Experiments in Financial Economics 9781783501410, 9781783501403

Citation preview

EXPERIMENTS IN FINANCIAL ECONOMICS

RESEARCH IN EXPERIMENTAL ECONOMICS Series Editors: R. Mark Isaac and Douglas A. Norton Recent Volumes: Volume 7:

Emissions Permit Experiments, 1999

Volume 8:

Research in Experimental Economics, 2001

Volume 9:

Experiments Investigating Market Power, 2002

Volume 10:

Field Experiments in Economics, 2005

Volume 11:

Experiments Investigating Fundraising and Charitable Contributors, 2006

Volume 12:

Risk Aversion in Experiments, 2008

Volume 13:

Charity with Choice, 2010

Volume 14:

Experiments on Energy, the Environment, and Sustainability, 2011

Volume 15:

New Advances in Experimental Research on Corruption, 2012

RESEARCH IN EXPERIMENTAL ECONOMICS VOLUME 16

EXPERIMENTS IN FINANCIAL ECONOMICS EDITED BY

SEAN M. COLLINS Fordham University, New York, NY, USA

R. MARK ISAAC Florida State University, Tallahassee, FL, USA

DOUGLAS A. NORTON Florida State University, Tallahassee, FL, USA

United Kingdom
North America
Japan India
Malaysia
China

Emerald Group Publishing Limited Howard House, Wagon Lane, Bingley BD16 1WA, UK First edition 2013 Copyright r 2013 Emerald Group Publishing Limited Reprints and permission service Contact: [email protected] No part of this book may be reproduced, stored in a retrieval system, transmitted in any form or by any means electronic, mechanical, photocopying, recording or otherwise without either the prior written permission of the publisher or a licence permitting restricted copying issued in the UK by The Copyright Licensing Agency and in the USA by The Copyright Clearance Center. Any opinions expressed in the chapters are those of the authors. Whilst Emerald makes every effort to ensure the quality and accuracy of its content, Emerald makes no representation implied or otherwise, as to the chapters’ suitability and application and disclaims any warranties, express or implied, to their use. British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library ISBN: 978-1-78350-140-3 ISSN: 0193-2306 (Series)

ISOQAR certified Management System, awarded to Emerald for adherence to Environmental standard ISO 14001:2004. Certificate Number 1985 ISO 14001

CONTENTS LIST OF CONTRIBUTORS

vii

CHAPTER 1 EXPERIMENTS IN FINANCIAL ECONOMICS: AN INTRODUCTION Sean M. Collins

1

CHAPTER 2 ASYMMETRIC INFORMATION AND BANK LENDING: THE ROLE OF FORMAL AND INFORMAL INSTITUTIONS (A SURVEY OF LABORATORY RESEARCH) Angelina Nikitenko Christie

5

CHAPTER 3 STRATEGIC DEFAULT WITH SOCIAL INTERACTIONS: A LABORATORY EXPERIMENT Jean Paul Rabanal

31

CHAPTER 4 DIVISIBLE-GOOD UNIFORM PRICE AUCTIONS: THE ROLE OF ALLOCATION RULES AND COMMUNICATION AMONG BIDDERS Martin Sefton and Ping Zhang

53

CHAPTER 5 THE DIVIDEND PUZZLE: A LABORATORY INVESTIGATION Sascha Fu¨llbrunn and Ernan Haruvy

87

CHAPTER 6 AN EXPERIMENTAL ANALYSIS OF MYOPIC LOSS AVERSION Tomoki Kitamura and Munenori Nakasato

111

CHAPTER 7 DOES EVERYONE ACCEPT A FREE LUNCH? DECISION-MAKING UNDER (ALMOST) ZERO-COST BORROWING Michael Insler, James Compton and Pamela Schmitt

145

v

LIST OF CONTRIBUTORS Angelina Nikitenko Christie

The Catholic University of America, Washington, DC, USA

Sean M. Collins

Fordham University, New York, NY, USA

James Compton

United States Navy, Annapolis, MD, USA

Sascha Fu¨llbrunn

Radboud University Nijmegen, Nijmegen, Netherlands

Ernan Haruvy

University of Texas at Dallas, Richardson, TX, USA

Michael Insler

U.S. Naval Academy, Annapolis, MD, USA

Tomoki Kitamura

NLI-Research Institute, Tokyo, Japan

Munenori Nakasato

Aoyama Gakuin University, Tokyo, Japan

Jean Paul Rabanal

City University of Hong Kong, Hong Kong, China

Pamela Schmitt

U.S. Naval Academy, Annapolis, MD, USA

Martin Sefton

University of Nottingham, Nottingham, UK

Ping Zhang

China Construction Bank, Beijing, China

vii

CHAPTER 1 EXPERIMENTS IN FINANCIAL ECONOMICS: AN INTRODUCTION Sean M. Collins Financial economics is a broad branch of study encompassing both the theory and practical applications of the macro- and microeconomic foundations of finance. Financial economics is distinguished from economics generally in that it relates to questions of monetary exchange, in which money today is exchanged for money in the future. A (sometimes asymmetric) lack of information about the future drives uncertainty in such exchanges through time, this in turn necessitating the contractual foundations of financial securities. Economists interested in intertemporal behavior under uncertainty cannot ignore the role of heterogeneity in human behavior in how financial markets function, nor the interaction of behavior with economic institutions or the environment (in the sense of Smith, 1982). If the predictions of theory are to be falsifiable, they must pass tests at the small scale as well as at the large. This volume presents an abundance of evidence that these factors matter in the laboratory setting. It also illustrates the role that experiments may play in adding to our knowledge of financial decision making and these decisions are mapped to outcomes in different institutions and environments. Angelina Nikitenko Christie (Chapter 2) reviews the body of experimental evidence on bank lending and credit markets in the presence of

Experiments in Financial Economics Research in Experimental Economics, Volume 16, 1
3 Copyright r 2013 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0193-2306/doi:10.1108/S0193-2306(2013)0000016001

1

2

SEAN M. COLLINS

asymmetric information. Such asymmetry is a common feature of financial markets, in which the information differs between creditors and borrowers as well as buyers and sellers of financial securities more generally. Because lack of information aggravates uncertainty, asymmetry of information may lead to selection problems for lenders and moral hazard for borrowers, which may result in credit rationing, (strategic) default, or collapse of the market. Further, the problems of asymmetric information can also be aggravated by costly or insufficient contract enforcement. A central theme of her survey is role of informal and formal institutions, and the interaction of the two, in credit markets. Jean Paul Rabanal (Chapter 3) studies strategic default in credit markets. As he notes, the ability to default on credit obligations may be treated more generally as a type of put option on a stochastic asset process. Varying asset volatility and the absence and presence of social interactions (resembling, for instance, externalities generated by mortgage failure), the study demonstrates that high volatility the theoretical predictions track well and that the studied form of social interactions delays default beyond optimal levels. Martin Sefton and Ping Zhang (Chapter 4) compare allocation rules in uniform price divisible good auctions (common to securities markets) in presence and absence of explicit communication. Theoretically, “standard” and “uniform” allocation rules admit different types of low-price equilibria, which are eliminated by a “hybrid” rule proposed by Kremer and Nyborg (2004). However, they observe little evidence of revenue differences or collusive prices among the allocation rules when bidders cannot explicitly communicate. With explicit communication prices are collusive, and they observe collusive prices even when collusive agreements are broken. They find that collusive agreements are particularly fragile when the gain from a unilateral deviation is larger, and an implication of this is that collusive agreements are more robust under the standard allocation rule. Sascha Fu¨llbrunn and Ernan Haruvy (Chapter 5) study how Managers’ incentives may be aligned with shareholders’ interest through ownership stakes. In an experimental setting which captures the role of ownership in managerial considerations, they see the emergence of both investor-aligned outcomes and managerial self-dealing outcomes. They find that the outcome depends largely on the initial ownership stake, even though managers are able to freely buy and sell shares in the market. This, the authors propose, is because initial allocation affects the ability of managers to coordinate. Moreover, they find that allowing managers to reinvest unpaid dividends results in a transfer of wealth from outsiders to management.

Experiments in Financial Economics: An Introduction

3

Tomoki Kitamura and Munenori Nakasato (Chapter 6) study myopic loss aversion. Previous studies have shown mixed results as to what the main drivers of myopic loss aversion are. Following Langer and Weber’s (2008), they investigate flexibility of investment, frequency of information feedback, and the timing of decision under multi-period settings. They find that investor’s behavior is consistent with them, having completely shifted their reference points on the current stock price in the most flexible and highest frequency of information feedback pairing (and not at all in the least flexible and lowest frequency of information feedback pairing). Finally, Insler, Compton, and Schmitt [Chapter 7] offer evidence from a natural experiment in which a group of college students are given the opportunity to take out a large low-interest loan which would yield positive net interest earnings if simply invested in low-risk assets. They study the characteristics of those willing and unwilling to take the loan. They find evidence that those who do not accept the loan are debt averse, and for those who accept the loan, also consider whether the borrower anticipates repaying it early. They find no consistent relationships between debt aversion and intellectual ability or gender. They find that individuals willing to accept the loan tend to have prior debt, longer planning horizons, come from middle-income families, and may have higher cognitive ability.

REFERENCES Kremer, I., & Nyborg, K. (2004). Divisible good auctions: The role of allocation rules. Rand Journal of Economics, 35, 147
159. Langer, T., & Weber, M. (2008). Does commitment or feedback influence myopic loss aversion? an experimental analysis. Journal of Economic Behavior & Organization, 67, 810
819. Smith, V. (1982). Microeconomic systems as an experimental science. American Economic Review, 72, 923
955.

CHAPTER 2 ASYMMETRIC INFORMATION AND BANK LENDING: THE ROLE OF FORMAL AND INFORMAL INSTITUTIONS (A SURVEY OF LABORATORY RESEARCH) Angelina Nikitenko Christie ABSTRACT Purpose
To provide a selective review of most recent developments in experimental economics of banking and lending and to summarize and synthesize the experiment designs and results in banking under asymmetric information. Methodology
The review includes recently published or working papers (2006
2013) that exclusively employ experimental economics methodology, especially for studying the impact of formal or informal institutions on lending in credit markets. Findings
The results of the reviewed experimental studies provide support for the important role of both informal (e.g., relationship banking and reputation) and formal (e.g., third-party enforcement; collateral)

Experiments in Financial Economics Research in Experimental Economics, Volume 16, 5
30 Copyright r 2013 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0193-2306/doi:10.1108/S0193-2306(2013)0000016002

5

6

ANGELINA NIKITENKO CHRISTIE

institutions and their impact on credit market performance, as well as the importance of studying the interaction of the two types of institutions. Research limitations/implications
The number of studies reviewed is fairly small but growing, indicating that this is the area of growing significance. Practical implications
Controlled economic experiments are better able to address the questions regarding the direction of causality in empirical relationships. Economic experiments are particularly useful in studying complex markets like credit and capital and in eliciting specific effects of institutions on credit market performance. Such wellestablished empirical relationships will be able to provide guidance for policy making for financial market reform. Originality/value
This is the first review of laboratory research in banking and lending under asymmetric information that aims to call attention to this area of research and serves as a starting point for an interested researcher and provide future direction. Keywords: Lending; relationship banking; debt enforcement; moral hazard; asymmetric information; experiments

INTRODUCTION The recent banking crisis of 2007
2010 highlighted the importance of our understanding of bank lending and its impact on macroeconomic variables of employment and output. Our lack of understanding is not due to lack of theoretical or empirical studies. Rather, it is due to the fact that field data typically do not allow for the clean and clear identification of causal relationships among numerous and simultaneous sources of moral hazard problems inherent in bank lending and credit markets in general due to asymmetric information. The fundamental problem of finance is that people have different information and are motivated differently to use their information. For example, natural information asymmetry may reflect that borrowers typically have more and better information about their investment opportunities, the risk involved, their own character, their credit history, and the amount

