes, you may need to return to the earlier stages of the process.

I Don’t Have Enough Items in My Initial Item Set, but I Can’t Think of Any More Try playing around with the phrasing of words and try restating some of the existing items you have so that they focus on slightly different aspects of the construct. Perhaps also try writing some negatively keyed versions of the items you have. If you’re really stuck, it is perfectly allowable to consult some SMEs to ask for their input to the process. Scale design needn’t be a solitary pursuit.

Software In this section, the authors will discuss some of the software packages available to aid in the scale development process. While some of these packages will be familiar to the reader, it may be the case that they are not aware of some of the shortcomings of particular software packages. Any researcher wishing to carry out a scale development project should be aware of these pitfalls, as they have the potential to influence the results of the analyses carried out in the process of scale development. A statistical package that will be familiar to most—if not all—students and researchers in psychology is SPSS (IBM, 2016). For the most part, SPSS is able to conduct the analyses relevant to scale development. SPSS will allow you to compute Cronbach’s α to examine a scale’s internal consistency, to run PAF and/ or another form of EFA to test its structural validity, and to compute the correlations used to explore its criterion-related validity and—at least to an extent—its convergent and discriminant validity. However, SPSS is not without its shortcomings. The first, and most striking, of these is that it is unable to run CFA for the purposes of confirmation of factor structure in the second item trialling phase of scale development, or for the more modern approaches to establishing convergent and divergent validity. To achieve this, a more specialist statistical package, such as those described next, is required. A more serious shortcoming—at least from an academic point of view—is in the way in which SPSS treats categorical data within PAF and other dimension reduction procedures. Even though SPSS allows the user to manually flag variables as being nominal, ordinal, or scale data, the algorithms it uses to produce its solution treats all variables as scale level (Basto & Pereira, 2012). As such, any dimension reduction procedures run in SPSS are based on Pearson correlations, as opposed to the tetrachoric correlations suitable for dichotomous data, or polychoric correlations for ordinal data. Holgado-Tello et al. (2010) have demonstrated that using Pearson correlations in instances such as these tends to produce less accurate factor solutions than with polychoric or

82  Daniel P. Hinton and Tracey Platt

tetrachoric correlations. However, one potential way to address some of these shortcomings is through the use of a free-to-use plugin that allows the use of the R programming language (The R Foundation, n.d.) within SPSS. Although fairly complex for the average user to negotiate, R allows users to conduct analysis such as CFA in SPSS, and for tetrachoric and polychoric correlations to be computed for dichotomous and ordinal data respectively. SPSS Amos (Analysis of moment structures; Arbuckle, 2016), is a statistical software package that is able to perform CFA and Structural Equation Modelling path analyses. The key advantage of Amos, particularly for those who are relatively experienced in conducting CFA, is that it includes a graphical interface that allows the user to draw out their structural models.This is a hugely attractive quality and one that sets it apart from other CFA software. However, although a powerful statistical programme with a user-friendly interface, one of the shortcomings of Amos is that it cannot generate tetrachoric or polychoric correlations for CFA. One alternative to both SPSS and Amos for the more experienced researcher is Mplus (Muthén & Muthén, 2017). Mplus is a very powerful and versatile statistical package that is able to run CFA as well as all the other procedures that SPSS will do. The key advantage of Mplus is that it allows the user to define which variables within a dataset are categorical variables. Having done so, it will then base subsequent analyses on polychoric and/or tetrachoric correlations, depending on the nature of the variables defined as categorical. In addition to this, a set of add-ons is available for the basic Mplus software package that allow it to be used to run IRT-based analyses, such as to detect DIF/DTF for the exploration of possible measurement bias, or for the generation of item and test information curves to aid in the item trialling process. The disadvantage of Mplus is that it is based largely on syntax, so can be off-putting for students who are used to SPSS’ menu system or the graphical system used by Amos. One little-known statistical programme that is of particular note for scale development is FACTOR (Lorenzo-Seva & Ferrando, 2006). FACTOR offers a number of alternative procedures for parallel analysis to the traditional one described by O’Connor (2000). In particular, traditional methods tend not to handle items with dichotomous response formats particularly well (Tran & ­Formann, 2009). FACTOR is able to run parallel analysis based on tetrachoric correlations such as minimum rank factor analysis (PA-MRFA), which provides much more reliable indications of the likely number of factors underlying an item set (Timmerman & Lorenzo-Seva, 2011). FACTOR is freeware, so it costs nothing to download and use.

C-OAR-SE: A Critique of the Psychometric Approach One relatively recent response to the psychometric approach to scale development within marketing that has received considerable attention comes in the form of C-OAR-SE (Rossiter, 2002, 2011). The C-OAR-SE method revolves around the principle that the only real scale development of any value is that of

Measurement Theory and Psychological Scaling  83

the initial establishment of content validity. C-OAR-SE itself is an acronym that reflects the six stages of scale development it recommends—namely, construct definition, object representation, attribute classification, rater-entity identification, scale selection, and enumeration (Rossiter, 2011). Rossiter reasons that, if the content validity of a scale’s items and response scale is effectively established, then there is no need to explore construct validity nor criterion-related validity. As one might expect, several critiques have been made by leading psychometricians and experts in marketing research of C-OAR-SE and of Rossiter’s assumptions in developing it (Lee & Cadogan, 2016; Ahuvia, Bagozzi, & Batra, 2013; Rigdon et al., 2011). Most of the criticisms levelled at C-OAR-SE say that, while encouraging good content validation practices is to be applauded, the wholesale rejection of the subsequent steps in the scale development process sets a dangerous precedent, one that has the potential to return scale development in consumer psychology and related fields to the dark ages described by Churchill (1979). Rather unsurprisingly, as psychometricians, the authors agree with this assessment.

Exercises EXERCISE ONE: ITEM GENERATION AND CONTENT VALIDITY Think about a psychological construct that is well understood and well researched. Search the literature for a brief, clear definition of this construct. Try writing ten items that are designed to tap into this construct, then investigate your item set’s content validity by asking five friends or family members to rate—from 1 to 5—the degree to which each item taps into the construct as you have defined it. Do any items look as though they might not be suitable to be included in a scale to measure this construct? EXERCISE TWO: RIGOROUS RESEARCH DESIGN Advertisers need to find out the potential problems they are facing when they try to sell what could be perceived to be an embarrassing product. What process would you choose in order to find out about the product’s level of embarrassability and whether this has an impact on the consumer’s willingness to buy it? How could you ensure that the way you gather the initial information and beliefs around this product is systematic and has rigour? How could you ensure the items you build to measure this phenomenon are psychometrically sound?

84  Daniel P. Hinton and Tracey Platt

References Aaker, J. L. (1997). Dimensions of brand personality. Journal of Marketing Research, 34(3), 347–356. doi:10.2307/3151897 Ahuvia, A., Bagozzi, R. P., & Batra, R. (2014). Psychometric vs. C-OAR-SE measures of brand love: A reply to Rossiter. Marketing Letters, 25(2), 235–243. Albert, N., & Valette-Florence, P. (2010). Measuring the love feeling for a brand using interpersonal love items. Journal of Marketing Development and Competitiveness, 5(1), 57–63. Arbuckle, J. (2016). IBM® SPSS® Amos™ 24 user’s guide. Retrieved from ftp://public.dhe. ibm.com/software/analytics/spss/documentation/statistics/24.0/en/amos/Manuals/ IBM_SPSS_Amos_User_Guide.pdf Bäckström, M., Björklund, F., & Larsson, M. R. (2009). Five-factor inventories have a major general factor related to social desirability which can be reduced by framing items neutrally. Journal of Research in Personality, 43(3), 335–344. doi:10.1016/j.jrp.2008.12.013 Bandalos, D. L. (2014). Relative performance of categorical diagonally weighted least squares and robust maximum likelihood estimation. Structural Equation Modeling: A Multidisciplinary Journal, 21(1), 102−116. doi:10.1080/10705511.2014.859510 Bartram, D. (2009). The international test commission guidelines on computer-based and internet-delivered testing. Industrial and Organizational Psychology, 2(1), 11–13. Basto, M., & Pereira, J. M. (2012). An SPSS R-menu for ordinal factor analysis. Journal of Statistical Software, 46(4), 1–29. Bauer, H. H., Reichardt, T., Barnes, S. J., & Neumann, M. M. (2005). Driving consumer acceptance of mobile marketing: A theoretical framework and empirical study. Journal of Electronic Commerce Research, 6(3), 181–191. Berry, L. (2000). Cultivating service brand equity. Journal of the Academy of Marketing Science, 28, 128–137. doi:10.1177/0092070300281012 Burisch, M. (1984). Approaches to personality inventory construction: A comparison of merits. American Psychologist, 39(3), 214–227. doi:10.1037/0003-066X.39.3.214 Cattell, R. B. (1966).The scree test for the number of factors. Multivariate Behavioral Research, 1(2), 245–276. doi:10.1207/s15327906mbr0102_10 Churchill, Jr, G. A. (1979). A paradigm for developing better measures of marketing constructs. Journal of Marketing Research, 16, 64–73. doi:10.2307/3150876 Cohen, J. B., Pham, M. T., & Andrade, E. B. (2008). The nature and role of affect in consumer judgement and decision making. In C. P. Haugtvedt, P. M. Herr, & F. R. Kardes (Eds.), Handbook of consumer psychology (pp. 297–348). Mahwah, NJ: Erlbaum. Costello, A. B., & Osborne, J. W. (2005). Best practices in exploratory factor analysis: Four recommendations for getting the most from your analysis. Practical Assessment, Research & Evaluation, 10(7), 1–9. DeVellis, R. F. (2016). Scale development: Theory and applications (Vol. 26). London: Sage ­Publications Ltd. Gattol, V., Sääksjärvi, M., & Carbon, C. C. (2011). Extending the Implicit Association Test (IAT): assessing consumer attitudes based on multi-dimensional implicit associations. PLoS One, 6(1), e15849. doi:10.1371/journal.pone.0015849 Glaser, B., & Strauss, A. (1967). The discovery of grounded theory. Hawthorne, NY: Aldine Publishing Company. Gibbert, M., & Ruigrok, W. (2010). The “what” and “how” of case study rigor: Three strategies based on published work. Organizational Research Methods, 13(4), 710–737. doi:10.1177/1094428109351319

Measurement Theory and Psychological Scaling  85

Gioia, D. A., Corley, K. G., & Hamilton, A. L. (2013). Seeking qualitative rigor in inductive research: Notes on the Gioia methodology. Organizational Research Methods, 16(1), 15–31. Goldberg, L. R., Johnson, J. A., Eber, H. W., Hogan, R., Ashton, M. C., Cloninger, C. R., & Gough, H. G. (2006). The international personality item pool and the future of public-domain personality measures. Journal of Research in Personality, 40(1), 84–96. doi:10.1016/j.jrp.2005.08.007 Hair, J. F., Tatham, R. L., Anderson, R. E., & Black, W. (1998). Multivariate data analysis (5th ed.). London: Prentice Hall. Harkness, J. A. (2003). Questionnaire translation. Cross-Cultural Survey Methods, 1, 35–56. Heath, H., & Cowley, S. (2004). Developing a grounded theory approach: A comparison of Glaser and Strauss. International Journal of Nursing Studies, 41(2), 141–150. doi:10.1016/ S0020-7489(03)00113-5 Hemphill, J. F. (2003). Interpreting the magnitudes of correlation coefficients. American Psychologist, 58, 78–80. doi:10.1037/0003-066X.58.1.78 Henrich, J., Heine, S. J., & Norenzayan, A. (2010). Beyond Weird: Towards a broad-based behavioral science. Behavioral and Brain Sciences, 33(2–3), 111–135. doi:10.1017/ S0140525X10000725 Hicks, L. E. (1970). Some properties of ipsative, normative, and forced-choice normative measures. Psychological Bulletin, 74(3), 167–184. doi:10.1037/h0029780 Hinkin,T. R. (1995). A review of scale development practices in the study of organizations. Journal of Management, 21(5), 967–988. doi:10.1177/014920639502100509 Hinkin, T. R. (2005). Scale development principles and practices. In R. A. Swanson & E. F. Holton III (Eds.), Research in organizations: Foundations and methods of inquiry. San ­Francisco, CA: Berrett-Koeller Publishers Inc. Hofmann, J., Platt,T., Ruch,W., & Proyer, R.T. (2015). Individual differences in gelotophobia predict responses to joy and contempt. Sage Open, 1–12. doi:10.1177/2158244015581191 Hogan, R. (1995). Hogan personality inventory. Tulsa, OK: Hogan Assessment Systems. Hogan, J., Barrett, P., & Hogan, R. (2007). Personality measurement, faking, and employment selection. Journal of Applied Psychology, 92(5), 1270–1285. doi:10.1037/ 0021-9010.92.5.1270 Holgado-Tello, F. P., Chacón-Moscoso, S., Barbero-García, I., & Vila-Abad, E. (2010). Polychoric versus Pearson correlations in exploratory and confirmatory factor analysis of ordinal variables. Quality & Quantity, 44(1), 153–166. doi:10.1007/s11135-008–9190-y Hu, L., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation Modeling, 6, 1–55. doi:10.1080/10705519909540118 IBM. (2016). IBM SPSS statistics 24 documentation. Retrieved from http://www-01.ibm. com/support/docview.wss?uid=swg27047033 Jacoby, J., & Kyner, D. B. (1973). Brand loyalty vs. repeat purchasing behavior. Journal of Marketing Research, 10(1), 1–9. doi:10.2307/3149402 Kidwell, B., Hardesty, D. M., & Childers, T. L. (2007). Consumer emotional intelligence: Conceptualization, measurement, and the prediction of consumer decision making. Journal of Consumer Research, 35(1), 154–166. doi:10.1086/524417 Klein, J. G., Ettenson, R., & Krishnan, B. C. (2005). Extending the construct of consumer ethnocentrism: When foreign products are preferred. International Marketing Review, 23(3), 304–321. doi:10.1108/02651330610670460

86  Daniel P. Hinton and Tracey Platt

Kotler, P. (1997). Marketing management: Analysis, planning, implementation and control (9th ed.). Englewood Cliffs, NJ: Prentice Hall International. Kurz, R., & MacIver, R. (2008). Coaching with Saville consulting wave™. In J. Passmore (Ed.), Psychometrics in coaching: Using psychological and psychometric tools for development. London: Kogan Page. Lee, N., & Cadogan, J. (2016). Welcome to the desert of the real: Reality, realism, measurement, and C-OAR-SE. European Journal of Marketing, 50(11), 1959–1968. doi:10.1108/ EJM-10-2016-0549 Lee, N., & Hooley, G. (2005). The evolution of “classical mythology” within marketing measure development. European Journal of Marketing, 39(3/4), 365–385. doi:10.1108/ 03090560510581827 Lewis, C., & Anderson, P. (1998). PAPI technical manual. London: Cubiks. Lin, C. F. (2002). Segmenting customer brand preference: Demographic or psychographic. Journal of Product & Brand Management, 11(4), 249–268. doi:10.1108/10610420210435443 Lorenzo-Seva, U., & Ferrando, P. J. (2006). Factor: A computer program to fit the exploratory factor analysis model. Behavior Research Methods, 38(1), 88–91. doi:10.3758/ BF03192753 Maraz, A., Eisinger, A., Hende, Urbán, R., Paksi, B., Kun, B., Kökönyei, G., . . . Demetrovics, Z. (2015). Measuring compulsive buying behaviour: Psychometric validity of three different scales and prevalence in the general population and in shopping centres. Psychiatry Research, 225, 326–334. doi:10.1016/j.psychres.2014.11.080 McCrae, R. R., Costa, Jr, P. T., & Martin, T. A. (2005). The NEO—PI—3: A more readable revised NEO personality inventory. Journal of Personality Assessment, 84(3), 261–270. doi:10.1207/s15327752jpa8403_05 Muthén, L. K., & Muthén, B. O. (2017). Mplus user’s guide: Statistical analysis with latent variables: User’s guide (8th ed.). Los Angeles, CA: Muthén & Muthén. Nenkov, G.Y., Morrin, M.,Ward, A., Schwartz, B., & Hulland, J. (2009). Re-examination of maximization: Psychometric assessment and derivation of a short form of the maximization scale. ACR North American Advances, 36, 734–735. doi:10.1177/0013164413495237 Netemeyer, R. G., Durvasula, S., & Lichtenstein, D. R. (1991). A cross-national assessment of the reliability and validity of the CETSCALE. Journal of Marketing Research, 320–327. doi:10.2307/3172867 Novick, M. R. (1966). The axioms and principal results of classical test theory. Journal of Mathematical Psychology, 3(1), 1–18. doi:10.1002/j.2333-8504.1965.tb00132.x Nunnally, J. (1978). Psychometric theory (2nd ed.). New York, NY: McGraw-Hill. O’Connor, B. P. (2000). SPSS and SAS programs for determining the number of components using parallel analysis and Velicer’s MAP test. Behavior Research Methods, Instrumentation, and Computers, 32, 396–402. doi:10.3758/BF03200807 Parry, C. D., & McArdle, J. J. (1991). An applied comparison of methods for least-squares factor analysis of dichotomous variables. Applied Psychological Measurement, 15(1), 35–46. doi:10.1177/014662169101500105 Price, J. L., & Mueller, C. W. (1986). Handbook of organizational measurement. Marshfield, MA: Pitman. The R Foundation. (n.d.). What is R? Retrieved from www.r-project.org/about.html. Rigdon, E. E., Preacher, K. J., Lee, N., Howell, R. D., Franke, G. R., & Borsboom, D. (2011). Avoiding measurement dogma: A response to Rossiter. European Journal of Marketing, 45(11–12), 1589–1600. doi:10.1108/03090561111167306 Rossiter, J. R. (2002). The C-OAR-SE procedure for scale development in marketing. International Journal of Research in Marketing, 19(4), 305–335. doi:10.1016/ S0167-8116(02)00097-6

Measurement Theory and Psychological Scaling  87

Rossiter, J. R. (2011). Marketing measurement revolution: The C-OAR-SE method and why it must replace psychometrics. European Journal of Marketing, 45(11/12), 1561– 1588. doi:10.1108/03090561111167298 Ruch, W., Hofmann, J., & Platt, T. (2015). Individual differences in gelotophobia and responses to laughter-eliciting emotions. Personality and Individual Differences, 72, 117– 121. doi:10.1016/j.paid.2014.08.034 Schmitt, B. (1999). Experiential marketing. Journal of Marketing Management, 15(1–3), 53–67. doi:10.1362/026725799784870496 Schwartz, B., Ward, A., Monterosso, J., Lyubomirsky, S., White, K., & Lehman, D. R. (2002). Maximizing versus satisficing: Happiness is a matter of choice. Journal of Personality and Social Psychology, 83(5), 1178–1197. doi: 10.1037//0022-3514.83.5.1178 Schweizer, K., & DiStefano, C. (2016). Principles and methods of test construction. Oxford: Hogrefe. Smith, W. R. (1956). Product differentiation and market segmentation as alternative marketing strategies. Journal of Marketing, 21(1), 3–8. Söderlund, M., & Rosengren, S. (2008). Revisiting the smiling service worker and customer satisfaction. International Journal of Service Industry Management, 19(5), 552–574. doi:10.1108/09564230810903460 Timmerman, M. E., & Lorenzo-Seva, U. (2011). Dimensionality assessment of ordered polytomous items with parallel analysis. Psychological Methods, 16(2), 209–220. doi:10.1037/ a0023353 Tran, U. S., & Formann, A. K. (2009). Performance of parallel analysis in retrieving unidimensionality in the presence of binary data. Educational and Psychological Measurement, 69, 50–61. doi:10.1177/0013164408318761 van de Vijver, F. J., & Leung, K. (1997). Methods and data analysis for cross-cultural research (Vol. 1). Thousand Oaks, CA: Sage Publications Ltd. Wolf, E. J., Harrington, K. M., Clark, S. L., & Miller, M.W. (2013). Sample size requirements for structural equation models:An evaluation of power, bias, and solution propriety. Educational and Psychological Measurement, 73(6), 913–934. doi:10.1177/0013164413495237

4 IDENTIFY, INTERPRET, MONITOR, AND RESPOND TO QUANTITATIVE CONSUMER DATA ON SOCIAL MEDIA Dr. Amy Jauman, SMS

Audience research, the exploration of attitudes, interests, and behaviors of social media community members, has evolved significantly with the advent of social media. While everyone would certainly agree we now have access to more information, social media strategists would be quick to tell you this is the perfect example of more not necessarily being better.As business owners, marketers, and researchers enjoy the benefits of easy access to large amounts of data, they have to learn how to separate the useful information from the useless distractions. The steady stream of data available through social media also means updated information is often continuously available and is yet another consideration that needs to be tracked and managed. We’ll explore four steps that will guide any professional through a routine that will allow the collection, analysis, monitoring, and maximization of social media. We’ll also look at the risks associated with interpreting data, including challenges with specific platforms, the frequent absence of tone, and researcher influence.

Learning Objectives • • • • •

Understand the four steps of consumer data collection and analysis Review quantitative data options on social media sites Understand the ongoing nature of social media research and its four steps Identify the most common risks associated with interpreting data gathered through social media channels Identify and categorize various content types

Selecting Content Early in your research process, you’ll need to create a list of exactly what it is you’re searching for—and that answer may not be as obvious as you think.

Identify, Interpret, Monitor, and Respond

89

To learn about your audience, do you need to research your own brand? A specific product or hobby? An individual? In most cases, you’ll end up searching for a combination of items that will allow you to paint a full picture of what’s happening with your audience. But where should you start?

Brands When you apply the techniques outlined in this article, you may find yourself researching a specific brand (product or service associated with one company) by name. In this type of search, it’s important to remember to include the brand name, any brand nicknames, related product names (especially if they would be used instead of the brand name), and any alternate spellings of the brand name. It’s also recommended that you look at what the brand publishes as well as what the audience shares without prompting from the brand.

Key Terms It can feel overwhelming to start, but identifying a list of key terms associated with your research can help you quickly identify the conversations that are meaningful to your data collection process. This list requires constant monitoring to determine which key terms are most relevant, which are unrelated, and which terms emerge as meaningful and need to be added to the list.

Individuals and Groups In some cases, a specific person or group of people may be directly related to the success of your brand. Whether Martha Stewart is your spokesperson, or you know vegans are your primary audience, sometimes the most informative conversations you can observe are only tangentially related to your product. It is often

90  Dr. Amy Jauman, SMS

those conversations that reveal new opportunities or areas of interest your brand should emphasize.

Additional Content Considerations When researching content on social media, there are considerations that should be addressed in addition to the actual content. Context and delivery can heavily influence the actual value of content.

Sentiment Analysis There are several ways of assessing the intended and perceived meaning of content on social media. The practice is commonly referred to sentiment analysis, the process of systematically identifying a person’s feelings (typically positive, negative, or neutral) about a topic through text analysis, and can be completed through tools, manually, or a combination of the two.

Tools Providing a list of the top sentiment analysis tools in an article like this is challenging because new tools emerge, existing social media tools add sentiment analysis, and existing tools improve their offerings on a regular basis. Depending on the goals of the business and budget, there are a lot of great options that can support any marketing plan. Before choosing a tool, ask the following questions: 1. What features do I need in the sentiment analysis tool? 2. What features do I want in the sentiment analysis tool? 3. What do current users think of each tool? This can be determined from reviews or conversations with people in your professional network. 4. What is my budget? Press the answers to these questions against the tools currently available to uncover the sentiment analysis tool that is the best fit for you. Keep in mind, too, that the best fit may be part of another social media tool; your first choice may not be a tool built exclusively for sentiment analysis.

Manual Completing a manual review of sentiment regarding your product or organization may feel overwhelming, and no matter what manual approach you use, it will take time. However, following a systematic approach to assessing feedback from your audience can make the process very manageable for someone with any level of marketing experience. Also, there is a tremendous amount of value

Identify, Interpret, Monitor, and Respond  91

in examining audience feedback to this degree. In addition to the general themes you’ll uncover, you’re also likely to identify specific details that can be applied in other areas of the business. 1. Compile data from your audience in a spreadsheet by pasting one piece of data per cell in the far-left column. If you’d like to focus on a specific product, this could be reviews or comments about that product on social media. If you’re assessing an entire organization, you may include content from multiple social media platforms. 2. Within each piece of data, highlight positive (green) and negative (red) sentiment. 3. For each piece of data, identify if you believe it is positive (1), neutral (2), or negative (3). In some cases, you may have red and green highlights.You will need to determine if the data is more positive or negative or if the contrast in the sentiment makes the content neutral. Place that ranking in the second column. 4. As you review your data, create a list of the keywords and how often they are mentioned. Consider combining synonyms when appropriate. For example, you may count 4 comments about a long hold time and 7 comments about a long wait time as 11 comments about long hold/wait times because they address the same concern. 5. Code the keywords as primarily positive (green), neutral (yellow), or negative (red). 6. Identify the following metrics: a. Average sentiment for the data collected. (Divide the sum of the second column by the number of pieces of data.) b. Most common positive, neutral, and negative keywords c. Once you have completed more than one sentiment analysis, you will also be able to track the differences in the metrics and how they change over time.

Blended Approach In some cases, you may have a tool that can collect data and perform some aspects of sentiment analysis, but it may not do everything you’re trying to accomplish. In those cases, consider a blended approach, using the tool you have available to its capacity and completing additional necessary steps manually. For example, you may be able to export reviews into a spreadsheet automatically and identify how often specific words are used, but there may not be any way to assess sentiment. The two steps that can be completed automatically are an easy and accurate way to start your analysis, and you can start the manual analysis by finishing the spreadsheet.

92  Dr. Amy Jauman, SMS

Interpretation No matter what approach you choose, there will be some level of interpretation required (see Figure 4.1). Because of this, there is a significant margin for error in sentiment analysis techniques.The most common errors include inconsistent identification of positive, neutral, and negative sentiment (e.g., “dinner was fine” identified as positive once and neutral in another instance) and the misinterpretation of tone.While both will remain a risk for even the most cautious researcher, there are steps that can be taken to mitigate errors. 1. Arrange for multiple researchers to review the data and determine sentiment independent of one another before discussing results 2. Have the same researcher complete the manual interpretation of the data in longitudinal studies 3. Ensure researchers understand the industry language, culture, references, and anything else that may influence content While you may not be able to meet all of the aforementioned criteria, an effort to meet as many as possible can reduce the risk of errors.

Context Sentiment analysis tools are improving, but the use of slang, intended tone, and the context (additional information that may influence your interpretation) of messaging can be misleading when assessing social media content and activity. For example, a sarcastic post may be perceived by an audience as humorous and may therefore garner a lot of positive engagement. A tool or even a person unfamiliar with the context of the post may interpret the message literally and inaccurately identify specific content as meaningful.

Non-text Contributions Your audience may share feedback in more ways than written content. Star ratings, emojis, GIFs and other non-text indicators are common ways to express sentiment. Those responses require separate analysis.

Ratings Particularly if you are assessing a product or service, you will likely have the opportunity to review some form of ranking, such as stars or a scale assessment. In those cases, the feedback may seem objective and easy to measure, but it’s critical to review any of the accompanying written feedback that is available. That feedback may provide you with the specific pain points that need to be addressed or the aspects of the product or service that are most important to the user.

Identify, Interpret, Monitor, and Respond  93

It’s also important to identify any ratings that aren’t relevant. For example, someone may accidentally leave a review on your product page and specifically state that they are talking about your competitor’s product. While their error may tell you it’s important to differentiate your product from your competitor, the actual feedback doesn’t apply.

Emoticons and Emojis On social media in particular, it’s common for some audiences to include some kind of emoji, an image that represents a thought or idea, in their comment or even to comment with only an emoji. In many cases, the sentiment is still clear, but it is important to recognize that the same emoji in different situations could have a different sentiment. A laughing emoji, for example, could be very positive when indicating laughter and happiness, or negative if used to mock or make fun of a post.

Images In some cases, you may request images (e.g., share a picture of you using our product) or users may simply include them to demonstrate their message (e.g., a picture of a broken product). Like emojis, sentiment is usually fairly easy to assess in these situations. The most challenging is often when an image is being used to support a question. It is up to the researcher to determine if they feel the image is meant to help find a resolution or an expression of frustration.

Video The ability for people to record and watch video on their mobile devices with ease is increasing and, therefore, you are likely to find ratings or comments shared as a video file. The same considerations and challenges for images can be applied to video. There is also the option to transcribe the audio from the video file and apply the text analysis approach discussed previously.

Interpret Findings

Identify Relevant Information

FIGURE 4.1 

Interpret Findings

The Research Process

Interpret Findings

Monitor Data

Monitor Data

Respond

94  Dr. Amy Jauman, SMS

Identify Relevant Information When using social media to research your audience, it’s critical to know what information you’re looking for—at least to start. As you pull information and identify themes, your list of relevant information is guaranteed to grow, and you’ll want to monitor your progress.

Finding Participants Whether you take an active role and look for people to respond or a passive role and attempt to uncover where people have already left information, finding participants can be the most important and the most difficult part of your data collection process. The challenge of both is rooted in the same fact: there is an overwhelming amount of activity in the world today, so getting anyone’s time or attention is difficult. There are several ways you can encourage people to share a single response, complete a survey, or be active in a group.

Responding to Your Questions and Surveys Timing is going to play a significant role in the success of your survey, so when sharing, keep in mind the time of day as well as the day of the week and any holidays that may positively or negatively influence survey participation. Even the most interested audience member is unlikely to engage if they don’t believe they have the time. Aside from timing, there are a few other things you can do to get the most possible responses.

Offer an Incentive Small businesses especially tend to shy away from offering incentives, concessions, gifts, or payments offered to increase motivation, because they think they can’t afford them, but that isn’t necessarily true. In fact, incentives don’t even have to be monetary. Depending on your audience, you may be able to entice people to participate by explaining how valuable the research will be or by offering to share the results. And if you do decide to offer an incentive, it doesn’t necessarily have to be something of a high dollar value. Consider your audience and offer something that would be valuable to them.

Tell Them How and Why Participation Is Quick and Easy One of the reasons it’s so important to keep surveys and even single questions short and easy to respond to is so that you can truthfully tell audience members the process won’t be difficult. If you’re explanation is persuasive enough, your

Identify, Interpret, Monitor, and Respond  95

participation will increase greatly because—even if they’re only marginally committed to the goal of the survey—most people will help if it’s little or no trouble for them.

Make the Results Intriguing (and Offer to Share them with Participants) Unless it puts a trade secret at risk, don’t be afraid to share why you’re asking a question and/or what the results are. Your audience may be interested enough in seeing what others think to play along. Consider a shoe company targeting working mothers who wear dress shoes to work every day. A simple description of why your company cares followed by a request to rank style, comfort, and cost may pique the interest of current and future buyers. After all, people are drawn to people who share a similar background or have shared interests. It’s only natural they’d want to see what their peers are saying. In some cases (like polls on Twitter (see Figure 4.2)), poll responses can be revealed instantly when someone participates. Other approaches may include using a survey platform and emailing the results to those interested, posting updates on your website, or doing a live reveal of the results at a pre-scheduled time on Facebook Live.

Share It in a Group Find a group interested in the topic and ask members to participate. The more closely related your topic is to the interests of the group and the more active or known you are in the group, the better your response rate will be.

FIGURE 4.2 

Sample Twitter Poll

96  Dr. Amy Jauman, SMS

Boost Surveys on Social Media Almost every social media has a feature that allows individual posts to be “boosted” to reach more people in your network and beyond. Brief and compelling language is especially important in this case because you’re more likely to be connecting with people who have never heard of you and are unlikely to be compelled to participate by anything other than the introduction to your survey. This approach boils down to one concept: the law of large numbers. If you can reach enough potential participants, you’re likely to get the number of responses you’re looking for.

Pay Qualified Participants If it’s in your budget and the results are worth an investment, one of the easiest options for finding survey participants is to hire an agency who can find qualified participants for you.

Assessing Activity in a Community If you’re interpreting data shared by members of a group, your approach is significantly different. However, there are still things you can do to gain valuable insights.

Start and Foster an Active Group One way to gather valuable insights is to create an environment for people to gather and share ideas. This may take some time to develop, but once it’s up and running, it can be a resource for you as you work toward multiple professional goals. Before beginning, it’s important to honestly consider if you have the time to and interest in running the group, though. A half-hearted effort won’t yield the results you’re looking for and will only serve to damage your reputation.

Join an Active Group and Contribute Regularly Unless your audience is very new, you can probably find and join a group. This is considerably less work than creating and maintaining a group, but there are a few limitations. As a member, you wouldn’t be able to access analytics (if any are available) or control who has access to the group.You may also find that your requests carry less weight with group members or are simply lost in everyone else’s activity. Though in most cases you would have the opportunity to encourage people to join the group, you wouldn’t be in control of adding or removing people.You also wouldn’t have the opportunity to write the group description and ground rules for participation.

Identify, Interpret, Monitor, and Respond  97

In some industries, being able to separate professionals from novices is important. With some hobbies, there are many branches and areas of interest to be defined. However, if you’re fortunate enough to be seeking connections without many specifications, becoming an active member of one or several groups can be a very enlightening and effective use of your time.

Observe a Group with Known, Active Members There is also a case to be made for observing online groups (Figure 4.3) without interacting with members. If your primary goal is to observe without influencing, you may be able to find an opportunity to observe an open group (no membership required) or a membership that you can keep without participating (either because participation isn’t required, the group isn’t heavily monitored, or the group manager gives you permission to observe without engaging). In those cases, you are limited as to what you will learn because you won’t have the ability to ask follow-up questions or steer the conversation. But not interfering and therefore not adding your own bias to the conversation could provide you with even better results than you had hoped.

Proactive Searching One way you can identify what your audience is talking about and where they’re spending their time is by asking them. Whether you have an idea and want to ask leading questions to confirm what you suspect or you’re in the dark and need open-ended questions to help you get started, there are several techniques that can be used to proactively collect responses from your audience.

FIGURE 4.3 

Sample Online Poll

98  Dr. Amy Jauman, SMS

Polls The simplest technique for gathering data from your audience—both in participation and in the quality of the data collected—is polling. Polling refers to asking a simple question with a finite number of possible answers that people can answer quickly and easily. In person, this may be asking a room full of people to raise their hand if they’ve ever subscribed to an online newspaper. On social media, this typically involves typing a question and possible answers into the platform and publishing it to your network. As polling is useful in many ways, most major platforms have added a polling option for their users. The benefits of polling include an increased rate in participation due to the minimal time commitment as well as the ease of collecting and reporting data. The disadvantage of polling is the risk of influencing participants by sharing answer options as well as the fact that limiting responses to the list provided may mean they are unable to clarify or share a more accurate response.

Surveys Similar to polls, surveys provide a uniform way to collect data from your audience, but you have a much wider range of choices for how you’ll collect the data and how you’ll analyze responses. As of the writing of this piece, the survey development company SurveyMonkey (Figure 4.4) offers users 14 different question types. While there are many different survey tools available, we’ll use SurveyMonkey as an example of what someone creating a survey may choose based on what they hoped to learn from their audience.

FIGURE 4.4 

SurveyMonkey Survey Question Types

Identify, Interpret, Monitor, and Respond  99

Before exploring the question options in detail, it should also be noted that how participants will be completing a survey is a significant influencing factor in survey creation. Are they likely to be on a mobile device, tablet, or PC? Is this audience comfortable typing or using a voice-to-text option on their device? If not, open-ended questions should be avoided, as they may lead to survey abandonment. Another important consideration is what you’ll be able to do with the data. If you can anticipate 1,000 responses, will open-ended questions yield data you will be able to interpret? Collecting data is only beneficial if you have the tools and time to interpret it. If it seems like these considerations indicate open-ended questions aren’t a good option, it should be noted that closed-ended questions have challenges, too. For example, you may not know what options you should offer survey participants. Or you may think you know what options you should offer, and you may be incorrect.

Multiple Choice Multiple choice questions are an excellent way to ask your audience to choose a favorite answer. Because they restrict the survey participant to one option, they can help you understand not only what your audience values but also what they value the most.

Checkboxes (Multiple Choice, Multiple Answers) Giving survey respondents the option to choose multiple answers is helpful when you’re seeking a more general interest level or when you’re trying to rule an option out. For example, you may list ten color options and learn that every color was chosen by 30%–50% of survey participants with the exception of orange and white, which was selected by less than 5% of participants. You don’t know what color is the most popular, but you know all but two colors are appealing to your audience members.

Star Rating Star feedback is ideal for gathering feedback about a product or service if you’re looking to demonstrate to future buyers that it’s a good buy or gather ideas about how it can be improved. For example, you may publish a survey on social media or use an existing star system on a platform (like Facebook’s star system on fan pages) to show a high star rating.You may also use it like polling as a simple way to collect quick thoughts. However you use the rating, provide a comment box. Often times people will share things like “I would have given it four stars if . . .” that can be some of the most valuable feedback you’ll receive simply because it’s so specific. It’s

100  Dr. Amy Jauman, SMS

also recommended to provide an even amount of star options (four stars instead of five, for example) if you’re concerned you may have too many people who are neutral, and you want a clearer picture of either a positive or a negative response.

File Upload Asking participants to upload a file is definitely one of the survey options that requires the most audience commitment, but it can also be the most fun for them and informative for you. Especially in an era where nearly everyone is carrying a camera phone, it’s reasonable to ask people to capture a screenshot or photograph of something related to your survey. This option typically also allows for files like Word documents as well, so longer written entries could be shared easily. Whether through a survey tool or directly through a platform, file uploads are an excellent way to get visual feedback from your audience and, if positioned correctly, can double as a marketing opportunity. For example, an author with an active Instagram feed may invite her readers to share pictures of the book in various places around the world with a note about what they learned while they were reading. This would bring positive attention to the book while simultaneously gathering data about what content was perceived as most valuable.

Single Textbox Providing respondents with a textbox for responses is the best way to gather information without influencing the results. Asking your audience to share what comes to their minds without any prompting will help you identify exactly what language they use (providing future search terms) as well as their unbiased first responses. The most common challenge is, as we have already mentioned, if you are surveying a large audience, you will need to have a plan to read and interpret the data. In many cases, following a single textbox question with a multiple choice or dropdown menu can be an effective way to prove what you think you know. For example, if you wanted to know the main reason people eat in restaurants, you may begin with a single textbox.You may then follow that question with a similar multiple-choice question listing the most common reasons you are aware of. The participant may duplicate his or her answer, or he or she may include other motivators (especially if you provide them with the option to choose multiple options).

Comment Box The same benefits, challenges, and even opportunities to follow-up with a similar question apply to using a comment box as a single textbox. The difference with

Identify, Interpret, Monitor, and Respond  101

a comment box is that you are signaling to the participant that you’d like to hear as much from them as they are willing to share. This is particularly useful if you’re asking for examples or detailed ideas. In addition to how you’ll interpret the data, this approach has two other important considerations. The first is whether or not your audience will be on a device that will allow them to comfortably reply with a lengthy response. If you’re surveying millennials likely to be completing their surveys on their phones, they are unlikely to have any problem sharing a long answer. If your audience belongs to an older generation and you can reasonably assume they won’t be responding from a computer, they may not be patient enough to leave a long response. On a related note, the second consideration is how much time your audience is willing to contribute to your survey. Whether it’s 60 seconds or 60 minutes, the first decision they make when they see your survey is whether or not they want to take the time to complete it. Especially with surveys that contain a lot of questions with comment boxes, you may risk losing participants with the perception that the time required is more than they have or are willing to give.

Matrix of Dropdown Menus Designing a survey question with multiple considerations that have multiple possible answers may initially feel a little overwhelming. But if your goal is to get feedback from your audience on a big picture topic, putting time into the design is well worth the work involved in the planning process (Figure 4.5). It is especially important with this design to begin with the end in mind. In the example shown, we can see that one of the goals of the survey is clearly to fully understand how a family is spending their free time. While this is simplified, it

FIGURE 4.5 

SurveyMonkey Matrix of Dropdown Menus

102  Dr. Amy Jauman, SMS

demonstrates that the survey creator is interested in helping the participants think about their entire family all at once and how they spend their time. This type of question is the best way to gather data when there are multiple significant considerations.

Dropdown Selecting responses from a dropdown menu is a quick and easy way to respond to questions, but it’s only a viable option if you know all of the options your participants will need. Dropdown menus have the same benefits and considerations as multiple choice, though they are typically used if the list of options is longer.

Matrix/Rating Scale The example in Figure 4.6 is a sample matrix/rating scale question provided by SurveyMonkey. As you can see, this is an opportunity for you to ask your audience to rank an organization or specific product and compare it to comparable items. This can be especially useful information when added to a competitor analysis.

Ranking If you provide an opportunity for your audience to rank items, you essentially see what they value most and—if you had to—what you could do without.

FIGURE 4.6 

Sample SurveyMonkey Ranking Question

Identify, Interpret, Monitor, and Respond  103

There are typically two challenges that emerge when surveys use a ranking method. The first is that in some situations, you risk paralyzing your respondents because they genuinely can’t choose between their top choices. (People typically feel less anxiety over any kind of tie associated with their least favorite options.) The second is that there may be external considerations that prevent them from choosing. If, for example, you asked who in your household is most likely to take out the garbage in the morning, the participant may have two answers—it’s the oldest child’s responsibility during the school year, but the parents complete the task during holidays and the summer months. Typically, qualifiers in the question like “most often” or “usually” can help participants answer the question, but it’s important to consider the potential of this challenge when crafting the question.

Slider Questions with a slider response are great for collecting concrete data, as demonstrated in Figure 4.7, or helping people assign a numeric value to something like an opinion or feeling. Whether you’re asking for data like a score or having them choose a number that represents how important something is or how satisfied they are, benefits of a sliding scale include easy-use and a visual component that for many will make it easier to answer the question accurately.

Multiple Textboxes In some cases, you may want to highlight specific differences that may change the answer to the question you have posed. In the example (Figure 4.8) you can see that they are interested in what activities the survey participant engages in, but they also want to know frequency. With questions like this where you are gathering basic information and then drilling down for more details, you may find some overlap, but it rarely affects participants. For example, someone many say, “If I participate in something weekly, don’t I also participate in it monthly, quarterly, and annually?” While these kinds of details are true, and you run the risk of that

FIGURE 4.7 

Sample SurveyMonkey Slider Question

104  Dr. Amy Jauman, SMS

FIGURE 4.8 

Sample SurveyMonkey Multiple Textbox Option

kind of feedback for any of your survey questions, you’ll find that most people apply common sense and respond without hesitation.

Contact Information and Date/Time Though it shouldn’t be viewed lightly because the information gathered is important, contact information, date/time, and demographic information are often collected with surveys. But before asking them to share additional information, make sure it’s relevant to your survey. Do you want to follow up with participants? Is their marital status important to your survey? If not, don’t waste their time and increase your risk of survey abandonment by adding questions that aren’t necessary.

Writing Effective Questions When writing any type of question in any survey tool, it’s highly recommended to utilize the resources offered to assist in designing effective survey questions. Often survey tools will provide questions you can use or recommendations to enhance questions you’ve written. These ideas are often designed to get you the best results and, especially if you’re new to writing survey questions, should be considered carefully.

Experiential Questions While it may be depressing to admit, it’s important to remember that most people tend to lie—even if just a little bit and often without realizing it. We often remember facts and events in a light that is most flattering to us. We tend to recall evidence that helps us while forgetting damning facts that may hurt our case. And the lies that most often cause complications for people conducting surveys often involve what we value, our habits, and how we will spend our money.When we’re asked these questions, we often respond with how we’d like to behave, but not always how we’re actually most likely to behave.

Identify, Interpret, Monitor, and Respond  105

For example, if a person was asked what was more important to them—­ contributing to the world through the work they do or making a large amount of money—they may quickly answer that doing meaningful work is more important than a big paycheck. However, that same person may be offered a 20% pay increase the next week if they’re willing to take a less meaningful position in the organization, and they may accept it without hesitation. Were they intentionally misleading you when they answered your hypothetical question? That’s possible, but it’s unlikely. It’s more often the case that people respond as their ideal self when considering hypothetical questions and evaluate the situation very differently when it’s presented as a reality. Suddenly they realize that pay bump could mean a new car and moving into a better apartment, and the scales tip in the other direction. That meaningful job just isn’t as important anymore. The reason factors like this are important for you to be aware of is because no matter what question you’re asking, you probably want to know what a person will actually do, not what his or her aspirations are. But if they don’t even know the truth, how can you get them to share it? If you ask an experiential question, you ask people to share about their past experiences as a way of predicting their future behavior. For example, instead of asking how often someone goes out to eat, you ask how many times they went out to eat last week. By asking about an actual measurable experience, you’re learning about their behavior, not their best guess of future behavior. It isn’t always an option. But consider the following alternatives and how they might produce more accurate results. Question:

How often are you willing to spend more than $1,000 on a single item? Experiential In the last year, how many times did you spend more than $1,000 on a single item? Version: Q: What are the most important considerations when selecting child care? If you had to choose one factor that made you choose your current E: child care, what would it be? How many people do you anticipate growing or shrinking your Q: office staff by this year? Over the past five years, how many people has your office staff E: grown or shrunk by each year (multiple answers required)?

Interviews Interviewing current or potential audience members allows you to gather specific information, adjust questions as needed to discover more information, and develop relationships within your community. There are several approaches to interviewing, but they all share a common challenge. The data collection and

106  Dr. Amy Jauman, SMS

analysis process have the potential to be more time consuming than other audience research methods. Various methods and recommended techniques are explored next, but it should be noted that this not an exhaustive list. These ­interview techniques are the ones most commonly used in conjunction with individuals conducting social media research.

Individual Interviews (Phone, Video Chat, In-Person) If you have access to audience members willing to be interviewed, you may consider a traditional in-person, phone, or video chat interview.This allows you to ask questions as planned as well as additional clarifying questions as needed.

Recorded Interviews You can invite your audience to complete recorded interviews through various social media platforms. You may, for example, post a single question and have people record a video response as a comment. This approach is valuable to you because you collect not only the words they choose in their response but also their tone and facial expressions.You may also find some people will share more information and without as much editing by video because it feels to them like a conversation (as opposed to a written response).

Email Interviews If you email interview questions, your respondent will know that—unlike a ­survey—their responses aren’t anonymous. (If that is how they will be reported, be sure to clarify that.) Unlike a phone interview, though, they will have the opportunity to think through and carefully craft their answers. This can work for or against you, depending on whether or not the interview participant perceives this approach as too time consuming or a welcome opportunity to carefully craft their message. While you can use the same interview questions, it’s a good idea to take a look at them with fresh eyes to make sure they are clear and prompt a detailed response. If, for example, you ask someone in a live interview, “Have you ever found yourself in a situation where . . .?” they would likely pick up on the conversational cue and answer in the affirmative by sharing a detailed story. In writing, the same question could be misinterpreted as a simple yes/no question. An easy solution is to add phrases like, “Please share an example that illustrates your response” to the end of each of the questions submitted.

Peer-to-Peer Interviews A common challenge with interviewing is time management. Having peers interview audience members or even each other can be an efficient way to gather data

Identify, Interpret, Monitor, and Respond  107

while simultaneously building strength within your community. If you explore this approach, the primary key to success is providing all of the participants with a clear interview guide, easy way to connect, and simple instructions for sharing their results. Whatever medium you select, preparation is similar. Much like questions you would craft for any other data collection, you want to determine what you want to learn from the interview process and ask only those questions critical to your research. The advantage you have with most of these interviewing techniques is the ability to ask additional questions for clarity or further information as needed. And, finally, by using the same interview questions, you’ll be able to more easily capture and report the findings in a consistent fashion. For example, you can take interview responses and assign them a number based on whether or not you felt they were positive, negative, or neutral, like you would do for a sentiment analysis.

Competitor Analysis A competitor analysis is no more complicated than it sounds. It begins with a systematic exploration of what your competition is doing—both successfully and not. After the appropriate data is collected and organized, the social media strategist has the opportunity to examine the findings and extract meaning.