Asymmetric Information and Bank Lending

7

of effort they are willing to provide compared to the information available to lenders. This asymmetry of information may lead to selection problems for lenders and moral hazard for borrowers, which may result in credit rationing (Jaffee & Russell, 1976; Myers, 1977; Stiglitz & Weiss, 1981). The problems of asymmetric information can also be aggravated by costly and insufficient loan-contract enforcement (Jappelli, Pagano, & Bianco, 2005; Levine, 1998). A large body of theoretical literature has proposed solutions that can mitigate these problems. One of the discussed mechanisms in the context of credit markets is reputation formation in bilateral relationships of borrowers and lenders (Sobel, 1985; Stiglitz & Weiss, 1983). Another response to asymmetric information and costly enforcement in credit markets is information sharing among lenders about the characteristics and behavior of their borrowers. The theoretical work of Pagano and Jappelli (1993) demonstrates that information sharing can reduce adverse selection in markets where borrowers approach different lenders sequentially. It can also motivate and discipline borrowers to choose their projects prudently and with agreed-upon terms (see Diamond, 1989). It can also alleviate moral hazard problems associated with the level of unseen effort exerted by borrowers in their investment projects (Padilla & Pagano, 2000; Vercammen, 1995) and with loan repayment (Klein, 1992). Historically, lending developed first as bilateral relationships between borrowers and lenders (Moulton, 1918a, 1918b, 1918c, 1918d). In such banking relationships, the borrower’s reputation for repayment is at stake, and the lender’s future extension of the loan is contingent on the borrower’s timely repayment and effort in investment project. Boot and Thakor (1994) put forth a theoretical model of relationship banking that is construed as an implicit contract (informal institution) between lenders and borrowers to motivate high effort and repayment. Some empirical studies provide support for the existence of such relationship banking, in particular in small business lending, and their positive role for borrowers’ access to credit (Elsas & Krahnen, 1998; Petersen & Rajan, 1994). Why should we turn our research attention to experimental studies of the role of formal and informal institutions for credit market performance? As the authors of the experimental studies reviewed here point out, it is a matter of establishing the direction of causality between the various institutions that deal with problems associated with asymmetric information or incomplete contracts and the credit market performance. For example, Brown and Zehnder (2007) raise the issue that empirical studies on information sharing cannot clearly identify the direction of causality between

8

ANGELINA NIKITENKO CHRISTIE

information sharing and credit volume. Aggregate bank credit to the private sector is higher in countries with developed information sharing (Djankov, McLiesh, & Shleifer, 2007; Jappelli & Pagano, 2002), while access to bank credit is easier in countries with credit bureaus or registries (Brown, Jappelli, & Pagano, 2009; Galindo & Miller, 2001; Love & Mylenko, 2003). However, theory also suggests that information sharing is likely to emerge where lenders clearly benefit from it (Pagano & Jappelli, 1993): the higher the credit volume, the higher the demand for information sharing. Consequently, the use of experimental methods can allow researchers to identify the impact of exogenously introduced information sharing on credit market performance. Brown and Zehnder (2007) also posit that the experimental methods are better suited for studying the impact of information sharing on borrower’s behavior since field observational data do not provide the necessary counterfactual. While some empirical studies show that credit reports reduce selection costs for lenders and allow them to predict loan default more accurately (Kallberg & Udell, 2003), Jappelli and Pagano (2002) show that loan defaults are lower in countries with developed credit registries; it is not clear whether these outcomes are the result of better selection of borrowers due to information sharing, or the result of the disciplining effect of information sharing on borrowers loan repayment. The available field data makes it hard to separate incentive effects from selection effects in relationship lending (Fehr & Zehnder, 2009). On the one hand, there is evidence that borrowers who have long-term relationships with a lender get better credit terms and have better access to external financing. On the other hand, lenders may expect loans to borrowers with whom they interact frequently to be less risky than loans to other borrowers (e.g., newcomers). Consequently, whether these effects result because relationships enable banks to select better firms with viable business opportunities or because firms (borrowers) in such relationships have stronger incentives to meet their obligations remains an open research question (Fehr & Zender). Moreover, to distinguish the disciplining effect of repeated interactions from selection effects would require access to detailed internal firm-level data (project characteristics and realized outcomes). The confidential nature of such data prevents its gathering. At the aggregate level, the investigation of the impact of reputation formation on market outcomes would require a comparison of similar markets with and without the possibility of reputation formation, its counterfactual. Finding such data in the field is extremely difficult. Additionally, the presence of repeated interactions in general leads to a plethora of equilibria

Asymmetric Information and Bank Lending

9

(e.g., Fudenberg & Maskin, 1986). As a result, theory alone provides little guidance for the likely impact of reputation formation on credit markets. For such situations, with poor or no field data and competing theoretical arguments, laboratory experiments present a valuable alternative. Experiments allow us to study the interaction of lenders and borrowers in a clear and controlled set-up, where many of the measurement and endogeneity problems present in the field data can be overcome. Smith (2008), in his chapter on asymmetric information, asserts that while there have been very important theoretical advances in the field of information economics by Akerloff, Stiglitz and Spence, the conclusions that the problems of asymmetric information necessarily require a thirdparty intervention, usually by government, are not empirically tested. Since it is not easy to find natural counterfactuals to regulatory regime implementation, the laboratory study is the only alternative. Smith further stresses in his discussion of capital markets and interest rates that the capital and credit markets have survived and developed in early days due to various informal institutions that alleviated, if not solved, the problems of asymmetric information. The importance of such institutions should not be taken lightly and should be studied carefully before we embark on any reform or regulation path. His review of experimental markets with asymmetric information naturally includes only the then-available studies on markets for quality and labor markets. There is a small but growing body of experimental studies devoted to the study of banking, lending, and financing.1 Duffy (2012) and Dufwenberg (2012) review in more detail experiments investigating the phenomenon of bank runs, their cause, and persistence. To my knowledge, there is no review of experimental studies focusing on bank lending in credit markets with asymmetric information. This review aims to fill this gap in the literature, to serve as a starting point for an interested researcher, and to summarize and synthesize the currently available experimental designs and results. Given the focus of these first studies in bank lending, the overarching theme is the role of formal and informal institutions for solving the problems of asymmetric information in credit markets. The review is organized in the following way: the review starts with a detailed discussion of Brown and Zehnder (2007) as this is the first study to use a credit market setting under asymmetric information in the lending game, followed by a detailed discussion of Fehr and Zehnder (2009), the design of which closely builds on the previous study, followed by discussions of Brown and Serra-Garcia (2012, earlier version is 2010) and Serra-Garcia (2010) who also employ a similar experimental lending game

10

ANGELINA NIKITENKO CHRISTIE

and study different types of moral hazard than the two previous studies. The review section concludes with a brief discussion of the preliminary results from Barboni and Treibich (2012) whose design extends the lending game to allow for multiple-bank relationships. A few other related studies are also mentioned at the end.

REVIEW OF EXPERIMENTAL STUDIES On Credit Reporting and Relationship Banking by Brown and Zehnder (2007) I first discuss the experimental study by Brown and Zehnder (2007) as it is, to my knowledge, the first one to apply experimental methods to study credit markets (lender
borrower framework) under asymmetric information in particular. Brown and Zehnder investigate how information sharing (credit reporting, or credit registry) affects a borrower’s loan repayment and credit market performance. The loan repayment by the borrower is, however, not enforceable, and consequently represents an incomplete debt contract. The lender cannot observe the potential borrower’s creditworthiness prior to extending the first loan. However, through repeated transactions with the same borrower, the lender can establish a banking relationship with that borrower and condition future loans on past repayment behavior. The authors study credit markets under four credit environments, with and without information sharing (credit reporting) by lenders, and with and without relationship banking (one-shot vs. repeated lending). The authors implement a simple lending game in a competitive trading environment. Their lending game is based on the investment game introduced by Berg, Dickhaut, and McCabe (1995). The basic features of the investment/lending game are as follows: the first movers are the lenders with an endowment which they can transfer to the second mover, the borrower; the amount transferred has a fixed certain return R (in their experiments, R = 2); the borrower then decides how much of this doubled amount to return back to the lender (investor). Thus the investment game is designed to measure the borrower’s (investee’s) trustworthiness. In the present context of the lending game, this trustworthiness is analogous to creditworthiness. The authors then modify this game in two respects in order to address their research questions. First, they introduce the possibility of information sharing about the borrower’s prior repayment decisions

Asymmetric Information and Bank Lending

11

(implementing a public credit registry) as a treatment variable. Second, they allow lenders and borrowers to endogenously choose their trading partners in a competitive trading environment. They achieve a competitive trading environment in two ways. First, there are more lenders (10) than borrowers (7) in their experiments: each lender and borrower can conclude only one credit transaction per period, which means some lenders (at least 3 per period) will not have a credit transaction. This competition will put a downward pressure on the amount of repayment (or interest rate) each lender requests from the borrower in return for loaned funds. Second, the offering of credit by lenders is a continuous one-sided auction, and the credit offers are of two types: public and private. During the auction each lender can make as many public and private offers as he wants, but each lender can conclude only one credit contract per period. Once a borrower accepts a credit offer from a given lender, any remaining offers from that lender expire and disappear from the market (disappear from the offer screen). Private credit offers can only be made to one specific borrower, and they cannot be seen or accepted by any other borrowers. This offer is also not visible to other lenders. Private credit offers allow the building of a credit relationship between a particular lender and borrower. Every participant’s experiment ID is fixed in the treatment for “Banking Relationship,” while in the treatment without it, each borrower’s ID is randomly assigned each period leading to one-shot interaction. A public offer, on the other hand, is always shown to all borrowers and all other lenders. Even with public offers the lender has to specify which borrowers are authorized to accept the offer. So while all borrowers can see the public credit offer, not all of them may be able to accept it. The public offers can theoretically influence the offers made by other lenders if they are competing for a borrower’s contract. Seeing other lenders’ offer terms (repayment amount), they may lower their own repayment amount to attract a borrower. The only caveat is that in doing so they may inadvertently attract a borrower who is not intending to repay his loan, unless the lenders can see the repayment history as in credit reporting treatment. Another important characteristic of the design is that the borrowers either repay the full amount requested by the lender (no negotiation about the repayment amount) or repay nothing at all; no partial repayments are allowed. To summarize, the authors are interested to investigate the impact of information sharing (treatments with Credit Reporting (CR) and no Credit Reporting (NO)) on loan repayment and credit market performance, the impact of relationship banking (fixed ID (FID) vs. random ID (RID)) on

12

ANGELINA NIKITENKO CHRISTIE

loan repayment and credit market performance, and the interaction of the two disciplining mechanisms in the treatment where both are possible (FID
CR). For the treatment of RID
NO, the theoretical predictions are similar to those for a one-shot investment game interaction: selfish,2 own-payoffmaximizing borrowers do not repay their debt,3 while “social” (or conscientious) borrowers repay as long as they are offered fair financing conditions. If the lenders anticipate facing only selfish or opportunistic borrowers, there will be little to no provision of credit in a one-shot interaction. If the lenders anticipate (or hope) that some (enough) borrowers are conscientious about repaying the loan, there may be some provision and extension of credit.4 Overall prediction for this treatment is that the credit market collapses. In the RID
CR treatment, the lenders receive a credit report at the beginning of each period for each borrower detailing their past repayment behavior. Consequently, while the lenders cannot build banking relationships with the borrowers, they can at least learn about their past repayment choices and consequently condition their current offers on the report. Hence even the selfish borrowers have an incentive to repay in early periods of the lending game in order to appear creditworthy. The overall prediction is that in early periods there is a substantial credit market due to reputational incentives but the credit volume would decrease in later periods as selfish borrowers begin to default. The predictions for FID treatments are that repayment rates and credit volume will be identical in both FID
NO and FID
CR: high rate in initial periods and decreases toward the end of the experiment. For FID
CR, due to credit reporting the long-term relationships may be less frequent and relied upon for repayment motivation. It may also influence the profit extraction by lenders since credit reporting can substitute for banking relationship. First, the results for RID treatments (RID
NO and RID
CR) display strong differences in market outcomes. On average 80% of all loans are repaid in RID
CR, while only 28% of loans are repaid in RID
NO. The mean repayment5 rate exceeds 70% in all sessions in RID
CR, providing support for the disciplining power of credit reporting, while the mean repayment rate in RID
NO never exceeds 40% in any session (the difference between RID
CR and RID
NO is statistically significant, p = .004). There is also a substantial difference in lending activity between these treatments. In RID
CR, 94% of all potential lending contracts are realized with the average size of the loan at 41 (out of 50 possible) points. By