Who: Choosing Your Competitors Professionals often misunderstand who their competitors are, overlooking those who are not offering the exact same product. A competitor is anyone your potential customer spends their money with instead of you. If you manage a nail salon, other nail salons are your competitors—but so hair salons, massage clinics, and any other place customers may go to pamper themselves. It’s critical to remember customer motives when examining their behaviors. The old adage that “no one wants to buy a vacuum cleaner; they want to buy a clean house” is still as applicable today as it was in the earliest days of advertising. It’s also important to acknowledge circumstances that are competitors in the sense that they prevent your customer from purchasing your product. For example, if road construction has decreased parking availability near your restaurant by 80%, that circumstance is your competitor. How do you go about finding your competitors? Ask yourself the following series of questions: 1. What is the primary value our product offers the customer? a. Who else offers a similar value? b. In what way(s) are we different from the competition (better and worse)? Answer this question for each competitor.

108  Dr. Amy Jauman, SMS

2. What is the secondary value(s) my product offers our customer? a. Who else offers a similar value? b. In what way(s) are we different from the competition (better and worse)? Answer this question for each competitor. 3. What keeps my customer from buying each of my products? Initially, don’t be afraid to explore first-, second-, and third-degree connections.You may uncover important information that had previously been off your radar. Eventually, so you’re working with a manageable amount of information, you’ll choose to focus on some competitors and not others, but in the beginning, it’s best to be open to as many possibilities as you can manage. The competitors who emerge when you complete this process should not lead to adversarial relationships or disappointment as you consider your success compared to theirs. This should be a very positive exercise that allows you to highlight your unique selling proposition while observing audience engagement techniques in your industry.

What: Identifying the Information Shared With Current and Potential Customers Once you have an understanding of who your competitors are, you need to determine what information they’re promoting heavily, posting routinely, and even just casually discussing.The next two questions will help you better envision your competitors’ digital presences by identifying where they market (which platforms, alternative sources, etc.) and how often they engage with their audience, but it’s critical to understand the actual messages first. 1. Beginning with a known competitor, group, or spokesperson associated with your industry, identify what information is most commonly shared. For example, you may review the Facebook page of your most successful competitor and identify that 80% of their posts in the last 30 days were related to current events. 2. Once you have identified what information is being shared, note what content receives the most engagement. For example, of the current event posts, you may note that each time information related to a new law that’s being discussed was shared, positive comments on the post increased.You may also note that humorous quotes were the posts shared most often and ­promotions/ discounts received the least amount of engagement of any kind. The process of identifying what is being shared by your competitors may seem like an objective data collection process, but once you begin to categorize content, you may find notable differences that make this kind of assessment difficult.

Identify, Interpret, Monitor, and Respond  109

What if you encounter a post that is both a current event and discount offer, and you had previously been tracking those types separately? If you encounter these types of challenges, don’t let them derail you. Make adjustments as needed and be consistent in your data collection.

Where: Finding Where Your Competitors Are—and Where They Aren’t As you learn more about your competitors, you’ll uncover their connections to groups, resources, individuals, and events that may be useful for you to explore as well. Often when we explore our marketing opportunities, we focus on what we need to create. In actuality, it’s usually more effective for you to join conversations. If you can identify a place where your audience has already gathered, why wouldn’t you join them? Asking the following questions can help you uncover those opportunities. 1. What locations or events are your competitors engaged with? Include references in discussions as well as posts promoting events and sponsorships. 2. How are they participating? Two significant pieces of information can be gleaned from this assessment. The first is a list of ideas of where you should consider spending your time. Just because your competitor is there, doesn’t mean you should be, but it’s at least worth exploring. If, after careful consideration, you decide the location would be a good use of your time, you can join the conversation. The second thing you can potentially learn is how to successfully engage in events by watching what does and doesn’t work for your competitors. Adjustments related to your own brand are important, but it’s typically much easier to build off of what others are doing to start.

When: Tracking Patterns of Competitor and Customer Activity The next consideration is when your competitors and your audience are active. This refers to every time measurement you can find that’s relevant. Time of day, day of the week, amount of time before an event, etc., should all be considered to identify any possible patterns. 1. When reviewing a competitor’s social media platform, how often do they post content and what day of the week/time of day do they choose? 2. When and how quickly do they respond to audience member comments (if applicable)? 3. Reviewing events associated with your audience, when do your competitors participate?

110  Dr. Amy Jauman, SMS

This is also an excellent time to observe any time-bound offers or content shared intentionally at a specific time. For example, did your competitor offer a deal that was only good for 24 hours one week before an event started?

Why: Understanding the Reasons for the Successes and Failures of What You Observe Each of the previous steps—whether it was successful or a failure—should be followed by the question, “Why?” The answer to that question will naturally lead to conclusions about how to more effectively reach consumers. Consider the following examples. We see a decrease in business every April.Why? There is an organization that isn’t a direct competitor that runs a promotion every April. I’ve now made the connection that, even though they offer a different product, their annual promotion is something our audience waits for, and it competes with our campaigns. My competitor is sharing a lot of posts about current events.Why? Upon consideration, it’s clear that there are a lot of significant decisions that affect our industry being made by the government right now. Our audience isn’t as focused on personal development as they usually are because these decisions are more important to them. Knowing that, we should replace some of our professional development content with current event information. My competitors engage daily with one specific event.Why? Upon consideration, I can see people interested in that event are interested in their product. I’ve also observed that many of the people they engage with express that they were not previously aware of their product but are interested in learning more. This event seems to not only be a great way to promote their product, it seems to be particularly beneficial for identifying new customers who have never heard of their brand or product before. My competitors always seem to post early in the morning, and posts in the afternoon and evening receive significantly less engagement.There is also noticeably less activity on the weekend.Why? After considering our market, it makes sense that we’d see more engagement in the morning. That is when our audience is on their way to work. Looking at patterns of activity, it seems they check their social media platforms as soon as they start their work day. With each of these examples, asking, “Why?”, again can provide even further clarification. Now that I know why that new competitor is someone I should be observing, I can ask why their offer is more appealing than ours. Now that I know why my competitor participates in an event, I can ask why my audience finds that important.

Identify, Interpret, Monitor, and Respond  111

Reactive Content In addition to exploring your competitors’ sites, you can also observe how they engage outside of their own platforms. If, for example, they engage as an organization, or if a well-known member of an organization chooses to respond on platforms outside of their own, you can see how they react to content shared by others.

Adding Relevant Information to Your Search As a result of your competitor analysis, you will likely uncover new keywords and ideas you want to continue to explore in the future. Add these terms and ideas to your current search. As you add content, review your existing list and remove anything that hasn’t proven valuable.

FIGURE 4.9 

Facebook post with various types of engagement

112  Dr. Amy Jauman, SMS

Avoiding Bias Bias is typically considered in the data interpretation phase of research, but it is also a risk during data collection. A systematic approach to data collection will help keep you focused, but the practice of adding and removing content as appropriate opens you up to the risk of making those decisions with bias. Also, like interpreting findings, working with others—particularly if they have significantly different experiences and opinions than you—can be an easy way to maintain balance in your research.

Interpret Findings Once you’ve collected your data, you’re faced with the challenge of interpreting your findings. When reviewing data you collected, you’ll be exploring objective data as well as subjective data. Both contribute to your research in important ways and both come with inherent benefits and risks. Before exploring the two in detail, consider the following.

Even Objective Measurements Can Be Misleading It’s natural to assume that objective data is clear-cut and easy to interpret—and sometimes it is. However, it’s likely that objective measurements you review have more of a story to tell than you may have originally realized. For example, you may look at your competitor’s Facebook page and see that they have 10,000 followers. That objective measurement would indicate they have been very successful reaching their community. But your thoughts about their success would vary if you knew they had been on social media for ten years or six months. Ten thousand followers is always a great number, but that additional information significantly alters your view of their success. Or what if you learned they paid to have their page boosted through Facebook and that is how they got 9,500 of their followers? While there’s nothing wrong with that approach, understanding that they didn’t grow their following organically would likely influence how effective you perceive their approach to be.

People Don’t Always Tell the Truth—Even if They Mean To We do have to be aware of trolls and people using other people’s platforms to push their own agendas, but an even greater risk is people who have good intentions and yet share bad data. A common example is a business that asks their audience if they’d participate in a three-day educational conference. The business would be trying to determine if there is enough interest to go through the time and expense of organizing the event. The audience may have an overwhelmingly positive response, leading the business to move ahead with the event. But when registration opens, they may find only a small percentage of people who expressed an interest register.Why would that happen? An emotional business owner assessing the situation may think

Identify, Interpret, Monitor, and Respond  113

it was an attempt to sabotage their success, but it’s more likely that—because they asked an aspirational question—the answers they got were what people wanted to believe about themselves, but never what they would be willing to do.

Peer Influence Can Have a Tremendous Effect on What People Share I asked the same question in two communities once and learned very quickly how powerful peer influence can be. I suspected the two communities would have similar answers because the individuals had similar backgrounds, I thought there would be a healthy debate within each group (i.e., 70% of group 1 and group 2 would have the same opinion, but the other 30% in group 1 and group 2 would disagree). Much to my surprise, there was nearly 100% agreement in both groups, but the groups had completely opposing views! When I reviewed the conversations that were had online, I noticed early on one contributor in each group expressed a strong opinion and the rest of the group followed—or at least no one with an opposing view chimed in. Which opinion really reflected how the majority felt? I had to admit, after interpreting the data, I couldn’t say for sure, because I could see the evidence of peer influence. While this is a dramatic example of the idea, peer influence can happen on a small scale as well and is something every researcher should monitor.

External Factors Can Skew Data To the extent possible, it’s important to watch for external factors that may be skewing your results. For example, if you ask people how much they are willing to invest in their professional development in a year, you would likely get very different answers in January and October. In January, most people see opportunity ahead of them and would report a higher number, whereas by October, they’ve probably experienced challenges throughout the year and are feeling a little short on funds. Their answer at the end of the year would probably be more conservative. While neither group is trying to mislead you, the external factor of time of year would have an influence on their answers.

Objective Measurement Nearly every social media platform offers users analytics to help them understand how their audience is engaging with the content they’re sharing. These analytics are generally considered objective measurements, but they are still subject to interpretation.

Platform Analytics The following examples of analytics demonstrate how a person can take objective measurement and begin to draw conclusions about the brand, content, and

114  Dr. Amy Jauman, SMS

FIGURE 4.10 

Twitter 28-day summary

even the audience as a whole.We’ll look at several examples from the standpoint of a platform owner/manager with access to detailed analytics as well as what we can learn from the engagement and analytics from others that are visible to the public. The 28-day summary offered by Twitter (Figure 4.10) is a high-level set of analytics to help account owners focus on areas of their digital presence that matter. This information is only available to the account owner. There are two primary ways the data is measured—raw numbers (e.g., the number of tweets or mentions) and that number compared to the previous month (e.g., an increase of 71.8%). The following is how each of the five categories provided by Twitter may be interpreted.

Tweets The number of tweets published is a very simple measurement with an easy interpretation. It illustrates how active you are on Twitter. Those conducting research through social media are able to use this number to determine if more or less activity would be beneficial. This is often determined by considering the return on investment (ROI) of the time spent on Twitter, comparing activity to similar accounts, and monitoring how many followers are gained (indicating you’re reaching more people) or lost (indicating you’re annoying people with too much activity).

Tweet Impressions Twitter defines impressions as the number of times a user is served a Tweet in a timeline or search results. Essentially, how often does someone have the opportunity to see your Tweets. A simple objective assessment of this measurement would be that a higher number is always better. But when researching through social media platforms, this measurement provides additional insight. An increase in impressions could be caused by being shared at an ideal time of day, the use of a popular hashtag, one or more influencers sharing the Tweet to their own network, or a number of other factors that would cause the information to appear in more feeds and therefore be seen by more people.

Identify, Interpret, Monitor, and Respond  115

Profile Visits The number of times your profile was visited is an informative measurement, but when paired with other considerations, it provides insight into a number of other areas. For example, if the number of people who visit your profile increases, it may be an indication that not only has your messaging been interesting, it’s been so interesting people were more likely to want to learn more about you.

Mentions If people are talking to or about you, you’ll receive a mention. This is a classic example of a metric that isn’t necessarily good or bad. For example, if people are complaining about you or mentioning you on Twitter because you aren’t responding on other platforms, it would be an indication of a negative metric. However, to paraphrase an old advertising adage, as long as they spell your name right, it often doesn’t matter what people say about you.

Followers The most frequently discussed metric on Twitter is followers—the number of people who have taken an active enough interest in the message you’re sharing to follow your account and receive your future messaging. This is the base measurement for how many people you have the potential to reach with a single message. The previous partial list of metrics for Twitter is just a starting point for analyzing your audience. Not only are their other metrics available through Twitter, there are completely different metrics for the other social media platforms (see Figure 4.11)! However, there is a great deal of similarity between metrics when you compare platforms. For example, followers on Twitter and Instagram are essentially the same as page likes on Facebook. Even when you tackle a new platform, you aren’t necessarily starting from scratch. There are some metrics that are more noteworthy on some platforms than others. For example, the average time someone views your video is important on any platform, but outside of YouTube, the majority of videos shared on social media are extremely short, making it much more likely that people view each video in its entirety. If you’re launching a campaign through YouTube that involves longer videos, you may be closely monitoring average view duration. You have to consider that some people will start your video and be interrupted (meaning their disengagement couldn’t have been helped) and others will play your video without actually watching it (giving you a false positive), but overall, a pattern of people disengaging at the same point can indicate that your video should be shorter or that the content consistently becomes less valuable to your viewers at the same point (see Figure 4.12).

FIGURE 4.11 

Twitter analytics page

FIGURE 4.12 

YouTube analytics page

Identify, Interpret, Monitor, and Respond  117

Comparative Studies When making decisions about what content is valuable, comparative studies can be very insightful. Especially in cases where your initial goals were based on an educated guess more than data, the experiences of similar organizations, individuals, or products can be the best place to start or a valuable tool to enhance the next steps of your plan. If a comparative study could be a good fit for your research, consider the following questions. What organization, individual, or product is similar but more successful than mine? When that organization, individual, or product was in the position I’m in now, what did they do that made them successful?

Subjective Measurement Even with all of the objective data available, there will still be instances where you’ll be interested in subjective measurement. This kind of data is typically more susceptible to errors because there is some interpretation in the process of gathering the data as well as interpreting it, but there are still many scenarios where this is an excellent addition to your data collection process. Whatever approach is taken for reviewing data, adding multiple independent reviewers is recommended. The benefits include the opportunity to explore and discuss multiple interpretations as well as the opportunity to uncover data missed by one person and found by another.

Monitor Data There are some instances where collecting data from social media could be a one-time event. You may launch an event campaign, or you may be exploring how people responded within 24 hours of a natural disaster. But most of the time, you’ll gain your greatest insights monitoring data over time.

Guidelines When monitoring social media content over time, there are guidelines that will help you observe the most meaningful information.

Consistent Metrics For your first data collection, decide what metrics will best help you understand your audience. Taking time to determine this early in your research project will pay off, though if you miss gathering a piece of data, there is a good chance you can still find it even at a later date. For example, if you’re looking at your activity

118  Dr. Amy Jauman, SMS

on Facebook, you may decide in month three of data collection that the gender of your followers is an important metric.You can go back to month one and month two and collect that data and add it to your research.

Consistent Measurement of Activity As you gather data, the raw measurement itself will be important, but in most cases, what will be the most telling will be the movement you record in the data. However you gather and report data, be consistent in tracking both the raw data and its relationship to previously collected data or industry norms. A few common examples of measurement of activity include the following: • • •

Percentage of increase or decrease since the previous month Amount (raw number of percentage) over or under an industry norm or competitor Comparison of activity of two related metrics (e.g., the number of active followers increased by 7%, and the percentage of engagement increased by 11%.)

Like the decision of what metrics to measure, choosing what activity to measure requires careful consideration. In most cases, historical data can be recovered, but your research will be best supported if, from the start, you’re monitoring the most meaningful data.

Influencers Pay close attention to any people or organizations that may be affecting your social media activity, either positively or negatively. This kind of information can tell you several important pieces of information. 1. New engagement from a social media influencer can explain an unexpected increase in activity. It may be short-lived, or you may have attracted a permanent and beneficial addition. Either way, identifying an influencer as the reason for your spike in activity is important. 2. A social media influencer’s interest may be fleeting, but if you notice their activity and reach out to them, you have the potential to engage them for a longer period of time. This could be good for your brand as more time with an influencer generally translates to more of their audience finding you. 3. The sudden absence of an influencer may cause an alarming dip in activity. Before deconstructing everything you’ve done recently, check to see if the decrease could possibly be related to one person or organization no longer actively engaging with your content.

Identify, Interpret, Monitor, and Respond  119

Social media influencers are important to identify because they can help you interpret your data accurately and they can provide insight into what your audience really values.

Notable Events Small or large, planned, or unplanned events have a significant impact on your social media. If these events are isolated (e.g., a natural disaster or a movie premiere), the effect on your social media will likely be isolated as well. However, if the event has a longer or even indefinite timeline (e.g., a new loyalty program from your competitor or a construction project that will make your store location harder or easier to get to), you may notice a change in activity on your social media over time. Monitoring notable events can help clarify what’s important to your audience (e.g., you may realize you need a new loyalty program based on how drawn your audience is to what your competitor is offering) as well as provide an accurate description of why you observed changes.

Respond The final step is to respond to your audience and to the data you’ve collected. Individual audience member responses allow you to make connections and build credibility as a brand. The strategic responses to data collected are your opportunity to make informed, real-time decisions to improve your digital presence.

Responding to Your Audience Time is often cited as the primary challenge associated with engaging with your audience online, but knowing what to say is a close second. Consider the following categories to help you quickly and easily keep the conversation moving.

FIGURE 4.13 

Four Steps for Identifying Relevant Information

120  Dr. Amy Jauman, SMS

Agree and Expand on the Content On any platform, it’s common to see businesses and individuals respond to a post with a short “Thanks!” or “We agree!” before moving on to the next comment. While this is usually better than no response at all, it is a missed opportunity to engage this follower in further conversation by adding more of your own thoughts. Thanks! That’s exactly why we added the color orange to our fall line. Can’t wait to share it with everyone! This requires more effort on the part of the business, but it is likely to pay off quickly. Consumers are so used to one-way conversations, a positive, individualized response stands out and is likely to improve your brand worth in their eyes and in the opinions of those reading your online conversation.

Disagree and Provide an Alternative or Solution It is often wise to avoid conflict on your social media sites at all cost. However, there are times when you may be able to respectfully disagree or provide a solution related to what they have shared.

Ask a Question It’s risky to assign work to one of your community members, so be careful how much you follow up on your request, but you may consider asking them for additional information—either expanding on their original contribution or telling you more on an indirectly related topic. These responses will lead you back to the third and fourth steps of the research process by providing you with additional findings you’ll need to interpret. There will be another party directly involved at this point, so you will need to consider that in your interpretation.

Strategic Changes The strategic changes you may decide to make when analyzing data include content type (e.g., videos vs. images), text, or even platform. One of the biggest benefits of social media research is agility and the ability to change what you are doing quickly and inexpensively as often as is needed.

Summary The recurring four steps of identifying relevant information (Figure 4.13), interpreting findings, monitoring data, and responding to an audience on social media

Identify, Interpret, Monitor, and Respond  121

as outlined in the Comprehensive Guide for Social Media Strategists (2017) is a systematic approach to understanding consumer data on social media. Without a system, analyzing content on social media can quickly become overwhelming. Even with a focus on quantitative data, there is also a risk of misinterpreting the data unless strict definitions are put into place.

Student Exercises EXERCISE #1 Collect and analyze objective data collected from a social media platform tool. 1. Choose a business whose data you can access for one social media platform. The platform must have a significant following (on most platforms, at least 500 connections) and recent, frequent activity. If the platform does not have enough activity, it will be difficult for you to review the activity. 2. Identify three objective measurements to review and briefly describe why you believe they are valuable. These options will vary by platform and their value should be determined by the individual business. For example, if you were analyzing a new organization primarily targeting women in their late 30s that is using a Facebook page, you’d be most interested in page likes (for a new product, growth is critical), post reach (again, since growth/connecting with new potential customers is so important), and which individual posts were most successful (this will tell you what your audience is responding to). 3. Summarize your findings by sharing your observations and recommendations for future activity. EXERCISE #2 Complete a manual sentiment analysis. 1. Choose a product, service, or organization for which you’d like to determine how users feel. 2. Compile data from reviews or comments on social media from your audience in a spreadsheet by pasting one piece of data per cell in the far-left column. 3. Within each piece of data, highlight positive (green) and negative (red) sentiment.

122  Dr. Amy Jauman, SMS

4. For each piece of data, identify if you believe it is positive (1), neutral (2), or negative (3). In some cases, you may have red and green highlights. You will need to determine if the data is more positive or negative, or if the contrast in the sentiment makes the content neutral. Place that ranking in the second column. 5. As you review your data, create a list of the keywords and how often they are mentioned. Consider combining synonyms when appropriate. For example, you may count 4 comments about a long hold time and 7 comments about a long wait time as 11 comments about long hold/wait times because they address the same concern. 6. Code the keywords as primarily positive (green), neutral (yellow), or negative (red). 7. Identify the following metrics. a. Average sentiment for the data collected. (Divide the sum of second column by the number of pieces of data.) b. Most common positive, neutral, and negative keywords c. Once you have completed more than one sentiment analysis, you will also be able to track the differences in the metrics and how they change over time. EXERCISE #3 Conduct a small data collection project from beginning (data collection) to end (analysis of findings). 1. Choose a group of 10–15 people from the social media platform you are most active on. Try to include a variety of demographics (male and female, varying ages, different employment, etc.). a. Create a document (electronic or a paper notebook) and classify each of the people you’ve chosen based on your ­ relationship with them. For example, you may know them as family, coworkers, neighbors, etc. If a person can be classified multiple ways, identify a primary classification and note additional classifications for future reference. b. Copy their ten most recent posts from one platform, and categorize them as one of the following: i. Personal—updates, check-ins, private life information ii. Professional—work-related promotion

Identify, Interpret, Monitor, and Respond  123

iii. Humorous iv. Educational—articles, information without an “ask” or personal/professional gain v. Inspirational—moral meme vi. Other c. For each of the ten posts, record the engagement by post and then by the category listed in the previous element. For example, two total Facebook posts in the inspirational category (one with 1 share/2 comments/20 likes and the other with 11 likes) would be recorded as Inspirational 1 share/2 comments/20 likes 11 likes 34 engagements *Note any additional information you find relevant. For example, you may be collecting data during an election or after a significant world event. If there is something that you feel may be influencing posts, include it as a note. 2. Analyze the data you have collected and answer the following questions. a. What patterns can you see in the initial classifications (a) or categories (b) you designated at the start of your research? b. How do you explain your findings? Consider external events and information you know about the research subjects in your results.

5 ALTERNATIVE RESEARCH METHODS Introducing Market Sensing—A Qualitative and Interpretive Perspective on Research David Longbottom and Alison Lawson

Learning Objectives In this chapter, we will present alternative research methods, which follow a qualitative approach. In part 1, we will explain the context and philosophy behind this type of research. In part 2, we will take you through the processes of strategy, planning, data collection, data analysis and data presentation.

The Nature of Qualitative Research We present that the purpose of research is to find out the truth—what and why things occur in order to make a meaningful contribution to our understanding and body of knowledge. In a perfect environment (where we can control all of the factors), we would wish our research to be carried out following scientific methods. This would ensure our research: • • • •

follows scientifically based procedures and protocols achieves objective results (free from bias) uses quantitative techniques to measure results produces valid, generalisable and statistically significant results.

Such methods are often associated with the research philosophy of the positivist. They follow a deductive approach and are quantitative in nature. The epistemology proposes that research can be scientifically tested and measured. The ontology proposes that there is a single reality that can be proven. These methods imply that researchers may have some measure of control over the research environment, that we can control and monitor inputs and outputs.

Alternative Research Methods  125

Such conditions may well apply for some academic disciplines (for example, where experimentation may be carried out and observed in controlled environments). Unfortunately, however, for many academic disciplines, such conditions do not apply. This chapter is written from the perspective of the social scientist, and in particular our own discipline: marketing. In our environments, we are concerned with human interactions in public spaces, markets, organisations. We are concerned with understanding human behaviour, what people do and why they do it. Controlled experiments are rarely possible, and quantitative surveys can at best produce shallow results or at worst be misleading. In these conditions, this presents researchers with difficult choices and the need to explore alternative research strategies and methods in order to gain more meaningful and deeper understanding. In this chapter, we will describe and examine alternative approaches to research which follow an interpretive research philosophy.

Using Interpretive Methods to Support a Positivist Research Philosophy Some researchers advocate a combining of different research methods in order to strengthen the overall research design (Cresswell and Plano Clark, 2017; Teddlie and Tashakkori, 2008). Such approaches are often described as mixed methods, where researchers use both quantitative and qualitative methods. Advocates argue that strengths and weaknesses of approaches are in this scenario counter balanced. Some refer to this as the principle of triangulation; results obtained from quantitative and qualitative research studies can be compared and contrasted to create this third dimension or triangular design. Some have presented that combining methods in this way may be aligned with a pragmatic research philosophy (Bryman and Bell, 2015; Saunders, 2015). A pragmatist will base the research design according to the needs and research questions. We present from the perspective of the ‘professional researcher’. We define a professional researcher as follows: ‘A professional researcher approaches the research project from the perspective of a problem to be solved and this involves: • Defining the aim • Establishing the objectives • Establishing indicative research questions (what is the problem to be solved).’ The professional researcher has no philosophical bias, but has a deep understanding of the differences, and the strengths and limitations associated with each

126  David Longbottom and Alison Lawson

perspective. In this scenario, the professional researcher selects a combination of specific research methods to formulate a research design that best addresses the problem to be solved.

Exploratory and Explanatory In Figure 5.1, we illustrate how interpretive methods may be combined with positivist methods. In exploratory research, we use qualitative methods before quantitative methods; for example, we may engage in-depth interviews in order to seek clarification and gain a better understanding in order to formulate a hypothesis for testing. In explanatory research, we use qualitative methods after quantitative methods—for example, to help understand at a deeper level a result derived from survey or experiment.

Part 1 Qualitative Research: Context and Philosophy Qualitative Research: Context There has been a major growth in qualitative inquiry within social sciences over the past two decades. Miles and Huberman (2014) suggest a growth of qualitativebased research papers being accepted for academic journal publication and find evidence of significant growth in academic textbooks.The reasons for this may be clear, as Cassell and Symon (1995, p. 2) point out that research in social sciences is mostly concerned with people, organisations and social interactions, and is not well disposed to positivist/scientific philosophy (for example, which are based on quantitative survey or experiment). This is also apparent in marketing research where increasingly marketers are striving to achieve a depth of understanding of consumers for important decisions in marketing strategy, brand development, internal marketing and marketing communications. In summary, the characteristics of qualitative research are as follows: • • • •

social context; people, behaviour, organisations and the environment depth, meaning, in social settings not conducive to experiments within controlled environments not conducive to testing of pre-determined hypothesis and survey-based methods.

From our own studies and observations, we find a significant growth of qualitative-based studies within the marketing subject area. Traditional methods include case study, depth interview, focus group and observation. A range of alternative methods are proposed in Longbottom and Lawson (2017) under the title

INTERPRETIVE PERSPECTIVE Depth and

Exploratory Research:

meaning:

•

What can we learn from literature and secondary data?

•

What gaps occur in our knowledge and understanding?

•

What do we need to find out in order to formulate a theory (for testing)?

building our understanding of the research question

POSITIVIST PERSPECTIVE Hypothesis development Hypothesis testing Generating statistically significant and generalisable results

INTERPRETIVE PERSPECTIVE Depth and meaning:

Explanatory Research:

building our

•

Why has this result occurred?

understanding of

•

Why do people think, behave, and react in this way?

•

What are the underlying critical factors?

the research outcomes FIGURE 5.1 

Exploratory and Explanatory Research Designs

128  David Longbottom and Alison Lawson

of Market Sensing. The objective of Market Sensing is to seek out qualitative methods that achieve greater depth and meaning from research, and this may not be achieved using some traditional methods of research (often described in mainstream research texts), such as surveys, focus groups and surface-level cases and interviews. In Market Sensing, we are seeking out alternative research methods which have the objective of going deeper, to explore areas which may be more difficult to uncover, but which provide marketers with very rich insights into consumers and markets.

A Word of Caution on Qualitative Research Methods The qualitative methods we will describe in this chapter have the objective of achieving rich and deep insights. In our context we use them to understand consumers and markets, but the methods are equally relevant for other studies in the social sciences. We will present and argue the strong case, importance and relevance of qualitative research, and that if conducted in a rigorous and robust way, the outcomes can make a meaningful contribution to knowledge in the field of social science studies.We should acknowledge, however, at the outset that there are some risks and challenges facing the researcher adopting this methodological approach. Some academics may challenge the credibility of the methodology, for example, on grounds of the validity of the approach and consequently the reliability of the findings. There are, for example, arguments that the methods are non-scientific, open to personal opinions, biases and subjective interpretations. These are all challenges that the qualitative researcher will have to face (and we will examine the implications within this chapter, suggesting ways to add rigour and robustness into design which will significantly counter such issues). Not everything that counts can be counted Not everything that can be counted counts Albert Einstein (1879–1955)

Qualitative Research: Philosophy Research philosophy is about examining beliefs (our own and those of others) on how knowledge is developed (for example, what is valid in adding knowledge in the marketing/social sciences field of study). How do we know what we know and what will be regarded as adding acceptable knowledge within a particular field of study? A discussion on research philosophy usually commences with a consideration of two opposite research perspectives (or sometimes referred to as paradigms): positivist and interpretive. Qualitative research falls within the interpretive philosophy. Interpretive philosophy may also be described by some authors as phenomenology.

Alternative Research Methods  129

There are good chapters on this in Bryman and Bell (2015), and in Saunders et al. (2015). Bryman and Bell describe the two philosophies as contrasting extremes, visualised as a sort of Likert-type scale with positivist at one extreme and interpretive at the other. Saunders et al., illustrates these concepts in the form of a ‘Research Onion Model’, where choosing a research philosophy represents the first important stage in constructing a research design (before peeling back further layers of the onion to construct a design). The model clearly implies that although there are two extreme philosophies, there may also be positions in between (or that have elements of each philosophy), and it may be that individuals have a research perspective somewhere within the scale (rather than at the extremes). It is, for example, quite common for researchers these days to use a combination of research methods: a mixed methods approach. For the purposes of this text, we take the position which we will define as the professional researcher. In the role of professional researcher, the researcher takes no predetermined philosophical approach and subsequent research pathway based on the requirements of philosophical position (as implied by some models). Rather, the researcher determines the particular study and the defined research objectives. In this chapter, we are starting from a common position that our research objectives require depth and meaning in social contexts, and therefore a predominantly interpretive philosophy is taken. The Market Sensing methods we have chosen will seek to describe how this depth and meaning might be achieved. A good discussion and comparison of positivist and interpretive philosophies can be found in Chapter 1 of Cassell and Symon (1995). Note that these authors have also produced an edited series of textbooks covering a variety of methods in this field. Positivism, they describe, is based on the assumption that there is an objective truth (a reality) existing in the world which can be revealed through scientific methods where the focus is on measuring relationships between variables systematically and statistically. That quantification lies at the heart of scientific methods. The key concerns are that measurement is reliable, valid and generalisable. Studies frequently involve the determining of a research hypothesis which may be subsequently tested for validity. Methods follow strict scientific and statistical protocols. Whilst such approaches may be predominant in scientific fields (where laboratory conditions may be applied to control variables and the environment), we will argue that such conditions may not be applied meaningfully in social sciences (such as marketing) where participants and the environment are naturally occurring and interacting in a social setting. In contrast, the interpretive philosophy is largely concerned with words and meaning arising in social contexts. There is an assumption that there is no single objective truth or reality, rather that relationships are socially constructed

130  David Longbottom and Alison Lawson

and rely on the subjective interpretations of the actors. Or as Fryer (1991, p. 3) presents, Qualitative researchers are characteristically concerned in their research with attempting to accurately describe, decode and interpret the precise meanings to persons of phenomena occurring in their normal social contexts and are typically pre-occupied with complexity, authenticity, contextualisation, shared subjectivity of researcher and researched and minimisation of illusion. The Market Sensing methods we have chosen to present in this chapter align with the interpretive perspective. Aligned with decisions on where research fits and philosophical perspective are issues of epistemology and ontology which are discussed next.

Epistemology Bryman and Bell (2015) define epistemology as a theory of knowledge, used to describe a philosophical stance. Within each stance lie underpinning principles and values, and associated procedures for capturing what may be considered acceptable or new knowledge. So far, for example, we have identified two extreme and opposing research philosophies: positivist and interpretive.Table 5.1 illustrates the contrasting epistemological principles on which these are based.

TABLE 5.1  Comparing Research Philosophy: Positivist and Interpretive

Research Philosophy Positivist

Interpretive

Socially constructed For example: Results are interpreted from words, images and observations in a social setting Inductive Deductive For example: A hypothesis is developed For example:There is no predetermined hypothesis, results (from literature and secondary data) are interpreted from emerging and is tested using scientific principles themes Subjective Objective For example: Results are subject For example: Results can be reported to interpretation and may not scientifically, free from bias and maybe be generalised but rather are capable of generalisation (from a particular to the study sample to a wider population)

Epistemological Scientifically tested principles: For example: Results can be measured and compared using quantitative analysis and statistical techniques

Alternative Research Methods  131

Ontology Ontology is concerned with the nature of social entities and questions about what constitutes reality. For example, within the interpretive research philosophy, the epistemological principles on which it is based are that knowledge is socially constructed, inductive and subjective. From this perspective, an ontological interpretation would suggest a belief that there is no single reality or truth, but that reality is subjective and bound by the interpretations of the social actors (for example, the researcher and the participants). Table 5.2 continues the development contrasting the two philosophical extremes. We can see from the earlier table that designing a qualitative methodology will present some issues and challenges for the researcher. With quantitative techniques, there are clear rules and statistical protocols and procedures which can be followed to produce objective and generalisable findings. Qualitative research, however, has to tackle issues of ‘interpretation’ ‘subjectivity’ and ‘social construction’, which are concepts far more difficult to pin down with simple rules and procedures. As Miles and Huberman (2014) write, The most serious and central difficulty in the use of qualitative data is that methods of analysis are not well formulated. For quantitative data there are clear conventions the researcher can use. But the analyst faced with a bank of qualitative data has very few guidelines for protection against self delusion, let alone the presentation of unreliable or invalid conclusions to scientific or policy making audiences. How can we be sure that our findings are not in fact wrong? These issues lead us to the important areas that a qualitative research design must tackle—namely, validity, reliability and bias, which we discuss in Table 5.3. TABLE 5.2  Comparing Research Philosophy: Positivist and Interpretive (Epistemology

and Ontology) Research Philosophy

Positivist

Interpretive

Epistemological principles:

Scientifically tested

Socially constructed

Deductive Objective There is a single reality or truth which may be tested and proven

Inductive Subjective There is no single reality or truth, but rather understanding and meaning is subject to the interpretation of actors in different social contexts

Ontological assumptions:

132  David Longbottom and Alison Lawson TABLE 5.3 Validity, Reliability and Bias

Validity Definitions • Validity is concerned with the integrity of the data gathered and the conclusions developed from it (Bryman & Bell, 2012). • The extent to which data collection methods accurately record and measure what was intended (Saunders et al., 2012). Reliability Definitions • Reliability is concerned with the question of whether the results of a study are repeatable (Bryman & Bell, 2012). • The degree to which data collection methods and analysis will yield consistent findings, similar observations would be made and conclusions reached by other researchers repeating the process (Saunders et al., 2012). Bias Definition • Bias refers to the extent to which the researcher or researched may seek to influence the process of data collection, analysis and findings.

So questions that the qualitative researcher must be concerned with are as follows: • • •

Validity: How do we make and demonstrate that the results are credible and valid (to what extent do our findings present a true picture of the situation)? Reliability: How can we be sure that if the same research was carried out independently by different researchers similar results would be achieved? Bias: How do we identify and eliminate our own and others’ personal agendas, preferences and biases to find the truth?

According to Miles and Huberman (2014), the task for the qualitative researcher is to build in robustness (will the design stand up in different and difficult situations) and rigour (will the design demonstrate comprehensive attention to detail) in the research design. This will involve careful attention to process, planning, data collection, data analysis and data presentation. These issues will be considered within part 2 when we go on to look at the processes and procedures involved in planning, data collection, analysis and presentation. In quantitative-based studies, a key objective is often to design the study to precise statistical rules so that results from samples can be generalised to a wider population. In qualitative research, generalisation can not be claimed. Results are subject to interpretation. The real value of qualitative research lies not in generalisation but in particularisation. In other words, the richness and knowledge are derived from understanding a particular situation or case study in greater depth.

Alternative Research Methods  133

Research Approach So far, we have presented that qualitative research is aligned with an interpretive philosophy and that it follows epistemological principles such as social construction, inductivism and subjectivism. It is based on the ontological assumptions that there is no single truth or reality.The implications for the research process are now examined with a look at the processes for inductive and deductive approaches to research.

Inductive Table 5.4 illustrates the two contrasting approaches of deductive (sometimes described as theory then research) and inductive (research then theory). Both processes normally begin with a review of current theory. In the context of developing a thesis for an undergraduate or postgraduate independent study (dissertation), this is the review of literature and other data, with a view to

TABLE 5.4  Approaches to Research: Inductive and Deductive

Stage

Positivist: Deductive Process (Theory Then Research)

Interpretive: Inductive Process (Research Then Theory)

1.

Theory Develop theory from literature and secondary data Hypothesis Develop a hypothesis (single or multiple) which may be tested using statistical protocols Data collection Quantitative data is collected, for example, using surveys, questionnaires

Theory Develop understanding of critical factors from literature and secondary data Themes and factors Identify critical factors which may be used as a semi-structure to frame the research themes and research questions Data collection Qualitative data is collected, for example, using interview and/or observational techniques Analysis and findings Findings are presented by interpreting and summarising words from interview transcripts, images, observations, recordings Conclusions Conclusions may be developed for example identifying existing, new or emerging critical factors and themes Develop theory/not generalizable Theory may be developed from interpretation of results Results are not generalisable but are particular to the study

2.

3.

4.

Findings Findings are presented using numbers following statistical analysis protocols

5.

Accept or reject hypothesis Hypothesis may be accepted or rejected following statistical protocols (to a level of statistical significance) Generalise findings/theory Provided statistical protocols are followed, results may be generalised (from sample to wider population) to a level of statistical significance

6.

134  David Longbottom and Alison Lawson

determining current theory and identifying trends, and, importantly, any gaps in knowledge. Within the deductive approach, this assumes that research will flow from developing a hypothesis (single or multiple) concerning the association or relationship between different variables (for example, does increased spend on advertising associate with higher customer spend?). Often such studies will seek to establish cause-and-effect relationships between data variables. Statistical tests or controlled experiments follow, with a view to confirming or rejecting the hypothesis. Following statistical protocols enables the researcher to present conclusions which claim to be predictive, or generalisable to a wider population (from the results of the sample). Within the inductive approach, the process assumes that the outcomes from the theory review will present a focus not on statistical relationships between data but more on the nature and understanding of the subject. Themes and factors (often referred to as critical factors) may emerge that are identified as important, and these form the basis for the research. The emphasis is on depth and understanding in social contexts, not statistical measurement of relationships. Data collection often proceeds in the form of interview and observation techniques (structured or semi-structured around emerging themes and critical factors). The strategies associated with data collection are considered in the final section of this chapter.

Research Strategy Research strategy is about choice, to determine appropriate methods to answer the emerging research questions and to form these into a research design and develop research instruments. We have presented that in qualitative research, the research objectives and research questions emerge from a detailed review of literature and secondary data in the form of themes and critical factors, and an awareness of gaps in knowledge. Research questions in qualitative studies will be concerned with depth, meaning and understanding in social contexts, and will follow an inductive approach within an interpretive philosophy. Saunders et al. (2015) provides us with a useful list of possible strategies and methods, and a brief explanation of each. Commonly used qualitative strategies they identify are as follows: • case study • grounded theory • ethnography • action research. Within the context of an inductive approach, they also raise the possibility of utilising mixed methods, employing one or more qualitative techniques, or

Alternative Research Methods  135

including some quantitative methods—for example, surveys or experiments. ­Bryman (2015) suggests in this way triangulation of results may be achieved, where findings from one source or method may be cross checked and contrasted with those from another. Some authors have suggested that utilising a mixedmethod strategy can add to the robustness of the research design, with strengths and weaknesses being better balanced and complimented. Cassell and Symon (from 1995 onwards) present a very useful series of ‘Reader’ textbooks (where a collection of highly regarded specialists contribute chapters on their specific area of expertise) with a good range of qualitative strategies and more detailed explanation covering issues of data collection, analysis and presentation. They cover, for example, • • • • • • • • •

case studies in organisational contexts discourse analysis repertory grid action research qualitative interviewing twenty statements test participant observation tracer studies intervention techniques. Some of these methods are considered further in part 2.

Part 2 The Qualitative Research Process Strategy, Planning, Data Collection, Data Analysis and Presentation We start part 2 with an overview of the qualitative research process, and the stages and processes are illustrated in Table 5.5. We have compiled the table from a number of sources, the generic outline you will find in most research methods textbooks.

Stage 1: Strategies for Qualitative Research At the end of part 1, we identified some potential research strategies for qualitative research. This part begins with a brief overview of some of the most commonly employed strategies for studies in marketing and organisation contexts, with a brief consideration of characteristics and suggested further readings. Table 5.6 identifies some traditional research strategies for qualitative research.

TABLE 5.5 The Qualitative Research Process

Stage

Process

1. Strategy 2. Planning

Choice of method Sample selection Background data gathering Gaining access Design of research instruments Conducting research: interview and observation methods Methods and tools for analysis Data reduction; recording, transcription, coding and analysis Methods and examples

3. Data collection

4. Data analysis

5. Presentation

TABLE 5.6 Traditional Research Strategies for Qualitative Research

Case Studies Key features The focus is on selection of one, or a few, organisations, groups or individuals identified as having particular relevance and importance to the area of study. Advantages Intensive examination of a single or small group of cases with particular relevance to the study. Can incorporate history and background for context. Disadvantages Internal validity: bias can be difficult to overcome. External validity: to what extent can results be utilised outside of the specific case context? Access to particular organisations and individuals might be difficult. Identity of case might need to be restricted for reasons of confidentiality (and this may result in some loss of impact). Reading Hartley, in Cassell and Symon (1995), provides a good overview and example in one concise chapter. Major authors include Yin (2013), Silverman (2012), Glaser and Strauss (1967) and Eisenhardt (1989). Grounded Theory Definition An approach to the analysis of qualitative data that aims to generate new theory from the data. The method is often associated with case study research. The method is very prescriptive and systematic requiring the researcher to carefully observe and follow a series of steps to gather, analyse and present the data. Key features Sets out a series of steps that can be followed to add rigour and robustness into the research process. It is regarded as a landmark theory in the field of qualitative research.

Advantages Logical sequence of steps to be followed, constantly checking and cross-checking data with theory. Disadvantages Can be seen as a very prescriptive process. Has become subject to abuse and modification with researchers claiming grounded theory but often not closely following the protocols and procedures. Reading Glaser and Strauss (1967) and Strauss and Corbin (1998). For a more general brief overview, see Bryman (2012). Ethnography Definition The term ethnography refers to folk (ethno) description (graphy). The ethnographer’s method is to live among the people in the study and record their way of life (frequently using modern media such as digital audio/visual recording). Key features Immersion into the lives of the subjects. Frequently, designs are unstructured. Use of observation, field notes and modern media to record events. Advantages Rich insights into life and unfolding events as they occur. Data is recorded for later analysis. Disadvantages Immersion (over an extended period) can be difficult to achieve in practice. Time consuming. Difficult to prejudge whether anything of merit will emerge. Reading Hammersley and Atkinson (2007). Ethnography is used extensively within marketing and further theory and example studies can be found within marketing and marketing research texts. Action Research Definition Action research occurs where researcher and participants collaborate to solve problems and implement solutions. Examples might arise in organisational contexts in consultancy or change programmes. Key features The researcher collaborates with participants and so gains practical insights into the process and procedures and problem-solving activities. Advantages Hands-on experience of working within the environment and tackling real problems that arise. Disadvantages Can be time consuming. Can be difficult to gain access/sponsorship. Conflicts of interest might arise between researcher and sponsor/participants. Reading McNiff and Whitehead (2005). (Continued)

138  David Longbottom and Alison Lawson TABLE 5.6 (Continued)

Discourse analysis Definition Discourse analysis focuses attention on language in social settings, both verbal and written. Key features Focus shifts away from individuals and cases as the unit of analysis, to language and context. Language is seen as a variable means of communication and expression of feelings, and not a standardised transparent medium. Advantages Allows the researcher to pick up on informal messages and feelings within communications. The use of language in particular settings and cultures. To better understand emerging and developing cultures. To understand how language (use and interpretation) influences behaviours in a variety of settings. The method may be particularly relevant to analysing talk and text in modern environments using digital and social media. Disadvantages Involves collection and study of large amounts of data/transcripts. Analysis of large amounts of data is time consuming, and ambiguities may arise which require clarification. Presentation of findings is difficult to keep concise (whole extracts are often needed, plus narrative, to understand the context). Reading Marshall, in Cassell and Symon (1995), presents a good concise chapter and example. Potter and Wetherell (1987), Paltridge (2012), Jones (2012) and Bryman (2012). Benchmarking Definition Seeks to establish emerging critical factors and best practices using a process-based approach. Key features Focuses on understanding how systems and processes work in organisational contexts. Compares and contrasts methods to identify and develop critical factors and emerging best practices. Uses flowcharts to illustrate and analyse processes, activities and procedures. Used largely in organisational studies but can be adapted for use in consumer studies. Advantages Semi-structured method built around process. Uses interview, observation and some quantitative data (measures of performance for example). Can involve collaboration with participants in problem solving and creative elements. Disadvantages Issues concern access (particularly where data is competitive or sensitive). Time-consuming process. Reading Zairi (1999).

In our own field of marketing research, we have identified a range of methods for Market Sensing. In these methods, we have the objective of researching at a deeper level. Table 5.7 summarises some of the methods.

TABLE 5.7  Market Sensing Methods

Using Images and Emotional Scaling Definition Aims to gain depth and meaning in interviews using images, metaphorical expression and emotion. Key features Respondents are briefed to prepare images they associate with an event or experience prior to interview. During the interview, they will be asked to describe and explain the images, and a range of questioning strategies will be used for depth and meaning. The researcher will be particularly concerned with probing metaphorical expressions and emotions. In the textbook, we show how the method can be used for understanding consumer behaviour in order to develop marketing strategy and brand. Key references Hancock and Longbottom (2017), Zaltman (2008). Discourse Analysis Definition Analyses talk and text in order to establish themes and patterns in data. Key features Focuses on discourse (words and expression) in social environments to establish understanding of trends and behaviours in particular social settings. In the textbook, we show how the method can be used to interpret conversations, meetings, speeches and presentations and exchanges in social media. Key references Crane (2017), Potter and Wetherell (1987). Consumer Ethnography Definition Focuses on understanding behaviour, customs and cultures, through immersion into the consumer environment. Key features In consumer ethnography the researcher makes use of modern media (audio and video) to aid the process of immersion and manage intrusion. In the textbook, a number of examples are provided which show how modern products and services have been designed successfully from understanding customer lifestyles. Key references Churm (2017), Hammersley and Atkinson (2007). Social Media Networks Definition Seeks to analyse exchanges on social media. Sometimes described as Netnography, where the researcher joins, participates or observes selected groups. Key features Netnography can be used successfully to pick up on emerging trends (a process in marketing referred to as granulation). Organisations may be very keen to manage social media exchanges in order to cultivate positive opinion towards their brand and manage any negative perceptions (and complaints). (Continued)

140  David Longbottom and Alison Lawson TABLE 5.7 (Continued)

In the textbook, examples are provided of positive and negative exchanges between customer groups and organisations and critical factors for success are identified. Key references Hanlon (2017), Kozinets (2015). Narrative and Storytelling Definition The focus is on gaining insights by asking respondents to recall significant events or experiences and explain them by ‘telling their story’ or describing ‘stories they remember’, which have had a significant impact on them (invoking feelings of nostalgia and emotion). Key features Storytelling often reveals deeply held values, beliefs and associated behaviours. Marketers often use the method to understand feelings of nostalgia. Key references Lawson (2017), Chase (2005). Gamification Definition Uses a framework derived from an understanding of why certain games have been successful and sustainable (and some become ingrained into cultures and society). Key features Gaming and interplay with respondents can reveal rich insights—for example, how a person reacts or behaves in a range of different scenarios. In the textbook, examples are provided of major brands and how they have engaged customers to gain loyalty. Key references Longbottom and Banwait (2017), Chou (2015). Service Design Methodology Definition Used to analyse a process, from beginning to end, identifying the principle stages and activities involved. Key features Makes use of flowcharting and blueprinting procedures to break down and visualise processes stage by stage. Widely used in services to improve the customer experience. Key references Baranova (2017), Bitner et al. (2008)

There are of course many different strategies and methods that have been developed by researchers, and those noted earlier are just a snapshot of common strategies. Often methods or constructs have particular relevance in specific subject areas and will emerge as common strategies from the review of subject literature. Some other strategies are listed very briefly in Table 5.8.

TABLE 5.8  Further Research Strategies for Qualitative Research

Strategies

Used In

Characteristics

Repertory grid

Psychology, Organisation Behaviour, Marketing

Attempts to identify personal constructs and arrange into a grid/matrix for analysis Based on 20 areas commonly used by individuals to assess the self and values

Psychology, Organisation Behaviour, Human Resource Management, Marketing Participant observation Psychology, Organisation Behaviour, Human Resource Management, Marketing Tracer studies Organisation Behaviour, Decision Making, Operations Intervention techniques Psychology, Organisation Behaviour, Marketing

Twenty statements test

Similar to ethnography involves immersion in the field of study and involvement in the process Critical events or tags are used to trace events in processes Similar to action research researcher involved with the process and problem solving Survey based instrument for assessing performance in service organisations Fly-on-the-wall method placing the researcher in the role of the customer Recording of events as they are experienced over a prolonged period

SERVQUAL

Services Marketing

Mystery shopper

Services Marketing

Personal diary

Psychology, Organisation Behaviour, Human Resource Management, Marketing Marketing, Human Resource Questioning technique based Management on selecting a specific event or issue for close examination Marketing, Psychology, Questioning technique based Organisation Behaviour on tracing terminal values from observed behaviours Marketing, Multi-disciplines Technique involving group interview and observation where interaction important, often to generate or explore new ideas Marketing, Multi-disciplines Technique involving extrapolation or creation of new ideas

Critical incident

Means end chain

Focus group

Projective techniques

142  David Longbottom and Alison Lawson

Stage 2: Planning Sample Selection In quantitative research, there are very clear and specific methods, rules and protocols which can be followed. These present the researcher doing a quantitativebased study with a set of rules and protocols which can be followed and provide a good basis to support and defend the accuracy and reliability of the findings. However, as we have argued previously, such methods may not produce the depth and meaning that we are seeking in Market Sensing, and so we have to look at alternative qualitative-based methods either used separately or as part of a combined (or mixed methods) design. In qualitative research, careful consideration and justification is necessary for each individual study, to build a position which may later be defended if challenged on the accuracy of the findings on issues of reliability, validity and bias. Qualitative researchers are not concerned with rules for probability and non-probability, association and statistical significance, and the outcomes in qualitative research will not be generalisable (predictive of a population). In qualitative research, the results will be what Thomas (2004, p. 131) describes as particularisable—the results are specific to the context, case study or individual. Application to other contexts may be inferred by the researcher (where similar characteristics are evident), but this is subject to interpretation (the researcher makes inferences, or others interested in the research outcomes interpret relevance and meaning to their own situations). Sampling in qualitative research is therefore normally based on selection (not randomised), and this brings into discussion the merits of the selection, the selection process and the judgement of the researcher. In short, the researcher will need to justify the sample selection (who is included and why) and sample size (how many are included and why). This task is further complicated by the fact that we must be aware that in qualitative research, emphasis is on depth and meaning, and this can mean that interventions are necessarily time consuming and relatively costly. It is, for example, unlikely we can achieve sample sizes of similar proportions to that which may be achieved by survey. It also brings into consideration issues of timing—i.e. are we able to collect our data using crosssectional interventions (a single sample at a single point in time), or does the study require longitudinal interventions (multiple or repeated samples spread over a long time period). Bryman and Bell (2015) state that in seeking to justify the sample selection, researchers need to refer to the relevance and context of the study (for example, close attention to the objectives of the study). The outcomes from the literature review may be a good source for justification—for example, in case selection, reported examples in the literature are frequently identified examples of good (or bad) practice. Another source may be through expert interview (the opinions of those considered to be expert or close to the field of study).

Alternative Research Methods  143

In terms of sample size, an often-used principal is that of ‘saturation’. Here the researcher proceeds with a set of semi-structured themes or research questions and conducts interviews and observations until a point is reached where it may be argued that ‘no new information is being revealed, and similar findings are being repeated or confirmed’. A paper by Guest et al. (2006) considers this further, posing the question, ‘How many interviews are enough?’ The authors examine a number of qualitative studies and draw some general guidelines—for example, samples of 5–25 are normally adequate for interpretive studies, 20–30 for grounded theory studies. From our own experiences, we would say as a general rule of thumb that the following, Table 5.9 ‘How Many Depth Interviews Are Enough?’, might be used as a guide. However, these must be viewed as general guidelines only, and it will be for the researcher to present and justify sample selection and size in the context of the specific study. TABLE 5.9  How Many Depth Interviews Are Enough?

Undergraduate Independent Study: Four to six in-depth interviews. Studies at this level are designed so that the student can demonstrate competence in applying appropriate research methods. Some interesting findings may emerge, but these are not likely (or expected) to be conclusive (the main purpose is to demonstrate competence in carrying out the research process). Postgraduate Independent Study: 6–12 in-depth interviews. Studies are designed principally so that the student can demonstrate competence in applying appropriate research methods. Some interesting findings may emerge, and further themes may be identified which would be suitable for further study at PhD or commercial levels. PhD: 20–30 in-depth interviews, following the principles of saturation. Studies are designed to develop new theory in a defined and focused area (new knowledge emerges in a defined field of study which will inform academics and practitioners). Some students may wish to include some elements of quantitative data or measures to support key areas (thus taking a ‘mixed methods’ approach). This also achieves ‘triangulation’ of findings, where results from qualitative and quantitative analysis may be compared and contrasted (for similarities or differences). Commercial studies: 50+ in-depth interviews. This of course is very much a rule of thumb and depends on the size of organisation, scope of the project and available funding and resources. Large organisations will often require a sizeable database of data to be gathered and stored in a database (for example, using software packages such as NVIVO). The advantage of this data store may be in ‘holding stock’ for future studies and references. Note to table: These are rule of thumb guidelines only based on our own collective experiences in this area. Researchers must check the requirements of their own institution to ensure that they are consistent and meet the standards expected.

144  David Longbottom and Alison Lawson

A major concern for the researcher is to be aware of the risk of bias in the selection process and take steps to ensure that the sample can be justified as representative in the study context. Before entering the field, it is vital that the researcher is confident of achieving a defendable and representative sample. Sample selection is often a ‘hot topic’ in presenting qualitative studies. Gathering Background Data It is good practice, and a vital part of preparation before entering the field, to trace the history, background and context to the case, individual or unit to be researched. This saves unnecessary time in the field being taken up seeking information which is already published or generally available. The researcher should exhaust all sources before entering the field, for example: • • • • • •

literature review secondary data company documentation (example published accounts, website, etc.) expert interviews recent conferences, events, exhibitions, trade fairs industry and professional bodies. Good preparation is vital for the following:

• • • •

saving time in the field on unnecessary activity giving greater credibility and professionalism to the research understanding of the wider context of the study may be important to include within the final thesis to give the reader an overall context for the study.

Gaining Access This can be difficult and needs to be discussed very early within the context of the study and study objectives. Qualitative research is very time demanding, and there are clearly major issues in engaging participants. Aligned with availability and access the researcher also needs to consider for example how ethical arrangements might impact the study. Within the UKʼs National Health Service (NHS) (and other public sector organisations), for example, there are very precise guidelines that must be followed, and increasingly large corporations in the public and private sectors have well-developed procedures that must be followed.The process for getting approval for access may be long and difficult, and so must be identified early. Students at university or similar institutions will find detailed guidance which needs to be followed. Clarity on ethical processes, such as confidentiality and right to withdraw, can be very important

Alternative Research Methods  145

credentials that the researcher in the field can assert. This has the advantage of giving the participants in the study confidence and trust in the professional nature of the researcher and the study. Saunders et al. (2015) make some suggestions for negotiating appropriate access: • • • • • • • •

make early contact make initial contacts short and easy build up networks have contingency plans use attendance at conferences, events, exhibitions, etc. use other research groups and experts active in the field use professional and trade bodies arrange presentations and disseminations of your work.

From our own experiences we have found attendance at conferences and exhibitions to be particularly useful in building more personalised networks.

Stage Three: Data Collection Interviewing A good chapter entitled ‘The Qualitative Research Interview’ written by Nigel King can be found in Cassell and Symon (1995). According to King, the qualitative interview will generally have the following characteristics: • • •

a low degree of structure imposed by the researcher a preponderance of open style questions a focus on situations and actions in the world of the interviewee.

These characteristics give weight to the interview taking place within the interviewee’s natural environment, thus affording the researcher the additional benefits of understanding context, observing behaviours and the environment, and facilitating the use of examples and evidence gathering. King argues that even natural interruptions that occur within the field can be useful for the researcher to better understand the environment and context. King’s view of a low degree of structure is supported by Thomas (2004, pp. 162–170). He discusses the issues involved in preparing an interview plan or schedule. He suggests that it should not be in the form of a checklist of questions (more relevant to survey work) but designed to facilitate a conversation with purpose. He advocates the use of key themes (emerging as the critical factors for the research from the literature review), supported by prompts and probes (to guide and assist the researcher). Similar ideas for developing frameworks can be found

146  David Longbottom and Alison Lawson

in Miles and Huberman (1994, pp. 16–38). We present in Table 5.10 an example of an interview plan and interview questioning techniques which we have developed from our own collective experiences of interviewing over many years. The work also draws on interviewing methods described by Cooper and Branthwaite (1997), and Hancock and Longbottom (2017), which we go on to describe further in the next section of this chapter when we present questioning strategies. TABLE 5.10  Example Interview Plan

Interview Plan: Study of Knowledge Management (KM) Henry Pewter Consulting (HPC) Name: AB Position: Senior Consultant, HP Consulting Background: Joined HP last three years from MBA Cranfield, age 45, specialises in KM, IT for KM. Previous experience with Price Waterhouse on KM IT development. Has a team of four consultants. Looking to generate own HP specialist section. Client base covers mostly UK (60%), France (20%) and Germany (20%). Purpose: Establish the status of KM within HPC to include critical factors, implementation process, key performance measures and future directions. Date: 24 April 2010 (Easter visit) Time: 11am–5pm (including lunch and site visit, IT facility, HP museum) Place/Location: HPC HQ Present: Interviewers Dave Longbottom, PC and RE. Interviewed AB Transcript: Formal tape recording two hours for full transcription, plus notes as supplementary evidence.Various policy documents and internal training materials gathered reference Appendix 1–5 Themes 1, 2 of 34 Theme 1: Discuss the Status of KM Within HPC • Definition (1) • Start date (1) • Time (1) • Resources (2) • Responsibility (2) • Effectiveness (3) • Competencies (3/4) Theme 2: Discuss Approach to Implementation and Methods used • Describe the approach (2) • Evaluate the approach (3) • Describe methods (2) • Lessons learned (3) • Key performance measures and impact (3/4) Notes to table: Numbers in brackets for each theme denote ‘difficulty levels’, which we discuss in the next section. For each difficulty level, we try to anticipate potential issues which may arise—for example, •  trust and confidentiality •  expressions of personal opinion or evaluation •  sensitive information •  possibility of bias • evidence.

Alternative Research Methods  147

In this particular example case study, we were engaged with representatives from Henry Pewter and our research was investigating critical factors in building knowledge management systems. Henry Pewter had been referenced many times within the literature review we conducted and was considered by many authors to be at the forefront of knowledge management systems (a justification for including them within the primary data collection phase). We then consider what questioning strategies might be most useful in overcoming any issues that might arise and that we need to secure the depth of understanding that is the purpose of our study. Questioning strategies are considered in the following sections. We might also identify sub-themes or probes (probing questions) to help if responses seem incomplete or unclear. We also will normally indicate on the plan projected timings to help keep the interview within the time scale agreed (though researchers have to be alert to emerging issues which may extend times in certain areas or introduce new themes not anticipated, as is the nature of this inductive learning process). The important issue with the plan is to use it as a guide to process, to help anticipate events and to ensure some measure of consistency across a range of participants.Taking detailed notes during the interview is not recommended. Rather, we prefer to use audio or video to record events. This allows the researcher to concentrate fully on the conversation and take important readings of body language and emotion (which may be cues for further questions and probes). Questioning Strategies Aligned with the interview plan we suggest that some attention to preparation of question strategies is important. Here we are trying to anticipate difficulty levels and choose appropriate questioning strategies to elicit meaningful responses. An example is illustrated in Table 5.11, ‘Interview Difficulty Levels’. TABLE 5.11  Interview Difficulty Levels

Awareness

Response

Level 1 Aware, communicable, public Level 2 Aware, communicable, private or personal, not public Level 3 Aware, difficult to communicate, private, personal or confidential Level 4 Not aware, unable to communicate

Level 1 Spontaneous, reasoned, self-explanatory Level 2 Protective, reserved, may need clarification Level 3 Sensitive, difficult, may be partial or slow release or withheld. Level 4 Confused, incoherent, withdrawn, open to misrepresentation

Source: Adapted from the work of Cooper and Branthwaite (1997).

148  David Longbottom and Alison Lawson

In Figure 5.2, Hancock and Longbottom (2017) provide a modified and updated version of this model which they have called the Deep Value Mining Depth Gauge and which is particularly aligned to market sensing methods. The challenge for the qualitative researcher is to uncover deeper insights by careful design of research instruments (typically a research interview plan) and skill in the use of questioning strategies. At levels 3 and 4 deeper insights may be revealed. Level 4 is described as operating at ‘below the level of the conscious mind’, where insights may be uncovered from participants of which they were not themselves previously aware (for example, in examining meaning and emotion behind selected images). Table 5.12 illustrates the relationship between time and trust in gaining interview data. Developing trust over extended time periods may simply be impractical in many research situations which are time constrained, so the researcher has to consider strategies to capture the information needed within the limitations of the interview. It suggests that the researcher can prepare a questioning and interpretation strategy based on four levels, with each incremental level becoming progressively more difficult. Table 5.13 suggests some interviewing strategies that may help, presenting that questioning style, degree of probing and interviewer/interviewee speaking ratios may change depending on the question level. For example, at level 3, you would expect much greater use of prompts and probes, and the interviewer might seek to employ particular questioning techniques (for example: critical incident, means end chain). Prolonged use of silence may also increase as levels increase, a useful

Extraction Method Survey Questionairre

Grade of Data

1

Low

Sufface

2

Shallow

Depth Interview Interview Planning Question strategies

3

Deeper

DVM©

4

Deep

Increasing Depth and Value

Exploratory Interview Focus Group

Level

Surface Fair

Quality

Gemstones Hancock & Longbottom (2014)

FIGURE 5.2 

Deep Value Mining Depth Gauge (Hancock & Longbottom, 2017)

Alternative Research Methods  149 TABLE 5.12  Levels of Questioning: Trust and Time Relationships

Trust High ^^^ ^^^ Low

Level 4 Level 3 Level 2 Level 1 Short

>>>

>>>

Long

TIME

TABLE 5.13  Levels of Questioning: Style Implications

Level

Dominant question style

Probing

Interviewee: interviewer ratio

Level 1 Level 2

Open Open and closed

80: 20 50:50

Level 3

Open, closed and cross examining

Level 4

Open, closed, cross examining and observation

Low level, clarification seeking Moderate level, clarification seeking and explanation building High level, clarification seeking, explanation building, evidence gathering High level, clarification seeking, explanation building, evidence gathering, moving to cross checking and observation methods

20:80

Moving to observation

technique for the interviewer to attempt to engage the interviewee further by being suggestive that more information is desired (whist allowing thinking and reflective time). Silences should be noted within the interview transcripts to help the reader understand the context and tone of the conversation. Table 5.14 suggests some questioning strategies that might be used. In particular, these may be helpful in eliciting responses at the more difficult levels (levels 3/4). Skill and Practice Whilst we have shown techniques for interview planning and developing appropriate questioning strategies we must stress that interviewing is a very difficult skill to master. In our experience, it requires considerable practice and evaluation. We would always recommend a rehearsal, with self and peer review of performance.

150  David Longbottom and Alison Lawson TABLE 5.14  Questioning Strategies

Questioning Strategies

Characteristics

Critical incident

Questions focus on a specific event or incident to gain rich insights into an actual example Questions move from understanding actual behaviour through to terminal values (why fundamentally do you behave that way?) Introduces other evidence or opinion which may contradict what the respondent is saying Interviewer is well rehearsed on opposing point of view and takes on this role Seeks to check out interview answers by observation of actual behaviour Seeks to gain depth by asking the respondent to talk through the process stage by stage Similar to process where respondent is asked to talk through an example, real or simulated Use of prolonged and deliberate silences during an interview to tease out more comment Asks the respondent to look at the situation from another angle or point of view Asks the respondent to engage in creative thought or exploration of trends Introduces group opinion and dynamics into discussion Interviewer plays on building relationship with respondent and sympathising with situation or events Useful for positioning opinion when answers not clear-cut Based on theory that suggests ‘deep emotions’ are revealed in metaphorical expressions

Means end chain

Prisoner’s dilemma Devil’s advocate Chief executive diary Process Story telling Silence Perspective Projective Group (focus) Empathy

Rank/rating scales Metaphorical analysis

Data Analysis and Data Presentation In quantitative scientific enquiry, the researcher has statistical rules and protocols which may be followed in order to produce results which can be presented as accurate and meaningful. In qualitative research, the procedures that may be followed are not so clear and are subject to the interpretation and application by the researcher who will have to justify the design and process.This places more emphasis on the researcher to demonstrate robustness and rigour in the process chosen. The process of analysis for qualitative data requires what is often termed data reduction. Here the researcher is moving from a lengthy interview transcript,

Alternative Research Methods  151

to a more concise summary and analysis of key points. The challenge is to do this in a robust and rigorous way so that readers can see the raw data and how interpretations have been made. The raw data (often in the form of transcripts or images) represents the researcher’s evidence base. It is necessary, therefore, that the researcher takes care to make transcripts and images fully available to the reader and introduces some form of referencing system so that extracts can be easily located (in transcripts or in collages of images). We would identify the following key stages in the data reduction process: 1. 2. 3. 4.

recording and transcribing summarising data analysis and evaluation of data drawing conclusions, constructing solutions and further research. Figure 5.3 illustrates the main stages in the data reduction process.

3.1  Recording and Transcribing There are considerable advantages in using audio or video recording for interviews. Advantages relate to: • • • • •

time saving removes need for note taking allows the researcher to focus on the participant for important signals, language and discourse, body language, facial expression, emotions provides a permanent record if stored in a database audio or video may add value to presentations.

Disadvantages relate mostly to the risk that it may inhibit the interviewee. On balance, we favour recording in most situations, and through the use of unobtrusive modern technology and skilful interviewing, the drawbacks may be largely

Transcript Data Summary

Data reduction Data interpretation Data meaning FIGURE 5.3 

The Data Reduction Process

Data Analysis Data Evaluation

152  David Longbottom and Alison Lawson

overcome. Sometimes a mixture of formal recording and informal conversation may be employed if it is judged that there are concerns or sensitivities. Typically a one-hour interview will equate to around 6,000/8,000 words of transcribed data. It can be seen, therefore, that the issue of recording and transcription is a difficult and time consuming task, and presents the researcher with considerable issues when deciding on presentation (which we discuss in later sections).There are many good voice recognition software packages that are available commercially, which can drastically reduce the time needed to type up a full transcript. At the time of writing, the market leader within the UK is Dragon software, and we have found that with careful practice in dictation, an accuracy level of around 80% is possible (leaving only final grammar, spell check, and proofreading to complete the transcript). At the time of writing, we have not been successful in using any voice recognition satisfactorily for live recordings in the field (voice recognition between researcher and participants is just too different and complex for the software to handle). So at the moment, our practice is to dictate the audio file using speaker/earphone headsets. Transcripts are then annotated to show expressions of emotion, body language, etc., but can never replace the feel of the live recording. Future advances in technology which could accommodate multiple voice recognition would be a major breakthrough for qualitative researchers. There is no prescriptive or one best way of transcribing that we can advise the researcher to follow. Some useful hints and tips can be found in Bryman and Bell (2012) on interviewing in qualitative research. We present in Table 5.15 an example interview transcript illustrating some of the main issues. This is a continuation of the case study of Henry Pewter Consulting and knowledge management introduced earlier.

Notes to Transcript At this point in the interview, AB went into the project in some detail for over 20 minutes, describing the background and how events unfolded. The information was useful to us in the context of our study, and as we had good time available, we did not interrupt. Having a detailed interview plan (and using audio recording to capture the conversation) enabled the interviewee to ‘tick off ’ issues that were being addressed (naturally in conversation) by AB (wholly or partially) and without the need to be distracted from listening and observation (which can be the case if involved with extensive note taking). For example, looking back at the interview plan, we were able to tick that we had received in the initial exchange a definition (and explanation and documents to support), start date, time, resources, responsibilities. We had some responses on process (but needed to go back on this for much more detail) but no comment yet on effectiveness. Some elements were beginning to emerge as important—for example,

Alternative Research Methods  153 TABLE 5.15  Example Interview Transcript

Theme

Question/Response

Theme 1: Status of KM DL (interviewer) AB, in this first part, I would like to spend a little time trying to understand what KM is within your organisation. I have a number of questions, but perhaps we could start by you telling me a little bit about it, for instance, when did your project start, who is involved and so on. AB (participant) Well . . . . . . I suppose our first concerns started to come out after we separated from HP Corporate and became HP Consulting. AB hands me the ‘New Direction’ document April 1992, and we have a brief interlude on this. See Appendix 12.4. This was a big move for us but entirely consistent with our routes. I suppose you must know of our mission to keep the values of a small business within a major global organisation . . . DL Yes, please continue. AB Well, the problem as we saw it was splitting up the business gave us lots of opportunities: independence, competitive spirit, greater flexibility and so on . . . all vital I might add, and we had been wanting a long time . . . but we saw some risks too.There were some areas where we might be exposed, and we wanted to be able to bring in expertise from the wider group and did not want to lose access to this.We wanted to still retain elements of shared expertise, knowledge if you like.The way it worked was like this . . .

Ref.

T1.1

1.1

1.1.1

1.1.2

issues concerning competitiveness, relationships between units and individual roles, which we noted as later themes that we would need to explore further and address within our plan. This introductory question and conversation are at level 1: most of the information AB is aware of and familiar with, able to describe and explain, and his response is spontaneous (not reserved) and easy to interpret. Some of it confirms data that we already knew from background preparation—e.g. documents within the public domain (so the interviewer faces the decision of whether to interrupt the flow and move on, which may be important if there are time pressures). Allowing the interviewee to provide a full, if somewhat lengthy introduction, can, however, be important for setting the tone for the meeting, judging overall attitude to the subject and building of trust. It might also lead to the identification of important issues emerging that may be off-plan—for example, AB revealed early on that there were tensions between some departments (which we noted down as an issue for further investigation at a later stage in the research). Note that in the ‘Ref ’ column, we have started a theme and paragraph numbering system. This will enable us to reference the transcript within the

154  David Longbottom and Alison Lawson

main text of our final report (and thesis). This will be important in data reduction when we summarise and draw extracts from transcripts (still allowing the reader to locate the extract from the transcript if further context is required). Similarly, we will edit the audio tape recording so that themed sections can be referenced in our final report and located by a reader wanting to hear an extended version. Within the transcript, we identify the themes structuring the discussion as main headings, with prompts and questions as they occur. We provide a numbering system (right-hand column) so that when writing up summaries in the main thesis references can be used (transcripts and audio recordings are then normally included in an appendix, increasingly in digital format). In this case example, numbering is by paragraph, but some researchers use line by line (particularly if itemised or content analysis is used). This structure tends to assume that interviews flow logically and according to plan, which often is not the case. For this, we also suggest that commentaries are added, for example, to illustrate where the conversation has referred back to previous issues, or has raised new issues, or simply to note context (interruptions, observations, change in mood or behaviour, prolonged silences, etc.). There is a further issue of the coding of data which the researcher may wish to introduce. This can help to shorten the transcription process, aid recall and analysis, or highlight key themes and features that are emerging. It may be particularly useful if the researcher is planning to use formal content analysis procedures to analyse and present data. Content analysis can take extreme forms such as searching out and counting the number of times an issue is raised. Some qualitative researchers, however, would feel that such extreme approaches may be too simplistic and really just a means of imposing quantitative analysis not appropriate to an essentially qualitative study. The issue of coding, however, will also be important if computer aided qualitative data analysis (CAQDAS) is being employed. Within the UK, the market leader is NVivo, formerly NUD*IST (Non-numerical Unstructured Data Indexing Searching and Theorising).The software enables the researcher to design and construct a language-based database, which is useful for data storage, retrieval and analysis.This can be particularly useful for storage of large volumes and where the researcher wishes to create an archive (allowing for repeated analysis over long time periods). The drawback is that it can take some time to design and set up, and some researchers feel that it can lead to some detachment of the researcher from the data.There is a useful website available for NVivo, which includes sample software, an online tutorial and a short presentation of the main product features. In the example provided, we have presented transcripts with reference codes by paragraph. Some researchers prefer to present transcripts in line-by-line format. Here each line is numbered, allowing the researcher to reference very specific areas and keywords from the transcript (and this allows the reader to look up the points from the transcript if additional context is wanted).

Alternative Research Methods  155

S ummarising Data to be Presented in the Main Text from a Transcript Presenting qualitative data is difficult. Unlike when using numbers and statistics in quantitative research (where rules and conventions may be applied), there are no such rules and conventions that may be followed. So the qualitative researcher has to follow general guidance on good practice from a variety of sources, keeping in mind the study objectives, the research philosophy and strategy underpinning the study, and the need in all qualitative research to demonstrate rigour and robustness. As we have stated already, a good general discussion of presentation issues can be found in Bryman and Bell (2015), and Miles and Huberman (2014). For specific research strategies, you can also refer to specialist texts which cover the particular research strategy you have selected.

Case or Thematic Analysis A key decision to be taken is whether to present the analysis on a case-by-case basis, or by theme. In case-by-case analysis, the responses from each individual participant are summarised and presented with illustrative extracts from the case transcript.This may be the preferred option if the researcher feels that it is important to build individual case-by-case profiles.The alternative is to adopt a thematic approach where the analysis is presented by emerging key themes. In thematic analysis, the researcher summarises the key aspects of the theme and then draws on extracts from transcripts across the full range of participants to illustrate the views being expressed. An example of presenting a summary from the main transcript is presented in Table 5.16. In this example, we have chosen a thematic approach. Our justification for the thematic approach rested on the principle that our main objective was to gain an understanding of the knowledge management process (and that this might best be achieved by assessing views and themes across a range of diverse participants). Continuing our example of HPC, themes were presented following our semi-structured format but also noted emergent (inductive) themes. Here, for example, we discovered that an emerging theme concerned inter-departmental attitudes and arguments which were seen as important and a cause of problems in implementation of the knowledge management programmes. A short extract from our summary report is given in Table 5.16 to illustrate an example of a summary derived from transcript. Notes to Summary The summary presentation is structured by key themes and then issues and questions arising within themes. Note that some researchers will prefer to adopt a

156  David Longbottom and Alison Lawson TABLE 5.16  Example Summary Derived From Transcript

THEME: IT lead/dominance When we began discussing the roles of the main departments, issues began to emerge. It became apparent from a number of sources that the driving force for the new system was the IT department. Some managers were of the view that IT had become a dominant player at the corporate level and was able to secure substantial investment denied to others. There were concerns expressed about a dominant attitude from IT analysts, a lack of consultation on key implementation issues and an indifference to people and resource issues. • ‘We were being told that we must transfer all our client data on to the central system. All our personal job records, diaries and CV too.There was no real discussion, . . . just like do it . . . by next week.Well a lot of us were uneasy.We talked among ourselves. I guess if I’m honest we sort of decided to play a game. Most of us put data on . . . to be seen to be engaging . . . but in reality, most of the stuff was old, outdated and not much good. Most of the real stuff we kept back.’ MT Senior Consultant Ref: MT T5.2.2., p101 • ‘There was no real thought about planning or disruption.The timings were just unrealistic.We can’t just stop everything whilst a new system is going in.We have jobs to get done.There was a lot of resistance . . . and frankly many thought it was a waste of time anyway. I’m not sure they new [sic] what they were doing anyway.’ SKH Head of Operations Ref: SKH T5.2.5., p. 133 • ‘There was clearly a clash between DS [DS is head of HRM] and JOD [JOD is head of IT]. It got so bad that I don’t think they were speaking to each other. DS had a point I feel. It was all going ahead too quickly, and we could see trouble. Fact is DS just withdrew his support, but it didn’t seem to matter. I don’t think he has a big say on the Board any more . . . I think they just ignored him.’ RH Supervisor HRM Ref: RH T1.4.2., p24

less-structured approach and attempt to build the presentation from word repositories (abandoning the semi-structured themes in favour of emerging groups and characteristics). It can also be useful in adding rigour to the process to crosscheck interpretations, either by involving others in the process or going back to respondents to clarify meaning. The researcher might also seek out documents or other evidence to confirm an interpretation. Illustrative quotes are selected and used, and the author references the full transcript (allowing the reader to locate the extract within the full transcript or recording if further reading or context is wanted). A commentary is provided, which at this stage seeks to stress similarities and differences but avoids being over analytical, preferring to take the reader into the situation. The problem of data reduction and summarisation is particularly difficult for qualitative researchers presenting academic papers for conference or journal publication, where often editors will limit word counts, and this places a difficult

Alternative Research Methods  157

discipline on qualitative researchers. It is becoming more common to attach audio and video and to insert text extracts (and sometimes images and illustrations) within an appendix to the main report. Whilst data reduction through summarising is perhaps the most commonly applied presentation method for qualitative research, Miles and Huberman (2014) present that there are many different ways to present data and give examples of use of metrics, diagrams, structures and pattern matching (which can be used as alternatives or in conjunction with summarising methods). This introduces the opportunity for some element of creativity, which could be important in the context of the study and communicating the findings.

Analysis and Evaluation of Data For some researchers, the presenting of summaries with reference to transcripts may be the limit of the analysis, leaving the reader to make his or her own further interpretations and judge the relevance for their own contexts. In some disciplines however, it is becoming more normal to carry out further analysis and evaluation (business and marketing disciplines, for example). In these scenarios, the researcher will carry out further analysis and interpretation of data, leading to evaluation of data and in some instances development of new theory (or suggesting practical business solutions). There are good examples of different methods for further reducing and analysing and presenting data to be found in Miles and Huberman (2014). Again, you will find particular examples within specialist texts dealing with the various particular research strategies (for example, in case study, grounded theory, discourse analysis, etc.), and you may find these useful reference points. Next, we illustrate a matrix type of analysis based on cross tabulation of results. This works well in case or thematic analysis, multi-interview situations and for focus groups. Table 5.17 illustrates an example of cross-tab analysis (again extracted from our Henry Pewter example). Table 5.18 illustrates a snapshot of responses to key themes from respondents (in practice for this study, there were many more respondents and themes that are not shown here), enabling the data to be reduced down to single pages. Further analysis can take the form of cross-tab analysis (horizontal and vertical). Horizontal analysis enables patterns to be identified for each theme. For example, in this study, under the theme of ‘IT role’, there is an emerging consensus that the project is ‘IT dominated’. We also find that ‘Marketing’ are ‘not engaged’. Within the theme ‘Resources’ there are differences between departments, some being ‘under-resourced’, whist others are ‘well resourced’. Table 5.17 shows an example of horizontal/thematic analysis. Vertical analysis in this study may be important in identifying individual perspectives by case. For example, respondent ‘AB’ displays responses that are generally ‘positive and supportive’, and he is ‘optimistic’ about the future of the

158  David Longbottom and Alison Lawson TABLE 5.17  Example Further Summary Analysis Using Cross Tabular Methods

Themes

Respondents AB

DH

KM status Approach

High Participative

Resources

Good

Performance

Financial improvements Important Need more time Some input High involvement Need to engage

High Low Project structured Top down IT driven Good Under resourced Too early to say No evidence

Some improvements

Will continue Will be long term Part of plan Main driver Don’t see benefits

Needs longer Weak Needs more effort IT focus Not involved

Future Competencies Strategic role IT role Marketing role HRM role

Getting more involved

SKH

May drop Very weak Lip service IT focus Against the idea Need more input Ineffective

MT High Involving but IT focus Moderate/reviewing

Minor role

project. In contrast, ‘SKH’ is much more ‘negative’ and ‘pessimistic’. Table 5.19 shows an example of vertical/case analysis.

Drawing Conclusions, Constructing Solutions and Further Research Practice in qualitative research is again variable, and there are no strict rules that must be followed. Within business disciplines, it is normal practice to expect the researcher to develop the analysis further and suggest a summary of the main conclusions: what was discovered, what are the immediate implications for the study sample and what are the possible wider implications for others (subject to interpretation)? Some researchers may wish to further engage with respondents and other experts (before final writing up) to review and check the findings (within a focus group, presentation or discussion).This can add rigour to the overall process. In drawing conclusions, the researcher must be careful not to imply that the results are generalisable, but rather that the strength lies in the results being particularisable to the case and respondents in the study. Many students go on to suggest or create new models which address some of the issues identified and provide a basis for further discussion and a focus for future research. It is generally considered good practice at the end of a thesis to reflect on the outcomes and discuss further research areas.

DH High Project structured Good Too early to say Will continue Will be long term Part of plan Main driver ⇒ Don’t see benefits ⇒

Need more input

High Participative Good Financial improvements Important Need more time Some input High involvement ⇒ Need to engage ⇒

Getting more involved

KM status Approach Resources Performance Future Competencies Strategic role IT role ⇒ Marketing role ⇒

HRM role

Respondents

AB

Themes

TABLE 5.18  Example of Horizontal/Thematic Analysis

Ineffective

Low Top down IT driven Under resourced No evidence May drop Very weak Lip service IT focus ⇒ Against the idea ⇒

SKH

Minor role

High Involving but IT focus Moderate/reviewing Some improvements Needs longer Weak Needs more effort IT focus ⇒ Not involved ⇒

MT

Marketing not engaged

IT focus

Key outcome by theme

HRM role

Marketing role

IT role

Strategic role

Competencies

Future

Performance

Resources

Approach

⇑ High ⇑ Participative ⇑ Good ⇑ Financial improvements ⇑ Important ⇑ Need more time ⇑ Some input ⇑ High involvement ⇑ Need to engage ⇑ Getting more involved Need more input

Don’t see benefits

Main driver

Part of plan

Will be long term

Will continue

Too early to say

Good

Project structured

High

⇑ Low ⇑ Top down IT driven ⇑ Under resourced ⇑ No evidence ⇑ May drop ⇑ Very weak ⇑ Lip service ⇑ IT focus ⇑ Against the idea ⇑ Ineffective

AB

KM status

SKH

‘Positive supporter’

Key outcome by participant DH

‘Negative sceptic’

Respondents

Themes

TABLE 5.19  Example of   Vertical/Case Analysis

Minor role

Not involved

IT focus

Needs more effort

Weak

Needs longer

Some improvements

Moderate/reviewing

Involving but IT focus

High

MT

Alternative Research Methods  161

Chapter Summary In this chapter, we have taken an interpretive and qualitative perspective on research methods. We have presented that such an approach may well be used to complement and add depth to research studies, either as a stand-alone design or to complement and add depth of understanding in a mixed-methods design. Qualitative studies may particularly be used in the exploratory phase of research (for example, to help understand and define a hypothesis) or in the explanatory phase of research (for example, to gain deeper understanding and explanation of our research findings). We have presented that there are many methods which may be used, from the traditional case study, depth interview and focus group, to the many and varied specialised techniques which for our own purposes we have grouped as a collection under the title of Market Sensing Methods.These methods we have drawn from a range of academic disciplines and particularly for our own purposes have adapted them for a marketing context. In part one of the chapter, we discussed issues of context and philosophy. In part two of the chapter, we went through the research process presenting research strategy, planning, data collection, data analysis and data presentation.

Exercises for Students EXERCISE 1: DESIGNING AN INTERVIEW TO COLLECT QUALITATIVE DATA Design a list of ten open questions for use in a project investigating whether people lead a healthy lifestyle. Think of all the different subjects this might cover—food, drink, exercise, smoking, drug use, barriers to leading a healthy lifestyle, awareness of relevant issues, etc. There’s lots you could cover. Just choose one angle if you prefer and concentrate on that. Then work in pairs to interview each other and make some notes on what you find. Group the notes into themes. Collate the themes from the whole class to find which are the most common. Consider how this information would illuminate the results of a survey that collected only quantitative data. How might the qualitative data be helpful? EXERCISE 2: USING THE LADDERING TECHNIQUE TO ACHIEVE DEPTH Design an interview schedule to investigate why people choose to go to university. Use the laddering technique (in which each question delves

162  David Longbottom and Alison Lawson

deeper than the previous one) to get to the deeper motives that drive people to make decisions. Working in pairs, interview each other using your interview schedule. Beware that interviews seeking depth may sometimes be uncomfortable or uncover troubling memories for interviewees, so be prepared to stop your line of questioning if it is clear that your interviewee is becoming upset or uncomfortable. At the end of the interview, make a note of the final reason arrived at for your interviewee’s decision to go to university. Is it a much deeper response than you got in response to your first question? Is it deeper than you thought it would be? How could results of this sort of research be used by universities? What are the advantages and disadvantages of this approach? EXERCISE 3: USING CRITICAL INCIDENT TECHNIQUE Working in groups of six, use the Critical Incident Technique to examine customers’ experiences of a shop, brand, outlet or service that is well known to you both. Three must tell their stories in answer to the other three’s questions. One group of three must tell stories of good/satisfying experiences they have had, while the other group must tell stories of bad/unsatisfying experiences they have had. When you have finished hearing each other’s stories, discuss them to find the themes that emerge. How could this information be used to improve the customers’ experience? What is the advantage of this approach over a customer satisfaction survey? What are the disadvantages? EXERCISE 4: USING IMAGES IN RESEARCH This takes some preparation! All students in the group should select six to eight images that express how they feel about a particular brand, shop, outlet, product, service, concept or celebrity. None of the images must show the brand, etc., but must instead represent how the student feels about the brand and his/her connection to it. It’s best if this is done a week in advance, as it will take some thought and some time to find the images. In the session, work in groups of three, with one taking on the role of interviewer, one being interviewed and one observing. Before the interview, it would be useful for the interviewer to prepare an interview plan and identify appropriate questioning strategies using the Deep Value Depth Gauge as a guide. The observer should take notes during the interview, for use in discussion later. If there is time,

Alternative Research Methods  163

repeat the exercise until all students have had the opportunity to take on each role. A group or class discussion about use of this technique should be conducted at the end of the interviews. How successful were the interviews? Did the approach lead to a greater depth of insight? If the same subject was used for several interviews, did any common themes emerge? What are the advantages and disadvantages of this technique?

References Baranova, P. (2017). Understanding the customer journey through the prism of service design methodology. In D. Longbottom & A. Lawson (Eds.), Alternative market research methods: Market sensing. London: Routledge. Bitner, M. J., Ostrom, A. L., & Morgan, F. N. (2008, Spring). Service blueprinting: A practical technique for service innovation. Californian Management Review, 3(3), 66–93. Bryman, A. (2012). Social research methods. Oxford: Oxford University Press. Bryman, A., & Bell, E. (2012). Business research methods. Oxford: Oxford University Press. Bryman, A., & Bell, E. (2015). Business research methods. Oxford: Oxford University Press. Cassell, C., & Symon, G. (Eds.). (1995). Qualitative methods in organizational research. London: Sage Publications Ltd. Chase, S. E. (2005). Narrative inquiry: Multiple lenses, approaches, voices. In N. K. Denzin & Y. S. Lincoln (Eds.), The Sage book of qualitative research (3rd ed.). Thousand Oaks, CA: Sage Publications Ltd. Chou,Y. (2015). Actionable gamification: Beyond points badges and leaderboards. Retrieved from Yu-Kai Chou.com Churm, I. (2017). Consumer ethnography. In D. Longbottom & A. Lawson (Eds.), Alternative market research methods: Market sensing. London: Routledge. Cooper, P., & Branthwaite, A. (1997). Qualitative technology: New perspectives on measurement and meaning through qualitative research. Proceedings of the Market Research Society Conference. London. Crane, L. (2017). Discourse analysis: Using talk and text. In D. Longbottom & A. Lawson (Eds.), Alternative market research methods: Market sensing. London: Routledge. Creswell, J. W., & Plano Clark, V. L. (2017). Designing and conducting mixed methods research. Thousand Oaks, CA: Sage Publications Ltd. Eisenhardt, K. M. (1989). Building theories from case study research. Academy of Management Review, 14, 532–550. Fryer, D. (1991). Qualitative methods in occupational psychology: Reflections on why they are so useful but so little used. The Occupational Psychologist, 14, 3–6. Glaser, B., & Strauss, A. L. (1967). The discovery of grounded theory: Strategies for qualitative research. Chicago, IL: Aldine Publishing Company. Guest, G., Bunce, A., & Johnson, L. (2006). How many interviews are enough? An experiment with data saturation and variability. Field Methods, 18(1), 59–82. Hammersley, M., & Atkinson, P. (2007). Ethnography: Principles in practice. London: Sage Publications Ltd.

164  David Longbottom and Alison Lawson

Hancock, C., & Longbottom, D. (2017). Market sensing using images and emotions. In D. Longbottom & A. Lawson (Eds.), Alternative market research methods: Market sensing. London: Routledge. Hanlon, M. (2017). Social media networks: Rich on-line data sources. In D. Longbottom & A. Lawson (Eds.), Alternative market research methods: Market sensing. London: Routledge. Jones, R. H. (2012). Discourse analysis: A resource book for students. London: Routledge. King, N. (1995).The qualitative research interview. In C. Cassell & G. Symon (Eds.), Qualitative methods in organizational research. London: Sage Publications Ltd. Kozinets, R.V. (2015). Netnography: Redefined (2nd ed.). London: Sage Publications Ltd. Lawson, A. (2017). Using narrative and story telling in research. In D. Longbottom & A. Lawson (Eds.), Alternative market research methods: Market sensing. London: Routledge. Longbottom, D., & Banwait, K. (2017). Gamification: Using game technology for marketing research. In D. Longbottom & A. Lawson (Eds.), Alternative market research methods: Market sensing. London: Routledge. Longbottom, D., & A. Lawson (Eds.). (2017). Alternative market research methods: Market sensing. London: Routledge. Marshall, H. (1995). Discourse analysis in occupational context. In C. Cassell & G. Symon (Eds.), Qualitative methods in organizational research.Thousand Oaks, CA: Sage Publications Ltd. McNiff, J., & Whitehead, J. (2005). All you need to know about action research. London: Sage Publications Ltd. Miles, B. M., & Huberman, A. M. (1994). Qualitative data analysis. London: Sage Publications Ltd. Miles, B. M., & Huberman, A. M. (2014). Qualitative data analysis. London: Sage Publications Ltd. Paltridge, B. (2012). Discourse analysis: An introduction. London: Bloomsbury. Potter, J., & Wetherell, M. (1987). Discourse and social psychology. London: Sage Publications Ltd. Saunders, M., Lewis, P., & Thornhill, A. (2012). Research methods for business students. London, FT: Prentice Hall. Saunders, M., Lewis, P., & Thornhill, A. (2015). Research methods for business students. London, FT: Prentice Hall. Silverman, D. (2012). Doing qualitative research: A practical handbook. London: Sage Publications Ltd. Strauss A., & Corbin, J. (1998). Basics of qualitative Research – Techniques and procedures for developing grounded theory (2nd ed.). London: Sage Publications Ltd. Teddlie, C., & Tashakkori, A. (2008). Foundations of mixed methods research: Integrating quantitative and qualitative approaches in the social and behavioural sciences. Thousand Oaks, CA: Sage Publications Ltd. Thomas, A. B. (2004). Research skills for management studies. London: Routledge. Yin, R. K. (2013). Case study research: Design and methods. Beverley Hills, CA: Sage Publications Ltd. Zairi, M. (1999). Benchmarking for best practice. London: Butterworth Heinemann. Zaltman, G., & Zaltman, L. (2008). Marketing metaphoria:What deep metaphors reveal about the minds of consumers. Harvard: Harvard Business Press.

Further Reading Atkinson, P., & Hammersley, M. (1994). Ethnography and participant observation. In N. K. Denzin & Y. S. Lincoln (Eds.), Handbook of qualitative research. London: Sage ­Publications Ltd.

Alternative Research Methods  165

Bogdan, R., & Taylor, S. J. (1975). Introduction to qualitative research methods. New York, NY: Wiley. Bryman, A. (1988). Quantity and quality in social research. London: Unwin Hyman. Cassell, C., & Symon, G. (Eds.). (2009). Qualitative methods in organizational research. London: Sage Publications Ltd. Denzin, N. K. (1971). The logic of naturalistic enquiry. Social Forces Journal, 50, 166–182. Denzin, N. K., & Lincoln, Y. S. (Eds.). (1994). Handbook of qualitative research. Thousand Oaks, CA: Sage Publications Ltd. Denzin, N. K., & Lincoln, Y. S. (2005). The Sage handbook of qualitative research. Thousand Oaks, CA: Sage Publications Ltd. Denzin, N. K., & Lincoln, Y. (2012). The Sage handbook of qualitative research. London: Sage Publications Ltd. Hartley, J. (1995). Case studies in organisational research. In C. Cassell & G. Symon (Eds.), Qualitative methods in organizational research. London: Sage Publications Ltd. Holden, M.T., & Lynch, P. (2004). Choosing the appropriate methodology: Understanding research philosophy. The Marketing Review, 4, 397–409. Lincoln, Y. S., & Guba, E. G. (1985). Naturalistic enquiry. Beverley Hills, CA: Sage Publications Ltd. Lincoln,Y. S., & Guba, E. G. (2012). Naturalistic enquiry. London: Sage Publications Ltd. Marshall, C., & Rossman, G. B. (1989). Designing qualitative research. Newbury Park, CA: Sage Publications Ltd. Marshall, C., & Rossman, G. B. (2012). Designing qualitative research. London: Sage Publications Ltd. Patton, M. Q. (1980). Qualitative evaluation methods. Beverley Hills, CA: Sage Publications Ltd. Patton, M. Q. (1988). Paradigms and pragmatism. In D. Fetterman (Ed.), Qualitative approaches to evaluation in education. New York, NY: Praeger. Silverman, D. (2017). Doing qualitative research: A practical handbook. London: Sage Publications Ltd. Symon, G., & Cassell, K. (Eds.). (1998). Qualitative methods in organizational research. London: Sage Publications Ltd. Symon, G., & Cassell, K. (Eds.). (2012). Qualitative organizational research. London: Sage Publications Ltd.

6 BIG DATA Data Visualization and Quantitative Research Apps Vaidas Lukošius and Michael R. Hyman

Learning Objectives The learning objectives for this chapter are as follows: Define big data and describe assumptions about it. Describe processes for extracting insights from big data. Explain and illustrate current big data use, especially in marketing research. Appraise emerging research in big data analytics. Recognize dimensions for troubleshooting the transition from small data to big data. 6. Apply chapter content to conceptual exercises about big data. 1. 2. 3. 4. 5.