Asymmetric Information and Bank Lending

13

contrast, in RID
NO only 59% of all potential contracts are realized with the average size of the loan at 23 points (the difference between the two treatments is significant, p = .004). The mean interest rates ([desired repayment-credit size]/credit size) between the two treatments are almost identical, even though the interest rates vary more across sessions (the difference is not statistically significant, p = .69). Last, the results for fixed ID treatments (FID
NO and FID
CR) with bilateral banking relationships display only negligible differences in market outcomes between the two treatments. The repayment rate is 79% in FID
CR and 74% in FID
NO, with little variation across sessions. There is also little difference in lending activity between these treatments: in FID
CR, 94% of all potential lending contracts are realized with the average loan size of 42, while in FID
NO the respective numbers are 92% of contracts and 40 points loan size. The interest rates and the gains from trade are again very similar in both treatments. In all four treatments, most gains from trade are reaped by the borrowers, which confirms that the excess supply of credit in their experiment induced a highly competitive credit market.

On Relational Reputation Mechanism and Debt Enforcement by Fehr and Zehnder (2009) The next study I review is by Fehr and Zehnder (2009), in which the authors study the interaction of reputation and third-party debt enforcement mechanisms and their impact on loan repayment and project choice by the borrower. In relational reputation mechanism, lenders and borrowers form bilateral long-term relations. The repeated interactions in these relationships allow lenders to condition their credit terms on the past repayment behavior of the borrower. Repayment provides future benefits for the borrower by positively influencing the lender’s beliefs about the creditworthiness of the borrower. The higher the borrower’s creditworthiness is, the better the credit terms he receives in the future from his lender and the less likely he is to have his credit rationed (reduced or eliminated). Such repeated bilateral relationships work as a disciplining mechanism for the borrower. They also improve credit market outcomes compared to spot-market (one-shot) transactions. Fehr and Zehnder study the causal impact of legal third-party enforcement of credit contracts and the interaction between legal enforcement and the endogenous enforcement of contracts in long-term credit relations. In

14

ANGELINA NIKITENKO CHRISTIE

their laboratory environment there are two potential sources of moral hazard. The first is the presence of asymmetric information about project characteristics. Lenders do not observe the project choice, and they cannot prevent borrowers from choosing inefficient, high-risk projects. The second source of moral hazard is the absence of legal enforcement of debt repayment. The borrowers may escape the repayment of their loans even if the investment project is successful. Lack or absence of third-party enforcement is representative of the institutional weaknesses observed in many developing and emerging credit markets and occurs in advanced economies too (Djankov, Hart, McLiesh, & Shleifer, 2008). To isolate the pure effect of individual reputation formation in endogenously built long-term relationships, Fehr and Zehnder compare a treatment in which they rule out any information about the identity of trading partners (no reputation is formed) with a treatment in which individual borrowers can acquire a reputation. After identifying the pure reputation effect on credit market functioning, they study the interaction between legal enforcement of credit repayment and reputational incentives. They implement third-party enforcement of credit repayments under conditions of limited liability and wealth constraints for the borrowers. The third party can force the borrower to repay his loan only if the borrower’s project is successful. If the project fails, no repayment can be enforced because the borrowers have no wealth to take away. Hence in this situation, the borrowers may have the incentive to choose more risky projects (high-risk projects). The limited liability implies some sharing of the project risk between a lender and a borrower (this may be characteristic to advanced Western economies). Previous work by Brown, Falk, and Fehr (2004) finds that individual reputation enhances efficiency in endogenously formed employment relationships. Brown and Zehnder (2007), discussed above, find that relationship banking provides strong incentives for borrowers to repay their loans. The papers by Fehr and Zehnder differ from these experimental studies in important aspects. First, previous experiments make reputation formation easier because they use deterministic returns on investment such that the principal (lender) can know for sure the rate of return on a given loan. Fehr and Zender’s design, on the other hand, captures key characteristics of credit markets
the asymmetric information between borrowers and lenders and uncertainty about the project success. Consequently, these previous experimental studies, by the nature of their designs, cannot address the occurrence and interaction of the two key moral hazard problems in credit markets
the repayment problem and the project choice problem.

Asymmetric Information and Bank Lending

15

They also do not examine the interaction of third-party enforcement mechanisms and relational, reputation-based mechanism in credit markets. Fehr and Zehnder pose two research questions: (1) In the absence of any third-party enforcement of repayments, to what extent do market participants make use of the relational reputation mechanism to lower the borrowers’ incentives to take the money and run? (2) How does the introduction of third-party enforcement interact with the role of bilateral relationships and what are the consequences for the performance of credit markets? The baseline condition is a credit market where debt (loan) repayments are not third-party enforceable, and the market participants have no opportunity to engage in repeated interactions in order to build reputation. The lenders face anonymous borrowers in every period. Experimentally, anonymity is guaranteed by reassigning a new identification (ID) number to each participant at the beginning of every period. By doing so, this treatment implements a series of one-shot credit markets (hence the name of this treatment
OC). Their prediction is that, under the standard assumption of rational own-profit maximization, the lenders do not extend credit; they anticipate that borrowers never repay their loans. However, since many experimental studies have provided evidence that not all people simply maximize own monetary payoffs (for overview, see Berg et al., 1995; Camerer, 2003), the authors also consider the impact of social lenders on credit markets even in a one-shot transaction. In the presence of stochastic project success and asymmetric information, it is not obvious whether repeated interactions are capable of sustaining cooperation between borrowers and lenders. The lenders cannot observe the borrower’s project choice nor can they observe whether the project has been successful. They can observe only whether the borrower repays the credit extended to him. Thus, if the borrower does not repay his credit in the market without third-party enforcement, the lender does not know whether it is because the borrower’s project has failed or because the borrower is not willing to repay his loan. As a result, the lender doesn’t have certainty about whether he is facing an opportunistic borrower or the borrower who just had bad luck. The double moral hazard problem in these credit markets makes it more difficult for a borrower to acquire a reputation as a good borrower. Thus, in the design of Fehr and Zender, a good borrower is the one who repays the loan, thereby forming a good reputation with the lender. In order to increase the chance of repaying the

16

ANGELINA NIKITENKO CHRISTIE

loan, the borrower should also choose a low-risk project which would increase the chance of the loan being repaid. Concern for good reputation should further motivate a good borrower to choose projects prudently or more “efficiently,” to use the authors’ term. The second condition is identical to the baseline except that it allows for the endogenous formation of bilateral relationships. The authors assign a FID number to every participant that remains constant for all periods of the experiment session. In this condition, the lender can then repeatedly offer credit to the same borrower (i.e., the same ID number) and, if the borrower accepts these offers, establish a long-term relationship. The authors call it relationship condition (RC). In RC, the desired repayment is not binding and the borrower can repay as much as he wants. By comparing OC and RC treatments, the authors are able to study the impact of reputation formation in endogenously build relationships on credit market performance and efficiency. Their third treatment is legal enforcement of credit repayment by a third-party enforcement condition (TPC). TPC is identical to RC except that the borrower is now subject to the agreed upon repayment of the loan when the project is successful. In this treatment, the desired repayment is binding once the borrower accepts the offer from the lender. In all three treatments, the credit market lasts for 20 identical periods. The market consists of 17 participants: seven are lenders and ten are borrowers. Because each lender can at most fund one project per period, some borrowers will not receive funding for their project. In every period each borrower has two projects available: project A and project B, (p = {A, B}). Each borrower can at most realize one of the two projects. Both projects require the same investment, I = 32 capital units. The authors call project A an efficient, low-risk project with a high expected return but a moderate return in case of project success, R = RA. Project B is an inefficient, highrisk project with a lower expected return than project A but a higher return in case of project success, RB>RA. If either project, A or B, fails, the project return is always zero. The borrowers have no endowments (equity) and cannot carry excess returns into future periods. As a result, the borrowers need external funding to realize a project. If the borrower does not conclude a credit contract, he can choose a risk-free project with a fixed return of b = 10 per period. Each lender has endowment of K = 32 capital units at the beginning of every period and thus can fund at most one project per period. The lender also has an alternative investment choice: he can invest his whole endowment in an “endowment-storing technology” that yields a payoff of 32

Asymmetric Information and Bank Lending

17

units, analogous to a risk-free return. Otherwise, the lender can use his 32 capital units to extend credit to a borrower. Each period is a two-stage process. The first stage is a continuous, onesided auction in which lenders can make credit offers to borrowers. Each credit offer must specify three parts: (1) The desired project, pd = A or B. (2) The desired repayment amount in case of project success, rd. (3) Whether the offer is private or public. The process for posting credit offers (stage one) is as follows: public credit offers can be seen and accepted by any borrower (all borrowers can see a public offer); while private credit offers can only be seen by a specified borrower (lender specifies experimental ID of the borrower to whom he wants to make a credit offer). Public offers can also be seen by other lenders, while private offers can’t. Each lender can have as many outstanding public or private offers as he chooses. However, as soon as a borrower accepts one of the posted offers of a lender, the contract is concluded and all other outstanding offers of this lender disappear from the market and cannot be accepted. Stage 1 concludes as soon as all the lenders’ offers are accepted by borrowers, or when the time, three minutes, expires. In Stage 2 of each contracting period, all borrowers who have accepted a credit offer choose which project they want to invest in, A or B. After their investment decisions are made, a random device determines whether a project is a success or a failure (in this case project success is determined by publicly throwing a 10-sided die). Then each borrower, who invested in the project, decides on the amount of the repayment. The borrower can make repayments up to the level of the project return. The borrower’s repayment cannot exceed the lender’s requested repayment or the realized project yield. In OC and RC, where debt repayment is not third-party-enforceable, if the borrower does not repay then the lender does not know whether it was because of the project failure or the borrower’s unwillingness to repay. In the TPC, however, if the project is successful, the requested repayment amount is automatically enforced and realized such that r = min (R, rd), but the lenders are not informed about the borrower’s project choice. The results of the experiments are as follows: in the OC treatment (aggregated over 5 separate sessions), the main results point to the gradual collapse of credit market trading. In the initial periods (1
3), roughly 90% of all feasible trades take place, declining sharply by period 4, while in the final period (20) only 17% of all feasible trades are concluded. Borrowers’ average repayments are below 32 units, which is the lenders’ risk-free

18

ANGELINA NIKITENKO CHRISTIE

outside option. Consequently, on average, the lenders were worse off participating in the credit market than using their outside option. However, individual data on borrowers’ repayments displays some heterogeneity, showing that roughly 30% of the concluded contracts were profitable to the lender (when lender’s expected profit is at least 32).6 Given this nonnegligible presence of trustworthy borrowers, it might explain the gradual nature of credit market breakdown. The authors additionally report that the borrowers chose project A in 94% of the cases. Why might the borrowers choose a more efficient, low-risk project in OC credit transactions? The authors explain that choosing project A is in the self-interest of selfish borrowers who do not plan to repay their loans: they are maximizing their own expected returns by choosing a low-risk option. Trustworthy borrowers, who stick to the credit terms of the lender, should also choose project A if the lender chooses project A. Interestingly, the proportion of lenders choosing project A as the desired project is 85% (lower than 94% of the borrowers). This means that in some cases, the lenders indicated project B as the desired project and, consequently, based their requested repayment on the return in project B (at most 200 units, if successful, compared to at most 100 units for project A). I conjecture that due to higher competition among the borrowers, some borrowers accepted the terms of the contract without intending to repay from the start, especially if they do not intend to choose project A. Here the difference in traders’ appetites for risk might also play a role in the decision to repay or default. It is possible that some borrowers were punishing lenders’ appetites for risk or greed by not repaying the loan or repaying little. The results for RC sessions indicate that reputation formation through repeated interaction leads to a stable credit market, except for the final period (due to end of experiment effect).7 The average of all realized contracts over all 20 periods is 81%, in sharp contrast to the OC (p = .0004). Given the double moral hazard problem, the authors ask how exactly reputation formation works in successfully solving both repayment and efficient project choice, resulting in high repayment. The answer provided in the data is as follows: because the lenders condition their contract renewal on the borrower’s past repayment behavior, (1) it gives the borrowers a strong incentive to repay their debt upon successful realization of the project, and (2) the project is more likely to be successful if the borrowers choose a low-risk project A. In fact, the borrowers choose the efficient project A in 91% of all cases. The data on positive repayments shows that 90% of all observations are in the interval [40, 60], which are

Asymmetric Information and Bank Lending

19

profitable to the lender. In fact, in 65% of the cases the lender’s desired repayments are in the same interval [40, 60], meaning that the borrowers complied with the credit terms in 65% of the cases. In 35% the desired repayments are higher, but are very rarely above 70. While the authors do not provide data on how often the lenders choose project B as the desired project in RC treatment; perhaps the above data on the desired repayment can indicate that the lenders also go for project A in the majority of the cases. I find this an interesting result as it points to another possible consequence of relationship lending, specifically that it might also discipline lenders to choose a low-risk alternative (A) as the desired project. Together, the increase in the number of trades and the high frequency of efficient project choice, produce a high fraction of realized gains from trade in the RC treatment (66% in RC vs. 46% in OC, p = .047).8 Finally, the results for TPC display a further increase in the number of trades; however, the share of efficient projects (A) drops significantly leading to much lower gains from trade. The share of realized contracts in almost all periods is higher in TPC than in RC, (p = .008). The realized gains from trade are somewhat higher, 72% in TPC vs. 66% in RC (p = .247). In more than half of the trades (54%) in TPC, the borrowers choose the inefficient, high-risk project B while in RC the borrowers go for it in only 9% of the cases (p = .004). This is a curious result. The authors explain that the borrowers in TPC face much stronger short-term incentives to choose project B. Due to the legal enforcement of debt repayments, the choice of project B always maximized the borrower’s period profit in the TPC. The lenders in TPC also requested much higher repayments in case of a project success, which further increases the borrower’s short-term incentive to choose project B. I think that borrowers’ desire to punish the lenders for unfair offers may be the explanation. If borrowers do not view the credit terms as fair and they have no way to punish the lenders via nonrepayment, the only other choice for punishment is for the borrowers to choose the high-risk project. Simply not accepting the credit offer is not a punishment since, due to high competition among the borrowers, some other borrower might accept the project, or the lender might simply exercise his outside option. So accepting an unfair offer and then choosing a high-risk project, in which case the lender gets zero and loses his loan if the project fails, is a much more effective way to punish unfairness in this context. It is not unreasonable to think of borrower’s motivations for punishment. The ultimatum offer games demonstrate that the second movers are willing to punish the first movers at their own expense by rejecting unfair offers (see Camerer, 2003, for overview of results).

20

ANGELINA NIKITENKO CHRISTIE

The discussed results from TPC lead me to believe that the introduction of the formal institution for repayment enforcement completely changed the way the borrowers perceive relationship banking in repeated interactions. The formal institution completely replaced the trust-trustworthiness relationship. As a result, relationship banking is undermined since repayment is automatically enforced if the project is successful up to the maximum return of the project (repayment cannot exceed the realized return on the project). In my interpretation of the authors’ TPC results, the formal institution for repayment enforcement replaced relationship lending. Since repayment is automatic (if a project is successful), there is no incentive to build a reputation for repayment. The lenders’ offers are no longer contingent upon repayment per se. Default only occurs if the project is unsuccessful, and the lenders are never informed about the borrower’s project choice. The fact that the experiment results demonstrate repeated interactions between the same lender and borrower may be due to the fact that the borrower simply accepted the credit terms of the lender from the start, and the lender had no particular reason to switch to another borrower unless the lender suspected that the borrower was consistently choosing a high-risk project, which was likely to lead to default. For example, if the lender experiences two or more defaults in a row with the same borrower, he may then be inclined to switch to a different borrower. Default in this case does not mean that the borrower is not trustworthy; it might mean one of two things: unlucky (as in project A) or high-risk (choosing project B). Hence, as the authors also point out, the only reputational incentive a borrower might face is to choose the low-risk project (A) that is more likely to lead to repayment. As the results demonstrate, in relations lasting more than 10 periods, project A was chosen in more than 70% of the cases, leading to high average repayments. Although the causality could also go the other way: because borrowers were choosing project A often enough, leading to positive repayments, the lenders had no incentive to switch to another borrower. Competition among the borrowers (10 borrowers facing 7 lenders) is likely to be the additional and perhaps even stronger disciplining mechanism in RC, and TPC.

Debt Enforcement and Relational Contracting by Brown and Serra-Garcia (2012) Another study that investigates debt enforcement and relational contracting is by Brown and Serra-Garcia (2012). Starting out with empirical studies demonstrating that relational contracts can serve as a substitute

Asymmetric Information and Bank Lending

21

for third-party debt enforcement (Brown & Zehnder, 2007; McMillan & Woodruff, 1999), the authors ask why there might still exist a strong relationship between creditor protection and financial development. They provide experimental evidence demonstrating that in order for relational contracts in credit markets to emerge in the first place, a minimum level of third-party enforcement is necessary. Their working definition of weak debt enforcement is something akin to slow and inefficient bankruptcy procedures, which may allow borrowers who have defaulted on their loans to reinvest the funds remaining after default. If such reinvestment can occur then, the authors claim, it can seriously weaken the borrowers’ dynamic incentives to repay their loans. They implement a credit market experiment, in which a lender and a borrower interact for 7 periods, thus allowing them to build a credit relationship. After 7 periods, the lender is matched with another borrower for another 7-period credit relationship. Each experiment consists of 3 rounds. There are 3 lenders and 3 borrowers who maintain their roles for the entire duration of the experiment. No lender is matched with the same borrower twice. In the experiment, one lender will have a total of three relationshipbuilding rounds, each with a different borrower. The same is true for each borrower: in each round the borrower is matched with a different lender. This setup is qualitatively different from experiments by Brown and Zehnder (2007) and by Fehr and Zehnder (2009), who employ a continuous one-sided auction with simultaneous private and public offers by lenders, where they have to actively nominate a particular borrower for private offer. In Brown and Serra-Garcia, the lenders do not have to choose a particular borrower. So this decision-making is removed from lenders, and the authors can, therefore, concentrate on the development of a credit relationship given a randomly matched pair. The lending game is again a modified investment game (after Berg et al., 1995) with a deterministic return on investment, as in Brown and Zehnder (2007). In the weak enforcement (WE) condition, the borrower has a choice not to repay the loan S extended by the lender and reinvest the accumulated capital from unpaid loans. The loaned funds that the borrower keeps for himself become his capital C for future investment periods (he starts off with zero capital in the first period). Thus, in subsequent periods, the borrower can reinvest his capital along with a new loan, such that It = Ct+St, and It cannot exceed 10 experimental units per period with a fixed rate of return 3 in any period, known to both lenders and borrowers. The main difference from the previous credit experiments is that the borrower is allowed to reinvest his capital, which he has accumulated from previous

22

ANGELINA NIKITENKO CHRISTIE

periods. The credit amount that the borrower does not repay to the lender in the previous period becomes the borrower’s capital for the current period. Thus the borrower’s capital in periods t = {2,…,7} equals the sum of the loaned funds which he did not repay from all previous periods. Only at the end of period 7, the borrower can liquidate his capital for payoff. In contrast to the WE treatment, in strong enforcement (SE) condition, the defaulting borrower cannot keep the lender’s funds and reinvest them, such that Ct = 0 in each period. The borrower can default in both treatments, but in SE he cannot reinvest the loaned funds which he did not repay. Consequently he cannot continue earning investment returns from them. This means that in the SE treatment, borrowers who want to take advantage of an investment opportunity always need funds from lenders. As a result, they have a greater incentive to repay their loans. The WE treatment nicely captures the credit and investment environment, in which the defaulting borrowers do not necessarily rely on repeated interaction with the lender/investor to generate future income. The authors also gather data on individual characteristics from three, short, preexperiment games measuring each participant’s levels of risk-aversion, strategic reasoning, trust, and trustworthiness. As in previous credit market experiments discussed above, the experimental instructions are framed in a credit market language (see Brown & Zehnder, 2007, for discussion of this issue). Brown and Serra-Garcia report aggregate results first and find that in the WE treatment (with capital reinvestment), credit volume (average loan size per period) is lower than in SE (average per period 3.17 in WE vs. 5.67 in SE, p = .01), interest rates (desired repayment divided by the loan size) are slightly higher than in SE (2.13 in WE vs. 1.99 in SE, p = .05), and the default rate is higher in WE (average 36% in WE vs. 21% in SE, p = .05). The authors also find support for their predictions that lenders in WE treatment employ a strategy of “starting small” in terms of initial loan sizes in order to motivate repayment: more than 40% of lenders offer initial loans of sizes between 1 and 4 (on a scale of 1 to 10 points). Only 19% of lenders offer an initial loan larger than 8 (i.e., 10, since there were no offers of 9) in WE, while in SE treatment, more than 35% of lenders offer an initial loan size of 10.9 The interest rate for the surplus sharing in both treatments is 2 (meaning they divide surplus 50/50). The probability of default in period 1 of the WE treatment is substantially lower for loan sizes 1
4 (9%) than for loan sizes 5
8 (54%) and loan size 10 (44%). By contrast, the probability of default in SE in period 1 is low for any size of the loan, small or large. Defaults in SE occur in less than 20% of the cases in period 1 and 2
5. This pattern is suggestive of

Asymmetric Information and Bank Lending

23

a strong reputation incentive on default (similar to Brown & Zehnder, above). Overall, the authors conclude that their results suggest that weak debt enforcement (as in weak or slow bankruptcy procedures) reduce the number of relationship-lending contracts which mitigate moral hazard via reputation. Even when such relationships emerge in WE, they produce smaller credit volume because lenders start off with smaller loans.