What Is Big Data? Historically, marketing researchers and analysts have relied on two types of data: primary data and secondary data. Primary data are collected to answer specific research questions like “What new features do our customers want added to our product?” Although primary data are customizable, they also are expensive to obtain.They are collected from a small subset (i.e., sample) of the relevant population; for example, new parents are the relevant population for home childproofing products. Generally, primary data are collected via surveys or experiments (e.g., testing how sales increase or decrease in response to price changes). Unlike primary data, which is collected for a specific and immediate research need, secondary data already exist. Secondary data are data gathered and recorded by someone else prior to and for a purpose other than answering a current research question. The four types of such data, which are often historical and

Big Data  167

previously assembled, are bibliographic, numeric, directory, and full text. Although not self-contained, bibliographic sources provide listings of possible sources for further query. Numeric sources contain numbers; for example, census data provides numbers that can be summarized in tables and graphs. A directory could be a listing of people by profession or restaurants in Nashville, Tennessee. Full-text databases exist for entities like newspapers and magazines. In contrast, big data is an “imprecise description of a rich and complicated set of characteristics, practices, techniques, ethical issues, and outcomes all associated with data” (Japec et al., 2015, p. 839). A more technical and complementary definition is “datasets that could not be perceived, acquired, managed, and processed by traditional IT and software/hardware tools within a tolerable time” (Chen, Mao, & Liu, 2014, p. 173). For companies, big data represents a cornucopia of digitalized content about consumers’ thoughts, emotions, behaviors, and reactions that are critical to the ongoing data-driven industrial revolution (Lohr, 2015). However, unlike primary or secondary data, it may not lend itself to traditional statistical methods due to several inherent biases. (See “Troubleshooting” section for more detail about these biases.) However, big data’s development has been strictly unidirectional: data creation continues to accelerate (Press, 2012). Academic research and business applications for developing the tools to collect, transform, and analyze big data are in their infancy; hence, the public has just begun to understand and experience its effects. For decades, businesses assumed and relied on marginal improvements in computational power and storage capacities. Analogous to Thomas Kuhn’s paradigm shift, in which normal science is punctuated by periods of revolutionary science, business thinking, and innovation tied to big data are shifting radically (Kuhn, 1996). Leading organizations are transforming their thinking on data, transitioning from treating data as an operational cost to be minimized to a mentality that nurtures data as a strategic asset that needs to be acquired, cleansed, transformed, enriched, and analyzed to yield actionable insights. (Schmarzo, 2013, p. 7) Big data ingestion, integration, and evaluation processes are complex and require sophisticated supporting IT infrastructure. Information systems and marketing researchers frequently cite four fundamental characteristics that define big data. The three Vs have emerged as a common descriptive framework: volume— how much, velocity—how quickly acquired, and variety—different types (Chen, Chiang, & Storey, 2012; Sagiroglu & Sinanc, 2013). The fourth V—­veracity/ value—is of particular interest to marketers. Briefly, the four Vs of big data are as follows. 1. Volume. Big data are characterized by its huge quantity. For example, consumer and marketing data may be generated from a broad range of sources,

168  Vaidas Lukošius and Michael R. Hyman

such as point-of-sale (PoS) systems, social media sites and apps, customer service centers, panels. Big data are now measured in zetabytes (i.e., 1 billion terabytes), with 40 zetabytes of such data forecasted by 2020. Organizations and users worldwide create over 2.5 exabytes (i.e., 1 exabyte = 1 million terabytes) of data daily (Erl, Khattak, & Buhler, 2016). 2. Variety.The types of content that comprise big data vary substantially. Fundamentally, data are either structured (e.g., consumer transaction entry or bank statement) or unstructured (e.g., product image or a social media post). Consumer point-of-sale data (financial), tweets (text and unstructured), device usage (unstructured sensor data), and price-tag images (visual) all capture aspects of consumption and brand engagement. Because big data are both structured and unstructured, one IT goal is to use technologies capable of interpreting unstructured data while merging it with structured data. 3. Velocity.Velocity is the rate at which data transfer occurs, which can be measured by assessing the speed of data transmission, processing, and delivery. Big data platforms are attuned to parallel and distributed computing mechanisms that can process data in real or near real time. 4. Veracity/Value. To be useful to business academicians and practitioners, big data must be reliable and valid. They must be consistent in time (e.g., across distribution networks), content (e.g., same measurement units), meaning (e.g., internally consistent), and unique identifiability (e.g., known users) (Buhl, Röglinger, Moser, & Heidemann, 2013). Data reliability can be represented as signal-to-noise ratio, where higher (lower) ratios imply more (less) veracity/value. Some data, such as financial transactions (structured), may be converted into information through standard algorithms (e.g., credit scoring); in contrast, other data, such as social media posts (unstructured), may require substantial interpretation prior to transformation into information.1 Uses and outcomes of big data should help marketing practitioners make more informed decisions.

Assumptions About Big Data There is an ongoing multi-disciplinary debate about the value of big data (Boyd & Crawford, 2012). Recognizing the limiting assumptions about big data is critical to its proper application in business decision making. Four key assumptions about big data (Huberty, 2015) are as follows. 1. N = All. Small samples of people—even those collected ‘scientifically’ (i.e., results are statistically applicable to the larger population that includes them)— are unreliable (i.e., high variance across repeated samples), non-representative (of the larger population), statistically underpowered (i.e., reasonable inferences are limited), and often non-normally distributed (i.e., problematic for parametric statistical analysis). Some researchers argue big data can overcome

Big Data  169

these limitations and provide ready tests of population parameters because it is relatively inexpensive, relatively representative, continually replenished, and easily replicated (Mayer-Schönberger & Cukier, 2014). However, one substantial caveat pertains to this assumption: big data excludes people who do not use either the Internet or technology that processes digital information. Currently, Internet penetration is 52%, with four billion users worldwide (MiniwatsMarketingGroup, 2018). Online participation, whether commercial or not, differs dramatically by geographic location, age, gender, race, and other socio-demographic factors (www.statista.com). 2. Today  = Tomorrow. Because it is mobile and omnipresent, digital life is far more dynamic than corporeal life. Consumers continually switch, try, and abandon their digital presence. Traces of such behavior may become an object of forensic investigation. Initially, Facebook limited membership to educational users; hence, registration required only an .edu email account. As Facebook expanded, it was opened to non-educational users. Professional network LinkedIn has drawn some of Facebook’s audience by targeting a tighter-knit professional community. 3. Online behavior = Offline behavior. Do consumers exhibit the same behavioral patterns online and offline? Teenagers maintain separate identities in each domain (Boyd & Marwick, 2011). Although they can be treated as dichotomous and separate, interactions between these domains are complex (Eklund, 2015). Marketers are starting to understand how consumer presence in both domains shapes buying behavior for specific products (Zhai, Cao, ­Mokhtarian, & Zhen, 2017). 4. Behavior of all today = Behavior of all tomorrow. This assumption aggregates the aforementioned three assumptions. Big data are an imperfect digitized reflection of people’s online and offline lives. Models describing consumer behavior are changing to account for the influence of technology and the Internet (You,Vadakkepatt, & Joshi, 2015). These assumptions temper insights gained from analyzing big data. Many marketers may believe all their consumers are available online and will behave identically whether online and offline. Not questioning or undermining the aforementioned assumptions could hamper the viability of big data analyzes for businesses. Yet, some data-intense companies such as Google may enjoy the luxury of undermining these assumptions by only caring about and modeling their online users’ behavior.

Processes for Gaining Insights from Big Data Rapidly increasing institutional reliance on big data has challenged how marketers understand and evaluate marketplaces. Broadly speaking, the purpose of marketing research has been to convert data into information and insights useful

170  Vaidas Lukošius and Michael R. Hyman

to decision makers (Hyman & Sierra, 2010). Although the conceptualization and goal of marketing research remains unaltered in big data contexts, ever-greater reliance on big data analytics requires a broader perspective for marketing research. Big data analytics are “technologies (e.g., database and data mining tools) and techniques (e.g., analytical methods) that a company can employ to analyze largescale, complex data for various applications intended to augment firm performance in various dimensions” (Kwon, Lee, & Shin, 2014, p. 387). Many recent books and research articles on big data analytics (Anderson & Semmelroth, 2015; Berman, 2013; Dasgupta, 2018; Foreman, 2013; Huang et al., 2015; Jackson, 2015) concur that the emergence of big data is gradually reshaping the research process. One widely accepted model (Figure 6.1) captures the marketing research process’ five key stages (Gandomi & Haider, 2015; Japkowicz & Stefanowski, 2016; Labrinidis & Jagadish, 2012). These stages may be grouped into data management and analytics. Data management is “a set of processes and supporting technologies designed to acquire and store data as well as to prepare and retrieve them for subsequent analysis”; in contrast, analytics is “a set of techniques used to analyze, visualize and produce intelligence from big data.”

Acquisition and Recording “[T]he core of data acquisition boils down to gathering data from distributed information sources with the aim of storing them in scalable, big data-capable data storage” (Lyko, Nitzschke, & Ngomo, 2016, p. 41). Data acquisition typically begins with hardware and software tools meant to generate and input data for big data systems related to social media networks, Internet crawling algorithms, user device sensors, financial transactions, et cetera. New data are recorded and stored in hyperscaled computing environments by defined data acquisition frameworks, such as a protocols or software tools.

Extraction and Cleaning Data extraction is the process of altering the format of distributed data for further processing. It largely occurs with unstructured and diversely formatted data that may be unreliable; for example, survey participants may respond falsely

Acquisition Recording

FIGURE 6.1 

Extraction Cleaning

Integration Aggregation Representation

Modeling Analysis

Process for Extracting Insights From Big Data

Interpretation

Big Data  171

or inaccurately, tracking devices such smartphones (with GPS app engaged) or wearable fitness trackers may malfunction, social media profiles may be stale, and data files may be corrupted by faulty storage media. Data cleaning (scrubbing), which is the process of identifying and fixing data errors via error detection and data repairing algorithms, is essential to ensuring data files meet specified quality standards (Tang, 2014).

Integration, Aggregation, and Representation Big data analysis often requires merging multiple heterogeneous datasets with diverse dimensionality into an integrated dataset meant to fit standardized schemes and analysis needs (Jagadish et al., 2014). Identical data, such as financial transactions, may be stored in multiple datasets owned by independent entities. For example, a credit card company and a retailer may record the same transactions. As a prelude to reliable analysis, a dataset derived from the card company’s and retailer’s data must be reconciled.

Modeling and Analysis Several types of analytics are used to find meaningful patterns in big data. Text analytics can convert human generated text, such as consumers’ brand evaluation tweets, into meaningful summaries. Audio analytics extract information from unstructured audio such as audio-recorded consumer complaints. Similarly, video content analysis extracts information from video, such as consumer-generated unboxing video. Image content analysis enables researchers to classify images, detect objects, and extract printed words. Social media and network analytics are used to explore the content and structure of social networks. By subjecting user-generated content residing in social media channels to audio, video, and other analytics, researchers can build models of people’s attitudes and sentiments. The interactions and relationships between consumers, organizations, and brands are best understood by applying two types of analytics: community detection, which is a method for spotting a social community by analyzing user attributes and topic/brand features, and social influence analyzes, which can reveal how people recognize peers and adjust behaviors. Big data modeling and analysis, which differs from traditional small-sample statistics, presents four challenges: (1) presence of noise, such as Internet trolls and Twitter bots flooding social media with unrelated posts; (2) multi-dimensionality, such as social networks containing millions of nodes and several attributes per node (e.g., gender identity, marital status, number of friends); (3) cross-correlations, such as multiple similar datasets; and (4) questionable reliability, such as inability to assess sensor precision level and calibration (e.g., GPS) (e.g., Cordeiro et al., 2011; L’Heureux, Grolinger, Elyamany, & Capretz, 2017; Yang, He, Qiu, & Ling, 2017).

172  Vaidas Lukošius and Michael R. Hyman

Interpretation Once analysis and modeling are complete, decision makers must be able to understand and interpret results. “The net result of interpretation is often the formulation of opinions that annotate the base data” ( Jagadish et al., 2014, p. 90). One powerful and fitting technique is data visualization, which is “a way to convey concepts in a universal manner, allowing your audience or target to quickly get your point” (Miller, 2017, p. 8).Visualization can explain data, add clarity, highlight obscure data patterns, and predict behavior. Analysis end users should uncover and verify all model assumptions and parameters to fully appreciate all model attributes and to suggest alternative models.

Big Data and Marketing Research “[L]arge-scale distributed data streams that provide numerical location and timedependent data of varying quality related to physical-world phenomena, and discovery scenarios where the data are location and time-dependent and varies in quality, requires a different set of solutions” (Barnaghi, Sheth, & Henson, 2013, p. 8). Advances in technology allow enterprises to use big data by (a) deploying algorithms to manage massive data streams, (b) relying on cloud services to store big data, (c) moving those streams with ever-wider bandwidth, (d) allowing mobile access to increase data supply and demand, and (e) encouraging new forms of data—such as text, speech, and video—meant to create richer and more dynamic content (Betser & Belanger, 2013). Such advances permit peopleobjects-physical environment networks that can connect people (consumers and company representatives), objects, and physical environment. Hence, consumers are embedded in a vast digitally networked world to which they can connect in myriad ways (Verhoef et al., 2017). While companies generally accrue marketing-related benefits from the big data consumers typically generate, consumers generally benefit from the new experiences such data facilitate. To better understand this dynamic, consider Stacey, a consumer who test drives a new electric car with the intent of leasing such a vehicle eventually. Stacey immediately connects her smartphone via Bluetooth to stream her favorite radio station (cloud services). The friend accompanying her streams their drive on social media while posting selfies and tweeting (mobile access, everwider bandwidth, new form of data). At a car charging station, Stacey ‘checks in’ via Snapchat and comments about the facility’s services by posting an image (new form of data). From one of her phone apps, Stacey receives a video recommending a new flavored coffee available at the station (data stream). She pays for a cup of that coffee with her digital wallet (cloud service). To avoid a traffic jam on her return drive to the dealership, Stacey queries her phone’s voice-command-interacting

Big Data  173

mapping app for the quickest route (data stream, could services, band availability, mobile access, new forms of data). Upon returning the car, she tells the salesperson, “I’ll think about it and we’ll be in touch!” and ‘signs up’ for a weekly promotional e-newsletter via the dealership’s social media page.

Viewing Marketing Through Big Data Lenses2 To make better informed and faster decisions, marketers have increasingly relied on advanced analytics/data science relative to business intelligence. Data scientists differ from business intelligence specialists (Schmarzo, 2013). Data scientists build, test, and refine their analytic models with extensive and diverse data manipulated, transformed, analyzed, and reanalyzed in an analytic sandbox, which “enables high-performance computing using in-database processing—the analytics occur within the database itself ” (Dietrich, Heller, & Yang, 2015, p. 10). In contrast, business intelligence specialists base their research reports and dashboard inputs on highly structured data often collected from a small set of instruments (e.g., questionnaires, salesperson reports) and/or marginally generalizable sources (e.g., local consumer samples). Although big data does not render business intelligence specialists less useful, data scientists are becoming increasingly integral to the research process. Hence, “big data is the realm of computer science, not social science” (Ninja Metrics, 2014, p. 12). Marketing research, which is rooted in the scientific method, generally relies on the assumptions and uses of traditional statistical methods (Hunt, 2010). This theory-oriented approach answers market-related questions by creating empirical-mathematical models meant to approximate consumers’ behaviors. However, the data modeling culture that grounds current marketing research practice is yielding to the algorithm modeling culture (roughly 2% of statisticians belong to this camp) that grounds current IT practice (Breiman, 2001). A data modeling culture assumes stochastic data and uses traditional modeling approaches (e.g., regression, factor analysis) and validation techniques (e.g., goodness of fit, residuals, percent explained variance). In contrast, an algorithm modeling culture assumes highly complex data and uses advanced modeling (e.g., neural nets) approaches and validation techniques (e.g., predictive accuracy). Big data is more compatible with algorithm modeling, in which a set of x’s enters a model and a set of y’s emerges.The algorithm modeler’s goal is to create an algorithm f(x) such that for any subsequently entered x, f(x) is a good predictor of y. During the last century, marketing scholars developed various disciplinary frameworks. For example, to systematize marketing management thought and practice, McCarthy (1960) created the 4Ps pedagogical structure of product, price, promotion, and place/distribution. This marketing mix framework became the de facto standard for marketing practitioners and academicians. Subsequently, the fifth P of people was posited (Booms & Bitner, 1981).

174  Vaidas Lukošius and Michael R. Hyman

Assessing each P requires collecting related data—big or small—that subsequently are analyzed with appropriate statistical methods. Insights gained from such analyses may be applied to marketing practice. For example, companies may reduce their new product failure rate by interweaving traditional marketing methods with big data analytics (Xu, Frankwick, & Ramirez, 2016). However, their understanding of big data is quintesential to their success, as ‘small data’ tools cannot simply ‘be scaled up’ to tackle complex marketing problems (Sivarajah, Kamal, Irani, & Weerakkody, 2017). Table 6.1 provides a general framework for integrating big data analytics into marketing management. To ensure suitable sources and analysis methods for a P, marketing goals and data sources must be aligned. Resulting insights can then be applied to that P. More likely than not, marketers will consider multiple Ps in their strategic decisions. While big data is ideal for solving multi-dimensional marketing problems, there are significant pitfalls with their use. (See the “Troubleshooting” section for more detail.) Although academically sound, orthodox research approaches overlook the influence of big data on marketing and its research tools. In fact, there are three core-marketing perspectives—consumer behavior, consumer sentiment, and marketing strategy—for understanding big data usage (Sathi, 2013; Sheng, Amankwah-Amoah, & Wang, 2017; Wedel & Kannan, 2016). Each perspective, which is now discussed in turn, has one or more dimensions that relate to academic studies in and applications of big data (see Table 6.2).

TABLE 6.1  5Ps and Big Data Analytics (Fan, Lau, & Zhao, 2015)

Big data analytics

5Ps People

Data Demographics sources Social networks Customer review Clickstream Survey data Analysis Clustering method Classification

Insights

Segmentation Profiling

Product

Promotion

Price

Place

Characteristics Category Customer review Survey data Clustering Association Topic modeling Ontology Reputation

Promotional data Survey data

Transactional Locationbased social data networks Survey data Survey data

Regression Association Collaborative filtering Recommendation systems Promotional analysis

Regression Association

Regression Classification

Location-based Pricing advertising strategy Competitor Community dynamic analysis analysis

Big Data  175 TABLE 6.2  Marketing Perspective on Big Data (Sheng et al., 2017)

Marketing perspective

Dimension

Academic studies

Consumer behavior

User behavior Online community

Consumer sentiment

Online review and ratings

Customer engagement Online learning Online community detection Online community leadership Marketing and sales prediction Review helpfulness Social influence Customer satisfaction Text mining Opinion mining Sentiment detection Sentiment classification Mobile advertising and targeting Digital and social media advertising Market prediction e-commerce Segmentation Recommendation personalization Digital WOM Positioning and perception

Sentiment analysis

Marketing strategy

Advertising and targeting

Market analysis

Recommendations

Brand analysis

Consumer Behavior Consumer behaviorists attempt to understand the psychology of consumer thinking, environmental influences, decision-making processes, and strategies. Understanding consumer behavior helps companies build better business strategies. Using big data, companies can predict consumer behavior more accurately than with conventional statistical techniques alone (Balar, Malviya, Prasad, & Gangurde, 2013). Consumers’ information searches via mobile devices generate a deluge of log data. For example, 80 million observations gleaned from multiple sources helped reveal consumer search behavior via mobile devices. Regardless of proximity to brick-and-mortar retailers, consumers tend to use mobile technology to search continually for product-related information. Surprisingly, consumers’ physical mobility influences search intensity (Daurer, Molitor, Spann, & Manchanda, 2015). Several psychological factors determine how big data alters traditional consumer decision processes (Fang & Li, 2014; Hofacker, Malthouse, & Sultan, 2016;

176  Vaidas Lukošius and Michael R. Hyman

Le & Liaw, 2017). As consumers traverse the well-established, five-step decisionmaking process (Engel, Blackwell, & Kollat, 1978), they are subject to everincreasing information exposure. Meaningful product-related information can influence need recognition and reduce information search. Evaluation of alternatives can be automated with recommendation agents and collaborative filtering tools, where ranking systems analyze user-generated content to assess customers’ preferences. In a big data context, purchase decisions are colored by feedback provided by online communities, customer ratings, and reviews. The post-purchase evaluation stage fuels a social network feedback loop for other consumers to participate in the evaluation process. Consumers also participate in virtual communities, where members communicate with each other through social media platforms. For marketers, understanding group behaviors is as important as understanding individual behavior. In virtual communities, consumers engage in discussions about brands, create and disseminate user-generated content, form brand-focused sub-communities, become leaders, adopt linguistic mannerisms, and the like. Consumer and retailer loyalty relate positively to one another (Rapp, Beitelspacher, Grewal, & Hughes, 2013).

Consumer Sentiment Consumer sentiment reflects consumers’ product-related perceptions and evaluations. Consumers leave public comments expressing disdain or favor toward a company, product, sales interaction, customer support, et cetera. Online reviews influence how other consumers choose and adjust expectations (Browning, So, & Sparks, 2013; Huete-Alcocer, 2017). Sentiment analysis—a special application of text or opinion mining—is the process of extracting and analyzing structured and unstructured consumer response data (e.g., web comments, microblogs, reviews, and other user-­ generated content) to reveal attitudes about products and/or brands. Typically, qualitative sentiment data are quantified by assigning them a positive, neutral, or negative value (Tuten & Solomon, 2018). Various tools—such as Google Analytics, Hootsuite, and People Browser—can yield reports that improve business intelligence. In addition, such tools yield a broad range of information about emotions, opinions, subjectives, emoticons, booster words, spelling corrections, negative word usage meant to flip emotions, et cetera (He et al., 2017; Salehan & Kim, 2016). Combining sentiment analysis with a neural network created from a 15,000-record dataset revealed online reviews, consumer sentiments, and online promotional strategies are significant predictors of online retailers’ sales (Chong, Li, Ngai, Ch’ng, & Lee, 2016). Such retailers can boost sales by carefully deciding the ‘how’ and ‘where’ to display online reviews and increase their social consumer interactions.

Big Data  177

Marketing Strategy Historically, a company’s commercial success largely rested on a marketing strategy grounded in a well-conceived and regularly updated marketing mix. Effective mixes often were a by-product of sophisticated data analyses performed on thenconsidered large secondary datasets. For example, to identify optimal prices and ad budgets, marketers created simultaneous regression (i.e., multiple equations with overlapping variables) and distributed lag (i.e., rate efficacy of marketing mix— especially advertising—expenditures decay) models based on industry-wide longitudinal sales and related secondary data (Assmus, Farley, & Lehmann, 1984; Bass & Clarke, 1972; Bell, Chiang, & Padmanabhan, 1999; Clarke, 1973; Magat, McCann, & Morey, 1986; Simon, 1979; Walters & Bommer, 1996). They used information—such as mix-element duration, content, timing, and ­expenditures— to estimate mix-element elasticities (e.g., percent change in dollar sales associated with a 1% change in mix-element expenditure). To forecast future sales and identify consumer-level market structures (i.e., within-product-category purchase switching; for example, between product forms such as instant versus ground coffee or between brands such as Nescafe and Folgers), these researchers relied on stochastic brand switching models estimated from microcomputerexceeding quantities of purchase diary panel data (i.e., data provided by an ongoing set of consumers who continually record their purchase-related experiences or observations in a diary) (Bass, 1974; Colombo & Morrison, 1989; Cooper & Inoue, 1996; Erdem, 1996; Netzer, Feldman, Goldenberg, & Fresko, 2012). Relative to the aforementioned industry-wide performance and marketing mix or consumer panel data, big data contains much richer and less structured real-time representations of consumer reactions to marketing mix changes. By allowing instant 360° customer profiles, big data can help marketers finetune their customer interactions. Marketers can use buying behavior, preference, coupon redemption, demographic, social group membership, and interaction data to customize ad content and product offerings at a micro-level and in real time. Online retailers such as Amazon and Overstock rely on such data to recommend offerings and/or actions to its online shoppers. Insights into consumer behavior, especially product consumption and usage, help companies innovate. Product lifecycle management benefits from three key applications of big data (Li, Tao, Cheng, & Zhao, 2015). First, vertical information exchange and sophisticated data warehousing technology allow production facilities to increase their efficiency and effectiveness via batch meta- and task-scheduling. Second, supply chain management encourages lean and agile distribution chains, which are responsive to consumer demands while producing less waste and obsolescence (Christopher & Ryals, 2014). Third, big data enables companies to achieve mass customization at a consumer level through flexibility, integration, and advanced manufacturing techniques (Fogliatto, da Silveira, & Borenstein, 2012).

178  Vaidas Lukošius and Michael R. Hyman

The emergence of big data has given new meaning to mass customization, which is “a strategy that creates value by some form of company-customer interaction at the fabrication and assembly stage of the operations level to create customized products with production cost and monetary price similar to those of mass-produced products” (Kaplan & Haenlein, 2006, pp. 176–177). By allowing a feedback loop between production and use phases, big data encourages smart, artificial-intelligence-enabled products, and real-time mass customization (Schmidt et al., 2015). For example, products linked to integrated cloud computing technology can alert users about routine or predictive maintenance and subsequently inform product design adjustments (e.g., building cars with advanced diagnostics and requiring less frequent oil changes). Location-based services (LBS) use real-time geospatial data from smartphones, tablets, and mobile computers to control device features and to provide information. For example, drivers use the GasBuddy mobile app to discover local gasoline prices based on the geographic location of the Internet-connected device. Retailers use LBS to display pricing information, make product recommendations, and customize online ads (Tsai, Wang,Yan, & Chang, 2017). Understanding ‘where’ as well as ‘when’ and ‘how’, which is possible with big data, is crucial to optimizing mobile advertising and targeting strategies (Luo, Andrews, Fang, & Phang, 2014).Yet, advertisers incorporating big data into traditional business models face data validity challenges (Fulgoni, 2014). For example, there is little evidence digital ad ‘clicks’ and ad efficacy correlate positively. Nonetheless, advertisers continue to rely on this misleading metric because it is fast, simple, and inexpensive to compute. Correctly attributing the ad-related factors that encourage consumer purchase is paramount to long-term corporate success. Hence, advertisers should combine big data with consumer panel data to determine the value of digital events—such as exposure to online ads, mobile search queries, website visits, and social media posts—to assess the impact of their ads on online and offline buying behavior. Like companies, consumers benefit from the transparency encouraged by big data. For example, Google offers a specialized search engine for customizing and researching air travel (http://flights.google.com). Such search-only travel services can build complex travel itineraries and offer recommendations for optimizing expenses and enhancing travel experiences. Data from a billion mobile phone traces show drivers using a routing app set for ‘socially optimal’ rather than ‘selfish’ routing can ease urban traffic congestion up to 30% (Çolak, Lima, & González, 2016). Such apps can exploit smartphone-generated GPS data to create a sustainable competitive advantage for app developers (Matthias, Fouweather, Gregory, & Vernon, 2017). Marketers can leverage big data to better identify and understand current and emerging market segments (i.e., naturally occurring consumer clusters) (Netzer et al., 2012). As enterprises shift from a data-poor to data-abundant environment, new ways to segment markets will emerge. For example, advances in artificial

Big Data  179

intelligence would permit self-customizing products that better serve consumers (e.g., Nest’s learning thermostat adjusts residential temperature based on owners’ behavioral patterns). Intelligence systems increasingly will allow humans to interact with technology in more natural ways (e.g., SILVIA-enabled devices listen and talk like human beings (www.silvia4u.info) (Morabito, 2015)). As tools for analyzing big data evolve, models will be able to aggregate data streams that include personal information, social networking data, LBS, past consumer behavior, and environmental data. Granular audience targeting is based on the availability of consumer preferences gathered from multiple sources via big data analysis (Srinivasan, 2017). For example, some Red Roof Inn franchisees deployed an algorithm that processed multi-source data about flight delays, weather forecasts, airport information, and local hotel room availability. Based on the algorithm’s output, these franchisees used social media to make personalized recommendations to visitors stranded at the local airport. The result: revenue increased by 10%. The goal of a recommender system is to produce meaningful suggestions to consumers about products or brands that might interest them. The benefits of such systems are multi-faceted: for consumers, reduced purchasing time and improved decision making, and for companies, higher revenue and product/brand exposure.Typically, such systems use either content-based filtering or collaborative filtering (Aggarwal, Tomar, & Kathuria, 2017). Content-based filtering systems recommend products by comparing item content to consumer profiles. In contrast, collaborative filtering systems predict the value of a particular product to a particular consumer by incorporating ratings from other consumers. Current digital technology provides real-time and wide-ranging big data streams that can yield advanced analyzes and personalized recommendations (Brown, Chui, & Manyika, 2011). Consumers’ beliefs about the value of recommendation systems relate to these systems’ level of personalization (Cremonesi, Garzotto, & Turrin, 2012). Consumers both welcome and reject the effect of big data on their lives (Baruh & Popescu, 2017).They may rely exclusively on recommendation systems when buying (assimilation), yet they may reject suggestions generated by algorithm analyzes (avoidance). Brand management through positioning is an important marketing concept because it helps companies attain a competitive advantage. Brand management frameworks (Keller, 1993, 2001) account for macro changes in business environments, such as big data, by providing a path for optimizing market performance and brand equity. The ability to use big data shapes market structures and brand management practices, as power is shifted from marketers to consumers (Erdem, Keller, Kuksov, & Pieters, 2016). Research tools such as user-generated content analyzes are useful in assessing consumer satisfaction with brand quality (Tirunillai & Tellis, 2014), which is useful for optimizing brand position. Social tagging, a process by which users apply keywords to interpret and classify content, is valuable for assessing brand

180  Vaidas Lukošius and Michael R. Hyman

familiarity and associations (Nam & Kannan, 2014). Brand reputation monitoring can help companies assess consumers’ attitudes about their brand and competitive positioning (Ziegler & Skubacz, 2006).

Emerging Research in Big Data Analytics Big data analytics refers to technologies that are established in data mining and statistical analyzes (see Table 6.3).These techniques are well grounded and rely on existing foundational technologies, such as data warehousing and Online Analytical Processing (OLAP). Text analytics primarily relies on information retrieval and computational linguistics. Most current commercial search engines are rooted in text analytics. As a result, such text processing and indexing techniques have evolved quickly. Web analytics uses web data to optimize web usage, such as enhancing website usability. Web crawling algorithms use heuristic functions to increase efficiency of crawlers (e.g., Bingbot is the web crawler for Microsoft’s Bing). To alleviate information overload, recommender systems search large volumes of dynamically generated information to provide consumers with personalized content. Network analytics reveals the properties and relationships between and within networks. Finally, widespread adoption and use of mobile technology by consumers gave rise to mobile analytics for understanding how consumers use such technologies.

Cloud Computing Cloud computing is “a distributed computing paradigm in which all the resources, dynamically scalable and often virtualized, are provided as services over the Internet” (Belcastro, Marozzo,Talia, & Trunfio, 2017, p. 104). Its main purpose TABLE 6.3 Foundational Technologies and Emerging Research in Big Data (Chen et al.,

2012)

Foundational Technologies

Emerging Research

Category

Examples

Big data analytics Text analytics Web analytics Network analytics Mobile analytics Big data analytics Text analytics Web analytics Network analytics Mobile analytics

Data warehousing, OLAP Search engines Web crawling, recommender systems Network visualization Smartphone platforms Cloud computing, machine learning, cognitive computing Multimedia analytics Social media analytics Social network analytics Gamification

Big Data  181

is to use vast computational resources and capacities to store and process big data. Although big data and cloud computing are intertwined, they differ meaningfully. Cloud computing is part of IT infrastructure and architecture; in contrast, big data depends on cloud computing to provide decision makers with insights. In essence, big data is supported by cloud computing (M. Chen, Mao, Zhang, & Leung, 2014). Cloud computing offers flexibility and efficiency for accessing, processing, and analyzing data. For example, it can facilitate efforts to ‘scale up’ processing power (Kshetri, Fredriksson, & Torres, 2017). E-commerce activity and reliance on cloud-based services are closely related (Kshetri et al., 2017); as the amount of data on online shoppers, shopping, and promotions grow, companies will rely increasingly on cloud-based services to model market segments and consumer behavior (Wang, 2016).

Machine Learning Machine learning is the data analysis process that automates analytical model building and can be treated as an optimization problem. Manual data manipulation and processing to produce meaningful results is impossible with big data due to the massive quantity. To discover hidden information within a huge dataset without human intervention, computer algorithms are applied iteratively to learn from such data (L’Heureux et al., 2017). Machine learning systems operate under three basic principles: simplicity, multiple explanations, and evaluating learned hypotheses (Kononenko & Kukar, 2007). The first principle maximizes the hypothesis’ likelihood relative to background knowledge and input (learning) data (i.e., akin to maximum likelihood analyzes in statistics). Learning algorithms trade-off between hypothesis complexity and accuracy. The second principle requires retention of all hypotheses consistent with inputs; hence, algorithms should generate many data-­consistent hypotheses. The third principle encompasses estimates of hypothesis quality. Due to various problems with machine learning, quality estimates depend on the statistics used. For example, discrete function estimates—such as classification-segmentation problems—rely on classification accuracy measures (e.g., Brier score (Bahnsen, Stojanovic, Aouada, & Ottersten, 2014)). In contrast, continuous function estimates—such as regression problems—rely on different accuracy measures (e.g., mean square (MSE) (Das & Uddin, 2013)). For example, Amazon uses a machine learning algorithm of acoustic fingerprinting to prevent its cloud-connected Alexa device from confusing broadcasted or narrow-casted voice audio with human commands (Murphy, 2018).

Cognitive Computing Cognitive computing, which is the probabilistic process for revealing patterns via iterative and self-learning algorithms designed to explore and analyze

182  Vaidas Lukošius and Michael R. Hyman

data from ambiguous and uncertain contexts, is a next step in machine learning ­(Hurwitz, Kaufman, & Bowles, 2015). The application and use of cognitive computing extends beyond marketing practice into law enforcement, healthcare, and education (Contractor & Telang, 2017), with the promise it can replace manual cognitive tasks and analyses, such as restructuring organizational business units (Nelson, Clark, & Stewart, 2018). Cognitive computing systems rely on contextual insights, hypothesis-­generating ability, and continuous learning processes (Hurwitz et al., 2015). Such systems collate data from diverse sources—such as big data analytics, machine learning, Natural Language Processing, and data visualization (Hurwitz et al., 2015)—that contain images, audio, geo-location, voice, and other content. Based on accumulated knowledge, they generate multiple hypotheses relevant to business problems and provide answers with associated confidence levels.The aggregate models they yield ultimately are adjusted based on inputs from system users and new data. In marketing, cognitive computing could suggest how consumers respond to specific marketing appeals (Wilson, Hall-Phillips, & Djamasbi, 2015), how retail environments affect purchasing and sale situations (Sato & Huang, 2015), and how to create better recommendations for consumers (Forrest & Hoanca, 2015).

Multimedia Analytics Unlike simple and structured data, multimedia data mainly consists of images, audio, and video. Useful information extraction is a major challenge. Processing such data typically requires semantic data models that entail building and standardizing relationships between data elements. Audio data analysis involves word and phrase extraction. Video data analysis can work with either static or dynamic content. Topic-oriented multimedia summarization based on natural language generation can automatically generate a written paragraph summarizing important information in a video (Metze, Ding, Younessian, & Hauptmann, 2013). Visual content-based video indexing and retrieval methods are used in video structure analysis to identify semantic content based on video genres (e.g., instructional video) (Hu, Xie, Li, Zeng, & Maybank, 2011). Static visual images, such as user-generated content on Instagram, can be subjected to classification analysis. For example, the Clarifai image recognition API software tool consists of more than 11,000 classifiers in categories such as objects, ideas, and feelings (Jaakonmäki, Müller, & von Brocke, 2017). Once qualitative features from images are extracted and classified, the resulting dataset can be subjected to traditional quantitative analyses such as regression.

Social Media Analytics Social media analytics is “the practice of gathering data from social media platforms and analyzing the data to help decision makers address specific problems”

Big Data  183

(Lee, 2017, p. 1). With data collected from social media platforms, such analytics enable marketers to develop intelligence on consumers, competitors, and business environments (Arrigo, 2014). Companies should plan for and coordinate their social media analysis efforts. Rather than a single tool, social media analytics provide an array of tools (Lee, 2017). Real-time consumer analytics are reactive and allow marketers to understand consumers’ conversations. Via keyword analysis—quantity and quality of keywords about specific topics or brands—companies like Nike (e.g., #complaint #nike) can monitor online communications for trending topics. Non-real-time consumer analytics can help marketers build proactive strategies. Information, such as social media/platform user base growth (or shrinkage), demographics, and geo-location can reveal past consumer behavior that marketers can use to model future consumer behavior. Real-time competitive analytics can, for example, help companies monitor the new features, price changes, and recalls of competitors’ products. Non-real-time competitive analytics allow companies to track changing customer service structures, relationships between consumer-generated content and sales, et cetera.

Social Network Analytics Social network analytics are research tools meant to identify actors and interactions between those actors (Aggarwal, 2011). Social network models are built via two types of data: linkage based (i.e., understanding linkage behavior in networks) and content based (i.e., understanding how content, such as tagged photos, creates linkages in networks). Such models can help answer four pivotal questions about networks: 1. 2. 3. 4.

Group detection, or Who belongs to the same group? Group profiling, or What is this group about? Group evolution, or How do group values evolve? Link prediction, or When will new relationships form?

Community structures can be visualized graphically, where nodes are actors and lines connecting nodes represent linkages. Using social network analytics, marketers have predicted customer churn (Óskarsdóttir et al., 2017), assessed sponsorship efficacy (White, White, & White, 2017), and modeled spending and cross-buying behavior (Kumar, Bezawada, Rishika, Janakiraman, & Kannan, 2016) Content-based analytics of large social networking sites generate massive quantities of text, audio, and video content in unstructured formats. For information to be authentic, timely, and actionable, real-time data must be analyzed with real-time tools. Signal-to-noise ratio is ever increasing due to the opportunistic nature of network manipulation; hence, data analysis tools must be able to identify and delete fake site data produced by people paid to artificially boost the seeming

184  Vaidas Lukošius and Michael R. Hyman

popularity and influence of those sites. If U.S. national security agencies are correct, then Russian bots influenced the 2016 election results by enhancing the credibility of fake news purveyors (National Intelligence Council, 2017).

Gamification Gamification is “the use of game design elements to enhance non-game goods and services by increasing customer value and encouraging value-creating behaviors such as increased consumption, greater loyalty, engagement, or product advocacy” (Hofacker, Ruyter, Lurie, Manchanda, & Donaldson, 2016, p. 26). Marketers use ‘games’ such as loyalty cards and frequent flyer programs to increase consumers’ engagement with brands and physical retail environments. Gamification plus mobile technology can reveal consumers’ responses to mobile promotions, ads, and shopping. For example, badges are “medals which are awarded to users based on some predefined levels of engagement” (Khodadadi, Hosseini, Tavakoli, & Rabiee, 2017, p. 3) that encourage consumers to post online reviews or make website contributions. Mobile apps now serve as the primary gamification platform for consumers. Big data analytics such as machine learning (L’Heureux, 2015) can track how synergies among the core gaming elements of mechanics (i.e., how players are rewarded for achieving their goals), aesthetics (i.e., creative vision of the platform), and technology (i.e., mobile device and usage properties) drive consumer responses.

Troubleshooting Big Data Transitioning from small data to big data requires comprehensive understanding of new and updated assumptions. Data structure, content, measurements, reproducibility, and analyses have to be approached with different tools and assumptions. The big data literature highlights errors and biases irrelevant to small datasets. Such biases include overfitting bias, bigness bias, data subset bias, accuracy bias, complexity bias, statistical method bias, and ambiguity bias (Berman, 2013). 1. Overfitting bias. Overfitting occurs when a formula closely defines one dataset but fails to predict the behavior of comparable datasets. In such cases, noise rather than system behavior is being predicted. Techniques, such as neural networks, are prone to overfitting. Likelihood of overfitting increases as dataset size increases. 2. Bigness bias. Bigness bias occurs when a researcher erroneously believes, for example, a predictive market model is underperforming due to insufficient data. Marketers tend to accept results, regardless of how counterintuitive, when estimates are derived from super-large samples. Dataset size does not equate with data representativeness.

Big Data  185

3. Data subset bias. Big data subsets are neither additive nor transitive. This paradox occurs when “findings that apply to each of two data sets may be reversed when the two data sets are combined” (Berman, 2013, p. 153). 4. Accuracy bias. Big data are not identical to ‘good data’. Because big data are collected from various sources, it can contain reliable and valid subsets. Good analysts validate their analyses with alternate data sources and analytical methods. 5. Complexity bias. Big data is not easily reconcilable because it is created from multiple sources. No common denominator may exist for data selection, filtering, or transformation to perform triangulation and reliability checks. 6. Statistical method bias. Data analysts tend to specialize in and prefer statistical methods that confirm their data analysis prejudices. For example, biologists generally continue to accept results of highly cited studies despite contradictory claims appearing in other reputable studies (Tatsioni, Bonitsis, & Ioannidis, 2007). When standard statistical methods prove unsatisfactory for analyzing atypical big data, more suitable analytical tools must be used and developed. 7. Ambiguity bias. Akin to the Gestalt Principle—i.e., the whole is greater than the sum of its parts—ambiguity bias haunts big data analysts, who often build algorithms for assessing simple and low-level data systems (e.g., pedometer). The elements of complex systems, such as human bodies, may be difficult or impossible to describe. The additive properties of algorithms meant to explain complex systems by combining several low-level data systems are convoluted.

Conceptual Exercises for Students EXERCISE 1 The basic notion behind the Internet of Things (IoT) is to enable interdevice information exchange by connecting different ‘smart devices’, such as barcode readers, sensors, and mobile phones. By programing and enabling these devices to communicate with one another as well as with users, consumers’ routine tasks can be simplified and automated. Assume the following connected and programmed consumer household devices: smart car, mobile phone, alarm, refrigerator, and microwave. 1. What type of sensor data would each device generate? 2. Which data would be most useful for business insight generation?

186  Vaidas Lukošius and Michael R. Hyman

3. What type of consumer behavior can these data predict? 4. What, if any, pitfalls should be anticipated when combining data from multiple devices? EXERCISE 2 Pick five social media websites with different foci and user bases. Compare and contrast your picks by answering the following questions: 1. Which site is ‘biggest’? What metrics did you use to reach your conclusion? 2. What kind of data (i.e., structured, unstructured, text, images, video, et cetera) most typifies each website? 3. If you were required to use text, web, network, and mobile analytic tools for each social media website, what types of marketing information could be extracted? What are possible uses for such information? EXERCISE 3 Gephi is an open platform for visualizing and analyzing large network graphs. The web page https://gephi.org/users/ provides a tutorial for using Gephi. Find a tutorial developed by the user community. You may choose one based on your interests, level of difficulty, or type of visualization. 1. Prepare a presentation highlighting the process and results of your visualization. 2. What, if any, hidden patterns did you discover in your visualization? 3. What insights are suggested by your visualization? EXERCISE 4 Companies engage in environmental scanning, which is a data-driven research process to assess business opportunities and threats. Some companies are secretive and divulge little about their internal operations. Consider Tesla, an electric car manufacturer, and the unit-sales model developed by Tom Randall and Dean Halford to estimate the

Big Data  187

number of Tesla Model 3 cars (www.bloomberg.com/graphics/2018tesla-tracker/). They used complementary methods to acquire needed data: (a) Method 1, Tesla Registered VINs Model, and (b) Method 2, Tesla Spotted VINs Model. 1. Describe the data collection mechanism for Method 1 and Method 2. 2. To what biases is each method susceptible? 3. What issues complicate reconciling and merging data from Method 1 and Method 2? 4. What is the joint production estimate?

Notes 1. Data and information are not synonymous. Data + interpretation = information. Data often are unstructured, messy, and difficult to interpret. Information is orderly, clean, and aids decision making processes. 2. Knowledge and understanding of big data are in their infancy. Taxonomies and typologies of approaches for compartmentalizing tools and techniques come from academic disciplines such as statistics, computer science, marketing, information systems. Scholars and practitioners of each discipline are subject to domain-specific rules. Therefore, the exposition about many of the presented concepts, techniques, and tools appear in multiple sections. This organization is characteristic of emerging research fields rather than a taxonomic failure.

References Aggarwal, C. C. (2011). Social network data analytics. New York, NY: Springer. Aggarwal, P., Tomar,V., & Kathuria, A. (2017). Comparing content based and collaborative filtering in recommender systems. International Journal of New Technology and Research, 3(4), 65–67. Anderson, A., & Semmelroth, D. (2015). Statistics for big data for dummies. Hoboken, NJ: John Wiley & Sons, Inc. Arrigo, E. (2014). Social media opportunities for market-driven firms. In F. Lee (Ed.), Integrating social media into business practice, applications, management, and models (pp. 180–199). Hershey, PA: IGI Global. Assmus, G., Farley, J. U., & Lehmann, D. R. (1984). How advertising affects sales: Metaanalysis of economic results. Journal of Marketing Research, 21(1), 65–74. Bahnsen, A. C., Stojanovic, A., Aouada, D., & Ottersten, B. (2014). Improving credit card fraud detection with calibrated probabilities. In M. Zaki, Z. Obradovic, P. N. Tan, A. Banerjee, C. Kamath, & S. Parthasarathy (Eds.), Proceedings of the 2014 SIAM international conference on data mining (pp. 677–685). Philadelphia, PA: Society for Industrial and Applied Mathematics.