On Incentive Effect of Collateral in Credit Markets by Serra-Garcia (2010) So far we have discussed few types of hidden action that the borrower can take after receiving the loan: choosing a type of project, high-risk or lowrisk; choosing to default on the loan even if the project is successful; and defaulting on a loan and reinvesting the loaned funds. Another type of hidden action often studied in labor economics is choosing the level of effort, which is costly to the employee, or the borrower in case of credit markets. Serra-Garcia (2010) studies this type of moral hazard in the framework of credit markets. In particular, she studies the impact of collateral on the borrower’s costly effort after receiving a loan from the lender (based on theoretical work of Bester, 1987). She aims to answer the questions of whether collateral decreases the problems of moral hazard; whether collateral affects credit supply, demand, and volume; and whether there is an interaction effect between collateral and interest rate on borrower’s level of effort. To answer these questions, the author builds her experimental design on a lending model under moral hazard by Innes (1990). The experiment is a one-period lending game. Each lender is matched with a borrower and decides whether to offer a loan. If he offers a loan, the lender can also request collateral. If the borrower accepts the loan offer, she decides on her effort level. The effort is a costly action that increases the probability of success, but it is not contractible or observable by the lender. The effect of collateral is studied by varying its amount across three treatments: baseline with no collateral, 50% of the loan amount as collateral, and 100% collateral. The interest rate may or may not influence the effectiveness of collateral as a disciplining mechanism for effort level. To study the possible effect of the interest rate, Serra-Garcia varies interest rates exogenously as a treatment variable: low or high interest rate, i.e. low or high fixed repayment amount. Repayment decision is also automatically enforced in the experiment once

24

ANGELINA NIKITENKO CHRISTIE

the borrower accepts the terms of the loan. Thus, strategic default on a loan is ruled out by design in order to concentrate on the single effect of collateral on borrower’s effort in the investment project. To briefly summarize the parameters of the design, the lender starts with 150 experimental units (here, “points”) and can lend 100 points to the borrower for investment project that yields 300 points in case of success and 0 in case of failure. The lender can request, by treatment, no collateral, 50% collateral or 100% collateral to cover the cost of the loaned amount, which is 100 points. The repayment amount to the lender is determined and enforced exogenously: 200 points in the low-interest case and 250 points in the high-interest case. The borrower has an endowment of 100 points which cannot be used in project investment but can be pledged as collateral. The borrower can use the 100 points he gets from the lender to buy anywhere from 1 to 5 red balls. At the start of the project there are 6 black balls. Each red ball bought replaces one black ball. A project’s success is determined by drawing one ball randomly from the six that represent the project. A red draw means the project is successful; a black draw means it is a failure. Thus the greater the number of red balls in the project, the higher the probability of success. Buying red balls is costly: 1 ball requires 4 points, 2 balls
16 points, 3 balls
36 points, 4 balls
64 points, and 5 balls
100 points. Thus, buying red balls is the hidden action of the borrower, is costly, and is the source of moral hazard as it directly affects the probability of success of the project.10 Due to space limitations, I will discuss only the main results of the study as they relate to effort level. Recall that the effort level is the number of red balls purchased by the borrower. In the baseline treatment, No Collateral (fixed repayment of 200), the effort is 1.9 (see Serra-Garcia for theoretical prediction for it, that is, no loans extended and hence no effort). In treatments with low repayment (200 = loan principal + interest), an increase in collateral yields a significant increase in effort. The average effort is 2.8 when collateral is 50%, and it increases to an average of 3.9 when collateral is 100% (Mann-Whitney test, p < .01). However, if the repayment is high (250; high interest treatment), an increase in collateral from 50% to 100% does not yield a significant increase in effort (from 1.8 to 2.3; MannWhitney test, p = .19). In the treatment with high collateral and high interest rate, the average effort is 2.3 compared to the theoretical prediction of 3 and compared to 3.9 average effort in the low interest treatment (when collateral is requested by the lender). Not all lenders request collateral, perhaps preferring to rely on trust and signaling that they are rewarding trustworthiness by not requesting collateral. Credit volume is also lowest

Asymmetric Information and Bank Lending

25

among the interest rate treatments in the treatment with high collateral and high interest rate, as is the credit demand. To better understand the effect of a high interest rate on effort level in the high collateral case, the author conducts an additional treatment (Tax Treatment), which aims at separating fairness and loss aversion concerns as possible explanations. She concludes that the effort level in the high interest/high collateral case is likely driven by loss aversion (see Serra-Garcia, 2010, for details). In conclusion, in credit markets with low interest rates, the incentive effect of collateral is likely to reduce loan defaults as it increases the borrower’s efforts.

Other Related Studies The only other study that uses a modified lending game framework based on Berg et al. (1995), Brown and Zehnder (2007), and Fehr and Zehnder (2009) is by Barboni and Treibich (2012)11 who aim to study multiplebank lending relationships with the firm (the borrower). This study, however, while very important in its own focus, changes the environment from single to multiple bank lending such that each borrower moves first by requesting a full loan from one lender or half the needed loan amount from each of the two lenders in a sequential manner. There are two treatments to address which lender the borrower can contact first: a random treatment, in which the sequential order of the two lenders is randomly chosen each period; and a relationship lending treatment, in which the sequential order of the two lenders is decided by the borrower each period. The third treatment is Information Disclosure treatment, which is identical to the relationship-lending treatment, but also allows each lender to know whether the borrower’s default was a result of investment failure or free-riding. This information is available only to the lender who has lent in that particular period of play. If the lender was not chosen for lending or did not lend in that period, he cannot know information about the borrower’s reason for default. This opens room for strategic manipulation by the borrower to choose only one lender at a time to prevent information sharing. Lending contracts and relationships are endogenously formed, as is reputation. Interest rates and project types are exogenous and the project return is stochastic. The enforcement of loan repayment is incomplete and allows for strategic default. There is information sharing among lenders in the form of a public credit registry (as in Brown & Zehnder, 2007). The borrowers do not have endowment for investment and cannot use excess returns for future rounds of the

26

ANGELINA NIKITENKO CHRISTIE

game (as in Fehr & Zehnder, 2009). Neither lenders nor borrowers have an outside alternative or safe option to invest (outside option is normalized to zero). This induces competition among lenders to fund the borrower’s investment project. In the experiment, one borrower is grouped with two lenders for the duration of the session of T periods, such that the last period, the end of experiment, is randomly enforced (after 20 periods, there is 1/10 probability that the session continues for another round). There are some interesting preliminary results (total of 43 observations over 6 sessions; 2 sessions per treatment) that the authors find: the borrowers may prefer multiple-bank lending if they receive partial loans (credit rationed) and are unable to stabilize their lending source. Such credit rationing is correlated with borrowers’ lower repayment rates and a tendency to switch between lenders. The results for relationship lending are similar to previous studies: relationship lending motivates borrowers to repay and continue with the same lender (see Brown & Zehnder, 2007, above). Trautmann and Vlahu (2013) study strategic loan default and coordination. Their experimental setup is a coordination game rather than the lending game discussed in detail above. Their study is important in its own way: they investigate the interaction of transparency rules and economic environment and whether the two affect the incidence of strategic default. One other study investigating the role of regulatory rules for transparency using a modified pass-through investment game is by Rietz, Sheremeta, Shields, and Smith (2013). Additionally, the design of credit bureaus is studied in a public goods game by McIntosh, Sadoulet, Buck, and Rosada (2012). Lastly, Kassis, Hazlett, and Battisti (2012) design a classroom experiment to study banking. They use an experiment setup that is close to natural banking (household-bank-firm), which is worth further study.

CONCLUSION I have reviewed a number of important studies of credit markets with asymmetric information and incomplete debt (loan) contracts. These experimental studies are important steps toward a deeper understanding of how credit markets function both at the micro-level
the underlying incentives of lenders and borrowers, and at the macro-level
the resulting market outcomes of credit volume, credit size, gains from trade, and

Asymmetric Information and Bank Lending

27

number of realized contracts. These studies demonstrate that institutions matter, be they formal or informal, for the functioning of credit markets. For example, relationship banking, viewed as an informal institution that developed historically and ecologically, has a significant influence on the borrower’s repayment behavior and choice of risky projects. Reputation is an ecological institution (a result of human action but not of human design) that has solved many problems associated with asymmetric information (Fehr, Brown, & Zehnder, 2009). Information sharing is likely to have arisen evolutionarily and ecologically long before it became codified into a formal institution of credit bureaus and credit registries. Studies by Moulton (1918a, 1918b, 1918c, 1918d) show how early bank lending depended on information gathered about a borrower from a local network of contacts in order to determine creditworthiness. The experimental study on emergence of information sharing by Brown and Zehnder (2010) is an important contribution in this line of inquiry, showing that asymmetric information in credit markets increases the frequency of information sharing even in the presence of lender competition. The study by Fehr and Zehnder (2009) further shows that the interaction of informal and formal institutions (relationship lending and third-party debt enforcement) can have unexpected results for credit markets: repayment enforcement may weaken the incentive effect of relationship lending. It may also lead to a transfer of market power to lenders who charge higher interest rates and thereby exacerbating the borrowers’ choices of risky projects and worsening the moral hazard problem. Brown and Serra-Garcia (2012), on the other hand, show that weak contract enforcement, in terms of implementation of bankruptcy rules (i.e., allowing the borrower to reinvest the loaned funds after default), leads to dynamic changes in borrower’s repayment. Serra-Garcia (2010) studies the impact of collateral on another variant of moral hazard, the hidden effort, finding that while collateral increases effort level, this incentive effect is exacerbated by high interest rate. All the reviewed studies, some of them conducted prior to subprime mortgage crisis and ensuing banking crisis, have taken initial and important steps toward improving our understanding of bank lending and credit markets. The frequency of financial and banking crisis (Reinhart & Rogoff, 2009), the complexity of these markets, and their impact on macroeconomic activity make it all the more important to conduct careful, controlled studies. Future studies of bank lending will profit from investigating, for example, lenders’ incentives under various regulation regimes and public policy solutions currently considered as panaceas for financial and banking crises.

28

ANGELINA NIKITENKO CHRISTIE

NOTES 1. See Christie and Houser (2012) for experiments on financing under asymmetric information. 2. Brown and Zehnder classify borrowers as either selfish (guided by narrow material self-interest) or social (guided also by a concern for another’s payoff or by fairness considerations). 3. It is important to note that Brown and Zehnder frame the language of the experiment as a credit market lending game, making it very context specific. The advantage of such framing, according to the authors, is to increase the external validity of the experiment and to reduce noise associated with subjects trying to create their own interpretations of the environment they are in. Interestingly, then we can be certain that the subjects do understand the connotations of “loan” and “loan repayment,” as loan repayment is considered as correct or expected action socially. We can still ask questions as to why some borrowers are driven not to repay their loans but accept the contracts anyway, acting opportunistically. Could non-repayment serve as a form of implicit punishment by the borrowers if they consider the terms of repayment “unfair,” too high an interest rate? 4. See Brown and Zehnder’s Appendix for theoretical derivations of the lender’s and borrower’s optimal behavior. For example, they assume that a fair financing offer for social borrowers would be to split the gains from trade 50/50. Unfair offer would be one where the lender requests more than half of the gains from trade. 5. For additional regression analyses see Brown and Zehnder. 6. As the authors explain, in the case when the borrower chooses project A and repays at least 40 in case of project success, the expected profit for the lender = .8 × 40 = 32. In the case when the borrower chooses project B and repays at least 107, the expected profit for the lender is .3 x 107 equals roughly 32. 7. See Fehr and Zehnder (2009) for detailed and informative regression analyses. 8. The interested reader should consult Fehr and Zehnder (2009) for further analysis of long-term relationships and their impact on market trading. 9. In addition, Brown and Serra-Garcia (2012) also report multivariate-analysis OLS regressions of first-period loan offers on the treatment variable and the characteristics of the lender. The estimated coefficient of the dummy variable WE treatment suggests that first-period loans are 2 points lower in WE than in SE treatment. 10. For theoretical derivations of the model, risk aversion, predictions, and detailed instructions (the author uses neutral language in experiment), please consult Serra-Garcia (2010). 11. Barboni and Treibich (2012) is a working paper; the experiment instructions were not part of the online version of the paper. Thus, I do not have a way to review their instructions.

REFERENCES Barboni, G., & Treibich, T. (2012). The impact of single versus multiple bank lending relationships on firms and banks’ behavior. Working Paper. Berg, J., Dickhaut, J., & McCabe, K. (1995). Trust, reciprocity, and social history. Games and Economic Behavior, 10(1), 122
142.

Asymmetric Information and Bank Lending

29

Bester, H. (1987). The role of collateral in credit markets with imperfect information. European Economic Review, 31(4), 887
899. Boot, A. W., & Thakor, A. V. (1994). Moral hazard and secured lending in an infinitely repeated credit market game. International Economic Review, 35(4), 899
920. Brown, M., Falk, A., & Fehr, E. (2004). Relational contracts and the nature of market interactions. Econometrica, 72(3), 747
780. Brown, M., Jappelli, T., & Pagano, M. (2009). Information sharing and credit: Firmlevel evidence from transition countries. Journal of Financial Intermediation, 18(2), 151
172. Brown, M., & Serra-Garcia, M. (2012). Debt enforcement and relational contracting. Departament d’Economia del’Empresa, Universitat Auto`noma de Barcelona, Working paper 12/1. Brown, M., & Zehnder, C. (2007). Credit reporting, relationship banking, and loan repayment. Journal of Money, Credit and Banking, 39(8), 1883
1918. Brown, M., & Zehnder, C. (2010). The emergence of information sharing in credit markets. Journal of Financial Intermediation, 19(2), 255
278. Camerer, C. (2003). Behavioral game theory: Experiments in strategic interaction. Princeton, NJ: Princeton University Press. Christie, A. N., & Houser, D. (2012). Financing and signaling decisions under asymmetric information. GMU Working Paper in economics No. 12
36. Diamond, D. W. (1989). Reputation acquisition in debt markets. The Journal of Political Economy, 97(4), 828–862. Djankov, S., Hart, O., McLiesh, C., & Shleifer, A. (2008). Debt enforcement around the World. Journal of Political Economy, 116(6), 1105
1149. Djankov, S., McLiesh, C., & Shleifer, A. (2007). Private credit in 129 countries. Journal ofFinancial Economics, 84(2), 299
329. Duffy, J. (2012). Macroeconomics: A survey of laboratory research. In J. Kagel & A. E. Roth (Eds.), Handbook of experimental economics, vol. 2 (forthcoming). Dufwenberg, M. (2012). Banking on Experiments? Mimeo. Elsas, R., & Krahnen, J. P. (1998). Is relationship lending special? Evidence from credit-file data in Germany. Journal of Banking & Finance, 22(10), 1283
1316. Fehr, E., Brown, M., & Zehnder, C. (2009). On reputation: A microfoundation of contract enforcement and price rigidity. The Economic Journal, 119(536), 333
353. Fehr, E., & Zehnder, C. (2009). Reputation and credit market formation: How relational incentives and legal contract enforcement interact (No. 4351). IZA Discussion Papers. Fudenberg, D., & Maskin, E. (1986). The folk theorem in repeated games with discounting orwith incomplete information. Econometrica: Journal of the Econometric Society, 54(3), 533
554. Galindo, A., & Miller, M. (2001, March). Can credit registries reduce credit constraints? Empirical evidence on the role of credit registries in firm investment decisions. In Annual Meetings of the Inter-American Development Bank, Santiago Chile. Innes, R. D. (1990). Limited liability and incentive contracting with ex-ante action choices. Journal of economic theory, 52(1), 45
67. Jaffee, D. M., & Russell, T. (1976). Imperfect information, uncertainty, and credit rationing. The Quarterly Journal of Economics, 90(4), 651
666. Jappelli, T., & Pagano, M. (2002). Information sharing, lending and defaults: Cross-country evidence. Journal of Banking & Finance, 26(10), 2017
2045. Jappelli, T., Pagano, M., & Bianco, M. (2005). Courts and banks: Effects of Judicial enforcement on credit markets. Journal of Money, Credit and Banking, 37(2), 223
244.

30

ANGELINA NIKITENKO CHRISTIE

Kallberg, J. G., & Udell, G. F. (2003). The value of private sector business credit information sharing: The US case. Journal of Banking & Finance, 27(3), 449
469. Kassis, M. M., Hazlett, D., & Battisti, J. E. Y. (2012). A Classroom Experiment on Banking. The Journal of Economic Education, 43(2), 200
214. Klein, D. B. (1992). Promise keeping in the great society: A model of credit information sharing. Economics & Politics, 4(2), 117
136. Love, I., & Mylenko, N. (2003). Credit reporting and financing constraints. World Bank Policy Research Working Paper 3142. World Bank, Development Research Group, Finance, Washington, DC. McIntosh, C., Sadoulet, E., Buck, S., & Rosada, T. (forthcoming, in press). Reputation in a public goodsgame: Taking the design of credit bureaus to the lab. Journal of Economic Behavior & Organization. McMillan, J., & Woodruff, C. (1999). Dispute prevention without courts in Vietnam. Journalof Law, Economics, and Organization, 15(3), 637
658. Moulton, H. G. (1918a). Commercial banking and capital formation: I. The Journal of Political Economy, 26(5), 484
508. Moulton, H. G. (1918b). Commercial banking and capital formation: II. The Journal of Political Economy, 26(6), 705
731. Moulton, H. G. (1918c). Commercial banking and capital formation: III. The Journal of Political Economy, 26(7), 705
731. Moulton, H. G. (1918d). Commercial banking and capital formation: IV. The Journal of Political Economy, 26(9), 849
881. Myers, S. C. (1977). Determinants of corporate borrowing. Journal of financial economics, 5(2), 147
175. Padilla, A. J., & Pagano, M. (2000). Sharing default information as a borrower discipline device. European Economic Review, 44(10), 1951–1980. Pagano, M., & Jappelli, T. (1993). Information sharing in credit markets. The Journal of Finance, 48(5), 1693
1718. Petersen, M. A., & Rajan, R. G. (1994). The benefits of lending relationships: Evidence from small business data. The Journal of Finance, 49(1), 3
37. Reinhart, C. M., & Rogoff, K. (2009). This time is different: Eight centuries of financial folly. Princeton, NJ: Princeton University Press. Rietz, T. A., Sheremeta, R. M., Shields, T. W., & Smith, V. L. (2013). Transparency, efficiency and the distribution of economic welfare in pass-through investment trust games. Journal of Economic Behavior & Organization, 94, 257
267. Serra-Garcia, M. (2010). Moral hazard in credit markets: The incentive effect of collateral. Tilburg University, Working Paper. Sobel, J. (1985). A theory of credibility. The Review of Economic Studies, 52(4), 557
573. Smith, V. L. (2008). Rationality in economics. New York, NY: Cambridge University Press. Stiglitz, J. E., & Weiss, A. (1981). Credit rationing in markets with imperfect information. The American Economic Review, 71(3), 393
410. Stiglitz, J. E., & Weiss, A. (1983). Incentive effects of terminations: Applications to the credit and labor markets. The American Economic Review, 73(5), 912–927. Trautmann, S. T., & Vlahu, R. (2013). Strategic loan defaults and coordination: An experimental analysis. Journal of Banking & Finance, 37, 747
760. Vercammen, J. A. (1995). Credit bureau policy and sustainable reputation effects in credit markets. Economica, 62(248), 461
478.

CHAPTER 3 STRATEGIC DEFAULT WITH SOCIAL INTERACTIONS: A LABORATORY EXPERIMENT Jean Paul Rabanal ABSTRACT Purpose
The chapter studies strategic default using an experimental approach. Design/methodology/approach
The experiment considers a stochastic asset process and a loan with no down-payment. The treatments are two asset volatilities (high and low) and the absence and presence of social interactions via a direct effect on the subject’s payoff. Findings
I demonstrate that (i) people appear to follow the prediction of the strategic default model quite closely in the high asset volatility treatment, and that (ii) incorporating social interactions delays the strategic default beyond what is considered optimal.

Experiments in Financial Economics Research in Experimental Economics, Volume 16, 31
52 Copyright r 2013 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0193-2306/doi:10.1108/S0193-2306(2013)0000016003

31

32

JEAN PAUL RABANAL

Originality/value
The study tests adequately the strategic default using a novel experimental design and analyzes the neighbor’s effect on that decision. Keywords: Real options; optimal stopping; laboratory experiment JEL classifications: G13; D83; C91

INTRODUCTION The option to default is instrumental in protecting household wealth. However, under certain conditions, this option can quickly become toxic. The importance of understanding the forces that drive the decision to default has caught the attention of academia and the general public following the global financial crisis that originated in 2007
2008. During this time period, home prices plummeted and as a result, many households found themselves underwater. Once liabilities became greater than the actual asset values, numerous households chose to default on their mortgages. Inspired by these events, my work considers a situation in which the asset serves as collateral in case of default and attempts to identify the strategic forces that explain default in a simple setting, without liquidity constraints or preferences shocks. The novelty of my approach is that I use a laboratory experiment as a complement to the empirical fieldwork in order to properly examine the option value present in the default decision. In finance, default is modeled as a put option. One of the main problems when testing the financial model using field data is that it is necessary to estimate the asset process and it is hard to isolate a number of confounding factors that may explain default. In the laboratory, I have the advantage to set the asset process and loan characteristics. Thus, I can accurately compute the optimal strategy and test whether subjects wait to default according to the financial option. I also study possible social interactions in the decision to default. For example, using survey data, Guiso, Sapienza, and Zingales (2009) find that people who know someone who defaulted strategically are 82 percent more likely to declare their intention to do so. I incorporate the element of social interaction in my setup via a direct impact on the termination payoffs. In my experiment, 118 human subjects are endowed with an asset that has been acquired with a loan. The asset price follows a stochastic process and the subject has the option to default at any point in time.

Strategic Default with Social Interactions

33

The experimental design is between subjects and two-by-two. The first level corresponds to asset price volatility
low and high
and the second level considers the presence of the neighbor’s effect
independent and dependent. The first level helps me capture changes in option value meanwhile the second allows me to study the effect of social interaction. My findings suggest that the subjects act according to the theoretical predictions in a high volatility environment, given the absence of social interactions. Adding social interactions via a direct effect on the subject’s end of period payoff leads to a delay in the strategic default. The organization of the chapter is as follows: the second section elaborates on the contribution of the chapter in the context of existing literature. The third section lays out the default problem. The fourth section contains the testable predictions based on the conjecture that actual behavior will approximate optimal behavior, that is, subjects will defer their decision until they are sufficiently underwater. The fifth section presents the results and the sixth section concludes with a discussion of results and suggests possible avenues for future research.