188  Vaidas Lukošius and Michael R. Hyman

Balar, A., Malviya, N., Prasad, S., & Gangurde, A. (2013). Forecasting consumer behavior with innovative value proposition for organizations using big data analytics. In IEEE International Conference on Computational Intelligence and Computing Research (ICCIC). Red Hook, NY: Curran Associates, Inc. Barnaghi, P., Sheth, A., & Henson, C. (2013). From data to actionable knowledge: Big data challenges in the web of things [Guest Editors’ Introduction]. IEEE Intelligent Systems, 28(6), 6–11. Baruh, L., & Popescu, M. (2017). Big data analytics and the limits of privacy self-­ management. New Media & Society, 19(4), 579–596. Bass, F. M. (1974). The theory of stochastic preference and brand switching. Journal of ­Marketing Research, 11(1), 1–20. Bass, F. M., & Clarke, J. G. (1972).Testing distributed lag models of advertising effect. Journal of Marketing Research, 9(3), 298–308. Belcastro, L., Marozzo, F., Talia, D., & Trunfio, P. (2017). Big data analysis on clouds. In A.Y. Zomaya & S. Sakr (Eds.), Handbook of big data technologies (pp. 101–142). Cham, Switzerland: Springer International Publishing. Bell, D. R., Chiang, J., & Padmanabhan, V. (1999). The decomposition of promotional response: An empirical generalization. Marketing Science, 18(4), 504–526. Berman, J. J. (2013). Principles of big data: Preparing, sharing, and analyzing complex information. Waltham, MA: Morgan Kaufmann. Betser, J., & Belanger, D. (2013). Architecting the enterprise via big data analytics. In J. Liebowitz (Ed.), Big data and business analytics (pp. 1–20). Boca Raton, FL: CRC Press. Booms, B. H., & Bitner, M. J. (1981). Marketing strategies and organization structures for service firms in marketing of services. In J. H. Donnelly & W. R. George (Eds.), Marketing of services (pp. 47–51). Chicago, IL: American Marketing Association. Boyd, D., & Crawford, K. (2012). Critical questions for big data: Provocations for a cultural, technological, and scholarly phenomenon. Information, Communication, & Society, 15(5), 662–679. Boyd, D., & Marwick, A. E. (2011). Social privacy in networked publics: Teens’ attitudes, practices, and strategies. Working paper presented at A decade in internet time: Symposium on the dynamics of the internet and society. Oxford: Oxford Internet Institute. Breiman, L. (2001). Statistical modeling:The two cultures. Statistical Science, 16(3), 199–231. Brown, B., Chui, M., & Manyika, J. (2011). Are you ready for the era of “big data”. ­McKinsey Quarterly, 4, 24–35. Browning,V., So, K. K. F., & Sparks, B. (2013).The influence of online reviews on consumers’ attributions of service quality and control for service standards in hotels. Journal of Travel & Tourism Marketing, 30(1–2), 23–40. Buhl, H. U., Röglinger, M., Moser, F., & Heidemann, J. (2013). Big data: A fashionable topic with(out) sustainable relevance for research and practice? Business & Information Systems Engineering, 5(2), 65–69. Chen, H., Chiang, R. H. L., & Storey,V. C. (2012). Business intelligence and analytics: From big data to big impact. MIS Quarterly, 36(4), 1165–1188. Chen, M., Mao, S., & Liu,Y. (2014). Big data: A survey. Mobile Networks & Applications, 19(2), 171–209. Chen, M., Mao, S., Zhang,Y., & Leung,V. C. (2014). Big data: Related technologies, challenges and future prospects. Cham, Switzerland: Springer International Publishing. Chong, A., Li, B., Ngai, E., Ch’ng, E., & Lee, F. (2016). Predicting online product sales via online reviews, sentiments, and promotion strategies: A big data architecture and neural

Big Data  189

network approach. International Journal of Operations & Production Management, 36(4), 358–383. Christopher, M., & Ryals, L. J. (2014).The supply chain becomes the demand chain. Journal of Business Logistics, 35(1), 29–35. Clarke, D. G. (1973). Sales-advertising cross-elasticities and advertising competition. Journal of Marketing Research, 10(3), 250–261. Çolak, S., Lima, A., & González, M. C. (2016). Understanding congested travel in urban areas. Nature Communications, 7, 1–8. Colombo, R. A., & Morrison, D. G. (1989). A brand switching model with implications for marketing strategies. Marketing Science, 8(1), 89–99. Contractor, D., & Telang, A. (2017). Applications of cognitive computing systems and IBM Watson. Singapore: Springer Nature. Cooper, L. G., & Inoue, A. (1996). Building market structures from consumer preferences. Journal of Marketing Research, 33(3), 293–306. Cordeiro, R. L. F., Jr., C.T.,Traina, A. J. M., López, J., Kang, U., & Faloutsos, C. (2011). Clustering very large multi-dimensional datasets with MapReduce. Paper presented at the 17th ACM SIGKDD Conference on Knowledge Discovery (pp. 690–698). San Diego, CA. Cremonesi, P., Garzotto, F., & Turrin, R. (2012). Investigating the persuasion potential of recommender systems from a quality perspective. ACM Transactions on Interactive Intelligence Systems, 2(2), 1–41. Das, D., & Uddin, M. S. (2013). Data mining and neural network techniques in stock market prediction: A methodological review. International Journal of Artificial Intelligence & Applications, 4(9), 117–127. Dasgupta, N. (2018). Practical big data analytics: Hands-on techniques to implement enterprise analytics and machine learning using Hadoop, Spark, NoSQL and R. Birmingham: Packt Publishing. Daurer, S., Molitor, D., Spann, M., & Manchanda, P. (2015). Consumer search behavior on the mobile Internet: An empirical analysis. Working Paper No. 1275. Cambridge, MA: Ross School of Business. Dietrich, D., Heller, B., & Yang, B. (2015). Data science and big data analytics: Discovering, analyzing, visualizing and presenting data. Indianapolis, IN: John Wiley & Sons, Inc. Eklund, L. (2015). Bridging the online/offline divide:The example of digital gaming. Computers in Human Behavior, 53(C), 527–535. Engel, J. F., Blackwell, R. D., & Kollat, D. T. (1978). Consumer behavior. Hinsdale, IL: Dryden Press. Erdem, T. (1996). A dynamic analysis of market structure based on panel data. Marketing Science, 15(4), 359–378. Erdem, T., Keller, K. L., Kuksov, D., & Pieters, R. (2016). Understanding branding in a digitally empowered world. International Journal of Research in Marketing, 33(1), 3–10. Erl, T., Khattak, W., & Buhler, P. (2016). Big data fundamentals: Concepts, drivers & techniques. New York, NY: Prentice Hall. Fan, S., Lau, R.Y. K., & Zhao, J. L. (2015). Demystifying big data analytics for business intelligence through the lens of marketing mix. Big Data Research, 2(1), 28–32. Fang, Z., & Li, P. (2014). The mechanism of “big data” impact on consumer behavior. American Journal of Industrial and Business Management, 4(1), 45–50. Fogliatto, F. S., da Silveira, G. J., & Borenstein, D. (2012). The mass customization decade: An updated review of the literature. International Journal of Production Economics, 138(1), 14–25.

190  Vaidas Lukošius and Michael R. Hyman

Foreman, J. W. (2013). Data smart: Using data science to transform information into insight New York, NY: John Wiley & Sons, Inc. Forrest, E., & Hoanca, B. (2015). Artificial intelligence: Marketing’s game changer. In T. ­Tsiakis (Ed.), Trends and innovations in marketing information systems (pp. 45–64). ­Hershey, PA: IGI Global. Fulgoni, G. (2014). Big data: Friend or foe of digital advertising? Five ways marketers should use digital big data to their advantage. Journal of Advertising Research, 53(4), 91–95. Gandomi, A., & Haider, M. (2015). Beyond the hype: Big data concepts, methods, and analytics. International Journal of Information Management, 35(2), 137–144. He, W., Tian, X., Tao, R., Zhang, W.,Yan, G., & Akula,V. (2017). Application of social media analytics: A case of analyzing online hotel reviews. Online Information Review, 41(7), 921–935. Hofacker, C. F., Malthouse, E. C., & Sultan, F. (2016). Big data and consumer behavior: Imminent opportunities. Journal of Consumer Marketing, 33(2), 89–97. Hofacker, C. F., Ruyter, K. D., Lurie, N. H., Manchanda, P., & Donaldson, J. (2016). Gamification and mobile marketing effectiveness. Journal of Interactive Marketing, 34, 25–36. Hu, W., Xie, N., Li, L., Zeng, X., & Maybank, S. (2011). A survey on visual content-based video indexing and retrieval. IEEE Transactions on Systems, Man, and Cybernetics, 41(6), 797–819. Huang, T., Lan, L., Fang, X., An, P., Min, J., & Wang, F. (2015). Promises and challenges of big data computing in health sciences. Big Data Research, 2(1), 2–11. Huberty, M. (2015). Awaiting the second big data revolution: From digital noise to value creation. Journal of Industry, Competition and Trade, 15(1), 35–47. Huete-Alcocer, N. (2017). A literature review of word of mouth and electronic word of mouth: Implications for consumer behavior. Frontiers in Psychology, 8(Article 1256). Hunt, S. (2010). Marketing theory: Foundations, controversy, strategy, resource-advantage theory. Armonk, NY: M. E. Sharpe. Hurwitz, J., Kaufman, M., & Bowles, A. (2015). Cognitive computing and big data analytics. New York, NY: John Wiley & Sons, Inc. Hyman, M., & Sierra, J. (2010). Marketing research kit for dummies. Hoboken, NJ: John Wiley & Sons, Inc. Jaakonmäki, R., Müller, O., & von Brocke, J. (2017). The impact of content, context, and creator on user engagement in social media marketing. In Proceedings of the 50th Hawaii International Conference on System Sciences (pp. 1152–1160). Honolulu, HI. Jackson, S. (2015). Cult of analytics: Data analytics for marketing. New York, NY: Routledge. Jagadish, H. V., Gehrke, J., Labrinidis, A., Papakonstantinou, Y., Patel, J. M., Ramakrishnan, R., & Shahabi, C. (2014). Big data and its technical challenges. Communications of the ACM, 57(7), 86–94. Japec, L., Kreuter, F., Berg, M., Biemer, P., Decker, P., Lampe, C., . . . Usher, A. (2015). Big data in survey research: AAPOR task force report. Public Opinion Quarterly, 9(4), 839–880. Japkowicz, N., & Stefanowski, J. (2016). Big data analysis: New algorithms for a new society (Vol. 16). Cham, Switzerland: Springer International Publishing. Kaplan, A. M., & Haenlein, M. (2006). Toward a parsimonious definition of traditional and electronic mass customization. Journal of Product Innovation Management, 23(2), 168–182. Keller, K. L. (1993). Conceptualizing, measuring, and managing customer-based brand equity. Journal of Marketing, 57(1), 1–22. Keller, K. L. (2001). Building customer-based brand equity: A blueprint for creating strong brands. Marketing Management, 10(2), 15–19.

Big Data  191

Khodadadi, A., Hosseini, S. A., Tavakoli, E., & Rabiee, H. R. (2017). Continuous-time user modeling in the presence of badges:A probabilistic approach. CoRR, arXiv1702.01948v1. Kononenko, I., & Kukar, M. (2007). Machine learning and data mining. Chichester: Horwood Publishing Limited. Kshetri, N., Fredriksson, T., & Torres, D. C. R. (2017). Big data and cloud computing for development: Lessons from key industries and economies in the global south. New York, NY: Routledge. Kuhn, T. S. (1996). The structure of scientific revolutions. Chicago, IL: University of Chicago Press. Kumar, A., Bezawada, R., Rishika, R., Janakiraman, R., & Kannan, P. K. (2016). From social to sale: The effects of firm-generated content in social media on customer behavior. Journal of Marketing, 80(1), 7–25. Kwon, O., Lee, N., & Shin, B. (2014). Data quality management, data usage experience and acquisition intention of big data analytics. International Journal of Information Management, 34(3), 387–394. Labrinidis, A., & Jagadish, H. V. (2012). Challenges and opportunities with big data. In The 38th International conference on very large data bases (Vol. 5, pp. 2032–2033). Istanbul, ­Turkey: Proceedings of the VLDB Endowment. Le, T. M., & Liaw, S-Y. (2017). Effects of pros and cons of applying big data analytics to consumers’ responses in an e-commerce context. Sustainability, 9(5). Lee, I. (2017). Social media analytics for enterprises: Typology, methods, and processes. Business Horizons, 61(2), 199–210. L’Heureux, A. (2015). Gamifcation framework for sensor data analytics. Electronic Thesis and Dissertation Repository. Retrieved from htps://ir.lib.uwo.ca/etd/3200 L’Heureux, A., Grolinger, K., Elyamany, H. F., & Capretz, M. A. (2017). Machine learning with big data: Challenges and approaches. IEEE Access, 5, 7776–7797. Li, J., Tao, F., Cheng,Y., & Zhao, L. (2015). Big data in product lifecycle management. The International Journal of Advanced Manufacturing Technology, 81(1), 667–684. Lohr, S. (2015). Data-ism: The revolution transforming decision making, consumer behavior, and almost everything else. New York, NY: HarperCollins Publishers. Luo, X., Andrews, M., Fang, Z., & Phang, C. W. (2014). Mobile targeting. Management Science, 60(7), 1738–1756. Lyko, K., Nitzschke, M., & Ngomo, A-C. N. (2016). Big data acquisition. In J. M. Cavanillas, E. Curry, & W. Wahlster (Eds.), New horizons for a data-driven economy. Cham, Switzerland: Springer. Magat, W. A., McCann, J. M., & Morey, R. C. (1986). When does lag structure really matter in optimizing advertising expenditures? Management Science, 32(2), 182–193. Matthias, O., Fouweather, I., Gregory, I., & Vernon, A. (2017). Making sense of big data— can it transform operations management? International Journal of Operations and Production Management, 37(1), 37–55. Mayer-Schönberger,V., & Cukier, K. (2014). Big data: A revolution that will transform how we live, work, and think. New York, NY: Eamon Dolan/Mariner Books. McCarthy, E. J. (1960). Basic marketing: A managerial approach. Homewood, IL: Richard D. Irwin, Inc. Metze, F., Ding, D., Younessian, E., & Hauptmann, A. (2013). Beyond audio and video retrieval: Topic-oriented multimedia summarization. International Journal of Multimedia Information Retrieval, 2(2), 131–144. Miller, J. D. (2017). Big data visualization. Birmingham: Packt Publishing. Miniwats Marketing Group. (2018, February 18). World internet users and 2018 population stats. Retrieved from www.internetworldstats.com/stats.htm

192  Vaidas Lukošius and Michael R. Hyman

Morabito,V. (2015). Big data and analytics: Strategic and organizational impacts. Cham, Switzerland: Springer International Publishing Murphy, D. (2018). How Amazon keeps your TV from accidentally triggering your Echo. PC Magazine. Retrieved from www.pcmag.com/news/359004/how-amazon-keepsyour-tv-from-accidentally-triggering-your-e Nam, H., & Kannan, P. K. (2014). The informational value of social tagging networks. Journal of Marketing, 78(4), 21–40. National Intelligence Council. (2017). Assessing Russian activities and intentions in recent US elections: The analytic process and cyber incident attribution. Retrieved from www.dni.gov/ files/documents/ICA_2017_01.pdf Nelson, H. J., Clark, T., & Stewart, I. (2018). Hunch mining: Intuition augmented with cognitive computing, analytics, and artificial intelligence. Paper presented at 51st Hawaii International Conference on System Sciences (pp. 5065–5074). Grand Wailea, HI. Netzer, O., Feldman, R., Goldenberg, J., & Fresko, M. (2012). Mine your own business: Market-structure surveillance through text mining. Marketing Science, 31(3), 521–543. Ninja Metrics. (2014). Prediction:The future of game analytics. Retrieved from www.ninjamet rics.com/tag/game-analytics/ Óskarsdóttir, M., Bravo, C., Verbeke, W., Sarraute, C., Baesens, B., & Vanthienen, J. (2017). Social network analytics for churn prediction in telco: Model building, evaluation and network architecture. Expert Systems with Applications, 85(1), 204–220. Press, G. (2012). A very short history of big data. Retrieved from https://whatsthebigdata. com/2012/06/06/a-very-short-history-of-big-data/ Rapp, A., Beitelspacher, L., Grewal, D., & Hughes, D. (2013). Understanding social media effects across seller, retailer, and consumer interactions. Journal of the Academy of Marketing Science, 41(5), 547–566. Sagiroglu, S., & Sinanc, D. (2013). Big data: A review. Paper presented at International Conference on Collaboration Technologies and Systems (CTS’ 13) (pp. 42–47). San Diego, CA. Salehan, M., & Kim, D. J. (2016). Predicting the performance of online consumer reviews: A sentiment mining approach to big data analytics. Decision Support Systems, 81, 30–40. Sathi, A. (2013). Big data analytics: Disruptive technologies for changing the game. Boise, ID: Mc Press. Sato, A., & Huang, R. (2015). From data to knowledge: A cognitive approach to retail business intelligence. Paper presented at 2015 IEEE International Conference on Data Science and Data Intensive Systems (DSDIS) (pp. 210–217). IEEE. Schmarzo, B. (2013). Big data: Understanding how data powers big business. Indianapolis, IN: John Wiley & Sons, Inc. Schmidt, R., Möhring, M., Härting, R-C., Reichstein, C., Neumaier, P., & Jozinović, P. (2015). Industry 4.0-potentials for creating smart products: Empirical research results. Paper presented at 18th International Conference on Business Information Systems (pp. 16–27). Poznań, Poland. Sheng, J., Amankwah-Amoah, J., & Wang, X. (2017). A multidisciplinary perspective of big data in management research. International Journal of Production Economics, 191, 97–112. Simon, H. (1979). Dynamics of price elasticity and brand life cycles: An empirical study. Journal of Marketing Research, 16(11), 439–452. Sivarajah, U., Kamal, M. M., Irani, Z., & Weerakkody,V. (2017). Critical analysis of big data challenges and analytical methods. Journal of Business Research, 70, 263–286. Srinivasan, S. (2017). Guide to big data applications (Vol. 26). Cham, Switzerland: Springer International Publishing.

Big Data  193

Tang, N. (2014). Big data cleaning. In L. Chen,Y. Jia, T. Sellis, & G. Liu (Eds.), Web technologies and applications (pp. 13–24). Cham, Switzerland: Springer International Publishing. Tatsioni, A., Bonitsis, N. G., & Ioannidis, J. A. (2007). Persistence of contradicted claims in the literature. Journal of the American Medical Association, 298(21), 2517–2526. Tirunillai, S., & Tellis, G. (2014). Mining marketing meaning from online chatter: Strategic brand analysis of big data using latent Dirichlet allocation. Journal of Marketing Research, 51(4), 463–479. Tsai,Y-T., Wang, S-C.,Yan, K-Q., & Chang, C-M. (2017). Precise positioning of marketing and behavior intentions of location-based mobile commerce in the Internet of Things. Symmetry, 9(8), 139. Tuten, T. L., & Solomon, M. R. (2018). Social media marketing. London: Sage Publications Ltd. Verhoef, P. C., Stephen, A. T., Kannan, P. K., Luo, X., Abhishek, V., Andrews, M., . . . Hu, M. M. (2017). Consumer connectivity in a complex, technology-enabled, and mobileoriented world with smart products. Journal of Interactive Marketing, 40, 1–8. Walters, R. G., & Bommer, W. (1996). Measuring the impact of product and promotionrelated factors on product category price elasticity. Journal of Business Research, 36(3), 203–216. Wang, K. W. (2016, June). Using analytics to turbocharge China’s e-commerce performance. McKinsey Quarterly. Wedel, M., & Kannan, P. K. (2016). Marketing analytics for data-rich environments. Journal of Marketing, 80(6), 97–121. White, D., White, C., & White, D. (2017). Sponsorship-linked marketing: Using social media brand affinity data to determine sport sponsorship effectiveness. Paper presented at the 2017 Atlantic Marketing Association Conference (pp. 369–370). Williamsburg,VA. Wilson, E. V., Hall-Phillips, A., & Djamasbi, S. (2015). Cognitive predictors of consumers’ intention to comply with social marketing email appeals. Computers in Human Behavior, 52(11), 307–314. Xu, Z., Frankwick, G. L., & Ramirez, E. (2016). Effects of big data analytics and traditional marketing analytics on new product success: A knowledge fusion perspective. Journal of Business Research, 69(5), 1562–1566. Yang, F., He, X., Qiu, R. C., & Ling, Z. (2017). A data-driven approach to multi-event analytics in large-scale power systems using factor model. Retrieved from arXiv preprint arXiv:1712.08871 You, Y., Vadakkepatt, G., & Joshi, A. (2015). A meta-analysis of electronic word-of-mouth elasticity. Journal of Marketing, 79(2), 19–39. Zhai, Q., Cao, X., Mokhtarian, P. L., & Zhen, F. (2017).The interactions between e-­shopping and store shopping in the shopping process for search goods and experience goods. Transportation, 44(5), 885–904. Ziegler, C-N., & Skubacz, M. (2006).Towards automated reputation and brand monitoring on the web. In IEEE/WIC/ACM International Conference on Web Intelligence (pp. 1066– 1072). IEEE.

7  XPLORING WAYS OF E EXTRACTING INSIGHTS FROM BIG DATA Peter Steidl

Key Points The focus of this chapter is not on technical aspects but on the prerequisites that facilitate the successful exploration and use of data. Rightly or wrongly, I thought telling my own story rather than rattling off a series of facts and assertions may be a more engaging and compelling approach. I believe the following points are essential for any reader who is at the beginning or early stages of his or her career: • Hardware and software are changing—not just in incremental ways, but sometimes in disruptive ways that render previously learned and applied methodologies or approaches obsolete. For example, AI will replace statistical methods for a range of optimisation challenges. Regardless of how adept you are at analysis, you will need to also learn about AI as an alternative approach. • When addressing major challenges, you need to have domain-specific experience or at least a sound understanding of the challenges you plan to address. If you don’t have that grounding, you may miss important patterns and even fail to see the meaning and significance of (some of) your findings. • To be successful, you must be able to assist those who make decisions on the basis of your analyses in understanding their challenges and the best way to address them. Conventions, a narrow experience base, and even habits may lead these decision makers to hold on to outdated or irrelevant views that, if accepted without question, could shape your analyses in a way that will simply confirm pre-conceptions rather than identify new and better ways to address challenges. • You will undoubtedly have heard this before, but it is a crucial point:You will hold many different jobs and will need to engage in life-long learning. What

Exploring Ways of Extracting Insights  195

I am attempting to illustrate is that this is in fact not new. I have gone through a massive learning and re-learning process over the last few decades. But you will work in an environment where change is far more rapid and disruptive than what I experienced.This will make for an exciting career, but it will also be demanding, never allowing you to sit back in the comfortable armchair of your knowledge but instead always keeping you at the edge of your seat.

Learning Objectives • • • •

Understand the need to have a sound concept in mind when analysing Big Data sets. Understand the importance of domain-specific knowledge when interpreting the meaning and significance of your findings. Recognise that the challenges, the data available, and the ways of analysing it are changing and will continue to change over time. Understand the need to adapt and continuously learn about new methodologies and approaches to data analysis to be able to make a meaningful contribution.

Theory and Application Given the approach I am taking in this chapter, I need to combine theory and application. My point is that theory is only useful when applied in a meaningful, domain-relevant way using the most appropriate methodologies available to you. It follows that theory is not something absolute that’s given but rather is informed by the applications you have in mind and your past experience with various applications. In other words, your application informs your theory, and your theory provides the parameters and guides the choice of methodologies when it comes to application.

The Excitement of the Hunt When I completed my PhD studies at the University of Vienna I was totally entranced by what could be called ‘exploratory statistics’—i.e., using statistical methods to find meaningful patterns in large data sets. Using exploratory statistics is a bit like going on a hunt: one hopes to find something exciting such as a new insight or explanation. As any hunter will testify, frequent hunting moves you up the learning curve, and as you have more successes, you become more obsessed with the activity. Of course, there were many others who loved the excitement of the hunt for meaningful data patterns, and soon I was part of a small group of committed hunters who developed, exchanged, and tested new software that allowed for deeper and wider exploration. Looking back, I have no idea how we found each

196  Peter Steidl

other given that this was long before the advent of the Internet. I do recall, however, that there were a number of academics at the University of Michigan who became our heroes. To put this effort in perspective, let me mention that we had to use punch cards. To explain this to younger readers, it means you had to punch your programme instructions and data onto cards using a machine not unlike a large typewriter, submit the whole stack of cards (often several cartons) to the university’s central data processing unit, and then wait for anything from a few hours to a couple of days to receive a massive printout. After checking the output, it was back to the punch cards to prepare the instructions for the next run. I can only assume that the dopamine release that is typically experienced by hunters of any kind kept us all going. But I am not sure that anyone today would be prepared to go through such an elaborate process in the search for meaningful patterns. The following years saw the publication of my first book, Experimentelle Marktforschung, which explored combining experimental designs with multivariate statistical analyses. With clients, I explored the use of multivariate scaling and cluster analysis to identify brand positioning and segmentation options. While this may have looked like ‘Big Data’ at the time, it certainly was nothing like it by today’s standards. Nor was it online data, as the Internet was in its infancy. Fast-forward many tumultuous years of consulting assignments that took me to some 20 different countries, and I received an invitation to serve on the Board of the Institute of Multi-sensor Processing and Content Analysis (IMPCA) at Curtin University in Perth, Western Australia. IMPCA was involved in a whole host of interesting Research & Development projects, ranging from developing software for teaching autistic children to smartphone programmes that could guide a blind person by warning of obstacles ahead of them. I know, this is not news today, but at the time it was breakthrough stuff. But what captured my imagination more than anything was the Institute’s work on Early Warning Systems. The idea was to use the then already sizeable online data generated by social media to search for early warnings of the advent of significant events. For example, one of the pilot studies found a discussion on vibrating walls a couple of days before an unexpected volcano eruption. Clearly, lives could have been saved had anyone taken this as a potential indicator of a catastrophe. This was exciting. I was back in the hunt, not as a hands-on hunter but nevertheless feeling the excitement of finding new and different ways of making sense of large data sets. The data used for the Early Warning System consisted of raw news and data feeds, online posts and discussions, and secondary data prepared by specialised organisations tracking particular types of potentially catastrophic events. Importantly, a mood monitor (an early form of sentiment analysis) was developed to allow for identifying the mood of people based on their communications. I had meetings with representatives of airlines, insurance, and re-insurance organisations and the media, and everyone I talked to was enthusiastic about this

Exploring Ways of Extracting Insights  197

exploratory use of Big Data—a long time before Big Data became a fashionable thing to talk about. Clearly, these were much larger data sets than I had experienced before and— let me say it—there were much smarter people than me working out how to extract patterns that might serve as warning signs. But even sitting at ringside was exciting to say the least.

Learning From Small Samples, Validating With Big Data Fast-forward another decade, and I find myself being seriously involved in neuromarketing, advising companies and writing books to advance the field. Neuromarketing is based on medical research, providing us with an in-depth understanding of how the human brain works. Academics and consultants extract relevant findings from this research and repurpose it to provide marketers with actionable insights and concepts that have the potential to lift the effectiveness of their strategies and tactics. In other words, the origin of neuromarketing lies very much in investigating individuals or, at the most, small samples of people, rather than delving into Big Data. However, my interest in exploring Big Data never left me, and I started to wonder how we could combine the insights we gained when assessing the brain activities of individuals with those that large data sets might provide. In other words, apply a neuroscience lens to the analysis and interpretation of large data sets. For example, we know that consumers are unable to predict their own behaviour, that attitudes are fleeting, and that stated needs are simply rationalisations. (All this and much more has been presented in a compelling work by Nobel Laureate Daniel Kahneman with his book Thinking Fast. And Slow—most likely the most quoted book by neuromarketers today—and proven again and again in a multitude of experiments and analyses.) It turns out that asking consumers directly about their purchase intentions or how they feel about a product or brand, as in traditional market research surveys, is a waste of time. The marketing fraternity did know this—sort of—for quite a while, as it was discovered some 40 years ago that around 85% of new products failed despite consumers in group discussions and surveys claiming that they would buy them once released. This statistic has hardly moved over the intervening decades, and we have to conclude that consumers are unable to predict their own purchase behaviour with any degree of accuracy. So what can a marketer do? The obvious answer is, focus on actual behaviour. Not surprisingly, observation studies are more common in consumer research today, including the installation of cameras and the remote capture and analysis of facial micro-muscle or eye movements, thus combining physiological measurements with actual behavioural data. But there are even larger behavioural data sets generated by social media sites and the Internet in general. We all leave a trail when we visit a website, become a

198  Peter Steidl

follower, ‘like’ something, leave a comment, make a purchase, and so forth. Then there is all the transactional data the Internet captures when consumers buy online or use a credit or membership card in a brick and mortar outlet. Let me report on two developmental studies I have recently been involved in. The first was carried out with a top-five global communications group. We extracted behavioural data from a data lake that was fed by an acquired third party, as well as proprietary data, and set out to meaningfully segment the market on the basis of behaviour.

Case Example 1: Implicit Goal Segmentation Neuroscience research has shown that consumers buy brands and products to address a goal (or multiple goals), and most of these goals are activated in the consumer’s nonconscious mind. It follows that there is no point in asking consumers to tell us why they are buying the brands and products they buy—they can only offer us rationalisations, while the real drivers of their purchases exert their influence on the nonconscious. To gain a deeper understanding of the goals driving consumers’ actions, we set out to conduct an Implicit Goal Segmentation. The underlying rationale was this: as goals are what drives the consumer’s behaviour, we should be able to infer goals from the behaviour we can observe. This implicit approach is not dissimilar to looking out the window and seeing tree branches sway, leading to the conclusion that it is a windy day. We have not experienced the wind firsthand but have used an indicator to come to that conclusion. Needless to say, to have any chance of success, we needed to consider a wide range of activities and actions as some of these behaviours won’t be aligned with the consumer’s goals. For example, the selection of a television programme or a holiday destination may not be determined by the consumers whose behaviour we analyse, but by family members or friends. Or a person may not be able or willing to pursue their goals in all spheres of life. However, by including a wide range of activities we can expect to find patterns that indicate a particular goal at work. The analysis consists of a factor analysis using principal component analysis and a varimax rotation method with Kaiser Normalization. The rotations converged in eight iterations. This was followed by a cluster analysis. Eight segments were identified, and our interpretation was informed by the results as well as our understanding of consumer behaviour from a neuroscience perspective. I will briefly describe a selection of the identified segments to illustrate how we utilised neuroscience concepts to make sense of emerging market segments. But before I do this, you may require an explanation of the role dopamine plays. Dopamine is a neurotransmitter the brain releases when we experience something positive. Importantly, the dopamine release does not depend on an actual experience but can also be triggered by simply expecting or imagining a positive experience. When we get a dopamine release, we feel good, but this only

Exploring Ways of Extracting Insights  199

Implicit Goal Seg m ent s

Dop am ine Hunt er s

Ext er nal Locus of Cont r ol FIGURE 7.1 

Cont ext Aw ar e

Socially Ret ard ed

Lif e is a V id eo

My Hom e is m y World

Identified Dopamine Segments

lasts for a short while, and as we want to repeat this positive experience we either repeat behaviour that has triggered a dopamine release, or we engage in behaviour or thoughts that are likely to deliver another hit. Not surprisingly, then, scientists believe that dopamine is responsible for most of what we do. The other interesting feature of dopamine is that we become habituated to the level of dopamine we experience, encouraging us to seek even stronger levels of dopamine. Dopamine drives consumers to buy things they don’t need, to replace perfectly working products with new ones, to strive for something better and more advanced, and so forth. It therefore makes good sense to consider dopamine when segmenting the market. Now we are ready to have a look at some of the market segments (see Figure 7.1).

Dopamine Hunters EXTERNAL LOCUS OF CONTROL 16.2%, 18–34, Slight Female Bias This segment seeks dopamine releases through social interactions and, due to their age, this happens largely on the Internet. We know that social media in particular can serve as a consumer’s private drug lab. When visiting ‘their’ social media site, they get a dopamine release

200  Peter Steidl

when expecting positive feedback from others such as views, likes, comments, or another invitation to become a ‘friend’ or to link up. When this happens, another dopamine release is triggered. When posting or sharing something, the consumer is likely to again get a dopamine hit, as this raises expectations of a response, or of at least being recognised. This segment is also socialising in the real world, as demonstrated by high loadings on the factors Good Life, Fun Life, and Social Life. Typical for this age group is a lack of ethical buying, a lack of interest in cultural activities, and a lack of interest in charities or spiritual involvement. They are more directly focused on themselves rather than worrying too much about others or engaging with others in contexts where there is no direct payoff (such as being socially recognised).

Dopamine Hunters CONTEXT AWARE 8.5%, 45+, Slight Female Bias This segment is also looking for dopamine releases, but, given the age group it captures, they seek their dopamine boosts in different ways compared to the segment outlined earlier. Their dominant interest is in cultural activities, and, given that they exhibit a high loading for Good Life (which is a highly social activity battery), we are assuming that these cultural activities are typically engaged in with others or provide a common interest with other people. We can also see the effect of ageing with their interest in healthrelated matters and ethical buying. This is mainly due to the effect of having accumulated rich and diverse memories that allow them to understand broader contexts. Neuroscience research has shown that older people are wiser—i.e., they have a more extensive memory bank, more connections between memories—and this provides a much richer context that allows them to understand and address complex issues. The fact that younger people are less interested in Ethical Buying and Culture can be (partly) explained by the fact that it may require many years to develop the broader contexts that make these activities meaningful. However, we can’t ignore that ethical buying and possibly cultural pursuits can also give the feeling of being connected with others who engage in the same activities.

Exploring Ways of Extracting Insights  201

They use traditional media rather than the Internet—but that’s just what they are used to, rather than related to their hunt for dopamine. While the Dopamine Hunters are outgoing (whether online or in the real world), we found two segments that could be described as ‘socially retarded’.

Socially Retarded LIFE IS A VIDEO; EXTERNAL LOCUS OF CONTROL; PASSIVE 15.3%, 16–34, Male Bias The two dominant characteristics of this segment are watching video content online and following others on various social media platforms. This is not a sensory-rich segment, with a retarded social activity calendar, yet a need to ‘do the right thing’ by others. They are influenced by others’ opinions and ask others for opinions more than they get asked by them. Their lack of involvement with the world they are living in is clearly demonstrated by a low loading on Good Life, Cooking, Gardening (this may be due to low home ownership), Health Conscious and Smart Shopper.

Socially Retarded MY HOME IS MY WORLD 18.7%, Broad Age Distribution, Slight Female Bias This is also a socially isolated segment, but there are differences to the one described earlier. These consumers use the Internet and engage to some degree in social networking. The red thread that runs through their activities is that they are mainly carried out from home. They use the Internet at home, relax at home, listen to music at home, and spend much of their time with their family, children, and pets (although some of this may include outings). They have low loadings on Social Life and Good Life, don’t invest in their health or fitness, and are not heavy users of public or private transport.

202  Peter Steidl

How Useful Is This Segmentation? Our segmentation makes perfect sense in light of neuroscience research telling us how the consumer’s mind works. This may give us some confidence that we have found some meaningful patterns in the data we used. But it does not suggest that our analyses have been useful. Rather, their usefulness depends on key decision makers—in this case marketers—being able to make better decisions that contribute to their success. In other words, we have to ask ourselves if a segmentation based on drivers of behaviour is actually useful when it comes to the development of marketing strategies.

Dopamine Hunters There are many ways marketers can boost dopamine release by creating expectations, from package design to competitions, rewards, and the development of rituals that precede the enjoyment of consumption. Dopamine release can be boosted in retail outlets by announcing—or simply using cues to trigger the expectation of—surprises, to raising expectations online. Dopamine management should be an integral part of face-to-face selling efforts as well as advertising copy. So dopamine is important, but can a segmentation identifying two specific types of Dopamine Hunter allow marketers to target specific segments more effectively? This is the key question we need to address. The young Dopamine Hunters primarily use digital media to get their fix. Here we note two important results: First, they are used to a dopamine high and thus likely to get bored quickly if there is no high to be had. This means that dopamine management is particularly important when it comes to communicating with this segment. Secondly, their extensive use of digital media suggests that memory formation is less likely to occur and that they may not enjoy exposures quite as much as other segments, even when these exposures offer an engagement opportunity. For example, research has shown that taking pictures during a concert or of a meal in a restaurant is likely to diminish memory formation and enjoyment. You will hopefully find this interesting—but what can a marketer do with these insights? What are the practical steps that can be taken based on the understanding of this segment gained through this research project? Here are some suggestions when aiming at the young Dopamine Hunters: •

•

For every marketing initiative, the marketing team needs to consider if they can adapt the message or how it is delivered to squeeze a bit more dopamine out of this target group’s brains. It is more effective to create expectations than to focus on facts, such as creating an intriguing brand myth rather than a factual, informative brand story.

Exploring Ways of Extracting Insights  203

•

• • •

There is no point in dishing up meandering stories and long copy that does not create expectations. If a long story needs to be told, it is important to create expectations with the first paragraph to ensure they will stick with it. Where possible this target group should be given positive feedback, making them feel they are respected and important. Intriguing scenarios that leave room for imagination are likely to work well. This segment loves interacting on social media—they love to connect and see every message, ‘like’, and so on as a validation of their self, so it is a sound strategy to help them to get validation from others by giving them ‘insider’ information, something they can tell others about, can post on their Facebook page, and so forth.

The older Dopamine Hunters are quite different. First, they are not used to frequent dopamine release and demand.Their ‘hit’ needs more elaborate planning and action. Here are some observations with respect to marketing communications: There are plenty of contexts this segment relates to—cultural events, spending time with friends, staying fit and healthy, and ethical buying. It should not be difficult to select a context that has brand relevance. • Social engagement requires mentalising—i.e., the ability to gauge and anticipate the reactions of other people. Mentalising is something our mind does naturally—in fact, research has shown that the mentalising state is the same as the daydreaming one, and when we are not thinking about anything in particular, our brains start mentalising (obviously in our nonconscious). The use of social media does not require much mentalising as it is a very basic interaction with only limited exposure to the other party. But face-toface mentalising is not only demanding, it also encourages the formation of extensive memory patterns.

•

What does this mean? It means that marketers should consider wherever possible a social scenario in their communication, ideally one that can trigger mirror neurons (the neurons that allow us to feel what other people feel). This can be done through TV ads, online videos, movies, great copy, etc. Unlike the younger segment which is likely to be in a rush and gets bored quickly, this segment is likely to engage with longer stories and complex situations as long as they trigger the mentalising function.

Socially Retarded Let’s now move to the exact opposite to the two segments we have covered so far. Let’s have a look at the Socially Retarded consumers.

204  Peter Steidl

Make no mistake, dopamine is also at play here, such as when the “Life Is a Video” segment expects to find an entertaining video or the “My Home Is My World” segment hopes to find some interesting content on the Internet or looks forward to playing some great music. But unlike the Dopamine Hunters, these segments seem to get a lot less of the juice. And we can assume that their isolated existence is only partly choice, but mainly due to a lack of self-esteem, a fear of the new, a (perceived) lack of social skills, or simply due to their having so far failed to find something that gives them some purpose and direction—something they really like, want to engage with, get enjoyment from, look forward to. Of course, this is a hypothesis, but it is arguably not without foundation. We do know that social connections are important to well-being (see a multitude of Gallup studies, for example), that it is natural for people (especially women) to connect and build networks, and that mentalising (the brain working through memory patterns to learn how other people are responding to various situations and actions) is a key brain function nature seems to have favoured above others, to name just a few reasons why we might assume that the isolation we observe here is not necessarily the individual’s choice. Again, the most important question is, “Can marketers make practical use of these findings?” If marketers simply looked at these segments as having few interests and lacking the up-and-go to engage, they would probably write them off as being least likely to respond to marketing efforts. But when they see these segments as being in—let’s call it ‘temporary suspense’—they will realise that these segments may in fact be highly receptive to messages of the right kind. What a brand needs to do here is to show an easy, non-demanding way to engage that can evolve into a long-term relationship. The goal to engage and be part of more may already be active—but no way of addressing this goal has been found—or it may need to be activated. Again, we could refer back to dopamine: the expectation of a positive and manageable interaction is likely to release a hit of dopamine. But we have to assume that there are specific barriers we need to overcome. With the “Life Is a Video” segment, it is quite clear how a marketer can reach them—but what should the communication be all about? First, it needs to be a message that is supported by others as this segment is influenced by the opinions of other people. This could range from the more clinical “the majority/x percent of people say that this product is . . . ” to having characters to tell the story. Several studies have shown that social validation can have a significant impact on purchases. Secondly, marketers could encourage this segment to engage with something they have not engaged with before. Here they do want a bit of dopamine release—something that promises to be exciting—but at the same time, they must keep in mind that these people are most likely isolated for a reason rather than by

Exploring Ways of Extracting Insights  205

choice. Small steps are most likely required, and these can most effectively be built into an online engagement. Thirdly, marketers can encourage them to buy into something that would trigger another goal, assuming that they are isolated because they haven’t yet found anything that truly engages them and that they feel comfortable doing. Clearly, it will depend on the brand and what it stands for as to what this activity might be.

Further Research There are some promising patterns emerging from the Implicit Goal Segmentation. There are clearly some parallels between insights gained into how the brain works through neuroscience research and the behaviour-based segments we identified. More analyses are required to validate some of the research hypotheses we have formulated on the basis of these initial results. However, the main question is whether Implicit Goal Segmentation can help marketers to more effectively shape consumer behaviour. In other words, is an Implicit Goal Segmentation likely to come up with relevant insights that can guide the development of marketing and communications strategies? I am optimistic. Segmentation needs to be an exercise in trial and error. There are infinite ways of segmenting a market. Marketers need to find a way of segmenting the market that creates meaningful and actionable segments their brands can target effectively. The Implicit Goal Segmentation represents a much better starting point, as it is not based on rationalisations: When we ask consumers about their likes, needs, attitudes, reasons for purchase, intention, and preferences, we simply get rationalisations. The consumer does not know what is going on in their nonconscious mind and thus cannot tell us. Of course, Implicit Goal Segmentation does not give us a direct insight into what is happening in the consumer’s nonconscious mind either: It simply allows us to infer the goals that are driving observable behaviour. However, at least we have some reliable data to start with and we can utilise the extensive neuroscience research available to us to interpret this data.

The Importance of Domain-Specific Knowledge My extensive coverage of neuroscience insights into how the brain works has not been accidental. Rather, I am attempting to highlight the importance of domain-specific knowledge. Without a sound understanding of marketing challenges, strategies, and concepts, I would not have been able to develop a protocol for analysing the data nor to interpret the findings in a relevant way that allowed for the operationalisation. In other words, domain-specific knowledge was a key contributor to the project’s success.

206  Peter Steidl

Sometimes, the challenge is not to explore a data set but rather to validate a hypothesis and operationalise the theory it is based on. ‘Operationalising’ means translating the theory into specific actions that can be followed—i.e., to turn the thinking concept into a practical approach that can be implemented. Again, this is not possible without domain-specific knowledge, as the following case illustration shows.

Using Big Data to Validate a Hypothesis and Operationalise the Underlying Theory The segmentation study described earlier is a typical example of exploring data to see if we can find patterns that deliver new insights or a new way of achieving an objective, in this case to meaningfully segment the market in a way that can provide a solid foundation for the development of marketing strategies. Another use of Big Data is to validate a hypothesis.

Case Example No 2: Re-targeting When Sharing A few years after working on the segmentation study, I had an opportunity to delve further into the role of dopamine and how it can be used by marketers to shape purchasing behaviour by analysing large data sets. This time I was working with a team at RadiumOne, a global digital agency that has one of the largest data lakes in the industry by accessing not only social media but also Dark Social. ‘Dark Social’ sharing is the copying and pasting of content and links from websites into email or instant message services and selectively sharing it with friends, family, and colleagues. For each Dark Social share, the content and the people it is being shared with are selected for a very specific reason. This is very different to the data collected from public social networks, where sharing centres around what an individual considers suitable for public examination. Most sharing of content—some 75%—happens in Dark Social, which includes sharing that occurs across private and more intimate channels, such as email, instant messenger, WhatsApp, and text. Public Social—including Facebook, Instagram, twitter, Pinterest, and other social media sites—accounts for only 25% of content shared. Importantly, we need to include not only messages but also links and articles shared via email and messaging services. How big a data lake would this be? There are 204,000,000 one-to-one or oneto-few implicit conversations every minute! Not all of these are being captured, but the data the case examples I will present consist of social (paid & organic), email blasts/CRM, blog posts, SMS, mobile, press releases, QR codes, talent and brand ambassadors, sponsorships and events, offline (outdoor and print), ad tags, and video embeds. In addition to that, there is also first-party data, including 300,000 publishers globally. Having all this data, what should one do to extract the most value, when value is defined as being able to shape purchase decisions? One way of dealing with it

Exploring Ways of Extracting Insights  207

would be to segment the market by identifying the purchase drivers shared by groups of consumers, allowing the marketer to target each segment with messages that activate the segment-specific purchase drivers. This would follow the path outlined earlier when working on the communications group’s challenge. However, there is a big difference: In the previous study, we downloaded a data set for analysis. In other words, the data we used was static—once downloaded for our analysis, it was frozen in time. Here we have dynamic data that is delivered in real time. Remember the 204 million conversations every minute I mentioned earlier? Not surprisingly, we wanted to take advantage of the immediacy our access to data offered. First, we had to develop a conceptual framework that would allow us to boost the effectiveness of marketing communications. We identified three drivers of behaviour that can be expected to affect the receptiveness of consumers to marketing communications: 1. Dopamine, the neurotransmitter you are already familiar with 2. Goal congruence—i.e., a situation where the content of our communications is aligned with the consumer’s goal 3. Endowment effect—i.e., the concept of consumers valuing something they have invested in and thus made ‘their own’ more than something they have not yet made a commitment to.

Dopamine You will recall that dopamine is a ‘feel good’ transmitter. Our brains reward us with a pleasurable release of dopamine when we successfully achieve something. Importantly, dopamine does not make us feel happy; rather, it pushes us to seek happiness. In other words, when we get a dopamine release, we feel great, but as the level of dopamine subsides, we want to get another dopamine hit. Neuroscience tells us that dopamine release in the brain is at its strongest when we are expecting something positive to occur. Our expectations—our ability to imagine the reward—typically have more impact than the actual experience itself. This is a key driver for people who buy lottery tickets. The expectation, in fact the mere possibility of a win, triggers a dopamine release. Of course, using social media as one’s personal dopamine factory is cheaper than buying lottery tickets. Every time we post, share, like, comment, or send an invitation online, we are creating an expectation.We expect a response; we expect recognition; we expect to be valued and acknowledged by others. We like to be involved in sharing and engaging with the latest news and the most interesting stories that are relevant to our lives. We feel like we ‘belong’ and advance our concept of self through sharing. But this is not where it ends. Every time a person responds, likes something, invites us to an event or comments on our post, we get another dopamine hit. This is why people share information online and post on social media. Social

208  Peter Steidl

sharing provides the same release of dopamine that we get from other pleasurable experiences like sex, eating, and exercise. Dopamine is the reason people visit their social site of choice as soon as they wake up in the morning, before they go to sleep at night, and hundreds of times in between. Fifty percent of Australians check social media first thing in the morning and last thing at night.1 Guided by a solid theoretical framework delivered by neuroscience research, we wondered if re-targeting consumers when they share would boost their response rate. Re-targeting consists of sending a consumer who has shared content a message that relates to that content. More specifically, our assumptions were the following: As the level of dopamine declines, the sharers are likely to seek another dopamine hit. If we send messages that are aligned with the content that has been shared, we have every chance that the recipient’s nonconscious mind will see the message as being a means of getting the desired dopamine hit (the same type of content would suggest a repeat of the dopamine release is likely), and this will make the recipient more receptive to our communication.

Goal Congruence There is another factor at work when re-targeting consumers when sharing: goal congruence. To be successful, marketers need to convince consumers that their offer addresses one or more of the consumer’s goals. This is typically referred to as goal congruence. This applies in the real world as much as in an online environment. Research has shown that consumers on social media sites are receptive to new news, gossip, stories they can share, and so on. On the other hand, when consumers search, their goal is to find an answer, and they therefore welcome anything that allows them to find what they are looking for faster or with less effort. They tend to be less open to diverting their attention to other matters.2 When they are on a shopping platform, such as eBay, Alibaba, or Amazon, they tend to be more receptive to communications that present information related to special offers because they are in buying mode. In summary, goal congruence can have a significant effect on the responsiveness of consumers to messages.

Endowment Principle Neuromarketing research has shown that people who take a small initial step are more likely to take a bigger step afterwards. For example, animal shelters have great difficulty in adopting out kittens. Uber ran a promotion on National Cat Day in February 2015, inviting customers to order a ‘Kitten Car’ through their app: a kitten from a local animal shelter would arrive at the consumer’s door to play for 15 minutes. The intentions around the promotion were to activate a fun way to promote feline welfare. However, for those people who took the ‘first step’

Exploring Ways of Extracting Insights  209

(bringing a kitten into their home or workplace), it increased the likelihood of attachment and, in turn, the successful adoption of a kitten in need. The campaign had exactly that effect. Globally, across all cities, participating shelters saw increased donations and adoptions. What this tells us is that taking a first step—regardless of how small it might be—will create an affinity with the event, brand, or activity. The consumer becomes more receptive to taking another bigger step in the same direction. Sharing content is a small initial step. It costs the consumer nothing, triggers a dopamine release, and offers another dopamine hit when someone responds in a positive way. At the same time, the act of sharing makes the consumer more receptive to taking a bigger step—namely, purchasing a brand’s product or service. So much for the theory. Let’s now have a look at some of the results of our analyses to see if these support our conceptual model.

One Direction We took advantage of an Australian tour by One Direction to see if there was any evidence that re-targeting when sharing would lead to a higher response rate and levels of engagement. After an initial burst of ticket sales when the tour was announced, and a new album released, Ticketek capitalised on the social buzz to sell the remaining tickets to One Direction shows by re-targeting consumers who had shared One Direction (related) content. The increase in sharing behaviour and the associated re-targeting correlated directly with a spike in ticket sales. Ticket transactions and sharing continued steadily in the lead up to the tour.

Conversions Shares of One Direcon related content Conversion Rate %

Time

210  Peter Steidl

Sharing opportunities represent the small steps we want consumers to take to be more receptive to taking a bigger step—to purchase or engage with the brand in a meaningful way.

Tigerair This next example is based on airline ticket sales for Tigerair, a budget carrier operating in Asia Pacific. We can safely assume that when intending to book a flight for non-business purposes, consumers already have a relevant goal in mind—a holiday, wanting to experience something new, repeating a positive travel or flight experience, wanting to meet up with people elsewhere whose company they enjoy, or having a romantic or exciting time away. Airfare sales can always be effectively promoted during certain periods, such as school holidays and long weekends. However, re-targeting is an opportunity to boost response rates during these and other times when consumers share content related to their travel plans.

Operationalising Our Theory of Targeting When Sharing This is an important finding but by itself it is not really helpful. To use this approach we need to know when we should target consumers: Immediately after they share? Ten minutes later? An hour later? Several hours? To operationalise our findings into the effectiveness of re-targeting we need to assess within which time frame the re-targeting needs to take place to have the strongest impact. An analysis of responses to re-targeting showed that the ‘sweet spot’ to conversion was within the first 20 minutes, where 30% of all sales occurred. Targeted advertising delivered within the first hour after a consumer shares relevant content results in conversion rates up to seven times higher than in the following hour.

Conversion latency: One-third of conversions happen within the first 20 minutes.

Exploring Ways of Extracting Insights  211

Conclusions These and similar analyses have delivered a new approach that promises to boost the effectiveness of marketing communications, thus boosting the Return on Marketing Investment. More work needs to be done of course. In particular, we need to see split-run studies using re-targeting with a random sample of the target group and a traditional approach (such as placing ads on Google and other search engines when the search contains relevant keywords) with another sample. However, the findings so far are most encouraging and re-targeting when sharing is, in fact, being used by a growing number of marketers competing in categories where frequent sharing is common such as sport, entertainment, and travel.

A Major Disruption Is Emerging: Artificial Intelligence There are typically a number of challenges involved in analysing complex systems. First, we may have sets of very different data.There may be numerical data based on rating scales, rankings, binary answers, multiple choice questions, and so forth. We may have answers to open-ended questions and even images. Perhaps we even have some biometric data derived from self-quantification gadgets.The list goes on. Take for example the health insurance industry. Organisations typically have huge data lakes but struggle to put this data to use. One of the challenges health insurance organisations face is to keep their members healthy, leading to lower payouts, which in turn facilitates lower membership fees, resulting in a stronger competitive position in the market place which will attract more members. This is quite obvious, but how can the health of members be boosted? What should we analyse, and how should we do this? Here I will introduce an approach that is very different to the statistical analyses you are most likely used to or learning about: The use of AI. This is the approach I would take: I would start with interventions (based on our best understanding of how we can shape behaviour delivered by neuroscience, neuropsychology, behavioural economics and other disciplines), testing alternative interventions with a large sample of customers—i.e., members of a health insurance firm. I would feed all the data I have on these customers into an AI engine, including self-monitoring data members have made available, demographic, and socioeconomic data, location, purchases, past health insurance claims, and so on. With respect to the AI engine, I would use a back-propagation artificial neural network. The AI engine will learn how different sets of background data best predict the outcome—i.e., a successful change in lifestyle. (I will keep this simple—in reality, we need to not just focus on an initial change, but on maintaining the desirable behaviour.) More specifically, the AI engine develops a quasi-neural network that explains why an intervention has been successful with the first health insurance member

212  Peter Steidl

going through the programme. It then realises that this neural network does not explain what happened with the second participant, and it goes back to the artificial neural network is has created (which is why it’s called back propagation) and adjusts it to make it work. And it keeps doing this with every additional health insurance member, creating a massively complex artificial neural network. The important point is that the AI engine will learn from each case and update its neural network to improve its predictive power. This is obviously quite different to carrying out a statistical analysis that creates a ‘fixed’ model designed to predict outcomes. This model does not learn, although the predictions will of course differ when fresh input data is being used. Here is the clincher.When you analyse a large data set using a statistical, multivariate approach, you are clear about the ‘model’ you are using: The role different variables play, the correlations and interactions that have created certain outcomes, the relative importance of a multitude of variables in delivering this outcome, and so forth. In other words, by examining your model a third party can learn how you have analysed the data, what assumptions you have made—implicitly or ­explicitly—and to what degree you have simplified to reduce complexity by making decisions on what is and isn’t important to include in your model. In other words, a third party can make a judgement as to how reliable they feel your results are. With AI, that’s not possible. First the complexity of the artificial neural network is too great to grasp it. But even if we could grasp it, it would be of little use as the neural network changes with every new case! This is, of course, the strength of AI—it learns all the time, improving the capability of the artificial neural network to generate optimal outcomes. So with AI, I can tell which interventions we should use with a particular member based on the huge data set to which it has access. But it won’t provide us with an understanding of how it worked this out. With trivial applications such as the one I have outlined here, this may not matter much, as we can see how successful the recommendations made by the AI engine are. But what if we want to use AI to make decisions on significant financial investments, to predict the likely performance of a job applicant or to recommend a treatment protocol for a patient? In the first case, we risk a significant financial loss; in the second, the AI engine may disadvantage applicants who would have done a great job if the recommendation is not optimal and in the third case we risk the health of our patient. This is why anyone investing in optimisation using AI needs to seriously consider if they are prepared to trust the resulting optimisations or not. Here is something else you need to be aware of: AI engines need to be trained using real-world data. Any biases inherent in this data will shape the recommendations made by the AI engine. This is why the use of AI has led to gender and race biases when used by social networks or search engines. As these biases are present in the ‘training data’, they flow into the artificial neural networks.

Exploring Ways of Extracting Insights  213

Having said all that, I note that the use of AI to optimise outcomes is spreading rapidly amongst major corporations and, not surprisingly, especially in sectors that are driven by data such as financial services, insurance, health services, logistics, transportation, and manufacturing. My view is that in the medium- to long-term, more and more optimisation tasks will be carried out by AI engines. But when it comes to evaluation, I can see a sound statistical approach being superior, as we do need to know not only what did and didn’t work but also why, and AI falls short when it comes to telling us about the why. This is my final report on the progress of a pilgrim seeking insights into why things happen and how we can shape these events. Currently, I am working with clients on the use of AI in various applications. This means shedding some of my curiosity as the AI engine is not going to reveal to me what the patterns and cause-effect relationships are. At the same time, I believe that in some applications the use of AI rather than a more traditional statistical approach will deliver better outcomes. And in my sleepless nights, I ponder what might happen when engineers and scientists finally develop an Artificial General Intelligence that will be smarter than all of us put together.

Troubleshooting Tips Context One of the best-selling management books of all time was In Search of Excellence: Lessons from America’s Best-Run Companies, by Thomas J. Peters and Robert H. Waterman Jr. (2002). In their book, the authors identified leading companies and set out to identify what made them famous. A few years later, many of the companies they identified as ‘excellent’ were struggling, had gone under, or been take over. The authors later defended their approach by saying that they did not write a book on ‘Forever Excellent’—but the fact remains that their guidelines were of limited use. The same applies to most case studies. Executives love case studies highlighting the success of a firm, product, market entry strategy, or whatever else. They often try to emulate these strategies only to find that success does not result. The problem is that the contexts are different. The same applies to successful executives. There are many examples of executives moving to a new position only to fail or end up being very average at best. They were the right person for the particular context they worked in previously, but they may not be the right person given the context of their new role. A question we always, always have to ask ourselves when analysing data is to what extent the findings are context specific when we apply insights, concepts, strategies, or interpretations we have extracted from analysing a particular data set. This requires judgement beyond analytical skills.

214  Peter Steidl

Historical Data Always keep in mind that the data you are working with is historical. It reflects what happened in the past, whether this be the recent past with real-time data that may be only seconds or minutes old, or the distant past when using data that reaches back many years. Analyses typically aim at delivering insights that inform future strategies and plans. In some instances, our algorithms do the work and project trends and predict developments. In other situations, we may get insights that explain relationships that help us to make more accurate predictions about the future. In either case, we are basing our views of the future and how we need to act in that future on what happened in the past. This is quite obvious, but it is equally obvious that we need to ask ourselves if our data is still representative when we project into the future. Unfortunately, at least when it comes to management decisions, this is rarely a consideration. Many companies routinely use outdated data that no longer represents today’s—and, even less so, future—conditions. They use legacy data they have used for a very long time. Familiarity and past success in using this data seems to be sufficient to convince them that it is appropriate to use these data sets in the future. As a general rule, it is a good idea to consider if a disruption may have rendered past data irrelevant. Of course, what is and isn’t a disruption depends on the focus and context of your study. When exploring the most effective marketing mix at retail level by conducting an econometric analysis that takes into account in-store promotions, pricing, product range, shelf facings, and other variables, you may decide that an aggressive new competitor entering the market, a radically new product being launched, or a shift towards online shopping may constitute a disruptive event that renders your historical data useless. This is a judgement you need to make as the most professional approach to data analysis is not going to deliver relevant outcomes if the data is not representative of the environment your outcomes will be applied to.

Meaning A few years ago, Google announced that it could track the spread of influenza across the United States without needing the results of medical checkups. Rather, they would rely on people’s online searches for flu symptoms and treatment recommendations. Google’s ingenious approach would be faster and less expensive than that used by the Centers for Disease Control and Prevention, which relied on surveillance reports from some 2,700 healthcare centres located around the United States. It did make sense that people would search for flu symptoms and treatment options, and that this could substitute for medical tests. However, Google’s assessment was a long way off reality. It overestimated flu prevalence by more than 50%.

Exploring Ways of Extracting Insights  215

What does it mean when people search for flu symptoms and treatment options? How many people usually search on behalf of a single flu victim? Would several family members search? Would the patient or others search multiple times? And how similar is the average person’s definition of ‘flu’ to that of a clinician? We don’t know the algorithms Google (presumably) used to adjust the raw data generated by searches, but we do know that the results were too unreliable to be of any practical value. Sometimes the meaning of data seems quite obvious, but it is always important to identify and document any assumptions we have made and to revisit these assumptions again and again, particularly when drawing conclusions from the findings.

Making Sure the Users of Your Results Understand Their Limitations Anyone who analyses Big Data sets on behalf of a third party needs to fully understand the decisions these parties plan to make on the basis of the results. Your analyses may only be one of several factors these decision makers will take into account, but it is still important for you to understand the challenges they are facing and what they hope to gain from the analyses they ask you to undertake. It is not unusual for these decision makers to ask for a specific focus or even a particular methodology. Their choice is typically based on a particular concept they have in their mind or past experiences with particular types of analyses. In other words, their stated preferences or instructions may not be the best way to address their requirements. It is important for you to help them explore options and to shape their views, so they settle on the most effective approach to the data that will be analysed and the methodologies that will be applied. It is equally important for you to understand the challenges the analyses you undertake are supposed to address, as this will allow you to interpret the results in a meaningful and relevant way. I have seen dozens of cases of research experts making recommendations that showed a total ignorance of how specific challenges could be addressed. In other words, when you undertake a major project, you will need some domain-specific knowledge to do it well. If you want to make a useful contribution and shape important decisions, you need to understand the relevant domain. Being a brilliant analyst doesn’t help, as you may analyse the wrong data, in an inappropriate way, and draw the wrong conclusions because you failed to understand the core issues that need to be addressed. Importantly, you can’t rely on the decision makers who brief you to fully understand what can and can’t be done, nor to even understand their challenges as they are often looking at these from a narrow perspective.

216  Peter Steidl

Reverse Correlations There is hardly a week that goes by without some agency or consulting firm suggesting that companies need to invest into a particular service or product as research has proven that companies that do so are more successful than others that don’t. Of course, more often than not, the assumed causality works the other way. For example, digital marketing agencies quite often claim that companies that use digital media more extensively are more successful and thus other companies need to follow suit.The reality is that companies that are more successful—due to a variety of reasons, such as dominant brands names, better distribution, aggressive advertising, long-standing relationships with retailers or other intermediaries, better products, lower prices—are more likely to spend money on a variety of different media options.They are spending more in digital because they are successful and can afford it. But their digital spending is not the cause of their success—it is the other way around!

Inability to Fully Understand Complex Systems Helga Novotny, eminent academic and president of the European Research Council during its formative years, had an opportunity to review and assess a multitude of studies spanning a wide range of disciplines and applications. In her recent book The Cunning of Uncertainty, Professor Novotny refers to an embarrassment of complexity, as the feeling of not being in control when circumstances indicate that one should be. “Even managers whose job it is to reduce complexity need to act as if they can grasp the whole when what they actually see is only one part” (Novotny, Helga, The Cunning of Uncertainty, 2016, p 129). It would be easy to reduce complexity if we understood the whole system and thus could make sound decisions on which aspects of the system are important and need to be considered, and which can be ignored. However, the very desire to reduce complexity is based on our inability to grasp the whole system. So we end up in a Catch-22 situation. As Novotny points out Complex systems are beset and energized by their nonlinear relationships between the variable interactive parts and dimensions of the whole. Therefore, what makes complex systems complex are their multiple feedback loops and indirect cause-effect relations, which play out at different timescales and speed. The numerical reduction of complexity offered as a management tool is seductive. It promises to reduce complexity and does so. But complexity does not disappear and predicting the behaviour of complex systems, let alone understanding them, is much harder than processing data with computing power. Data alone are not enough. Big Data are necessary to further the understanding of complexity, but not sufficient.

Exploring Ways of Extracting Insights  217

Much depends on the kinds of questions asked and on the place assigned to the human being in numerically run but complex systems. I leave you with these thoughts!

Student Exercises 1  ANALYSING A LARGE DATA SET Exercise 1 Define a specific challenge and develop three different approaches to addressing it. Identify the strengths and weaknesses of each approach with respect to the following: • • •

The appropriateness of the data The methodologies you suggest The relevance of your results, i.e., to what extent they will contribute effectively to addressing the challenge.

First, make a recommendation on which approach should be adopted and support your recommendation. Second, explain why the other methodologies you have considered fall short, leading to their elimination. Exercise 2 For the same data set, address the following questions: Do you understand the structure of this data set—i.e., how different subsets of data relate to each other? Do you believe you can shed light on these relationships with your analyses? How would you do this? Have you checked the relevance of the data given your research objectives? What are you trying to prove, or which decisions are you attempting to support? Does the data you have allow you to do this? Have you considered the timeliness of the data you are using? Given your research objectives, are you convinced that the data is relevant? Have there been any disruptive events that might render

218  Peter Steidl

some of the older data irrelevant? Has there been continuous incremental change that has added up to a very different scenario today, rendering some of your older data irrelevant? Have you considered if the relationships in the data you have identified are simply correlations or are they cause-and-effect relationships? Consistency is not a sign of the latter, as there could be a systemic error in your thinking. You have to use logic to test any potential cause-and-effect relationships. Look for reverse causation and for higher level relationships that explain what you have found. 2 DECIDING ON THE USE OF AI OR STATISTICAL ANALYSES Here are three challenges executives face. For each challenge recommend the use of AI or Statistical Analyses. When recommending Statistical Analyses specify the core methodology/ies, when recommending AI provide a brief outline of how AI would be used. 1. “We get hundreds of applications when we advertise a job. We have an elaborate interview process to make sure we offer the job to the right person—but experience has shown that we often get it wrong. How can we improve our selection process?” 2. “We have extensive historical in-store data covering a multitude of retail outlets and including data on our—and our competitor’s— range, number of shelf facings, pricing, in-store promotions, special offers and spending on advertising campaigns. We are not sure how we should allocate our limited budget. What should we spend more or less on to boost our revenue?” 3. “As a health sciences firm, we have gathered and bought selfmonitoring data for hundreds of thousands of individuals, showing how much they exercise, their weight, heart rate, fitness level, and more, matched with demographic and purchasing data covering a wide range of goods and services. The challenge we face now is to find ways of changing the behaviour of these people—i.e., to get them to shed ‘bad’ behaviour while developing good habits. How can we use the extensive data we have to support us in this endeavour?”

Exploring Ways of Extracting Insights  219

3  KEEPING TRACK OF YOUR PROGRESS While this is not an exercise you can complete and put aside, it is, in my mind, one of the most important exercises you can undertake. It is very easy to lose sight of how your strengths and weaknesses develop. Many people find out one day—typically much too late—that they have lost touch with developments and that their skills are now hopelessly out of date. Many would say that they are committed to continuous improvement, but few monitor their relative capabilities and experience and ensure that they improve these. Here is how you can avoid falling into this trap: After completing a significant project, you should undertake a self-assessment. This should not be a formal, rigid assessment, as there is every chance that the questions you raise in such an exercise will eventually no longer reflect what is important. Rather, focus on core issues and explore for each of these where you stand. For example, you might explore the following core issues which are not likely to change: •

My ability to truly understand the challenges my analyses need to address.

It is, of course, not necessary to always fully understand a complex issue, but you need to understand the specific challenge you are supposed to help address. If the application area is new to you, it may be useful to search for information, read a few articles or books, or simply spend extensive time with the person briefing you on the project to make sure you fully understand the challenge. Do not assume that asking many questions will diminish your standing. On the contrary, it will demonstrate that you really want to make a useful contribution. •

My ability to select the right methodology rather than being limited by only a working knowledge of a narrow set of methodologies.

Note that the question is not if you have used the right—or at least an appropriate—methodology, but rather if you have an extensive tool kit of methodologies you can utilise, and if you have explored various options rather than just repeating blindly what you have done in the past.

220  Peter Steidl

•

My ability to carry out an analysis, explore the meaning of each analysis, and adapt my approach in light of what I learn on the way.

Here your challenge is to explore and learn from early results, allowing you to refine and to explore interesting patterns or relationships further. •

My ability to interpret the results of my analysis in a meaningful, relevant way that allows the users of my results to make informed decisions or gain useful insights.

The circle closes with this final core issue: You will have a good start if you have ensured that you fully understand the challenge your analysis is supposed to address, which is the first core issue listed earlier. But the end result, of course, depends on getting the process, choice of methodology, and interpretation right as well. •

Conclusions: Finally, you need to draw some conclusions that will give you clear direction as to what you need to do before—and during—the next project to broaden and deepen your capabilities. What do I need to learn (more) about? What do I need to do differently? Are there bad habits or routines that hold me back? Where have I improved since my last review, and where have I regressed?

What you should do as a matter of routine is to answer these questions truthfully and record the key points—i.e., why you believe you have done well or what you have failed or been unable to do. Spending an hour after each major project will, over time, lead to a deep understanding of how to manage your career, how to capitalise on your strengths and how to correct weaknesses that hold you back. This may look like a simple exercise to do—and it is—but it is difficult to muster the discipline to keep undertaking a review after each major project. Sometimes you will find that you are extremely busy and at other times that you are feeling low or doubtful about your career, and these and other factors will be your enemies when you try to adopt this approach.

Exploring Ways of Extracting Insights  221

Notes 1. Sensis Social Media Report May 2015 www.sensis.com.au/content/dam/sas/PDF directory/Sensis_Social_Media_Report_2015.pdf 2. Note that consumers sometimes search to explore—for example, to explore holiday destinations or cars. In this case, they fall into the exploratory group as their goal is not to get a quick answer but to indulge in exploratory behaviour.

8 CONTEMPORARY APPROACHES TO MODELLING THE CONSUMER Debbie Isobel Keeling

Modelling plays an important role in consumer psychology and is used to understand consumers across diverse contexts. For example, modelling of consumer profiles, loyalty and switching behaviours can support marketing decisions around the launch of new brands. Consumer psychologists design campaigns to encourage healthy living routines based on modelling of consumer attitudes, peer influences and individual differences in personality and health status. Others use modelling to effectively segment their consumers into identifiable groups as a means of tailoring marketing communications, product offerings and recommendation systems. Further, modelling is fundamental to understanding consumer decision making, such as understanding risk-taking behaviours amongst consumers in investments, health, travel and so forth. Whatever the application, modelling as a term in consumer psychology should be understood as a whole process, encompassing defined objectives, design, data collection and analysis, for the development and/ or testing of theories relating to why consumers behave as they do. The result of this process is a ‘model’ that can be expressed in multiple forms, including path models, equations, charts, matrices etc., which provides a holistic overview of the relationships between key influencers (variables) on consumer behaviour. Why is modelling such an important process for understanding consumers? The generic value of modelling in consumer psychology is threefold. First, it provides vital insights into the complexity of the psychology of consumers’ behaviours. There are multiple variables that influence consumers’ behaviours, the role of modelling is to make sense of this complexity by identifying the most influential variables on behaviour and exploring how these variables work together to produce behavioural outcomes. Second, modelling offers as an outcome an integrative framework, namely a model, that both captures the relationships between variables and informs understanding of how to influence consumer behaviours.

Contemporary Approaches to Modelling  223

By identifying the most important variables and also the (combined) level and nature of their influence on behaviours, interventions can be designed that utilise these variables to optimum effect. Third, the resultant models offer insight into how to assess the success and/or impact of such interventions on consumers’ behaviours. This is achieved both by identifying the key variables to measure and assess, and providing guidance on the expected level of impact of each variable on behaviour. Whilst I have stressed the value of modelling in terms of understanding consumer behaviours, these behaviours can be diverse. For example, they can be direct observations of performance behaviours (e.g., sales, learning, fitness levels), use or consumption behaviours (e.g., purchase, repurchase, store visits, application of the product or experience) and choice behaviours (e.g., choosing between insurance policies, brand preferences). Proxy measures in the form of self-reported behaviours or intentions are used when observations of actual behaviour are not readily available. Of course, consumer psychology models usually include some combination of emotional, cognitive and behavioural variables alongside social, demographic and geographic variables. To demonstrate the three values of modelling, let us take a look at a classic conceptual model—the Theory of Planned Behaviour (TPB) devised by Ajzen (1991). Figure 8.1 shows the TPB in its simplest form. This is an example of how a model can be used as a visual representation of a theory, where the key influences on behaviour are indicated and the hypothesised relationships between those influencers represented. (N.B.: see later for a note on conceptual versus statistical models.) First, the TPB model identifies that four important variables (attitude towards the behaviour, subjective norm, perceived behavioural control and intention) are influential on behaviour. Through ordering these variables, the model also indicates the relative relationships between these variables and with behaviour. Second, the model visually represents that three variables (attitude towards the behaviour, subjective norm and perceived behavioural control) shape intention, which subsequently shapes behaviour. Perceived behavioural control also has a direct impact on behaviour. From this model, we know that to influence behaviour, we can narrow our interventions down to these four variables. For example, in campaigns to stop consumers smoking, our interventions would not only require efforts to change consumer attitudes towards smoking but also the social context and associated perceived norms, plus the individual’s perceptions of their ability to give up smoking. Third, in tandem, we can also use these variables as a means of evaluating the success of our interventions and where they succeeded (or otherwise). Contemporary developments in modelling have particularly focused on increasingly sophisticated methods of analysis and supporting software, including the increasing availability of specialist open access software. These developments in turn have facilitated advancements in the specification of models aiding the description, prediction and explanation of consumer behaviour. In this chapter,

224  Debbie Isobel Keeling

Attitude towards the behaviour

Subjective norm

Intention

Behaviour

Perceived behavioural control

FIGURE 8.1 

The (Simplified) Theory of Planned Behaviour (Ajzen, 1991)

we will focus on the broad category of standard statistical modelling (mainly written from a frequentist perspective) popular within the field of consumer psychology. The aim of the chapter is to give you an overview of the underpinning principles and concepts that guide modelling. As such, the diverse goals and key building blocks of standard statistical modelling are explained. Offering two practical examples: technology adoption and the health consumer, segmentation and profiling in a B2B context, plus troubleshooting tips for using modelling in consumer psychology and identifying some of the statistical packages available to support you.Where relevant you are guided to more in-depth readings to explore specific analyses or issues. The chapter ends by reflecting on the future directions of modelling in consumer psychology and how these new directions will challenge and shape our thinking about the nature and application of modelling.

Learning Objectives After reading this chapter, you should be able to: 1) 2) 3) 4) 5)

Explain the standard statistical approach to modelling Distinguish the diverse goals of standard statistical modelling Understand the building blocks of this modelling approach Address some of the pitfalls when using modelling in consumer psychology Understand the scope of software on offer to help you with the modelling process 6) Identify the future directions of modelling in consumer psychology and how these are challenging our perspectives on modelling

Contemporary Approaches to Modelling  225

Standard Statistical Modelling Standard statistical modelling is one of the most widely applied families of modelling in consumer psychology and one in which there is constant innovation. The general aim of modelling from a statistical point of view is to develop an understanding of the relationships between important variables and to represent these relationships within a defined framework (i.e., a model).This understanding is not limited to the bivariate relationships between two variables, but, most importantly, includes an exploration of the multivariate relationships between many variables. Multivariate relationships help us to understand how variables operate in combination with one another. Modelling aims to simplify this complexity through development of an organisational framework—the model. The resultant models can come in many forms, for example, graphs, equations, matrices and figures. As such, this modelling process uses a diverse range of data inputs (e.g., transactional datasets, clickstream data, consumer surveys, laboratory or field experiments, and observations of behaviours) and analyses (e.g., regression, structural equations modelling, Bayesian approaches). Note that in this chapter we are using the term modelling in its widest sense to refer to a process of model development, and not in its narrow sense referring to specific types of analysis. Typically, standard statistical modelling studies can be categorised as either exploratory or confirmatory. In exploratory modelling, we seek to identify and define possible relationships between the variables in which we utilise the data and the analytical method to help us to define the nature of those relationships. In confirmatory modelling, we seek to test pre-specified relationships between our variables. As such, we must have a defined theory of how variables will work together and, therefore, which relationships will be significant. Exploratory and confirmatory modelling should be viewed as part of a model development continuum. During the modelling process, consumer psychologists often use multiple steps that move from an exploratory (development) basis in the early steps, to a confirmatory (testing) basis in later steps. Thus the difference between exploratory and confirmatory analyses can be understood in terms of the role of modelling in relation to theory. A theory can be classically defined as “a set of interrelated constructs (concepts), definitions and propositions that present a systematic view of phenomena by specifying relationships among variables, with the purpose of explaining and predicting the phenomena” (Kerlinger, 1986, p. 9). In exploratory analyses, the goal is to develop a theory, which we achieve through formulating and building a model. We may derive our model variables from one or a number of sources, for example, managerial intelligence and observations. Existing theory may play a role, but it is used in combination with other sources. Our model developed using exploratory analyses helps us to derive a number of propositions about our model variables and hence contribute to development of a theory. A proposition is a broad statement describing a relationship between two (or more) variables. In confirmatory analyses, the goal is

226  Debbie Isobel Keeling

to test our theory, achieved through testing our model. Hence theory underpins the specifically defined hypothesised and testable relationships between concepts. A hypothesis takes propositions one step further and formulates a more specific statement that is empirically testable. A hypothesis operationalises the proposed relationship and puts it in an empirically testable form. In confirmatory analysis, by testing our model we are providing evidence for a proposed theory.

Goals of Standard Statistical Modelling: Description, Prediction and Explanation Perhaps the most important goal of modelling is the ability to succinctly account for relationship complexity. That is, modelling helps us to both make sense of the multiple relationships that exist between variables and to then derive interventions to influence these variables. For example, healthcare professionals need a better understanding of what drives patient behaviour in order to work with patients on developing health promotion programmes. As such, we may want to understand how to encourage patients to avoid smoking or drinking too much alcohol. Of course, patient behaviours are influenced by more than one variable. As healthcare professionals, we want to understand the most important influential variables and how they work together. For example, how do peers influence patients in social situations, such as, family parties or in restaurants? We could examine the bivariate relationships separately, but that would only provide a partial picture (often too simplistic) of the processes of consumer behaviour. In examining relationship complexity, we can distinguish three specific goals of modelling: description, prediction, explanation.

The Goal of Description Modelling may be used as a means of summarising or organising the key variables into a simplified structure, enabling us to describe the relationships between variables and consumer behaviour. This is most often used in the first steps of developing theory—that is, within exploratory modelling. The idea is not to suggest or test causality, but to provide a way of organising the potential relationships between variables and behaviour into an understandable pattern. That is, the goal is to capture the associations between the variables in a more ‘usable’ format. Regression techniques, often referred to as exploratory analytical techniques, are a common technique applied in the context of descriptive modelling. For example, as a first stage in developing a health promotion programme, we might first observe a set of consumers in different social settings focusing on their interaction with peers. From these observations we might derive a set of influential variables that impact on smoking and drinking behaviours and set these out in an exploratory model.

Contemporary Approaches to Modelling  227

The Goal of Prediction Modelling may be used to predict consumer behaviours given the input values of the independent variables. There are multiple analytical approaches to predictive modelling (e.g., PLS-SEM), all of which focus on predicting future behaviours. Predictive modelling sits at the intersection of exploratory and confirmatory modelling as it can contribute to both.Whilst it has been neglected as a technique to aid theory development, recent developments, especially in terms of analysis, have made this type of modelling both more accessible and more relevant to contemporary consumer psychology problems. For example, from an exploratory perspective, with the increasing availability of larger and richer datasets, predictive modelling techniques offer a way to develop theory by exploring the complex relationships and patterns that can be hard to initially observe and theorise through more traditional methods. This can be an excellent way of identifying and prioritising the best predictors of behaviour from a very large possible set. From a confirmatory perspective, predictive modelling offers an important test of developed models through their application to practice. By explicitly testing the ‘confirmed’ patterns between variables and behaviour in practice, predictive modelling can provide confidence in a model, allow comparisons between competing models and, in the case where models do not effectively predict behaviours, suggest improvements to existing explanatory models. That is, predictive modelling can be an important tool to establish the ecological validity of the model and its ongoing development. For example, as a second stage in developing a health promotion programme, we might seek a more parsimonious model in terms of the number of important variables. Using predictive modelling, from the set of variables identified in stage one, we can identify the most influential variables on our outcome behaviours and derive a simpler model.

The Goal of Explanation Arguably the largest application of modelling within consumer psychology is the explanation of consumer behaviour. As such, a large body of consumer psychology models fit within the realm of confirmatory modelling. That is, when the goal of modelling is explanation, we aim to test our hypotheses regarding how a variable or group of variables will shape behaviour. The justification for the relationships comes from the underpinning theory, which is clearly set out in a model. Through subsequent statistical techniques (such as SEM, referred to as confirmatory techniques), we seek to confirm (within statistical boundaries) whether variables (X1, X2, etc.) shape behaviour (Y) in the way that was hypothesised. Advances in statistical analysis have led to the development of increasingly sophisticated techniques that allow us to more easily analyse complex relationship patterns, through simultaneously estimating several interrelated dependence

228  Debbie Isobel Keeling

relationships. It is not that these techniques did not exist before, rather that they have become more accessible and manageable for non-statisticians. Note that we should be very careful within consumer psychology to use the term ‘causes’, as we cannot often justify this (depending on the methodology used and context of the modelling). For example, as a third stage in developing a health promotion programme, we might seek to confirm our refined model by putting this to the test in a lab/field study. At this stage, we have a defined set of variables that we can now measure and seek to confirm their influence on the outcome variables (e.g., smoking or drinking behaviours).We may test this through presenting stimuli in an experimental situation or use a combination of questionnaire and observation data.

The Building Blocks of Models Within modelling there is some specific terminology that you need to be aware of, referred to here as the key building blocks of models.

Model Formats Models come in various formats, with the two most dominant formats in consumer psychology being path models and equations. Path Models: Figure 8.2 is an example of a path model. A good path model should summarise the key variables and their hypothesised relationships with each other and with the dependent variable(s).Whilst Figure 8.1 may be more properly classed as a conceptual model, in that it offers proposed relationships, Figure 8.2 is more specific in that it indicates the relative value (e.g., nature, direction) of those relationships that we are able to statistically derive. Hayes (2013) makes an important distinction between the conceptual model (e.g., Figure 8.1) and path models (e.g., Figure 8.2), which represents the hypothesised relationships, and the statistical model, which represents the actual relationships being tested in an analysis (see Hayes, 2013, for a detailed discussion). You will also come across the terms inner model or structural model to refer specifically to the depiction of the key variables and their hypothesised relationships.The terms outer model or measurement model refers to relationships between the variables and indicator items (that is, the items used to measure each variable—see 4.3 on measurement). Equations as representations: A very familiar example of an equation used to represent the relationship between independent and dependent variables is the y = a + bx equation derived through regression analysis. For multivariate models this formula becomes y = a + b1x1 + b2x2 . . . bnxn. Where y is the value of the dependent variable, ‘a’ is a constant (i.e., the y-intercept), ‘b1’ is the regression coefficient for variable ‘x1’ (the first independent variable) and ‘b2’ is the regression coefficient for variable ‘x2’ (the second independent variable). We can continue to add ‘bnxn’ terms for ‘n’ variables. For example, we could derive a model

Contemporary Approaches to Modelling  229 Exogenous

Attitude towards the behaviour

Subjective norm

Endogenous

+

Mediating variable

+

Intention

Outcom variable +

Behaviour

+ Perceived behavioural control

Exogenous

Relationships between the variables

FIGURE 8.2  Delineating

the Model Building Blocks of the Simplified Theory of Planned Behaviour (Ajzen, 1991)

to understand influences on consumer earnings and represent this model as an equation, y = −5.504 + 0.495x1 + 0.210x2, where y represents monthly income, x1 represents years in education and x2 represents years in work. We could work out the value of a person’s monthly income using this formula (substituting the known values of x1 and x2). (Note that for completeness we should also include an error term in this formula, not indicated here.) Other representations: Whilst visual (path models) and equations representations are popular formats, you will come across other formats. For example, graphical or chart representations such as the Kano Model (Figure 8.3). This model identifies five categories of customer requirements and illustrates their different relationships with customer satisfaction.The idea being that if you know how your customer requirements affect their satisfaction then this informs development and management of products. A specially devised assessment tool accompanies this model (Kano Survey), the results of which can be directly plotted onto a graph.

Constructs and Variables Theories are based on identifying abstract constructs, for example, perceived behavioural control, self-efficacy, customer satisfaction and behavioural intentions. These constructs are often referred to as latent. That is, they cannot be observed or measured directly. In modelling terms, we use the terms unobserved variable or latent constructs to refer to a construct that is theorised to exist but is not

230  Debbie Isobel Keeling

s

Satisfaction

t

er

li De

gh

ce

an

rm

fo

r Pe

Poor Execution

Excellent Execution

eds ic Ne Bas

Dissatisfaction FIGURE 8.3 

A Simplified Version of the Kano Model (1984)

directly observable.With reference to Figure 8.2, examples of latent constructs are attitude towards the behaviour, subjective norm, perceived behavioural control and intention. We, therefore, in our analysis approximate the latent constructs by using observable or measurable indicator variables.These are called the observed variables, sometimes referred to as manifest or indicator variables. This type of data is collected from consumers through data collection methods such as questionnaires and observations in experiments. Of course, we can incorporate directly observable variables into our model too, such as salary, age and so forth. Furthermore, partly when referring to confirmatory analyses, we can distinguish between exogenous and endogenous variables based on their role within the model. Exogenous variables refer to the latent equivalent of independent variables that are not influenced by other variables in the model. They act as predictors of other variables in the model but are not predicted by other variables. Endogenous variables are the latent equivalent to dependent variables, they are affected by other variables in the model. These variables are the ‘outcome’ in at least one contributory relationship. As shown on Figure 8.2, the endogenous variables are both those based in the ‘middle’ of the model—e.g., intention is influenced by the three preceding variables—as well as incorporating the outcome variable—e.g., a specific behaviour.

Contemporary Approaches to Modelling  231

Measuring Model Variables When representing constructs within our model using observed variables, what we are doing is translating our abstract concepts into measurable (and therefore testable) variables through a process of operationalisation. Often latent constructs are complex and cannot be measured using a single indicator. Instead, we measure these constructs, indirectly, using multiple indicators (e.g., items on a questionnaire). This requires that we clearly define how we are going to operationalise (put into practice) and hence measure our model variables. Whilst it is beyond the scope of this chapter to discuss the construction of measures (or scales), the importance in terms of contemporary modelling stems from advances in analytical techniques. The development of analytics such as SEM allows us to more readily include the measurement items used for each latent construct within our modelling (i.e., not the scale total as with regression techniques), and as such, we can incorporate measurement error directly into the estimation process. This leads to better model specification. Error refers to the unexplained variance within a model, which we must acknowledge and account for. We can further distinguish between formative and reflective indicator variables used to measure the unobserved variables (latent construct). With formative indicator variables the arrows between the unobserved variable and the observed indicator variables go from the observed indicator variables to the unobserved variable.This indicates that the observed indicator variables ‘cause’ the unobserved variable. Error in this case can be defined as the inability of the observed indicators to fully explain the unobserved variable. Often a set of formative indicators used to measure an unobserved variable is called an index. (See Hair et al., 2016, for an important discussion on the distinction between composite and causal indicators in CB-SEM and PLS-SEM.) With reflective indicator variables, the arrows between the unobserved variable and the measurement indicators go from the unobserved variable to the observed indicator variables. This indicates that the unobserved variable ‘causes’ the observed indicator variables. A change in the unobserved variable would bring about a change in all of the observed indicator variables. Error in this circumstance can be defined as the inability of the unobserved variable to fully explain the observed indicators. Often a set of reflective indicators used to measure an unobserved variable is called a scale.

Representing Relationships We have emphasised that the generic goal of modelling is about understanding relationship complexity. Referring to path models, on Figure 8.2 we have represented the theorised relationships between the variables using arrows. We can distinguish between different types of relationships within a model. Recursive relationships (indicated on Figure 8.2 with an arrow going in one direction

232  Debbie Isobel Keeling

between two variables) are where one variable X is theorised to impact on another variable Y (therefore the arrow points from X to Y). Conversely, a non-recursive relationship, which would be represented by a double-headed arrow, indicates a relationship where X influences Y and Y influences X. A correlational relationship (indicated on Figure 8.2 by a curved double-ended arrow) does not indicate any specific direction of influence.

Variable Roles Variables also have different specific roles within a model. For example, the exogenous variables that we met earlier have a predictive or explanatory role within the model. On the other hand, endogenous variables can have both a p­ redictive/ explanatory role and can themselves have a dependent role on antecedent variables. As such, endogenous variables can take on the role of mediation. A mediating variable is a variable that intervenes between two other related variables.That is, the mediating variable explains (facilitates) the relationship between the two other variables. On Figure 8.2, intention is theorised as a mediating variable. For example, the relationship between subjective norm (X) and behaviour (Y) is facilitated by intention (M). Parallel mediation is where a number of variables (M1, M2, etc.) mediate the relationships between X and Y separately (i.e., they are unrelated) to each other. So the mediating variables (M1, M2, etc.) would all be influenced by X and then separately influence Y. Serial mediation can be thought of as a ‘causal chain’ such that X influences M1, which influences M2, which influences Y. Other variables may take peripheral, but nonetheless vital, roles within the model. Most important of these is moderation. A moderating variable is a variable that changes the relationship between two other related variables. That is, the existence of a moderator means that the relationship between two variables changes with the level of the moderating variable; i.e., the moderating variable influences either the strength and/or direction of the relationship. We refer to moderated mediation, where one variable modifies the effect of a mediating variable.

Two Examples of Modelling The two cases provided next are taken from real projects, although as indicated in the footnotes, the data has been simplified or modified to allow illustration of the modelling, whilst protecting confidentiality in the cases. The purpose of these cases is to demonstrate the varied contexts and methods that can be classed under the broad modelling label.

Case 1—Testing the UTAUT2 Model Case 1 provides an example of dependence techniques. Dependence techniques are where we use of a set of (independent) variables to help describe,

Contemporary Approaches to Modelling  233

explain and/or predict one or more (dependent) variables. Typical analytical techniques include Regression, Discriminant Analysis, Conjoint Analysis, Partial Least Squares-SEM (PLS-SEM) and Structural Equations Modelling (SEM). Further, macros such as PROCESS have been specifically designed to test mediation and moderation.

Problem Identification and Context A national health organisation wants to understand why consumers (in this case patients with specific conditions) would use the Internet as a means for selfmanaging their health.The goal is to identify the factors that could be targeted to encourage more use of the Internet to empower patients.

Context and Sample A market research company collected data from patients (N = 623, 53.3% female, average age = 44yrs, s.d. = 15yrs) with specified conditions using a structured questionnaire. The questionnaire was composed of a number of scales designed to measure the constructs within the UTAUT2 (Venkatesh et al., 2012) and use of the Internet for self-management of health. The simplified model for testing is presented in Figure 8.4.

Analysis Following a Confirmatory Factor Analysis (CFA) to confirm the reliability and validity of the measurements, Structural Equations Modelling (SEM) was used to estimate the model parameters. The AMOS software was used in this case.

What Can We Say About Testing the Model? Overall, we can observe that as hypothesized, all three (exogenous) variables have a significant relationship with intention. Effort Expectancy (β = 0.69, p < 0.01) has the strongest relationship with intention, followed by Facilitating Conditions (β = 0.51, p < 0.01) and then Performance Expectancy (β = 0.40, p < 0.01). All three relationships are positive, indicating that as perceptions of Performance Expectancy, Effort Expectancy and Facilitating Conditions become more positive, so intention to use the Internet for health purposes increases. Further, in its mediating role, intention has a significant effect on use (β = 0.41, p < 0.01) such that increasing intention to use is associated with higher actual use of the Internet for health purposes. Apart from being able to describe the relationships, we can also assess how good our model is at reproducing our test dataset. SEM does not have one statistical test that we can use to test the strength of the model’s predictions. Instead, we use multiple goodness-of-fit indices. Usually, this includes the χ2/df (referred to as

234  Debbie Isobel Keeling

Performance Expectancy

Effort Expectancy

Facilitation Conditions

FIGURE 8.4 

0.40

0.69

Intention to use the Internet for health purposes

0.41

Use of the Internet for health purposes

0.51

A Simplified UTAUT2 Model with Standardised Coefficients

CMIN in the AMOS software), a goodness-of-fit index (e.g., GFI, CFI, NFI,TLI) and a badness of fit index (e.g., RMSEA). We can evaluate our model by looking at these indicators of goodness of fit. We observe that CMIN/df = 3.206 [values between 2 and 4 are acceptable], CFI = 0.957 [values above 0.9 are acceptable], and RMSEA = 0.067 [values below 0.08 are acceptable].1 From this we can conclude that our model is an acceptable fit for our data. (See Hair et al., 2013 for an excellent discussion on the use of these indices in SEM.) As a note of caution, remember that establishing a good model fit (or establishing that a model is not ‘bad’) does not mean that this is the only model that could explain the patterns in our data, only that it is a good fit for reproducing our dataset. We can, of course, go on to refine our model by exploring (i) possible moderators (e.g., age and gender) and (ii) potential other variables that may impact on use. On a related note, there is also some controversy about the goodness (or badness) of fit indices and their use. (See, for example, Kenny et al. (2015) for a discussion on the use of RMSEA.) You will find that there are many available indices and that different disciplines favour different indices.You should be comfortable that you understand those relevant to your particular area, but always be aware of challenges to these.

Case 2—Segmentation and Profiling Case 2 provides an example of Interdependence technique. Interdependence techniques are where we are focus on understanding the underlying structure of data, assessing interdependence without any associated dependence relationships. That is, the goal is to understand how variables can be usefully grouped together. For example, consumer populations are often divided into subgroups sharing characteristics as in target markets. Typical analytical techniques include Cluster Analysis, Factor Analysis and Multidimensional Scaling.

Contemporary Approaches to Modelling  235

Problem Identification and Context A global commercial organisation wants to develop a diagnostic tool for profiling its business partners within a specific geographical region.The goal is to prioritise actionable strategies for those partners who have the highest likelihood of engaging with and climbing up the ladder of the organisation’s loyalty programme.

Context and Sample The dataset comprises transactional behaviours (655,000) over a one-year period, encompassing 98,000 partners. The total number of products sold over the year is 2,300,000, with a value of $600,000,000.2 Notice that this data is all observed, using automatic capture of transactions, rather than self-reported as in case 1.

Analysis Providing an example of an application of interdependence techniques, cluster analysis is used as a way to identify the underlying patterns within the data that can help reveal key segments. These identified segments are used in this case as a basis for defining the model loyalty programme member. Following this stage, propensity scoring is used to match non-member cases in the dataset to the identified model segment. In this way the company can not only understand the segments within their dataset but also apply actionable strategies for enrolling more members in its loyalty programme.

What Can We Say About the Cluster Identification? In stage 1, our analysis suggests that three segments will be useful as a way of succinctly understanding our sample (Table 8.1). There three groups can be identified by a combination of their measurements on four key indicator variables (following exploration with a wider dataset). The company identified segment 3 as their model elite profile. However, it was recognised that this elite profile was unrealistic for a large section of their partners to achieve. As such, segment 2 was identified as the model target profile. That is, this was the profile that they wanted to facilitate their partners to achieve.

How Does This Relate to Propensity Scoring? In stage 2, propensity scoring was undertaken to assess which of the non-members in the dataset were most closely matched to the segment 2 model target profile. Table 8.2 shows the number of matched cases according to the propensity scoring. From this the company were able to develop a prioritised target list of cases as part of their strategy for widening their loyalty programme membership. They started by recruiting the companies in the best matching group indicated earlier.

236  Debbie Isobel Keeling TABLE 8.1  Resultant Three Segments1

Three Main Segments

Av. No. Av. No. Units Av.Total Sales Av. No. of Portfolio Transactions Purchased (U.S. $) Divisions

1: Low value, low 3 transaction frequency, low variety 2: Moderate value, moderate 8 transaction frequency, moderate variety 3: High value, high 88 transaction frequency, high variety

6

1,500.00

1

21

6,000.00

2

380

120,000.00

4

1 The figures reported here are for illustration purposes only—for confidentiality purposes, the actual figures have been altered, but the relative scaling is preserved.

TABLE 8.2  Cases Matched to Model Group 2 by Propensity1

Propensity score range

Number of matched cases

Matched variables

Best matching companies The next most attractive segment Middle of the road

76%–100%

n = 20

51%–75%

n = 70

Good match across all four variables Good match on portfolio divisions

26%–50%

n = 350

Lowest engagement

0%–25%

n = 95,000

Good match on average total sales Low value transactions across all variables

1 The figures reported here are for illustration purposes only—for confidentiality purposes, the actual figures have been altered, but the relative scaling is preserved.

Troubleshooting Tips When undertaking any kind of modelling you will most likely encounter ­problems—even in the best design. Six common problems are identified with troubleshooting tips.

Tip 1: There Is Always Error! If we want to increase confidence in our work and its applicability, then we need to control error as much as we can. Error can occur at any stage of a project, but thoughtful project management can go a long way towards minimising this.When planning your project think about the potential sources of error and how these

Contemporary Approaches to Modelling  237

might be controlled right from the problem formulation stage. Common sources of error are: •

•

•

•

Neglecting to take account of non-response bias—that is, do not forget that it is as important to consider who is not in your study as well as who is. For example, is an important group not represented in your study? Ensure that you include a consideration of the non-response group in your write-up. Difficult to control but nevertheless important to consider is the impact of historical events during data collection. Has anything occurred during data collection that could have impacted on or biassed the data? In particular, did this affect one group or participants or all? Poor data collection practices, such as inadvertently influencing a participant in the study or not mitigating against missing data, which can invalidate your dataset and any subsequent analysis. Incorrect application of analyses, especially working outside the limitations of analytical methods (e.g., not observing good data screening practice).

Tip 2: Causality Is Difficult to Prove! Cross-sectional design, data collected at one point in time from a sample selected to represent the population of interest, is commonplace in consumer psychology. However, it is important to remember that the aim is to describe a population or document and test differences in a subset of the population at one point in time. As such, you are limited in what you can conclude about causality. Longitudinal design, where the aim is to examine a dynamic process that involves change over time and understand the sources and consequences of a phenomenon, is a desirable design but often not possible through resource constraints. The underlying principle is that we measure some dimensions of interest of a given entity before and after an intervening phenomenon to determine whether or not the phenomenon has some effects. As such, this gives greater confidence for causal inferences than cross-sectional designs. Remember that demonstrating a statistical fit for your proposed model does not immediately ‘prove’ your causal assumptions about the relationships between variables. Instead, it makes them more plausible. Replication studies—that is, repeating a study in another sample(s), or ruling out alternative models (often within the same study), that is, comparing a different model to your proposed model, providing further evidence and confidence in a model.