RELATED LITERATURE Strategic default, or the decision of a borrower to default despite having the capacity to make payments, can be modeled as a real put option. In such case, the borrower decides to give up the asset when the spot asset price is sufficiently lower than the financial obligation.1 Unfortunately, using field data, it is difficult if not impossible to obtain exact knowledge of borrower strategy since the option value is unobservable to the econometrician. For instance, in the mortgage literature researchers have used the loan-to-value (LTV) variable to predict when a borrower will default on the mortgage (see Bajari, Chu, & Park, 2008; Deng, Quigley, & Order, 2000; Foote, Gerardi, & Willen, 2008; Gerardi, Shapiro, & Willen, 2007; PenningtonCross & Ho, 2010). The theory predicts that the option value is present in the level of LTV at which the households default on their mortgage. However, it is difficult to predict the asset price process for each household as other confounding variables are present in the default decision. In an experimental environment, isolating and measuring the option value becomes feasible, and we are thus able to properly test the strategic default decision. Contagion effects are among the variables that may also influence the strategic default. These effects can be present via social norms or material

34

JEAN PAUL RABANAL

incentives. In terms of the mortgage default decision, Chan, Gedal, Been, and Haughwout (2010) find that mortgage holders, living in the New York City areas with high foreclosure rates and high real estate owned (REO) activity, have a substantially greater risk of falling behind on the mortgage, as do mortgage holders living in predominantly black neighborhoods. Using survey data, Guiso et al. (2009) find that people who know someone who defaulted strategically are 82 percent more likely to declare their intention to do so. Pertaining to material incentives, Campbell, Giglio, and Pathak (2009) show that foreclosure at a distance of 0.05 miles lowers the price of a house by approximately 1 percent in the state of Massachusetts. Identifying endogenous social interaction effects in the field data can be cumbersome. Applying the work of Manski (1993) to default in neighborhoods, there may be correlated effects that need to be considered. For instance, neighborhoods will have higher default rates due to a common aggregate shock to the economy rather than endogenous interactions. There are also contextual interactions at play when default rates vary with the socioeconomic composition of neighborhoods. Therefore, when empirical studies show that neighborhoods with high foreclosure rates are more likely to fall behind on their mortgages, it is not necessarily clear whether this is due to endogenous or correlated effects. Given these difficulties, a laboratory experiment appears to be a better environment for studying the effect of social interactions. The present chapter is not the first to study deferral options in the laboratory. A few recent studies have focused on testing call options inspired by Dixit and Pindyck (1994). In this framework, the project value follows a random process and the subject must incur a fixed cost to seize an irreversible investment opportunity. List and Haigh (2010) work with a two-period model using a simple design. A set of contracts is offered to the subjects, and each contract specifies two alternatives: invest today or tomorrow. Results indicate that undergraduate subjects, as well as professional traders, choose the correct alternatives, according to the theoretical predictions of the option model. In a different experiment, Oprea, Friedman, and Anderson (2009) design a stochastic model with three different option values (high, medium, and low), using a finite horizon environment. They effectively demonstrate that undergraduate subjects follow the option model in the low treatment. Subjects in other treatments invest at values below optimum, but with predicted ordering. The authors also find evidence of learning behavior when the option value is either low or medium. The present chapter has similar features but also incorporates social interactions. The next section describes the environment and the optimal default decision.

Strategic Default with Social Interactions

35

THE DEFAULT PROBLEM The problem assumes that the subject needs an asset to get a certain level of felicity. She starts the game with an asset H that has been acquired with a loan (L0) whose LTV is equal to one. The loan characteristics are as follows: (i) interest rate payments must be made each period and (ii) the principal is paid at the end. The asset H varies randomly. At any moment, the subject can decide either to fulfill the financial obligation or to default, which implies a fee C. After defaulting, the agent will rent the asset until the end of the period. Notice that C is interpreted as the upfront lease and possible transaction costs associated with defaulting. Alternatively, a subject that pays all interest rate payments and L0 will receive B as a reward for honoring the payments. B may depend on the decision of other subjects
but is assumed to be constant for now.2 Fig. 1 displays the environment that each subject faces. H varies randomly and the present value of liabilities is constant
depicted as a horizontal

Fig. 1. The Environment. The Initial Asset Value (H) is 400 and the Horizontal Line Represents the Liabilities L. Thus, the Vertical Axis Shows the Value and the Horizontal Axis the Time Period t ∈ [0; 360]. The bars at the Right Have the Same Length and Represent the Quit Fee C and the Reward B.

36

JEAN PAUL RABANAL

line
due to the loan specifications. The reward (B)
conditional on making all interest rates payments throughout the game
is represented as the bar above the horizontal line at the end of the period. The default fee C is shown as a bar below the horizontal line at the end of the period. The default problem can be expressed as a Bellman equation, which I solve using numerical methods. The Objective Function In order to solve the problem, I introduce new notation. FðH; tÞ is the value of the objective function that depends on the asset value H and time t ∈ [0;T]. The subject has a binary decision problem at every point in the game until time T: (a) default and take the early termination payoff or (b) continue to the next period where another binary decision will be present. The first alternative considers the fee C and the second includes the loan interest rate payments rL0 and the value of problem in the next period.3 Thus, the Bellman equation for any t is FðH; tÞ = maxf− C; − rL0 dt þ ð1 þ rdtÞ − 1 E[FðH þ dH; t þ dt|HÞ]g

ð1Þ

The subject will maximize the value of FðH; tÞ taking into account the following information: (i) final payoff, (ii) asset price dynamics, and (iii) the initial conditions. The payoff at the end of the game is equal to the asset value at time T, minus the principal repayment L0, plus the reward B or FðH; TÞ = HT − L0 þ B

ð2Þ

Also, the asset follows a standard Brownian motion, dH = αHdt þ σHdz

ð3Þ

where α is the asset price growth and σ is the asset price volatility.4 Finally, the interest rate (r), the initial asset value H0, B and C are given. Binomial Approximation Solution Eq. (1) is solved numerically. Appendix A shows the solution method. It assumes risk-neutrality. Following Cox, Ross, and Rubinstein (1979), the continuous asset process is approximated by a discrete random walk. Usually this solution method is represented as a tree. The solution begins at the terminal period in the tree and is obtained by solving backwards.

37

Strategic Default with Social Interactions

At each node, the subject decides to stop paying if the expected early termination payoff is greater than expected future value of continuing. Denote  H as the threshold price, defined as the asset price at which the subject is  indifferent between paying and defaulting. Asset values above H > H will induce the agent to stay in the game and pay the financial obligations. On  the contrary, for H ≤ H the agent would be better off defaulting.  Fig. 2 shows the levels of H at every step in the discrete model. In total there are 360 points or T = 360. The starting asset price is 400, the reward and cost are equal B = C = 60.5 The figure includes the difference between the initial loan and the default fee labeled as the “underwater” price or  H U = L0 − C. In general, H < H U because the subject places a value on the foregone opportunity of defaulting later. The difference in prices then reflects the option value, which is also affected by the volatility of the asset  price. The higher the volatility, the lower the value of H . Social Interactions The subject can be influenced by the decisions of others. The social interaction effect will be captured via the value of B received at the end of the 500

Dollars

400

300

200

100

Underwater Threshold

0 0

Fig. 2.

50

100

150 200 Time 

250

300

350

The Optimal Threshold (H ) and Underwater Price (H U ) Over Time t ∈ [0; 360].

38

JEAN PAUL RABANAL

period. This assumption follows some work on the impact of defaulting on the sale price or on the monetization of preferences toward asset ownership.6 Thus, the subject will maximize a problem similar to Eq. (1); however, B will be specified as follows
B=

0 B

if n ≥ 1 if n = 0

where n is the number of people that default in the group.

HYPOTHESES AND EXPERIMENTAL DESIGN Before presenting the hypotheses, I will first define the treatments used in the experimental design. There are four treatments in total (a 2 × 2 design). The first pair corresponds to the volatilities (low and high) and the second pair, to the presence of neighbor’s effect (dependent and independent).

Hypotheses Let us denote hjv as the asset price when subject defaults, where j refers to the treatment with dependent (DEP) or independent (IND) bonus and v refers to the volatility environment: low or high. Notice that default prices depend on time, however, the time subscripts are omitted for brevity. Hypothesis 1. Volatility order: observed default price has the same ordinal rank across treatments as does the theoretical price. Therefore, hjhigh < hjlow This hypothesis is drawn from the model. The higher the volatility of the asset price, the lower the asset price that triggers default. In other words, it is optimal to wait longer to default. Hypothesis 2. Point prediction: for each treatment, the observed default price is equal to the theoretical price. This hypothesis focuses on the price level that triggers default. It assumes that the subject defaults accordingly to the optimal solution of the model.

Strategic Default with Social Interactions

39

Hypothesis 3. Social interactions: observed default price in the independent treatment is equal to or lower than the observed default price in the dependent treatment. Therefore, hIND ≤ hDEP v v In this case, I allow for deviation from the optimal behavior. If a subject defaults under the dependent treatment, then it is also likely that her partner stops due to the absence of the bonus.

Implementation Fig. 1 shows the screen observed by each subject. At the beginning of the period7 each subject is endowed with a base payment and owns an asset with value H that has been acquired with an LTV equal to 100 percent. The asset evolves according to the discrete binomial approximation of the Brownian motion described above. The liabilities are constant (depicted with a horizontal line) and equal to the initial value of H. Throughout the game, the subject is able to observe whether the bonus is present (bar above the horizontal line) as well as see the value of early termination fee (bar below the horizontal line). The subjects default by pressing the space bar. In the case of default, the payoff will be equal to the base points minus the sum of the quit fee and the interest payments. Despite stopping early, the subject will still be able to observe the black H-line evolving until the end of the period. The payoff in the case that subject does not default are equal to asset value plus the bonus
conditional on the treatment group
minus the financial payments (principal and interest). The payoff is depicted with a green bar. Appendix B includes screenshots of both scenarios, as observed by the subject. It is also important to highlight that instructions do not use the term “default” rather they refer it as “stop” to avoid any framing effects on the subject’s decision. In addition to the graphical display, the experimental software shows the values to be gained from defaulting, the payoff received in the current period and the cumulative earnings. The gain is computed as the difference between the liabilities and the sum of asset, default fee and bonus. The baseline parameters are T = 360 ticks (36 seconds), H0 = 400; r = 0:0006, and C = 60 = B. In order to arrive at a binomial approximation of the asset price, we use p to denote the probability of the asset value

40

JEAN PAUL RABANAL

going up, and u to indicate the increment. Their values are treatment dependent ðp = 0:506; u = 1:025Þ and ðp = 0:498; u = 1:04Þ for the low and high volatilities, respectively.8 The subjects participating were 118 undergraduate students at the University of California, Santa Cruz. At the beginning of the experiment, each subject was seated at a visually isolated computer terminal and assigned to a treatment (e.g., low volatility and dependent bonus). The dependent treatment consists of a random pair matching per round (or asset realization). Instructions were read aloud and the software was displayed on a screen. The binomial parameters for the chosen treatment were explained and written on a whiteboard. Subjects participated in four practice periods. Each subject then participated in 70 paid periods with no change in treatment. Sessions lasted 80 minutes each. This experiment may be deemed complex for two main reasons, and as such, several remarks are necessary before proceeding. First, it is hard to figure out the optimal threshold without using numerical techniques. Second, many subjects are not financially literate. For these reasons, I emphasized the graphical display to show the dynamic of the price and the payoff value. Lastly, during practice periods, I make sure that the subjects clearly understand the purpose of the game. A total of 34 subjects participated in the low-dependent treatment, 36 in the low-independent treatment, 34 in the high-dependent treatment, and 14 in the high-independent treatment. No subject was allowed to participate in more than one session. A subject with cumulative payoff over all periods received cash at the end of the session. Subjects also received a $5 bonus show-up fee. On average, the subjects received $16.25 in low volatility treatment and $15.76 in high volatility treatment. The conversion rate was calculated so that if a subject decided to default in the beginning of the period, he received $10 (400 × 70 × 3:57E − 04).

RESULTS Table 1 summarizes the experimental results and provides the following information for the four treatments: the number of subjects, the number of observations
equivalent to the total number of periods or the total number of loans
and the percentage of periods (loans) in which subjects default. The data is presented in three formats, using: (i) all periods, (ii) periods in which the minimum price is lower than the optimal threshold,

41

Strategic Default with Social Interactions

Table 1. Summary: Default Decision. Low Dep.

Low Indep.

High Dep.

High Indep

34 22 2380

36 23 2198

34 24 2380

14 37 968

Sample: min. price < threshold Subjects 34 % Default 25 Observations 634

36 41 537

34 39 578

14 79 234

26 53 51

34 42 240

14 94 62

Sample: All Subjects % Default Observations

Sample: min. price < 60% of threshold Subjects 34 % Default 32 Observations 34

Note: Obs. represent the total loans. Threshold refers to the predicted asset value that triggers default.

and (iii) periods where minimum price is lower than 60 percent of the optimal threshold. The optimal threshold is provided by the numerical solution to Eq. (1). Looking at the complete sample, it is worth noting the low default rate in all treatments. The default rates are under 40 percent and can be explained by the following: (i) the random asset price realization does not cross the underwater line and (ii) default is more likely to be observed when subjects choose a low price threshold. For these reasons, I decided to work with a subsample of data where the asset crossed 60 percent of the optimal threshold value. In Appendix C, I present more subsamples. I choose the 60 percent level since this is a sufficiently small asset realization that allows me to observe greater default rate in all treatments, specially in the high-dependent treatment. The selection of subsamples does not introduce any bias to the presented results since neither I nor the subjects had any ex ante information regarding asset price realization. Working with the subsample data, we can clearly observe higher default rate in the high volatility treatments compared to the low volatility treatments. In particular, the high-independent treatment displays a default rate that is greater than 90 percent. However, the default rate is lower when the social interactions are incorporated. The rate is roughly doubled when the subjects make their decision without the presence of social interaction.