Tip 3: Larger Does Not Always Mean Better! Researchers can be tempted to collect as much data as possible. But the size of the dataset is not everything. Quality of the dataset should be the first consideration with a focus on collecting pertinent and well-measured data. With regard to

238  Debbie Isobel Keeling

sample size, some analyses can be too sensitive with larger samples. Before you start any data collection, consider the analysis to be used to test the model and what data is needed to test that model using that analysis. Ensure that you design your data collection techniques so that they are adequate for the purposes of that analysis. Also, consider the necessary size of the dataset for the analysis to be used. Some considerations that help you to determine sample size are the multivariate distribution of the data, the estimation technique to be used (e.g., Maximum Likelihood Estimation) and the model complexity (and the required number of indicators per construct where relevant). Most reference books on specific analytical techniques will provide guidance on sample sizes and/or variable to N ratios. With regard to the number of variables included in the model.Your aim should be to be reasonably parsimonious. That is, that you should identify the most pertinent variables in your model. Simply adding variables may increase your goodness of fit, but not necessarily the theoretical or practical relevance of your model. Do not allow yourself to be driven by achieving better statistical outcomes—instead, be driven by achieving a better quality, meaningful model.

Tip 4: Reliability Does Not Mean Validity! It is of course vitally important that you develop and/or use measures that are both reliable and valid. If you are using pre-existing scales, then check the reported reliability and validity, but do not just rely on these, check for yourself with your own data. Pay special attention when working in cross-cultural contexts (see tip 5). If you are developing your own measures, then remember that you will need time to develop these and to establish their reliability and validity. It is not sufficient to demonstrate reliability of your measures—you can have a perfectly reliable measure that is not valid (i.e., does not measure what you think it measures). As such, you must also establish validity. (See Oppenheim 1992 for a classic text on scale development.) Remember that even methods that directly observe behaviours need to be validated (including automated data capture).

Tip 5: Context and Culture Matter! Models cannot be separated from the context(s) or culture(s) within which they were developed.When applying a model to another context or culture remember to consider how this may impact on or challenge the underlying assumptions of the model.You need to consider a number of cross-cultural/contextual issues: •

•

Do the underlying constructs hold for the new context/culture—that is, do they make sense, do they need to be adapted, or do they need to be totally rethought? Do the measures used to test the model hold for the new context/culture? Consider not just language translations, but also semantics and use of scales or response formats (and the use of appropriate stimuli in different cultures).

Contemporary Approaches to Modelling  239

•

Do you need to adapt the modelling process? For example, access to participants, data management issues, literacy levels in the focal country/population are all important considerations.

Tip 6: Equivalence Is Illusive! One of the things that you may wish to establish is whether your model is equivalent or different across groups (e.g., gender/age groups, teams/departments, ­cultures/countries). Cross-validation is an attempt to reproduce the results found in one sample (or population) using data from a different sample—that is, to establish equivalence or invariance across groups. However, it is often the case that you will not achieve full equivalence (see, for example, Malhotra & McCort, 2001, for a comparison of behavioural intentions models across cultures).You must therefore decide what level of equivalence is adequate to address your objectives. For example, a partial cross-validation using SEM might establish that factor loadings are equivalent; this is typically considered adequate.Tight cross-validation—whilst ideal is rare—where you establish the equivalence of factor loadings, inter-factor covariances, and error variances. (See Hair et al., 2013 for a fuller discussion on statistical equivalence and Craig and Douglas for a fuller discussion on different methods in cross-cultural research.)

Future Directions of Modelling in Consumer Psychology We have focused on standard methods used in contemporary modelling in consumer psychology. However, this is a world of constant innovation, not only in terms of developments in methods, analytics and software but also in the problems faced within the world of consumer psychology that need increasingly sophisticated approaches to resolve. There are three identifiable trends in modelling.

A New Era for Bayesian Modelling? A recent review (van de Schoot et al., 2017) has shown a clear upward trend in the use of Bayesian approaches within psychology. Although not a new technique, the accessibility of this approach has been enhanced by both an increasing supply of communications on methods and developments in software applications. With specific relevance to consumer psychology, there has been increased use of Bayesian approaches within the areas of cognition, personality and neuropsychology, alongside the contexts of health, education and development (Depaoli et al., 2017; van de Schoot et al., 2017).There are parallel developments from a methods perspective. There are equivalent Bayesian approaches to the popular frequentist approaches—e.g., Bayesian SEM—and there is a trend in studies demonstrating Bayesian approaches as credible alternative estimation procedures. Simulation studies focus on the development of and comparison between alternative Bayesian methods (van de Schoot et al., 2017).

240  Debbie Isobel Keeling

Of the reasons for using Bayesian approaches identified by van de Schoot et al. (2017), those applicable to consumer psychologists are (1) the intuitive appeal of explicitly using prior knowledge to estimate models and (2) the ability of Bayesian models to handle more complex models. Examples of the application of diverse Bayesian approaches to consumer psychology demonstrate the potential of these approaches to address contemporary consumer problems. Lee et al. (2011) use Bayesian approaches to forecast the future digital video-player market, taking into account consumer preferences. Jerath et al. (2014) use Bayesian approaches to provide insight into consumers’ behaviours in online searching tasks. Jang et al. (2016) predict consumer spending patterns and share of wallet to inform consumer targeting strategies. Swani et al. (2017) demonstrate the propensity of users to popularise (i.e., like and/or comment) brand-related posts on social media. Marinova et al. (2016) investigate how physicians make trade-offs between medical and non-medical information to determine a patient’s access to life-enhancing products. Lin et al. (2017) use Bayesian approaches to investigate how to develop an understanding of a patient’s risk of future adverse health events to support clinical decision making in personalised and preventive care. Whilst there may be a divide between those favouring the frequentist approaches or Bayesian approaches, there is potential for applying the methods in a complementary fashion. This will allow advances in consumer psychology that are based on robust development and tests of models and address the complexities faced by consumers in the real world.

A Kickstart for Predictive Modelling? Another evident trend in modelling is a resurgence in the use of PLS-SEM (or PLS Path Modelling). Despite evolutions in this technique, there was a paucity of mainstream supporting texts until Hair et al. (2013). Emphasising the rapid developments in PLS-SEM in recent years, the second version closed followed the 2013 edition to document these developments (Hair et al., 2016). Whilst this analysis is ideally placed as a model development tool (Morard & Simonin, 2016), it has helped to kickstart studies in the marginalised area of predictive modelling (see section 3.1). PLS-SEM supports predictive modelling in complex situations, especially where there is little theory. Beyond this, PLS-SEM can cope with modelling when faced with problems such as the issue of small consumer populations or samples (Hair et al., 2014). As such, it is a useful management tool to also provide insight into model optimisation, free from theoretical and statistical assumptions (Morard & Simonin, 2016). Contemporary examples of consumer psychology relevant studies involving the use of PLS-SEM as a predictive modelling tool can be found across a diverse range of fields. Ayeh et al. (2013) developed a predictive model of intention to use consumer-generated media for travel planning. Lehto and Oinas-Kukkonen (2015) predicted continued engagement in a weightloss programme. Wakefield

Contemporary Approaches to Modelling  241

(2016) used fan passion to predict attendance, media consumption and social media engagement. Schubring et al. (2016) used PLS-SEM for predictive modelling of the diffusion process in consumer networks.

A Big Data Push for Modelling? The increasing availability of consumer data has driven sophisticated advances in data mining techniques and software, and enhanced computing power to handle storage, management and analysis of such data. These developments have resulted in more accessibility to ‘big data’ for modelling in consumer psychology to noncomputer science specialists. Cheung and Jak (2016) argue that psychologists may have historically lacked the programming or computational skills for data mining, but that they bring valuable insight into the interpretation of such analyses, particularly in terms of understanding and modelling of consumer behaviours and methods of assessing the reliability and validity of data. Indeed, Seitz et al. (2017) argue that the strength of simulation modelling is the ability to achieve parsimony, but this algorithmic approach needs to be embedded within an understanding of the psychology of consumers to avoid reductionism. The volume, variety, velocity and veracity of available data is driving new areas of research and the discovery of new models and analysis (Cheung & Jak, 2016; Goldstone & Lupyan, 2016). The variety and volume of data offers the opportunity for more nuanced versions of existing models. In particular, the ability to locate and test models of behaviour in diverse online (new) cultural and social environments (e.g., consumer engagement in online brand communities, Brodie et al., 2013). These environments may also be more suited to modelling of sensitive behaviours, and offer up new ways of establishing the ecological validity of the results from laboratory experiments (Goldstone & Lupyan, 2016). The velocity and veracity of available data means that we can not only collect data more rapidly but also have access to more observable naturally occurring data. Meaning that consumer psychologists can now be more agile in their response to developing solutions to managerial and societal problems. Indeed, technological advances mean that some of these phenomena and/or data did not exist before, opening up entirely new areas, and connections between consumers, brands, services and practitioners have increased exponentially. For example, patterns of online collaboration and networks, reactions to regional events on a global scale and new forms of celebrity are all areas of interest to consumer psychologists (Goldstone & Lupyan, 2016). These modelling efforts are informed and enhanced by advances in the development of statistical algorithms and artificial intelligence (AI). Opening up avenues for delivering services based on automation (e.g., the use of avatars, Kohler et al., 2011), aiding consumers in making decisions (e.g., recommendation systems, Yin et al., 2017), and informing how companies approach forecasting and prediction of consumer behaviour (e.g., neural networks, Greene et al., 2017).

242  Debbie Isobel Keeling

Moving forward, we are likely to see more and more consumer psychologists working within multidisciplinary teams (e.g., alongside informatics, computer scientists and AI experts) in order to push the boundaries of knowledge within our own discipline. As such, this emphasises the wider process of moving from data mining, through to modelling, through to application (often management).

Software There is a plethora of paid-for and free-to-download software packages to help you with your analysis. Here is a selection (and certainly not an exhaustive list): 1. General Packages a. AMOS (www.ibm.com/uk-en/marketplace/structural-equationmodeling-sem) b. SAS (www.sas.com/en_in/software/stat.html) c. SPSS (www.ibm.com/analytics/data-science/predictive-analytics/spssstatistical-software) d. Process (www.processmacro.org/index.html) e. R (www.r-project.org/about.html) 2. PLS-SEM a. SmartPLS 3 (www.smartpls.com) b. PLS Graph (www.statisticssolutions.com/pls-graph-software/) c. Also available with options in SAS, SPSS and R 3. Bayesian Approaches a. b. c. d.

Win BUGS (Lunn et al., 2000) MPlus (Muthen & Muthen, 2017) SAS, SPSS and R offer different options Other software is constantly being developed

Student Exercises These exercises can be applied to your specific area of interest in consumer psychology. You can treat the exercises as stand-alone tasks, or you can combine them to apply to a set of studies. These exercises will help you in your own modelling studies (perhaps in a dissertation or similar). The first exercise asks you to consider your own phenomenon

Contemporary Approaches to Modelling  243

of interest and think about this in modelling terms. The next three exercises ask you to examine the approaches that other studies have taken in detail and to critique these approaches. The final exercise asks you to focus on cross-cultural issues. 1. Choose a phenomenon related to consumer psychology that is in your area of interest. Think about the variables that could help you to understand the phenomenon. Which of these variables would you identify as the dependent variable? How would you organise these variables into a model? Would you identify mediating variable(s)? Are there any variables that could act as a moderator? Do you know enough to create a sufficient model, or do you need to find out more information? Is exploratory or explanatory research justified? 2. Do a review of published studies in modelling in your area of interest. Now create a table that lists the purpose of the study (exploratory, predictive or explanatory) and the types of model used (e.g., path model, equation etc.). What type of analysis is used in these studies—e.g., Bayesian techniques, SEM, PLS-SEM (including software packages)? Do you notice any commonality in these studies? Is there a dominant approach to modelling and analysis in this area? Do you agree with the approaches taken? If so or if not—why is this? 3. Identify three or four published studies based on modelling in your area of interest. Create a table. In column one, list the types of variables used. In column two, list what data has been collected to measure these variables. In column three, list the rationale given for use of such measures. In columns four and five, list what you think are the advantages and disadvantages of these measures. Do the authors of the studies pick out the same advantages and disadvantages as you? 4. Identify three or four published studies on modelling in your area of interest (you could use the same studies as in exercise 3). Now think about the following three categories of error: measurement, fieldwork, data analysis. From your chosen studies identify what sources of error there may within your chosen studies under these three categories. How might you mitigate against these sources of error?

244  Debbie Isobel Keeling

5. In their article “On Improving the Conceptual Foundations of International Marketing Research”, Douglas and Craig (2006) state, “All too frequently, cross-country research begins. . . .without consideration of the underlying conceptual framework and related constructs and their applicability in other research contexts”. List the issues from their conceptual framework. Think of examples for each of these issues from your own area of interest. What problems might these issues raise for you in modelling?

Notes 1. The figures reported here are for illustration purposes only. 2. The actual sales figures have been altered for confidentiality purposes—but the relative scaling is preserved.

Recommended Reading Craig, C. S., & Douglas, S. P. (2005). International marketing research. Chichester: John Wiley & Sons. Field, A. (2017). Discovering statistics using IBM SPSS statistics. Sage Publications Ltd: ­London, UK. Field, A., Miles, J., & Field, Z. (2012). Discovering statistics using R. Sage Publications Ltd: London, UK. Hair, J. F., Black, W. C., Babin, B. J., Anderson, R. E., & Tatham, R. L. (2013). Multivariate data analysis. Pearson Education Limited: London, UK. Hair Jr, J. F., Hult, G. T. M., Ringle, C., & Sarstedt, M. (2016). A Primer on Partial Least Squares Structural Equation Modeling (PLS-SEM). Sage Publications Ltd: Los Angeles, UK. Hayes, A. F. (2018). Introduction to mediation, moderation, and conditional process analysis: A regression-based approach (2nd ed.). New York, NY: Guilford. Kenny, D. A., & McCoach, D. B. (2003). Effect of the number of variables on measures of fit in structural equation modeling. Structural Equation Modeling, 10(3), 333–351. Lambert, B. (2018). A student’s guide to Bayesian statistics. Sage Publications Ltd: Los Angeles, CA. Oppenheim, A. N. (1992). Questionnaire design, interviewing and attitude measurement. (New Ed.). London and New York: Pinter Publishers. Sarstedt, M., Ringle, C. M., & Hair, J. F. (2014). PLS-SEM: Looking back and moving ­forward. Long Range Planning, 3(47), 132–137. van de Schoot, R., Winter, S. D., Ryan, O., Zondervan-Zwijnenburg, M., & Depaoli, S. (2017). A systematic review of Bayesian articles in psychology:The last 25 years. Psychological Methods, 22(2), 217.

Contemporary Approaches to Modelling  245

References Ajzen, I. (1991). The theory of planned behavior. Organizational Behavior and Human Decision Processes, 50(2), 179–211. Ayeh, J. K., Au, N., & Law, R. (2013). Predicting the intention to use consumer-generated media for travel planning. Tourism Management, 35, 132–143. Brodie, R. J., Ilic, A., Juric, B., & Hollebeek, L. (2013). Consumer engagement in a virtual brand community: An exploratory analysis. Journal of Business Research, 66(1), 105–114. Craig, C. S., & Douglas, S. P. (2005). International marketing research. Chichester: John Wiley & Sons, Inc. Cheung, M.W. L., & Jak, S. (2016). Analyzing big data in psychology: A split/analyze/metaanalyze approach. Frontiers in Psychology, 7. Depaoli, S., Rus, H. M., Clifton, J. P., van de Schoot, R., & Tiemensma, J. (2017). An introduction to Bayesian statistics in health psychology. Health Psychology Review, 11(3), 248–264. Douglas, S. P., & Craig, C. S. (2006). On improving the conceptual foundations of international marketing research. Journal of International Marketing, 14(1), 1–22. Goldstone, R. L., & Lupyan, G. (2016). Discovering psychological principles by mining naturally occurring data sets. Topics in Cognitive Science, 8(3), 548–568. Greene, M. N., Morgan, P. H., & Foxall, G. R. (2017). NEURAL networks and consumer behavior: NEURAL models, logistic regression, and the behavioral perspective model. The Behavior Analyst, 40(2), 393–418. Hair Jr, J. F., Hult, G. T. M., Ringle, C., & Sarstedt, M. (2013). A Primer on Partial Least Squares Structural Equation Modeling (PLS-SEM). Sage Publications Ltd: Los Angeles, CA. Hair Jr, J. F., Hult, G. T. M., Ringle, C., & Sarstedt, M. (2016). A Primer on Partial Least Squares Structural Equation Modeling (PLS-SEM). Los Angeles, CA: Sage Publications Ltd. Hair, Joe Jr, F., Sarstedt, M., Hopkins, L., & Kuppelwieser,V. G. (2014). Partial Least Squares Structural Equation Modeling (PLS-SEM): An emerging tool in business research. European Business Review, 26(2), 106–121. Hayes, A. F. (2013). Introduction to mediation, moderation, and conditional process analysis: A regression-based approach. New York, NY: Guilford, ISBN:1609182308. Jang, S., Prasad, A., & Ratchford, B.T. (2016). Consumer spending patterns across firms and categories: Application to the size-and share-of-wallet. International Journal of Research in Marketing, 33(1), 123–139. Jerath, K., Ma, L., & Park, Y. H. (2014). Consumer click behavior at a search engine: The role of keyword popularity. Journal of Marketing Research, 51(4), 480–486. Kano, N. (1984). Customer satisfaction model. Available at: http://www.12manage.com/ methods_kano_customer_satisfaction_model.html. [Accessed 28th August 2018]. Kenny, D. A., Kaniskan, B., & McCoach, D. B. (2015). The performance of RMSEA in models with small degrees of freedom. Sociological Methods & Research, 44(3), 486–507. Kerlinger, F. N. (1986). Fundamentals of behavioral research. Holt, Rinehart, Winston: New York. Köhler, C. F., Rohm, A. J., de Ruyter, K., & Wetzels, M. (2011). Return on interactivity:The impact of online agents on newcomer adjustment. Journal of Marketing, 75(2), 93–108. Lee, J., Choi, J. Y., & Cho, Y. (2011). A forecast simulation analysis of the next-generation DVD market based on consumer preference data. International Journal of Consumer Studies, 35, 448–457. doi:10.1111/j.1470-6431.2010.00958.x

246  Debbie Isobel Keeling

Lehto,T., & Oinas-Kukkonen, H. (2015). Explaining and predicting perceived effectiveness and use continuance intention of a behaviour change support system for weight loss. Behaviour & Information Technology, 34(2), 176–189. Lin, Y. K., Chen, H., Brown, R. A., Li, S. H., & Yang, H. J. (2017). Healthcare predictive analytics for risk profiling in chronic care: A Bayesian multitask learning approach. MIS Quarterly, 41(2). Lunn, D. J., Thomas, A., Best, N., & Spiegelhalter, D. (2000). WinBUGS – a Bayesian modelling framework: Concepts, structure, and extensibility. Statistics and Computing, 10, 325–337. Malhotra, N. K., & McCort, J. D. (2001). A cross-cultural comparison of behavioral intention models-theoretical consideration and an empirical investigation. International Marketing Review, 18(3), 235–269. Marinova, D., Kozlenkova, I. V., Cuttler, L., & Silvers, J. B. (2016). To prescribe or not to prescribe? Consumer access to life-enhancing products. Journal of Consumer Research, 43(5), 806–823. Morard, B., & Simonin, D. (2016). Partial Least Squares modeling in marketing research: A tailor-made model of wine e-Commerce consumers in Switzerland. Journal of Economics, Business and Management, 4(5), 410–417. Muthen, L. K., & Muthen, B. O. (2017). Mplus User’s Guide (8th ed.). Muthen & Muthen: Los Angeles, CA. Oppenheim, A. N. (1992). Questionnaire design, interviewing and attitude measurement (New ed.). London and New York, NY: Pinter Publishers. Schubring, S., Lorscheid, I., Meyer, M., & Ringle, C. M. (2016). The PLS agent: Predictive modeling with PLS-SEM and agent-based simulation. Journal of Business Research, 69(10), 4604–4612. Seitz, M. J., Templeton, A., Drury, J., Köster, G., & Philippides, A. (2017). Parsimony versus reductionism: How can crowd psychology be introduced into computer simulation? Review of General Psychology, 21(1), 95. Swani, K., Milne, G. R., Brown, B. P., Assaf, A. G., & Donthu, N. (2017). What messages to post? Evaluating the popularity of social media communications in business versus consumer markets. Industrial Marketing Management, 62, 77–87. van de Schoot, R., Winter, S. D., Ryan, O., Zondervan-Zwijnenburg, M., & Depaoli, S. (2017). A systematic review of Bayesian articles in psychology:The last 25 years. Psychological Methods, 22(2), 217. Venkatesh,V., Thong, J.Y., & Xu, X. (2012). Consumer acceptance and use of information technology: Extending the unified theory of acceptance and use of technology. MIS Quarterly, 157–178. Wakefield, K. (2016). Using fan passion to predict attendance, media consumption, and social media behaviors. Journal of Sport Management, 30(3), 229–247. Yin, C.,Wang, J., & Park, J. H. (2017). An improved recommendation algorithm for big data cloud service based on the trust in sociology. Neurocomputing, 256, 49–55.

9  ONNECTIONIST MODELLING OF C CONSUMER CHOICE Max N. Greene, Peter H. Morgan, and Gordon R. Foxall

Connectionism Originally, behavioural economics relied heavily on the insights of Simon (1982). Despite his pioneering work on the serial symbol processing hypothesis in cognitive psychology and artificial intelligence, Simon’s contributions (1982, 1987) are rather outdated in the face of the current focus on parallelism and connectionism (Sent, 2004). Connectionism is a philosophical framework within which mental and behavioural phenomena are modelled as properties emerging from networks of simple units. Some of the more commonly encountered forms of connectionism rely on the use of neural-network models. Central to connectionism is the principle that mental processes can be described by interconnected networks of simple uniform units that represent neurons and connections that represent axons and dendrites that link the neurons through synapses. Most networks tend to change over time and incorporate a concept of activation. A computational unit within the network has a numerical activation value, which could represent the probability of that neuron producing a response. Spreading activation models allow for activation to extend to the other interconnected units over time and are a common feature of the NN modelling discussed later.

The Connectionist Model Features The importance of learning and training the models is usually emphasised by the connectionist researchers, and numerous complex learning algorithms to train NN models have been devised. Learning does not, however, carry a guarantee of infallibility—human beings can, after all, learn incorrectly. In the NN models we discuss here, learning takes the form of modifying the connection weights

248  Max N. Greene et al.

by some algorithm to reduce (optimise) some error measure. Backpropagation, one of the most popular algorithms for feedforward neural networks operates by adjusting individual weights according to the partial derivatives of the imputed error at that weight but amounting to a gradient descent on the error surface overall. There is nothing magical about backpropagation however and conventional optimisation methods such as Davidon-Fletcher-Powell which alter all the weights simultaneously are often superior at locating a set of weights giving a minimum error (whether this is a global minimum1 is another story of course). From a connectionist point of view, the appeal of backpropagation is that it corrects weights one by one neuron by neuron and therefore devolves the learning to the neuronal level The mathematical framework that is the foundation of most connectionist models today was proposed as part of the parallel, distributed, processing approach (Rumelhart & McClelland, 1987) that emphasised neural processing nonlinearity. The most common connectionist models are feedforward neural networks using mathematical neurons (Figure 9.1) in connection (Figure 9.2). They operate under the assumptions that it is possible to describe the mental state as a multi-dimensional vector containing the numeric activation values for the computational units within the network, and that the gradually modified connection strength (weights) creates memory. The inputs can be interpreted as coming from the network system’s environment, being processed through the network and producing outcomes at the outputs. Target and input variables are then identified in the data and the weights are varied by a learning mechanism to ensure that the outputs match as closely as possible the target variable values for their given inputs.Variations in the models come from the interpretation of the neurons, the activation function, and the learning algorithm employed to train the network. Little is known about the actual functionality of the brain, NNs have traditionally been seen as simplified neural processing models. The degree of complexity and individual properties that computational units should have to accurately mimic the functionality of the brain for representative purposes is yet to be determined. From the computational view however, contrary to the traditionally inclined algorithms predominant in computer technologies that follows a sequential processing and instructions execution in an automated predefined manner, neural networks attempt to model the information processing in a way similar to biological systems that rely on parallel nonlinear processing and pattern recognition. As a result, the very core of a neural network is not just an algorithm tasked with sequential execution of predetermined commands but rather a complicated statistical processor.

Artificial Neural Networks Even though computational models based on NNs were developed many decades ago (Hebb, 1949), technological and computer science advances in recent

Connectionist Modelling of Consumer Choice  249

decades have facilitated the growing interest in using the NNs to study a number of diverse phenomena in statistics, cognitive psychology and artificial intelligence (Ripley, 1996). Originally developed (McCulloch & Pitts, 1943 for representational purposes to model the functionality of the human brain (Bishop, 1995), NNs have since lost that as a primary function and are increasingly utilised as a method of analysis in predictive modelling and forecasting (Adya & Collopy, 1998). Figure 9.1 compares the McCulloch and Pitts ‘mathematical neuron’ with a very simplified version of a biological neuron. An important difference between these two formulations is that the latter is a pulse code modulated entity whereas the former is essentially an analogue device even if it may be simulated on a digital computer. Inspired by structural and functional features of biological neural networks (nonlinear distributed information processing), NNs normally comprise a group of simple processing units, or artificial neurons, which, through their interconnection, are able to display a complex global behaviour determined by the connections between the processing units. Information is processed employing the connectionist approach as follows: functions are performed in parallel by the units, rather than clearly assigning subtasks to various unit groups. In most cases NNs are adaptive systems, able to adjust their structure by fine-tuning the strengths (weights) of the connections in the network according to external or internal information flow (Haykin, 1994). Today, NNs are often used as statistical techniques designed to find patterns in data or to model intricate relationships between dependent and independent variables. As the term training suggests, weights are adjusted by comparing actual target data values with predictions from the network outputs and trying to reduce the error between the two values is known as Weighted Inputs

w1X1 Dendrite from another neuron

w3X3

i=3

Y=ΣXi

Summaon

Y = g(Y)

Applicaon of Acvaon Funcon , g()

i=1

Nucleus

Synapc gap shaded in grey

Axon branching to other neurons

FIGURE 9.1 

w2 X 2

Biological Neuron

W1Y

W2Y

W3Y

Weighted Outputs

Schematic Biological and Mathematical Neurons

Mathemacal Neuron

250  Max N. Greene et al.

supervised learning. Other network architectures exist, notably the Kohonen SelfOrganizing Map (SOM) and the Hopfield Networks which operate as unsupervised learners. We will not be considering unsupervised learning here but, suffice it to say, that the SOM has been investigated as a means of market segmentation for example (Curry et al., 2003). An example of a simple three-layer feedforward neural network is shown in Figure 9.2. The first layer contains the input neurons, which send the data by means of weighted connections to the second layer of neurons, and the outputs from that layer go to the third layer of output neurons. This is the most commonly encountered architecture as it contains only one intermediary hidden layer (although even simpler input-output NNs with no hidden layers are possible) and no skip layer connections (connections that would in Figure 9.2 go from the input layer straight to the output layer bypassing the hidden layer), and more complex architectures would include more layers and an increased number of neurons within each layer. Three factors define the type of NN model: (1) the pattern of interconnection between the neurons, (2) the activation function to convert the weighted input to its output activation, and 3) the error function used—e.g. mean square error. Neural networks have been shown to be universal function approximators (Cybenko, 1989; Hornik, 1991) where networks are capable (in theory) of simulating any ‘reasonable’ function. The use of a neural network is often justified by

Bias Node with constant value of 1

Neurons with logisc acvaon funcons (shaded nodes) typically add a bias constant to their inputs via weighted connecons from a bias node

Linear Output Node

Linear Input Nodes

Hidden Layer of Logisc Nodes

FIGURE 9.2 NN

Architecture That Includes Three Layers: Two Nodes in the Input Layer, Five Nodes with Biases in the Hidden Intermediary Layer, and One Output Node

Connectionist Modelling of Consumer Choice  251

appeal to this universal approximation property, but there are several other considerations to be worked through. These are (1) how easy is it to fit the network to a function (dataset)? (2) How fast does the network approximation approach the target function with increasing network size? (3) How well will it predict once it has been fitted (trained)? This last issue, Point 3, is related to the idea of overfitting where, as the network gets bigger (and the number of adjustable weights increases very fast with the number of hidden nodes), the model learns more and more detail and ‘learns the noise as well as the signal’. (This expression, of course, assumes erroneously that noise is high frequency detail and has led to many a ‘discovery’ of a trend by essentially a low pass filtering of white noise.) We will return to Point 3 when we look at pruning methods. Point 1 comes into play with a vengeance in the fitting of large networks which is a notoriously ill-conditioned problem due to the fact that large numbers of weights create a high dimensional search space with local mimima behaving like bunkers on a golf course to trap ‘unwary’ algorithms and the weights are often highly correlated leading to narrow valleys and flat areas in the error surface which are hard for training algorithms to navigate. Point 2 is often a matter of network architecture. For example, a single hidden layer can approximate a peak function by using a much larger number of hidden nodes but the behaviour of the fitted function outside the area of fit is hopeless in terms of extrapolation. Adding a second hidden layer makes such functions much easier to fit (Chester, 1990; Morgan et al., 1999). This should not be a surprise as ‘real’ neural networks often adopt a multilayered approach. One of the major research directions in the field aims to establish NN models as a powerful and versatile method of analysis—often employing comparative design which contrasts neural networks with other traditionally employed methods (Bishop, 1995). As a result, it is often reported that neural networks not only perform as well as other methods considered, but also often outperform traditionally employed approaches tasked amongst other things with segmentation and targeting (Adya & Collopy, 1998). As it seems to be the case in consumer behaviour literature that ongoing research is largely concerned with identifying underlying patterns involving stimuli, it is only natural to attempt examination of consumer behaviour with NNs (Curry & Moutinho, 1993).

NNs and Consumer Behaviour Research on the application of neural networks to the analysis and modelling of the consumer response to advertising stimuli was published by Curry and Moutinho (1993), where a comprehensive discussion of theoretical implications of neural networks is followed by practical application considerations. These authors cautioned against limitations and overoptimistism in the field. A typical NN input-output structure supplemented by a number of intermediary hidden layers brings certain advantages through a more sophisticated platform for modelling consumer behaviour as intermediary levels have a tendency

252  Max N. Greene et al.

to be linked with important conceptual phenomena predisposed to indirect measurement (e.g. latent variables). NNs models are trained: either through a supervised learning process where example connections of input and output pairs are fed into the model, or otherwise through relying on clustering methods in unsupervised learning (Curry & Moutinho, 1993).The ability to extrapolate rules from training sample data puts NNs in superior position compared to many rulebased arrangements common in expert systems. Neural networks are not however exempt from the basic statistical concept that results from non-random samples may not be generalisable. NNs are particularly appropriate in tasks that involve a concept of cognitive behaviour or pattern identification that is similar to the examination of consumer economics. In order to consider the application of artificial neural networks to a dataset composed of fast-moving consumer goods, similar to data used by Foxall and colleagues (Foxall, 2003; Foxall, Oliveira-Castro, & Schrezenmaier, 2004; Foxall & Schrezenmaier, 2003), a number of relevant articles are reviewed in the following sections. Van Wezel and Baets (1995) test the predictive performance of neural networks and compare it with traditional techniques in their paper on evaluating market response through the examination of variables on fast-moving consumer goods. They suggest a number of different choices to tackle the complex market response estimation task, including more commonly employed statistical models such as multiple linear regression and multiplicative model, and compare their predictive power with what authors call the best-known type: the back-­propagation neural network approach. The innate configuration of neural networks does not require any prior knowledge about the model structure over and above the constraints imposed by the neural network on the functional form of its outputs (Morgan et al., 1999) as it is established through training, and therefore does not require any assumptions about the input and output relationship (Van Wezel & Baets, 1995). This ability (network structure does not need to be predetermined), also suggested by Curry and Moutinho (1993) in work previously discussed, provides a powerful modelling arrangement. Some of the problems with neural networks are discussed as well—e.g. ­overfitting—when the model fit to the training set is so high that the model does not perform well with external data. Using comparative analysis, models are evaluated and this shows that neural networks outperform other traditional methods in all the cases tested. However, the NN model is often viewed as a ‘black box’, where theoretical interpretation of the process might pose a difficulty (Van Wezel & Baets, 1995). Explaining complex phenomena with comparatively simple techniques such as linear regressions could oversimplify the interpretation, and neural networks could be a better option. This suggestion was also expressed by Curry and Moutinho (1993).Van Wezel and Baets (1995) suggest a possible extension of research into the use of recurrent neural networks to model market behaviours, as such networks are capable of working with effects that are not immediately occurring.

Connectionist Modelling of Consumer Choice  253

Another study reports the findings of two experiments into a comparison of neural networks with discriminant analysis and logistic regression in terms of their ability to predict consumer choice (West, Brockett, & Golden, 1997). They argue that, even though neural networks are designed to quantitatively imitate the neurophysical structures and decision-making ability of the human brain, they nevertheless show a statistical resemblance to linear modelling and could predict consumer choice. The application of neural networks to study consumer behaviour choice poses benefits unmatched by other statistical methodologies, such as the ability to detect nonlinear and noncompensatory processes without assuming parametric relationships between variables, already suggested by others (Curry & Moutinho, 1993;Van Wezel & Baets, 1995). Their work shows neural networks to consistently outperform traditional approaches in predicting the outcome of noncompensatory rules. The robustness of neural networks has also been discussed, and the issue of overfitting addressed through the use of a validation sample in determining training termination. One can conclude then by stating that the neural networks show great usefulness in predicting consumer choice based on product attributes (West et al., 1997). In the analysis of supermarket shopping behaviour, neural networks have been used to predict customer satisfaction, number of trips to the supermarket, and the amount spent (Davies, Goode, Moutinho, & Ogbonna, 2001). The advantages of using the neural network in this type of analytical work are stated by them as follows: the neural network’s learning capacity allows sophisticated approximation which does not require researchers to specify underlying relationships prior to research and values of hidden nodes that could be interpreted as unobservable consumer behaviour variables. Davies et al. (2001) proceeded by building a number of neural networks and found that broad product range and quality exhibits the highest influence on customer satisfaction. They also found that customers with higher incomes were among those most satisfied, as such customers could take full advantage of choices offered and could travel longer distances to reach those supermarkets with higher available selections (Davies et al., 2001). Other shoppers were found to be more concerned with reasonable prices and store atmosphere. It seems that customer dissatisfaction comes from a feeling of choice restriction, through limited range or restricted purchasing power, and these could often be interconnected. Others caution that customer satisfaction should not be linked with spending, as the model suggested that only disposable income impacts spending directly, with other factors playing a small part (Davies et al., 2001). Moore et al. explore attitudes towards a web site (Moore, Beauchamp, Barnes, & Stammerjohan, 2007). They note that choice is based on noncompensatory rules and emphasise nonlinearity of the preference concept. Moore et al. (2007) also state that referential decision making contributes to new methods such as neural networks being considered. Through a demonstration of a neural networks model, they suggest that the process of decision making is amenable to neural networks analysis considering environmental and personal influences. Innovative

254  Max N. Greene et al.

consumer behaviour theory could also benefit from neural network applications (Moore et al., 2007). Practical beneficial outcomes of an effective neural networks predictive model could include better inventory management and merchandising, improved online experiences for customers, successful employee profiling and training, and strategic decision making.

Discussion The substantial amount of information obtained as part of the analyses undertaken here provides a suitable starting point for theoretical contemplation. A number of important points need to be outlined.

Simple NNs and Logit Comparison A comparison between NNs and Logit models was made using a market research homescan dataset with a very large number of cases of biscuit purchase and many variables such as brand or price, with a high number of individual consumers over 52 weeks of purchasing behaviour. In accord with the BPM concepts, informational and utilitarian reinforcement variables were included amongst the inputs to act as a test of some of the methods for assessing the relative importance of these variables in predicting customer loyalty. Anything between 0 and 200 neurons in a single hidden layer were used (where zero neurons represents a single output neuron with no hidden units). This zero-hidden neuron case has connections purely between the inputs and the output. In each case, the output neuron was a logistic node (having a logistic activation function). If this minimal network were used with the so-called cross-entropy error function, it is formally identical to a multivariate logistic (logit) regression. Indeed, it was found that the NN models performed very similarly to logit models even when the more usual Mean Square Error (MSE) was employed. As there were no hidden layers, simple NNs models at this stage were not able to account for nonlinear relations within the data. As a result, simple input-output NNs models equal logit models’ performance not only at the level of overall predictive capacity but also at the level of variable contribution.This analysis provides a clear link between NNs and logit as a method of analysis while working with consumer data and establishes a point where it is possible to observe the deviation of models based on NNs from the traditionally employed methods such as logit. A number of comments could be offered to discuss this.

Predetermined vs. Flexible Models The NN has been described as a nonparametric method—in the same sense as a nonparametric regression. Logit models are more likely to have some theoretical justification behind them. If faced with the choice between a defined logit model

Connectionist Modelling of Consumer Choice  255

based on a theory (with linearity of the model in variables, some variable being in the model as a quadratic, etc.) then one should always stick to the model which comes out of theory. However, we may be looking for relationships between variables as yet undefined or suspect the action of some latent variables. In that case, a NN has a lot to commend it since the familiar logit model is nested in it as a special case and yet one may find a set of large neural network weights all leading to the same node giving us the idea that the inputs where these weights come from may form a good explanatory combination. Both models as discussed so far—NNs and logit—are rather simple, but it is possible to develop and expand not only the NNs models, but the logits as well. The logical way to do so with the logit model is to explore variable contributions and trim the model to include only the most predictive independent variables, and introduce interactive variables. This, however, is increasingly difficult with traditional methods like logit, as large datasets would have a high number of possible interactions that need to be examined separately and in combination, and would require considerable time and effort to explore. This is, as noted elsewhere (Bishop, 1995), an inherent weakness of traditional methods of analysis compared with the NNs: traditional methods of analysis require a predetermined model structure. With the NNs models however, the ability of the network to determine its own structure as a result of the modelling process and learning the data is an obvious benefit. Since the NNs modelling process determines the optimal structure and there is no need for a researcher to specify it beforehand, increasing the model’s complexity and capacity is just a matter of introducing an additional ­(hidden) layer of computational nodes into the model between the input and output layers. This introduces nonlinearity, which considerably increases the model ability with consumer data.

Model Specification Second, the theoretical considerations of predetermined model structure are discussed.Variable selection is a complex task and in some instances could be a research question in its own right (for example Greenland, 1989). In other instances, model structure could be dictated by an underlying theoretical or philosophical framework as described earlier. Nevertheless, in traditional methods, such as logit, it is necessary for the researcher to specify the model structure before the intended analyses are performed.This could carry both negative and positive consequences, depending on the theoretical perspective. On the one hand, the ability of linear methods to link the model structure with the underlying theoretical or philosophical framework with relative ease could prove to be quite useful as it allows for straightforward, easy to understand explanations of the examined phenomena. On the other hand, a predetermined structure would inevitably have an effect on the results obtained in such a way, and it could be argued that a researcher’s influence and presence within the research is unnecessarily increased as a result—to

256  Max N. Greene et al.

the point that obtained results are nothing but a statistical artefact in extreme cases (Harris & Hahn, 2011). Depending on the nature of the research, the aspect of predetermined structure could play an important role. If research directly deals with the analysis that examines the structure and form of the process or phenomenon in question, it would be quite sensitive to the structure of the model, as the very subject studied would be constrained by the choices a researcher makes during the modelling stages. It would be much more appropriate to allow for a model to develop its structure in a process most suitable to the data.

Are NNs ‘Black Boxes’? A frequently encountered claim in the marketing literature against NNs models suggests that an NNs model process resembles a ‘black box’ modelling process, where the NNs model takes in the data, and as a result of some internal manipulations, provides an output with no apparent way to examine the internal modelling process (Gevrey, Dimopoulos, & Lek, 2003; Olden & Jackson, 2002; Olden, Joy, & Death, 2004). Our experiments with the consumer loyalty data suggest that this is not true. Of course, for trivial cases with no hidden layer it must be untrue (the word ‘hidden’ in ‘hidden layer’ has a lot to do with this black box misconception). If the least complicated NNs input-output model is able to provide results that are identical to logit models, it should be possible to infer that the explanatory power attributed to logit models is equally applicable to the identical simple NNs model. The equality of the weights obtained by the NNs models and the coefficients values from the logit models suggest the same level of explanatory capacity of NNs models as that of regression. Highly complex NNs models that are often referred to as ‘black box’ models are based on the same theoretical and statistical assumption as are the simple input-output models. So there is a hope that complex NNs models with hidden layers should also provide high explanatory power. Perhaps the highly complicated nonlinear account that NNs models are able to offer is often just too complex to be represented by simpler tools often used to represent linear relations. A consideration of the explanatory capacity of NNs models in this light would hardly suggest a ‘black box’ analogy and we give a short account in Section 2.4 of some of the methods for overcoming the barrier of this seeming complexity.

Model Performance and Nonlinearity Analysis NNs models performed considerably better than logit models once a hidden layer was incorporated into the models.The network structure that incorporates hidden nodes between inputs and outputs allows exploration of nonlinear relationships.2 For example, an NN with no hidden layer is unable to provide anything other than a linear separatrix between groups in a classification problem.3 It is clear from the results, that when consumer data and consumer behaviour is the field of

Connectionist Modelling of Consumer Choice  257

study, nonlinearity provides a substantial advantage over the linear models. Relatively weak performance levels of logit models shown could be due to a number of factors. It is possible that the data does not contain the variables vital for the prediction of consumer loyalty information, or that the variables that are readily available for marketing researchers (and therefore most frequently collected and analysed) do not contain sufficient predictive power. Another possibility is that the relations of independent variables with the dependent variable that describes consumer loyalty are not linear. If this is indeed the case with the insufficiently predictive data, there is not much that can be done. ‘Rubbish in—rubbish out’ as they say, meaning no good models could be built using bad data. (It is easy enough to build bad models with good data, and an overreliance on the approximational promise of neural networks can also lead us to deceive ourselves into believing we have built good models from bad data!) The importance of model validation is therefore clear and the k-fold cross-validation technique is very useful in this context. The 10-fold cross-validation method, for example, was used to assess the model’s performance, the data being split into ten random subsets where nine subsets are combined to train the model and the remaining tenth subset is used to test the model. This was then replicated ten times, by rotating each subset in as a test set whilst using the other nine as a training set. This method offers a very powerful statistical analysis. If however, the problem lies in the nonlinearity, more appropriate methods of analysis would be able to extract the relations from the data. As results show, this is the case with the dataset used here, as NNs methods were able to extract a lot more information useful in prediction analyses. The NNs model complexity tests showed some promising results as well. While working with the smaller dataset, sufficiently sophisticated NNs models are capable of learning the entire dataset with the appropriate training and therefore make perfect predictions.The dataset used here though is sufficiently large to avoid such issues and allows the testing of networks that are particularly complex. Results obtained employing such test design provide information on the effects of model size on overall model performance. The performance of NNs models containing a number of hidden nodes that range from 1 to 100 (and several models that incorporate even higher numbers of hidden neurons as described earlier) compared with the performance of logit models, show continuous improvement in the NNs model ability as the model size increases—and as a result, shows the ability of the model to account for the nonlinear relations within the data. From this, it should be safe to suggest that NNs models are particularly suitable for the analysis of consumer behaviour data. Figure 9.3 results show receiver operating characteristic curves with 100 overlaid NN models in each figure with hidden nodes ranging from 1 to 100. Two random subsets (left and right) were used each time, and the whole procedure was replicated ten times (iterations) for validity and reliability purposes. Results look very similar across the iterations.

258  Max N. Greene et al.

FIGURE 9.3 

Receiver Operating Characteristic Curves

Overfitting It is expected that at some point NNs models’ performance will flatten out as model size is considered in the comparative analysis and large models are penalised. The consumer behaviour dataset used here however, may be large enough to allow for bigger NNs models to improve continuously, extracting even more information every time to increase model performance. The NNs model developed for the tests on the loyalty data was used for a series of single hidden layer networks with a number of neurons that ranges from 1 to 100 to examine to what extent an increase in the size and nonlinear capacity of the model improves the predictive power. It is however, possible in addition, by increasing the number of neurons within a hidden layer, to also include multiple hidden layers, thus increasing model complexity even further by adjusting a number of neurons within each hidden layer. This is often an unnecessary step as the NNs models with multiple hidden layers are prone to overfitting if the number of neurons in the multiple

Connectionist Modelling of Consumer Choice  259

hidden layers are appreciable since the number of weights between any two layers is related to the product of the number of nodes in each layer.,They are not always that easy to fit as well having, as mentioned earlier, more inherent nonlinearity than their single hidden layer counterparts.The problem of overfitting and model generalisability is a potential area of future research, but what multiple hidden layers enable us to do is to provide much smaller models in terms of the number of weights than for a single hidden layer (see the example of fitting a peak function in Section 1.3). To do so, the single hidden layer NNs model size needs to be increased by increasing the number of hidden neurons to the point where an increase in the model performance due to the model size is not sufficient to cover the size penalty applied to the model by the comparative mechanism. As a result, the optimal, single layer NNs model could be identified from the computational standpoint. It is then necessary to evaluate the performance of the models on outof-sample data to assess their generalisability and their ability to make predictions using new data and identify the optimal model architecture using these criteria as well, consequently comparing such models with the computationally optimal model. Mechanisms such as early stopping (where an out-of-sample test dataset is monitored during training on a separate training dataset for any substantial increase in error to stop training when prediction is about to get worse) could be examined as that could help avoid overfitting issues during the model training stages. Depending on the primary model application, be that either predictive capacity or explanatory power, a number of strategies could be implemented to improve model performance. The following sections focus on NNs models predictive and explanatory abilities in greater detail.

NNs Models Predictive Capacity It should be apparent from the results presented here that forecasting models based on NNs have the most promising perspectives with the consumer behaviour data. One of the possible applications of such models is the accurate prediction of future consumer behaviour. The number of units and the connection topology could exert a major influence over network performance. A number of varying techniques have been developed to optimise the network architecture—some are incorporated into the process of training the network. It is possible to divide the task of architecture selection into two distinct concepts. The first concept is the methodical procedure for exploring the range of potential architectures, whereas the second concept is concerned with the decision mechanism on which architectures to select. The issue of selection mainly deals with the problem of model generalisation, briefly discussed earlier, and the primary function of the network and what it is designed to accomplish. The most commonly used and obvious approach to network structure optimisation is a deliberate exploratory analysis of a limited

260  Max N. Greene et al.

subset of architectures, such as the fully connected feedforward network with a single hidden layer and direct input-output connectivity as described here. The aspect of network architecture that remains to be explored is the number of neurons within the hidden layer. So then to identify the optimal architecture we would train a number of networks as described earlier and select the one which meets our performance criterion most fully. This approach however presents a number of limitations as it could require a considerable computational effort (the models described here with 50 to 100 nodes incur long training times), to explore a limited group of networks. If the group inclusion criteria are broadened, as suggested earlier, to include additional hidden layers or exploring partial network connectivity, the feasibility of reaching the computational limit becomes increasingly real (Bishop, 1995). In the case where network predictive capacity is of primary significance, (in contrast to our experience with NNs fitting our consumer behaviour and marketing research data) the apparent drawback would be the necessity of training a large number of networks with varying architectures. An alternative approach would involve a relatively small initial network to be gradually developed through continuously added additional neurons. Methods of this nature are commonly referred to as growing algorithms and are a potential area of future research using consumer behaviour data that may be worth exploring in the future (Bishop, 1995). Yet another way to improve predictive capacity would be to gradually remove neurons and synapses from a large network such as the ones described here. These methods are referred to as pruning algorithms (Bishop, 1995). Pruning methods essentially reduce the fully connected network to a sparsely connected one (Morgan, 2008). (This is what happens in the early times of a developing brain where superfluous connections and neurons remove themselves.) One would hope that somehow the pattern of connection in the sparse network would somehow mirror the process which generated the data (the consumers’ decision process, for example). It is also possible to develop a complex network that consists of a number of simpler networks combined together. This method is commonly referred to as a network committee and is suitable for a problem decomposition approach where a task is subdivided into smaller tasks for separate examination (Bishop, 1995). These offer numerous opportunities to explore new methodologies such as training networks in parts and then retraining the whole network again once it has been assembled. All these methods of network architecture optimisation could potentially improve the model and provide superior predictive abilities with a simpler architecture. It is an obvious area for future research which would expand the abilities of the models presented here.

NNs Models Explanatory Capacity If one is convinced by the predictive abilities of nnS-BASED models, it is important to address their explanatory capacity. Largely omitted from this research

Connectionist Modelling of Consumer Choice  261

project for reasons of space limitation, an assessment of explanatory descriptive modelling abilities would complement the discussion of normative prescriptive power explored here.To address the ‘black box’ label frequently attributed to NNs models, exploration of variable contribution profile is logical. A framework successfully implemented in the ecological sciences (Gevrey et al., 2003; Olden & Jackson, 2002; Olden et al., 2004) could prove useful when used with the consumer behaviour data, where a number of methods are systematically compared and assessed. Continuing the discussion regarding the ‘black box’, it could even be said that traditional models are the ones unable to provide comprehensive variable contribution information. In regression, the common practice is to interpret variable influence through the assessment of partial regression coefficient, final values. However, this is only applicable to the variables with significant coefficients. Moreover, regression provides only coefficients with a sign and a value for each independent variable, signifying a relationship direction, but no other information could be extracted from the model output; whereas with NN models it is possible to employ explanatory methods to determine independent variable contributions and their effect the dependent variable. These algorithms have been developed to clarify the so-called black box modelling method of NNs models and are able to provide the relative variable contribution and contribution profile of the input factors. Unlike the pruning methods discussed earlier that aim to improve the predictive capacity of NNs models, variable contribution algorithms are designed to determine what contribution each variable makes. They are procedures designed to estimate the relative contribution of independent variables and determine the influence and contribution of each variable to the output. Again referring to the ecological modelling research (Gevrey et al., 2003), seven methods could potentially be employed to perform contribution analysis with consumer behaviour data: a) The ‘PaD’ method calculates the partial output derivatives according to the input variables (Dimopoulos, Chronopoulos, Chronopoulou-Sereli, & Lek, 1999); b) The ‘Weights’ method uses the connection weights (Garson, 1991); c) The ‘Perturb’ method performs input variable perturbation (Scardi, 1999); d) The ‘Profile’ method is a successive variation of one input variable while others are kept at a fixed value (Lek et al., 1996); e) The ‘Classical Stepwise’ method observes the change in the mean square error value by sequentially adding or removing input neurons (see Gevrey et al., 2003); f) The ‘Improved Stepwise Method A’ is the same as e), but the input eliminations occur while the network is trained (see Gevrey et al., 2003); and g) The ‘Improved Stepwise Method B’ evaluates the change in the mean square error by sequentially setting input neurons to the mean value (see Gevrey et al., 2003).

262  Max N. Greene et al.

A multi-layer feedforward network architecture was used (similar to the network architecture used here) to identify the most useful method of variable contribution analysis. PaD and Profile methods are able to provide two elements of variable contribution information: order of contribution and mode of action, along with the ability to order the variables by importance of their contribution to the output that the other methods are capable of. The PaD method is more coherent computationally, as it uses partial derivatives and real dataset variables values, whereas the Profile method constructs a representational matrix based on the data. As a result of the test, the PaD method was found to be the most useful, as it was able to provide the most complete results, followed by the Profile method that offers an input variables contribution profile. The Perturb and the simplified Weights methods achieve good input parameter classification but lack stability. The two Improved Stepwise Methods A and B and Classical Stepwise Method did not perform as well as the other methods on the contributions. Ecological data is highly nonlinear. So is consumer behaviour data since we find that the hidden layer models fit our consumer data so much better than the logit/no hidden layer NN. Such promising results when used with the ecological modelling may suggest that these variable contribution analysis methods could be equally applicable to the consumer behaviour data. It is important to note however that the sample Gevrey and colleagues (2003) used is considerably smaller than the data sample used here, and the methods that did not perform as well in their study could very well be quite powerful with larger samples. The 10–5–1 network structure that they used generates a model considerably different in size to the models tested here (even though 10–5–1 models are also tested here) which could be a factor as well—this however, could be a possible area of further research in order to examine how variable contribution analysis methods perform depending on the model size. Similar concerns were expressed elsewhere (Olden et al., 2004) where it was noted that true relationships and variable order in the empirical dataset used by Gevrey et al. (2003) are not known, and therefore results obtained could not be accurately assessed. Furthermore, a number of methodological issues with the research design were pointed out including the approach to assess method stability used by Gevrey et al. (2003), where Olden and colleagues (2004) argue that this is in fact an issue of optimisation stability. To address these concerns, simulated data was used to re-examine variable contribution methods discussed by Gevrey et al. (2003), with an additional method examined: the Connection Weights method (Olden & Jackson, 2002). As a result of the experiment, Connection Weights method (Olden & Jackson, 2002) showed best performance. Therefore, due to the fact that simulated data is used by Olden et al. (2004), it is suggested that the results obtained could be equally applicable to consumer behaviour data—and not only ecological research. This however needs to be further explored, and variable contribution analysis methods assessed in relation to consumer behaviour data, to identify the best methodology that gives the most robust and accurate results.

Connectionist Modelling of Consumer Choice  263

Theoretical Implications Once convinced by the predictive capacity of NNs when used with consumer data and by the appropriateness of using nonlinear modelling techniques, it is important to discuss what this means in terms of theory. NN modelling is based on the connectionist theoretical framework where simple computational units put together are capable of displaying high performance that is unattainable individually. The inherent mechanisms of activation within the connectionist models are rather simple. The fact that original NNs models were developed to imitate the processing capacity of the brain should suggest the connectionism would be particularly useful and appropriate to study human behaviour.The work done in the field of cognitive psychology could provide supporting evidence for such claims. While studying animal behaviour and cognition, the connectionist framework has been shown to be particularly appropriate in explaining certain aspects of discrimination learning. Pearce (1994, 2002) developed a comprehensive model based on the work of Herrnstein (1970) that could be used to predict animal behavioural responses to outside stimuli in a quantitative manner. It was shown how such behaviour could be elegantly explained through connectionism on a biological neurons and synapses level inside the brain. Many others also noted the appropriateness of NNs and connectionism to study human behaviour (Curry & Moutinho, 1993). Another field that embraces NNs models as powerful tools capable not only of explanation, but also of new promising lines of research discovery, is psycholinguistics. Acquisition and development of language is extremely complex and follows a long learning process. NNs models are developed through the process that is also often referred to as the learning process. It is not surprising then that psycholinguists increasingly turn to NNs models to help explain language acquisition. A somewhat different yet very successful application of NNs models could be observed in engineering. Here NNs-based models have been proven to be increasingly successful at computationally demanding tasks such as automated face recognition (Er, Wu, Lu, & Toh, 2002; Lawrence, Giles, Tsoi, & Back, 2002; Rowley, Baluja, & Kanade, 2002).

Concluding Remarks This chapter has given an account of and the reasons for the widespread use of NNs models to study a wide range of phenomena with great success. It is argued here that connectionism holds the finest instruments for explaining consumer behaviour. The predictive abilities of NNs models in explaining consumer choice have been investigated and the usefulness of the connectionist framework to the BPM discussed.

264  Max N. Greene et al.

This study set out to determine whether NNs models could be useful in explaining consumer behaviour following the established theoretical framework of the BPM. In the course of this project a large number of NNs models (2,000 models, number of nodes within the hidden layer ranging from 0 to 200) of varying complexity have been developed and assessed. This was done by comparing the NNs models with the traditional methods of analysis such as logistic regression, and through a comparison of NNs models with each other in the test that examined the predictive power and contribution value of informational and utilitarian reinforcement variables. Returning to the hypothesis posed at the beginning of this study, it is now possible to state that NN models with hidden layers showed a better performance than the traditional methods of statistical analysis (logit) did. This study has shown that NNs models show the capacity to help develop the understanding of consumer behaviour in the future. These findings suggest that in general the complex nonlinear nature of consumer behaviour data could be analysed with the parallel connectionist models relatively successfully. One of the more significant findings to emerge from this study is that the logistic regression that may very well be the preferred method of analysis in marketing literature was greatly outperformed even by the simplest of NNs models. The second major finding was that the performance of more complex NNs models just kept improving as the additional neurons were added into the models. This is likely to be explained by the relatively large dataset employed here that allowed each successively more complex model to find more significant relations within the data that contributed to the explanation of consumer choice. (The training time is clearly longer for larger networks.) The relevance of utilitarian and informational reinforcement variables in predicting consumer behaviour is clearly supported by the experiments we have carried out. A number of models of varying complexity (200 models, 1–100 hidden nodes) were developed to examine the utilitarian and informational reinforcement variable contribution, and results have shown that the models that included reinforcement variables consistently produced better NNs models as compared with the NNs models that excluded the informational and utilitarian reinforcement variables. Utilitarian and informational reinforcement emerged as reliable predictors of consumer choice. The evidence from this study suggests that consumer data contains nonlinear relations between the variables normally considered by marketing researchers (demographics, product details, consumer situation information). NNs models therefore could be increasingly useful in working and modelling such data. The results of this research support the idea that the proven framework of BPM could be considerably extended with the application of connectionist constructs to help explain consumer behaviour and consumer choice. The interdisciplinary nature of connectionism also complements the complexity of consumer behaviour research that often draws upon different disciplines such as psychology, economics,

Connectionist Modelling of Consumer Choice  265

marketing, and other to develop a complete account of consumer behaviour. In general, therefore, it seems that connectionist models are able to account for complexities within the data that linear models are unable to do. Not only is the predictive power that NN models are able to provide in many cases superior to traditionally employed statistical methods such as logistic regression, but also the explanatory power that NN models offer employing a number of algorithms greatly surpasses that of traditionally employed methods. Taken together, these results suggest that connectionism is one string of research that could be a logical continuation for the BPM framework promising good new findings in the future. This research will serve as a base for future studies into the application of connectionist models to consumer behaviour data. These findings enhance our understanding of the consumer situation and provide an alternative approach to examining the decision-making process that revolves around purchasing behaviour. The current findings add to a growing body of literature on the application of NNs to the study of complex cognitive phenomena and the examination of nonlinear data that is subsequently able to provide the predictive ability of the future events with a convincing degree of accuracy. The methods used for this consumer data and product category may be applicable to other data and product categories as well, which would allow for an assessment of the generalisability that the models developed are able to offer. The empirical findings in this study provide a new understanding of the application of these variables in predicting and forecasting consumer choice. Finally, a number of important limitations need to be considered. First, the most important limitation lies in the fact that due to space constraint the models have not been tested using the initially planned out-of-sample testing and k-fold cross-validation. Second, the current investigation was limited by the nature of the dataset. Even though the data employed here included a very large number of cases and many variables with a high number of individual consumers, it was limited to 52 weeks of purchasing behaviour and a single product category. The limited time span prevents certain test designs such as where the data is split chronologically, and models are trained on the first weeks (months, years) of the data and are subsequently tested on the last weeks of the data. Tests of such nature provide an obvious benefit of testing the model on the real data taking the experiment even further away from the laboratory into the real market situation. It is also not assumed at any point in the paper that these results should also be applicable to other product categories, which remain to be examined. Third, even though the current research was not specifically designed to evaluate the data with a continuous dependent variable, one source of weakness in this study, which could have affected the measurements of consumer loyalty, was that the probabilistic loyalty value had to be converted into binary. The nature of decision making is rarely represented in a form of a choice between few alternatives or a binary type of a problem (such as belonging to one of the two groups) as examined here, but

266  Max N. Greene et al.

rather is a probabilistic measure. As essential predictive information is lost during the transformation of probabilistic value into binary, the tests designed to employ probabilistic variables should offer better results. This then also is a promising area of future research as NNs models could be compared to other traditionally employed methods such as multiple regression. (After all, feedforward NNs like the ones described here are simply nonlinear regressions.) This research has thrown up a number of questions for further investigation, and a number of possible future studies using the same experimental set up are apparent. It is therefore recommended that further research be undertaken in the following areas: (1) the predictive and (2) explanatory capacity of the connectionist framework and NNs models working with the consumer behaviour data and the process of consumer decision-making needs to be further examined employing the variable contribution algorithms and out of sampling testing to accurately assess model ability working with the new data. Further experimental investigations are needed to estimate the extent of NNs model capacity to predict new previously unseen data. Further research might investigate the best algorithms to maximise model predictive ability relative to its size, and the relative ease of integrating the application of these models in the marketing industry. A promising line of research entails the use of pruning methods since these echo the mechanisms inherent in the development and abandonment of neurons and neural connections in the developing brain. Considerably more work will need to be done to determine the explanatory capacity of NNs models and their ability to explain the consumer decision-making process. The research here might take a more academic direction and involve significant interdisciplinary collaboration from such research fields as cognitive psychology, behaviourism, connectionism, economics, and marketing.

Student Exercises QUESTION 1 What are the essential features of a ‘mathematical neuron’ (McCulloughPitts neuron’)? Describe one significant difference between a biological and a mathematical neuron. Inputs to neurons can be described as ‘excitatory’ or ‘inhibitory’ when the weights on them are positive or negative, respectively. If a neural network were modelling a consumer’s decision to buy so that the

Connectionist Modelling of Consumer Choice  267

neuron would ‘fire’ if they purchased the item or not fire if the item were not purchased, discuss how the following inputs to the purchasing decision might be modelled by positive or negative weights. 1. 2. 3. 4. 5. 6.

The shopper has just had a gift voucher. The item is branded. The item is part of a ‘two-for-one’ offer. The item is priced surprisingly low. The item is attractively packaged. The item has been priced at twice the price of a number of nearby stores.

OUTLINE ANSWER The mathematical neuron has inputs which are summed, and this sum subjected to an activation function, the output of which is the output of the neuron which may then be used as an input to other neurons. The biological neuron uses pulses which travel along the dendrites to the body of the neuron whereas the mathematical neuron uses continuously valued signals. 1. This would be modelled by a positive weight since access to funds would enable rather than discourage a purchase. 2. As long as the brand was trusted, this would be modelled by a positive weight and enhance the probability of purchase. 3. Again this is likely to be represented by a positive weight as long as the consumer believes the offer to be genuine. 4. In this case, low price could act as an indicator of low quality if the good in question is a luxury item and hence inhibit purchase. 5. One would expect this feature to increase the probability of purchase. 6. Here the higher relative price could inhibit purchase. QUESTION 2 Sketch a feedforward neural network with biases, two linear input nodes, a single hidden layer of two logistic neurons, and a single logistic output node.

268  Max N. Greene et al.

OUTLINE ANSWER Bias Node with constant value

Linear Input Node

Logistic Output

Hidden Layer of Logistic Nodes

QUESTION 3 Consider the following statement. Connectionist models of consumer data, such as neural networks, offer superior performance because they can simulate human learning. To what extent does this statement have any merit? What flaws can you see in this position? OUTLINE ANSWER Some of the points which could be raised are 1) a connectionist model, such as a neural network, is adaptive because of the way that successive data fed into it can change the weights and hence fit the data. 2) The neural network is a nonlinear model and can therefore fit a greater range of datasets. 3) Large neural networks have a very large number of adjustable parameters (weights) and are hence prone to overfitting

Connectionist Modelling of Consumer Choice  269

(which can of course be countered by measures such as pruning). 4) Neural-network models with hidden layers are highly nonlinear and hence may converge to a suboptimal solution (a local minimum in the error surface). Just because a system is ‘learning’ does not mean that it cannot ‘make mistakes’. Human learners make mistakes all the time!

Notes 1. The development of methods such as differential evolution, genetic algorithms, and other evolutionary methods is an active area of research. 2. It is worth remembering that regressions and single hidden layer NNs employ a linear combination of possibly nonlinear (e.g. quadratic) functions. The output of such a network is a sum of scaled and shifted logistic functions. If the input-to-hidden layer weights were kept fixed the hidden-to-output weights would play the same part as regression coefficients. It is these input-to-hidden weights that make the network nonlinear. NNs with additional hidden layers are no longer as simple as that and are intensely nonlinear since they act nonlinearly on already nonlinear functions 3. This is the origin of the Minsky and Papert (1969) observation that two-layer perceptrons were incapable of solving the XOR problem which is reputed to have contributed to the temporary halt in NN research at the time.

References Adya, M., & Collopy, F. (1998). How effective are neural networks at forecasting and prediction? A review and evaluation. Journal of Forecasting, 17(5–6), 481–495. Bishop, C. M. (1995). Neural networks for pattern recognition. New York, NY: Oxford University Press. Chester, D. L. (1990). Why two hidden layers are better than one. IJCNN-90-WASH-DC, Lawrence Erlbaum, 1990, 1, 265–268. Curry, B., Davies, F., Evans, M., Moutinho, L., & Phillips, P. (2003). The Kohonen selforganising map as an alternative to cluster analysis: An application to direct marketing. International Journal of Market Research, 45(2), 191–211. Curry, B., & Moutinho, L. (1993). Neural networks in marketing: Modelling consumer responses to advertising stimuli. European Journal of Marketing, 27(7), 5–20. Cybenko, G. (1989). Approximations by superpositions of sigmoidal functions. Mathematics of Control, Signals, and Systems, 2(4), 303–314. Davies, F. M., Goode, M. M. H., Moutinho, L. A., & Ogbonna, E. (2001). Critical factors in consumer supermarket shopping behaviour: A neural network approach. Journal of Consumer Behaviour, 1(1), 35. Dimopoulos, I., Chronopoulos, J., Chronopoulou-Sereli, A., & Lek, S. (1999). Neural network models to study relationships between lead concentration in grasses and permanent urban descriptors in Athens city (Greece). Ecological Modelling, 120(2–3), 157–165. Er, M.,Wu, S., Lu, J., & Toh, H. (2002). Face recognition with Radial Basis Function (RBF) neural networks. Neural Networks, IEEE Transactions on, 13(3), 697–710.

270  Max N. Greene et al.

Foxall, G. R. (2003). The behavior analysis of consumer choice: An introduction to the special issue. Journal of Economic Psychology, 24(5), 581–588. doi:10.1016/s01674870(03)00002-3 Foxall, G. R., Oliveira-Castro, J. M., & Schrezenmaier, T. C. (2004). The behavioral economics of consumer brand choice: Patterns of reinforcement and utility maximization. Behavioural Processes, 66(3), 235–260. Foxall, G. R., & Schrezenmaier,T. C. (2003).The behavioral economics of consumer brand choice: Establishing a methodology. Journal of Economic Psychology, 24(5), 675–695. doi:10.1016/s0167-4870(03)00008-4 Garson, G. D. (1991). Interpreting neural-network connection weights. AI Expert, 6(4), 46–51. Gevrey, M., Dimopoulos, I., & Lek, S. (2003). Review and comparison of methods to study the contribution of variables in artificial neural network models. Ecological Modelling, 160(3), 249–264. Greenland, S. (1989). Modeling and variable selection in epidemiologic analysis. American Journal of Public Health, 79(3), 340. Harris, A. J., & Hahn, U. (2011). Unrealistic optimism about future life events: A cautionary note. Psychological Review, 118(1), 135. Haykin, S. (1994). Neural networks: A comprehensive foundation. Oxford: Maxwell Macmillan International. Hebb, D. O. (1949). The organisation of behaviour. New York, NY: Wiley. Herrnstein, R. J. (1970). On the law of effect. Journal of the Experimental Analysis of Behavior, 13, 243–266. Hornik, K. (1991). Approximation capabilities of multilayer feedforward networks. Neural Networks, 4(2), 251–257. Lawrence, S., Giles, C., Tsoi, A., & Back, A. (2002). Face recognition: A convolutional neural-network approach. Neural Networks, IEEE Transactions on, 8(1), 98–113. Lek, S., Delacoste, M., Baran, P., Dimopoulos, I., Lauga, J., & Aulagnier, S. (1996). Application of neural networks to modelling nonlinear relationships in ecology. Ecological Modelling, 90(1), 39–52. McCulloch, W., & Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. Bulletin of Mathematical Biophysics, 5, 115–133. Retrieved from www.cse.­ chalmers.se/~coquand/AUTOMATA/mcp.pdf Minsky, M., & Papert, S. (1972). Perceptrons: An introduction to computational geometry (2nd ed., 1st ed., 1969). Cambridge, MA: The MIT Press, ISBN 0-262-63022-2 Moore, R. S., Beauchamp, M. B., Barnes, C. Z., & Stammerjohan, C. A. (2007). Artificial neural networks: An introduction and application in a consumer behavior setting. Marketing Management Journal, 17(1), 27–38. Morgan, P. H. (2008). Differential evolution and sparse neural networks. Expert Systems, 25(4), 394–413. doi:10.1111/j.1468-0394.2008.00466.x Morgan, P. H., Curry, B., & Beynon, M. J. (1999). Comparing neural network approximations for different functional forms. Expert Systems, 16(2), 60–71. Olden, J. D., & Jackson, D. A. (2002). Illuminating the “black box”: A randomization approach for understanding variable contributions in artificial neural networks. Ecological Modelling, 154(1), 135–150. Olden, J. D., Joy, M. K., & Death, R. G. (2004). An accurate comparison of methods for quantifying variable importance in artificial neural networks using simulated data. Ecological Modelling, 178(3–4), 389–397.

Connectionist Modelling of Consumer Choice  271

Pearce, J. M. (1994). Similarity and discrimination: A selective review and a connectionist model. Psychological Review, 101(4), 587. Pearce, J. M. (2002). Evaluation and development of a connectionist theory of configural learning. Animal Learning & Behavior, 30(2), 73. Ripley, B. D. (1996). Pattern recognition and neural networks. Cambridge, New York, NY: Cambridge University Press. Rowley, H., Baluja, S., & Kanade, T. (2002). Neural network-based face detection. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 20(1), 23–38. Rumelhart, D. E., & McClelland, J. L. (1987). Parallel distributed processing, explotation in the microstructure of cognition-Vol. 1: Foundations. In Computational models of cognition and perception (p. 1). Cambridge, MA: MIT Press. Scardi, M. (1999). Developing an empirical model of phytoplankton primary production: A neural network case study. Ecological Modelling, 120(2–3), 213–223. Sent, E. M. (2004). Behavioral economics: How psychology made its (limited) way back into economics. History of Political Economy, 36(4), 735. Simon, H. A. (1982). Models of bounded rationality. Cambridge, MA: The MIT Press Simon, H. A. (1987). Behavioural economics. The New Palgrave: A Dictionary of Economics, 1, 221–225. Van Wezel, M. C., & Baets, W. R. J. (1995). Predicting market responses with a neural network: The case of fast moving consumer goods. Marketing Intelligence & Planning, 13(7), 23–30. West, P. M., Brockett, P. L., & Golden, L. L. (1997). A comparative analysis of neural networks and statistical methods for predicting consumer choice. Marketing Science, 16(4), 370.

10  NITING THEORY AND U EMPIRICAL RESEARCH Marketing Research and Market Sensing Melvin Prince, Gillie Gabay, Constantinos-Vasilios Priporas, and Howard Moskowitz

In the context of market sensing, marketing research plays a role in organizing and extending knowledge about markets and their consumer. Marketing research accomplishes these tasks by generating raw data to be placed into analyzable data files, unearthing useful facts, producing actionable knowledge, estimating market response, and recommending marketing actions and resource allocations. An understanding of the underlying dynamics may or may not drive marketing research, which can end up in effect being a ‘cartography’ of the consumer (Gabay & Moskowitz, 2012). The cognitive perspective of information as proposed by Targowski (2005) is an underlying theory which can provide this organizing mechanism, connecting marketing research and market sensing. In his theoretical structure Targowski posited an information hierarchy (cognition reservoir and semantic ladder) with five cognition units traversing the range from data to wisdom. The five units are, respectively, data, information, concept, knowledge, and wisdom. In general, data are processed to information, information is processed to concepts (sense of direction), concepts are evaluated by available knowledge (point of reference), and, finally, knowledge morphs or transforms into wisdom. Scholars (e.g. Awad & Ghaziri, 2004; Bjorkman, 2012) assert that information has value and it supports decision making and thus it must be meaningful, valuable, useful, and relevant (Rowley, 2007). Information starts out as data collected during research. Information is processed data that provide answers to who, what, where, when, and how questions (Ackoff, 1989). These questions are central in marketing research as well in market sensing. However, market sensing is not synonymous with market or marketing research. The former mainly focuses on data collection and interpretation through various techniques and tools (i.e., surveys, focus groups, e-surveys, experimentation,

Uniting Theory and Empirical Research  273

observation, projective techniques, etc).The latter is broadly defined as firm’s ability to acquire and use market information that can be obtained through formal and informal mechanisms to make sound decisions (Prince & Priporas, 2015). More explicitly, it is a process that not only allows organizations and professionals to enhance their understanding and gain deeper insight of its external environment and actors (i.e., customers, competitors, suppliers, etc.) but also to be proactive to market signals (changes) and make well informed and sound decisions (Day, 1994; Piercy, 2012; Prince & Priporas, 2015). In other words, market sensing is an approach of thinking about the market research data and not just a way to use them (Kinberg, 2014). The aim of this chapter is to integrate marketing research and market sensing theoretical aspects and empirical research using a case study on congestive heart failure patient re-admissions. This chapter addresses the issue of decision making regarding messaging which affects patient adherence consequent to congestive heart failure (HF) based on Targowski’s theory. Given market sensing is a continuous effort (Day, 1999) and crucial for every firm’s survival on the other hand marketing decisions are highly analytical and are assisted by marketing research, we contribute to the ongoing literature of market sensing in the following ways. First, although there is ample marketing literature focusing on different aspects of market sensing (Prince & Priporas, 2015), the comprehensive and holistic view as well as direct studies (theoretical and empirical) on the topic are still limited. Second, we present the conceptual linkages of the market sensing and marketing research. Finally, we shed light on decisions and practices associated with market sensing studies and marketing research techniques. In the following case, the synthesis of the two concepts is presented by demonstrating the scope of marketing research in market sensing. The case illustrates the central role of marketing research in conducting empirical investigations about consumers and markets to promote market sensing uses and value. Additionally, several methodological themes are shown to be associated with this role.

Market Sensing and Marketing Research Case: Reducing Congestive Heart Failure Patient Re-Admissions Abstract Our research suggests three different groups of heart failure patients, when we study how they react to messages about taking medication. Each mind-set responds positively to some messages and responds negatively to others. Certain different high-impact marketing messages appeal to each segment of medication takers but what appeals to one group may repel the other group. The mind-set viewpoint identifier developed in this research allows one-onone, facts-driven patient counseling, based on mind-sets. Segmentation increases adherence by focusing on the messages appropriate for the particular patient’s

274  Melvin Prince et al.

mind-set. As a result we improved patient health, reduced patient stays and readmittance due to chronic illnesses following HF. As clinicians attempt to implement messaging in order to reduce re-admissions of HF patients, it is important to discover and then to more deeply understand what messaging is effective for each mind-set. Some patients are more impacted by content-oriented messaging and some by education-oriented messaging. The establishment of rapport, the conveying of a genuine interest in patients while talking about specific health care goals are all recommended with all patients. Providers should assess the importance patients place on and the confidence they feel with respect to specific health behaviors to determine their readiness or motivation. It is important during the clinical encounter to explore the importance of regimen-related behaviors and build patient confidence. Assuming that patients do want to hear what providers want to tell them, exchanging information is a critical part of the behavior-change process. A rationale should be provided for the recommended treatments. However, it is important to remember that simply providing information to increase knowledge will not guarantee that behavior change occurs.

Background One of the most frequent causes of hospital readmission is congestive heart failure (HF). HF is a major concern for medical care, because of its widespread occurrence and because of the costs involved in the hospitalization and readmission. 1. Approximately 6 million Americans and 15 million Europeans live with HF (Dickstein et al., 2008; Lloyd-Jones et al., 2010). Many of these will require hospitalization at one time or another. 2. Hospitalizations account for nearly one-third of the total $2 trillion spent annually on U.S. health care (Minott, 2008). 3. In 2006, HF was the leading cause of U.S. hospitalizations with a rate of 23 per 1,000 for men over 64 and 20 per 100 for women over 64 ( Jencks, ­Williams & Coleman, 2009; Chan et al., 2011). 4. About 25% of those hospitalized were readmitted within 30 days post discharge (Jencks, Williams & Coleman, 2009). About 30% were readmitted within 60–90 days post discharge (Gheorghiade & Peterson, 2011). 5. Readmission rates have increased by 164% between 1998 and 2007 (Bueno et al., 2010; Chen, Normand, Wang, & Krumholz, 2011). 6. Re-admissions are not profitable to nor success stories for hospitals, leaving patients feeling lost and confused. 7. Hospitalizations and re-admissions adversely impact the economy, allocation of budgets, the costs for payers and providers, not to mention the morale of patients.

Uniting Theory and Empirical Research  275

Poor adherence to guidelines leads to worsening of disease and re-admissions. According to one estimate, 54% of re-admissions may be preventable. Inadequate discharge planning and education or lack of patient follow-up, are common factors in readmission (Grange, 2005). Lack of medication adherence, failure to follow a salt-restricted diet and delays in seeking medical attention are among the primary reasons for the high rate of re-admissions among HF patients. To reduce re-admissions, hospitals implement strategies such as: the optimization of evidence based drugs and device therapies addressing causes of HF, treating comorbidities, and improving management of care (Gheorghiade & Braunwald, 2011). Evidence from randomized trials has established the efficacy of certain drug and device therapies in reducing re-admissions. Unfortunately, those therapies are limited. Chronically ill patients usually have complex drug regimens, multiple concurrent diagnoses and resulting polypharmacy (Wong, Chaudhry, Desai & Krumholz, 2011). Thus, patient care requires more than straightforward, simplistic medical disease management.

Contributing Factors to Patient Adherence to Medical Regimes Chronic patients adhere well in at least five situations: 1. 2. 3. 4. 5.

When the treatment regimen makes sense to them When the treatment regimen seems effective When they believe the benefits exceed the costs When they feel they have the ability to succeed at the regimen When their environment supports regimen-related behaviors

Messaging to Promote Adherence Perceptions may also contribute to adherence. Perception is the ability to become aware of something, a cognitive awareness that something is relevant and possibly important. Perception, mental impressions, underlies the realization that brings one to a certain subjective way of regarding, understanding or interpreting things. Perceptions can strong drivers of behaviors (Mann et al., 2009; Mora, DiBonaventura, Idler, Leventhal, & Leventhal, 2008). The most frequently studied perception is control over health (Wallston, 1992). Perceptions have been reported to co-vary with one one’s health state, specifically whether or not remains healthy or becomes ill (Benyamini et al., 2000, 2003, 2011). Perceptions were related to more preventive health behaviors, less risky behaviors and less abnormal illness behaviors (Cross et al., 2007). Clinicians today face several challenges in enhancing adherence. It is unclear what content and education-oriented messaging to use. Also, the customizing of the right mix of messaging to each patient is complex.

276  Melvin Prince et al.

Third, the translation of the knowledge into day-to-day practice with patients is in void. In this study we identified mind-sets in the sample population and examined the most effective messaging for the total sample and the effective messaging for each emergent mind-set segment. We measured the rate of re-admissions after HF at a local hospital with 100 cardiology beds. The intervention compared readmission rates after the implementation of effective messaging by mind-sets to readmission rates before the wisdom was created.We continued this effort for two years.To best of our knowledge, no prior study uncovered mind-sets regarding adherence, applied this knowledge to HF with the objective to discover what educational and content messaging enhance the adherence of each mind-set segment, nor investigated how the application of a tool to classify one’s mind-set with respect to adherence covaries with readmission rates to the hospital after initial discharge.

Discovering What Messages Work and Identifying New-to-the-World Mind-Sets This project began with three objectives: 1. Assess the impact of messaging designed to encourage adherence 2. Identify mind-sets—i.e., homogeneous groups of people defined by similar pattern of responses to the messaging elements 3. Generate a typing tool, a viewpoint identifier—i.e., a set of assignment question, the pattern of responses to which allows clinicians to determine the mind-set of a patient, and thus determine which set of messages are likely to be most effective as drivers of adherence

Systematized Experimentation With Health Care Target Groups We proposed a different approach underlying the study, akin to the thinking of the physicist, specifically a color scientist who works with ‘primary colors,’ and for whom any color is a combination of the primary colors. For any topic we propose that there is a limited number of mind-sets that can be discovered empirically. In the case of adherence, therefore, we propose the ‘potential’ existence of several mind-sets to be discovered by experimentation, with the existence of these several mind-sets to be of practical consequence. Most likely a person will fall into one mind-set, but there is always the possibility that the person’ way of thinking may place that person equally into two mind-sets.The understanding of the dynamics of these mind-sets will allow clinicians to better communicate with patients. We look at six aspects based on the literature on adherence. These are health beliefs, medical checkups, medications, medication cost, self-regulation, perceived

Uniting Theory and Empirical Research  277

quality of life (severity of health condition) and patient physician trust (Gabay, 2016). Using conjoint analysis, we reveal how the different mind-sets respond to these aspects of adherence. Beyond the science stand the scientific and practical benefits. It is challenging to know how patients are distributed among these mind-sets, and whether a mind-set can be discovered, first by interacting with patients through a questionnaire, and ultimately on an epidemiological basis, through predictive types such as the scoring of databases of customers by companies interested in more commercial issues such as credit scoring. To this end we offer typing tool which assigns a patient into one of these mind-sets. The actual research for the study, therefore, comprises three sections: 1. The ‘micro-science’ to establish the nature of responses to the different messages and identify the mind-sets. 2. The creation of the typing tool, to allow use to assign any patient to one of the mind-sets 3. The implementation of the typing tool in a hospital setting, and the measurement of within-30-day readmission after discharge.

Unique Features and Advantages of Systematized Experimentation We propose a systematic science, Mind Genomics, to identify how people respond to messaging. Our approach comprises a series of experiments to establish patterns, to learn what messages work, and the nature of underlying ‘primary groups,’ mind-sets, like primary colors, which exist in the population. The Mind-Genomics science, armed with the typing tool for a particular topic, allows science to map all patients and even non-patients, and then present the distribution of mind-sets for the particular topic, looking at conventional measures by which people are classified—e.g., age, gender, lifestyle, education, disadvantaged vs. advantaged populations, and so forth.

Establishing the Basic Science Using a Convenience Sample The study was comprised of 244 respondents from the Rome and Amsterdam areas of mid-state New York, affiliated with St. Mary’s Hospital. In the spirit of our analogy with color science we use convenience samples of respondents to establish the mind-sets, just as a color scientist uses the available stimuli to make measurements, from which the primary colors are determined. Only later, when we work with the mind-set typing tool, will we confine our focus to the HF population. Right now we focus only on establishing mind-sets for adherence. Mind Genomics creates its own ‘micro-science’ at the start, reveals the mindsets, and then developers the tool to identify these mind-sets. In a sense Mind

278  Melvin Prince et al.

Genomics creates the science as it solves the problem. It separates out the creation of the micro-science (Step 1) from the measurement of the distribution of the mind-sets (Step 2). We are dealing here with the mind itself, and so any convenience sample can be used to establish the mind-sets, as long as there are a sufficient number of respondents on which to do so. Only afterward, when the mind-sets are established, is it necessary to work with the target population. Using Mind Genomics, we identify what messages appeal to people in specific topics and use experiments to discover these appealing messages. Mind Genomics builds a science of the ordinary, of the everyday behavior discovering what messages ‘work,’ for the total population, and further identifies different mind-set of segments and different points of view held by various individuals.The key discovery is not the understanding of behaviors but rather the understanding of mindsets. Mind Genomics senses the market by clearly differentiating among groups of people in any topic area by mind-sets—i.e., by responses to the messages (Gabay, Moskowitz, Silcher, & Galanter, 2017) We developed the raw materials, the silos (groups of related elements) and elements (specific messages). For this specific exercise we developed six silos or categories of ideas, each silo comprising six different elements that are based on the literature of adherence. Exhibit 1 shows the silos the elements from each silo. The Mind-Genomics approach mixes and matches elements from the silos, creating small, easy-to-read combinations, the vignettes which include both content and educational oriented elements. The respondent reads the vignette and rates the combination on the scale provided. Each element appears 5 times in the set of 48 vignettes. The vignettes themselves comprise 12 combinations containing 3 elements, and 36 combinations comprising 4 elements. In any vignette at most one element from a silo can appear, but often the silo is entirely absent. This structure enables the estimation of the performance of each element using OLS, ordinary least-squares regression.We also created a self-profiling classification questionnaire (see Exhibit 2), allowing the respondent to provide background about themselves, in terms of geo-demographics, relevant medical history, etc. The science of Mind Genomics emerges out of our systematic, orderly progress, when we identify, explore, and catalog the different mind-sets for topic after topic in the life of people, and then look for general patterns in those mind-sets. It is the structured accumulation of such information which constitutes the science. With the knowledge of the mind of people in the market, and a way to identify people in the different mind-sets for any topic, we move from theory and data to application. The respondent began the evaluation with an orientation page, shown in Exhibit 3. For each test vignette the respondents rated the entire vignette on two attributes, one assessing adherence intentions and the second regarding patient’s feelings which the respondents rated on a 9-point scale: a. How likely are you to take your medication based on this information? (Exhibit 4) The scale ranged from a low of 1 (not at all likely) to a high of 9 (very likely)

Uniting Theory and Empirical Research  279

b. Based on this screen ALONE . . . How do you feel when you read this screen? (Exhibit 5) 1 = Healthy 2 = Uncertain, 3 = Confident, 4 = Frightened and 5 = Hopeful For examples of vignettes, see Exhibits 4 and 5.

Analyzing the Results to Create a ‘Model’ Relating Elements to Ratings We analyzed the ratings following this sequence: 1. We transformed the ratings. The respondents rated their feelings about the test concept on a 9-point scale. We transformed the respondent ratings to a binary response of ‘no’ or ‘yes’ which simplifies the interpretation of the ratings. Rating of 1–6, representing ‘no,’ ‘low’ or ‘modest’ interest in taking their medications were transformed to 0. Ratings 7–9, representing ‘high’ or ‘very high’ interest in taking their medications were transformed to 100. 2. We created individual-level equations. Each equation, one per respondent, relates the presence/absence of the 36 elements to the binary rating (0/100) for likelihood to get screening test (0 = not likely, 100 = likely). The experimental design permits us to relate the 36 independent variables to the binary dependent variable.To make it easy to understand the results we use ordinary least-squares regression. The independent variables are coded 0 (absent from the test concept) or 1 (present in the test concept). The dependent variable is coded 0 or 100, depending upon the original rating. The equation is expressed as Likelihood (Binary) = k0 (Constant) + k1 (Element A1) + k2 (Element A2) . . . k36 (Element F6). 3. The regression model generated a “Constant,” which represents the unexplained variation in the equation. Regression analysis quantifies the incremental appeal of each message relative to a baseline level of response (i.e., the “Constant”). The constant is often thought of as a “baseline” indicator of the “call to action.” However, it is important to look at the constant in association with the utility or coefficient scores. As the constant represents the amount of “call to adhere” which exists but cannot be attributed to a particular message, it is often the case that a low constant is accompanied by very strong message scores. 4. The coefficient score assigned to a given message indicates the ability of that message to influence the product/service’s call to action independent of the influence of any other message. 5. The system is additive, beginning with the constant.To the constant (estimate conditional probability of a person adhering to the drug regiment without any messaging) is added the part-worth contribution of the message to be incorporated.The total percentage—i.e., the total appeal or likelihood to follow the regimen is:

280  Melvin Prince et al.



Constant + Message Score 1 + Message Score 2 . . . etc. (up to four elements) 6. Since the silos were free floating (not all silos appeared in each concept) the coefficient score of each message has ‘absolute value’ can be compared within a silo and across silos, as well as across experiments. The utility shows the incremental appeal of any given message relative to any other message. 7. We clustered the respondents based upon the pattern of coefficients.We look for the smallest number of clusters or segments (parsimony), which tells a story—i.e., which ‘makes sense’ (interpretability). For the clustering, we use a slightly different, more granular model, namely the equation relating the presence/absence of the elements to the original 9-point rating, rather than the equation run with the transformed (i.e., binary) data. 8. The clustering revealed three different groups, based both on the goal to work with a limited group of mind-sets (parsimony), but at the same time ensure that these mind-sets make intuitive sense (interpretability). The selection of mindsets is judgmental.The clustering program cannot easily determine the interpretation of the mind-sets, although it can return statistically meaningful clusters.

Aggregated Results Vs Mind-Set Results 1. Table 10.1 shows the results from the analysis presenting the set of 36 elements, sorted by performance of the element for the total panel. The elements shaded are those with coefficients from the binary model (transformed data) whose values exceed a high of +8 or a low of −6. These are high cutoff points for the coefficients. A +8 means that incorporating the element into a vignette will add about 8 more respondents who say that they can follow the regimen. In contrast, a −6 means that incorporating the element into a vignette will reduce the number of those ‘adhering’ respondents by 6%. 2. Table 10.1 suggests that when we focus only on the total sample, all 244 respondents, without any addition information, virtually all elements appear to have low impact, either positively or negatively—i.e., either driving the respondent to adhere to the regimen or to abandon the regimen. TABLE 10.1 The Performance of the Elements Based Upon Responses From Total Panel

and the Three Emergent Mind-Sets (Interpretable Clusters)  

  B6

Total S1: Health Sample Care Professional Base Constant You believe that your medication is helping your condition

244 77 6

75 72 12

S2: Cost S3: Medication Benefit Efficacy & Safety 66 81 5

103 77 3

 

D4 You can feel or see the benefits of taking your medication as prescribed A2 You understand the benefits of taking your medication D2 You believe that not taking your medication can be harmful in the long run F6 You trust your choice of physician, pharmacist, and health care provider D1 You believe that taking your medication helps you to live longer F1 You trust your physician to make the correct diagnosis F2 You trust your physician to prescribe the correct medication F3 You trust the advice giving by your pharmacist B5 Your health care plan makes your medication affordable A1 Your doctor has fully informed you of the benefits and side effects of taking your medication D5 You believe that taking your medication will help you enjoy more time with family and friends A4 Your doctor has adequately explained why you need to take your medication A5 Your pharmacist has adequately explained why you need to take your medication E1 You believe that taking your medication helps your condition E2 You believe that your medication is safe E6 You believe your condition warrants taking your medication D6 You want to live a longer and healthy life

Total S1: Health Sample Care Professional

S2: Cost S3: Medication Benefit Efficacy & Safety

5

6

7

3

5

10

−3

6

5

10

3

2

5

8

0

5

5

9

5

1

4

6

1

5

4

4

3

6

4

7

2

4

4

5

7

1

4

14

−5

2

4

5

3

3

3

8

−1

3

3

9

−4

3

3

2

−2

6

3

1

1

5

3

−2

1

7

3

4

4

0 (Continued)

TABLE 10.1 (Continued)

 

A3 C5

C6 D3

E3 C4

B1 E4 C3

C2

E5 A6

B2 F4 C1

B3 B4 F5

Total S1: Health Sample Care Professional You understand the possible side effects of taking your medication You have been taught how to make taking your medication a daily routine You have been taught why taking your medication is important You believe that taking your medication gives you a better quality of life You believe that the side effects of your medication are manageable You have pill organizers that help you to remember to take your medication The cost for your medication is affordable You believe that the side effects of your medication are tolerable You have a support group that encourages you to take your medication You receive phone calls to help you to remember to refill your medication You believe that your medications do not interfere with each other Online literature has adequately explained why you need to take your medication The cost for your medication is worth it You trust the advice of other health care providers that you deal with You receive text messages to help you to remember to take your medication You are able to cut your medication in half to reduce the cost You are able to cut your medication in half to make it last longer You trust the manufacturers of the medication that you take

Source: The Authors

S2: Cost S3: Medication Benefit Efficacy & Safety

2

7

−6

3

2

−1

5

2

2

−4

6

3

2

1

2

2

1

3

−8

5

1

−2

0

3

0

3

1

-2

0

−3

2

2

0

−2

5

−1

0

1

3

−2

0

3

−5

1

0

5

−11

3

−1

3

2

−4

−1

−3

−3

1

−2

−4

3

−4

−3

9

−4

−12

−5

2

−4

−11

−5

-6

−8

−2

Uniting Theory and Empirical Research  283

3. We see evidence for impact elements when we cluster the respondents based upon the pattern of their impact values. Looking at the different elements is difficult in the total table of results, however. It is hard to discern a pattern for cluster—i.e., mind-set. 4. We learn more by sorting the large data table by segment, as we see in Table 10.2.Table 10.2 shows only the strongest performing positive and negative elements, as we defined—i.e., exceeding +8 for the positive elements driving adherence, and exceeding −8 for diminishing. The mind-sets now begin to make sense. Note that before looking at three segments we looked at two segments, but the results were too confusing, and no clear story emerged. We had a parsimonious solution, two segments, but not an interpretable one. 5. Mind-set or viewpoint Group A: “It’s My Decision”—education-oriented messages that fit people with high perceived control over their health. Individuals in this group want to retain control.When they follow a doctor’s order it is because they, the patient, have been convinced in their mind to follow the prescribed medication administration because it is the right thing to do. 6. Mind-set or viewpoint Group B: “It Takes a Support Structure.” Patients belonging to this group need to feel that they are not going through this alone and need to have support to motivate them to take their medication. 7. Viewpoint Group C: “Faith in the Medical Field”—content-oriented messages that fit people with high trust in their doctors. HF patients in this group have great trust in their medical team that is usually based on a relationship. This trust usually extends to other members of the medical community especially when recommended by the patient’s medical trusted advisor. TABLE 10.2 Strongest Performing Elements for the Three Segments and the Performance

of the Same Element for the Total Panel  

S1: Health Care Professional

A1

Your doctor has fully informed you of the benefits and side effects of taking your medication You believe that your medication is helping your condition You understand the benefits of taking your medication You believe that not taking your medication can be harmful in the long run You believe that taking your medication helps you to live longer Your pharmacist has adequately explained why you need to take your medication You are able to cut your medication in half to reduce the cost Your doctor has adequately explained why you need to take your medication You trust your choice of physician, pharmacist, and health care provider You trust the manufacturers of the medication that you take

B6 A2 D2 D1 A5 B3 A4 F6 F5

Total

S1

4

14

6 5 5

12 10 10

5 3

9 9

−3 3

9 8

−5

8

−5

−6 (Continued)

284  Melvin Prince et al. TABLE 10.2 (Continued)

S2: Cost Benefit B5 D4 A3 E3 F5 F5 A6

E6 E1 A2 F2 B4 B3

Your health care plan makes your medication affordable You can feel or see the benefits of taking your medication as prescribed You understand the possible side effects of taking your medication You believe that the side effects of your medication are manageable You trust the manufacturers of the medication that you take You trust the manufacturers of the medication that you take Online literature has adequately explained why you need to take your medication

Total

S2

4 5

7 7

2

−6

1

−8

−5 −5 0

−8 −8 −11

S3: Medication, Efficacy, and Safety

Total

S3

You believe your condition warrants taking your medication You believe that taking your medication helps your condition You understand the benefits of taking your medication You trust your physician to prescribe the correct medication You are able to cut your medication in half to make it last longer You are able to cut your medication in half to reduce the cost

3 3 5 4 −5

7 6 6 6 −11

−3

−12

Source: The Authors

Associating Messages With Feelings Consumer research continues to build upon the importance of emotions as defining brands. Rather than simply communicating superiority, the advertising is designed, or hopefully designed to create a link between the product or service and the heartstrings of the consumer. How that happens, is course, the presumed purview of the creative. As scientists and researchers we ought to be able to put numbers on these feelings and emotions which are created and offered in advertising. Question 2 instructed the respondent to a single feeling experienced when reading the vignette. There are many feelings and emotions one experiences when reading these vignettes, but for the purposes of this study we looked at emotions which in our judgment were relevant to adherence. The nature of the Mind-Genomics study is such that with 48 vignettes it is important to make the experience NOT onerous, and thus the requirement to select only one emotion from a limited set of five. Question 2 is known as a ‘nominal scale’ in psychometrics. The answers are selections and have no quantitative relation to each other. The answers are

Uniting Theory and Empirical Research  285

placeholders. To convert the responses to usable data we followed a simple recoding and analysis: 1. Create five new variables, one for each emotion. 2. For each vignette assign a value of 100 to the emotion selected and a value of 0 to the four emotions not selected. 3. Add a small random number (