42

JEAN PAUL RABANAL

An alternative way to illustrate the experimental results is to use the cumulative distribution of default. I work with the Product Limit (PL) estimator, which is used to establish the hazard distribution of default and it is an appropriate method for censoring data (see Kaplan & Meier, 1958). Fig. 3 shows the PL estimator for each treatment using pooled data. In order to properly read the graph, it is important to highlight that we are working with a subsample consisting of low asset realizations. Therefore, the x-axis displays values lower than the initial loan value of 400, while y-axis reveals the percentage of people that have defaulted at a given asset value. Recall that the initial loan value is equal to 400. As the price decreases, it is natural to observe more defaults. For the high independent treatment, the median default price is close to the theoretical prediction. However for the other treatments, the median price is much lower than the prediction or it cannot be computed due to the severe left-censoring. To control for the within-subject dependence, I also work with a “bysubject” sample which is constructed as the mean default price for the

1.0 Low dependent Low independent High dependent High independent Median optimal high Median optimal low

1−Survival

0.8

0.6

0.4

0.2

0 18 39 60 81 10 5 13 2 15 9 18 6 21 3 24 0 26 7 29 4 32 1 34 8 37 5

0.0

Default price (H)

Fig. 3.

Default Price Distribution
PL Estimator for Subsample (Pooled Data).

43

Strategic Default with Social Interactions

periods in which the asset price crosses the 60 percent threshold. In the case that subjects do not default, I use the censored value which is the minimum price observed in the corresponding period. For the treatment in which the bonus is not affected by others’ decision, the observations are completely independent. For the dependent treatment, however, this does not hold. Therefore, I treat the random neighbor-pairs in each session as a single observation. It is also appropriate to work with the “by-subject” sample since the pooled data over-samples the subjects that have a higher threshold value of default. Table 2 presents means and standard deviations of default prices for the high-dependent and high-independent treatments. It does not consider the low volatility treatments, due to severe left-censoring. Notice that the mean default price approximates the prediction for the high-independent treatment. The prediction is computed as the median optimal threshold. According to the solution of the stopping problem, the threshold is relatively constant over time except at the end of the period, when it approaches the underwater line. As a complement to the nonparametric analysis, I estimate a Tobit regression using the subsample pooled data (see Table 3) that uses clustered standard errors by subject for the independent treatment and by group for the dependent treatment. The dependent variable is the default price and the regressors are dummies that capture treatment effects. The estimation shows that the only significant treatment effect observed is the presence of social interactions. The lack of statistical significance of the asset volatility treatment could be explained by the severe left censoring. Below, I present the nonparametric tests of the point predictions and social interactions, followed by a summary of the parametric results and then conclude with a brief discussion. I test the point prediction using the Wilcoxon T-test on the “by-subject” sample and fail to reject that the median default price is equal to the theoretical prediction for the high independent treatment at 5 percent significance

Table 2. Default Prices: Mean and Standard Deviation. Pooled By subject Prediction

High Dependent

High Independent

105.2 ± 3.0 109.8 ± 9.9 166

130.4 ± 5.3 150.7 ± 13.4 166

Note: The prediction is the median optimal threshold.

44

JEAN PAUL RABANAL

Table 3.

Parametric Estimation (Tobit Model). Dependent Variable: Default Price (h)

Coefficient

Value (SD) 65.97*** (6.33) 28.73 (11.90) 80.82*** (14.12) −23.92 (21.50)

Constant Low Independent Low × Indep Note: Sample: Subsample (387 Obs). Clustered standard errors by group. *** Significant at 5% significance.

level. For the other treatments, the observed default price lies well below the predicted price. Result 1 Point predictions: The data supports only part of the point predictions. In the high-independent treatment, the observed default price approaches the predicted level. In other treatments, the observed default prices are significantly lower than predicted. I also use the “by-subject” sample to test the hypothesis of equal distributions of the high dependent and independent treatments using the Mann
Whitney test. The null hypothesis is rejected at 5 percent significance level. Likewise, using the parametric regression, the treatment effect of dependent/ independent is relevant with a positive coefficient at a 5 percent significance level. Therefore, the subjects in the dependent treatment wait longer to default. Result 2 Social interactions: The data shows that subjects significantly delay their decision to default when experiencing the social interaction treatment via a contingent reward.

DISCUSSION In this chapter, I have analyzed the strategic default in a loan using an experimental approach. My findings suggest that the subjects act according to the theoretical predictions in a high volatility environment, given the absence of social interactions. Adding social interactions via a direct effect

Strategic Default with Social Interactions

45

on the subject’s end of period payoff leads to a delay in strategic default. This behavior persists even when the asset price falls below the 60 percent threshold level. The theoretical predictions are drawn from a simple model that captures the subject’s decision between holding an asset acquired with a loan and defaulting. The asset follows a stochastic process and the subject receives its value at the end of the game, in addition to a bonus for paying interest. When the subjects defaults, she loses the asset, pays a quit fee, and is then free from any future financial obligation. Under the dependent treatment, social interactions are tied to the bonus. A possible explanation of the low default rate under social interactions treatment is that the subjects would like to avoid harming others. Stigma costs should also be considered, even though this is a blind pair-match between subjects. More work is needed to fully distinguish the factors that drive the strategic default. Having a better approximation of the real costs of default, would allow us to better understand how the default option is exercised. If this cost is indeed small, then it must be social norms that are driving the willingness to wait longer to default. Furthermore, social interaction can also affect the learning process regarding asset realization. In my environment, all parameters values are constant over the span of the game. However, a more realistic scenario may consider how a player might also learn from the behavior of others and could be an interesting extension. However, such modifications are beyond the present scope of the chapter, which aims to test how strategic are the agents when deciding to default.

NOTES 1. Dixit and Pindyck (1994) elaborate on this interpretation by analyzing investment decisions
call options. In their framework, the agent makes an irreversible investment decision once the project value surpasses the total cost, which includes the material fixed cost plus deferral option value. Deferring the decision has a positive value because exercising the option today implies that it cannot be exercised later. 2. A plausible argument in the present environment is that the felicity of renting the asset is different than owing it. Notice that in my experiment, the agent is forced to acquire the asset and I assume that there is no difference between renting it or owing it in the laboratory. Furthermore, I assume a fixed upfront lease in order to simplify the presentation of the problem to the subjects.

46

JEAN PAUL RABANAL

3. In any credit market, the lender may ask for a different rate for the interest payments, a rate that considers the loss associated with default. Since the main objective of this chapter is to analyze the borrower’s behavior, I assume that interest rate is risk-free. 4. Notice that the asset process does not consider any dividend service for holding the asset. This assumption is meant to simplify the model and to encourage higher default rates. Furthermore, to avoid any uncertainties regarding the parameter values, I assume that they are constant and the subjects have complete knowledge about their values. See the experimental design section for a better description of the information given to the subjects. 5. I also assume r = 0.0006 and σ = 0.025 as part of the parameters for the binomial approximation explained in Appendix A. 6. The social interaction can be captured by two other variables within the model: (i) the asset diffusion process and (ii) the cost of default. These alternatives are not considered in my experiment, as I aim to keep the dynamic of the game and the graphic display as simple as possible. To the extent of my knowledge, there is no existing evidence on how these interactions take place via these mechanisms. 7. This can be interpreted as a reincarnation, as the subject is able to observe random asset realization during each “lifetime.” Because a number of periods is being played, this allows for some learning in the default decision. However, the nature of the experimental data (high censoring) impedes me from analyzing such behavior. 8. The two criteria for the selection of parameters on the binomial approximates is to have a clear distinction between low and high volatilities in the graphical display and to observe enough underwater situations. The level of C and B are equal in order to simplify the math in computing the payoffs for the subjects. Given that interest rate is really small, I do not provide it to the subjects. Rather, I emphasize its importance in the graphical display of the game.

ACKNOWLEDGMENTS I would like to thank Dan Friedman, Ryan Oprea, Ai-ru Cheng, James Pettit, Olga Rabanal, Alessandra Cassar, Nirvikar Singh, and Sean Tanoos, and seminar participants at UCSC, USF, Banxico, CREED at University of Amsterdam, Max Planck Institute of Economics, University of Munich, and VGSE at University of Vienna for their invaluable comments and support. This research was supported by funds granted by SIGFIRM at the University of California, Santa Cruz.

REFERENCES Bajari, P., Chu, C. S., & Park, M. (2008). An empirical model of subprime mortgage default from 2000 to 2007. Working Paper No. 14625, National Bureau of Economic Research.

Strategic Default with Social Interactions

47

Campbell, J. Y., Giglio, S., & Pathak, P. (2009). Forced sales and house prices. Working Paper No. 14866, National Bureau of Economic Research. Chan, S., Gedal, M., Been, V., & Haughwout, A. (2010). The role of neighborhood characteristics in mortgage default risk: Evidence from New York city. Working Paper, Furman Center for Real State and Urban Policy, NYU. Cox, J., Ross, S., & Rubinstein, M. (1979). Option pricing: A simplified approach. Journal of Financial Economics, 7, 229
263. Deng, Y., Quigley, J. M., & Order, R. V. (2000). Mortgage terminations, heterogeneity and the exercise of mortgage options. Econometrica, 68(2), 275
308. Dixit, A. K., & Pindyck, R. S. (1994). Investment under uncertainty. Princeton, NJ: Princeton University Press. Foote, C. L., Gerardi, K., & Willen, P. S. (2008). Negative equity and foreclosure: Theory and evidence. Journal of Urban Economics, 64(2), 234
245. Gerardi, K., Shapiro, A. H., & Willen, P. S. (2007). Subprime outcomes: risky mortgages, homeownership experiences, and foreclosures. Working Paper No. 07-15, Federal Reserve Bank of Boston. Guiso, L., Sapienza, P., & Zingales, L. (2009, July). Moral and social constraints to strategic default on mortgages. Working Paper No. 15145. National Bureau of Economic Research. Retrieved from http://www.nber.org/papers/w15145 Kaplan, E., & Meier, P. (1958). Nonparametric estimation from incomplete observations. Journal of American Statistical Association, 53(282), 457
481. List, J. A., & Haigh, M. S. (2010). Investment under uncertainty: Testing the option model with professional traders. The Review of Economics and Statistics, 92(4), 974
984. Manski, C. F. (1993). Identification of endogenous social effects: the reflection problem. Review of Economic Studies, 60(3), 531
542. Oprea, R., Friedman, D., & Anderson, S. T. (2009). Learning to wait: A laboratory investigation. Review of Economic Studies, 76(3), 1103
1124. Pennington-Cross, A., & Ho, G. (2010). The termination of subprime hybrid and fixed-rate mortgages. Real Estate Economics, 38(3), 399
426.

48

JEAN PAUL RABANAL

APPENDIX A: NUMERICAL SOLUTION Under the binomial approximation, the asset return may take one of two possible values ðu; dÞ with the probability p ð1 − pÞ of the asset moving up (down). The parameters are defined as 1 u pffiffiffiffi u = expðσ δtÞ d=

p=

expðr⋅δtÞ − d u−d

where δt is the time step. In the discretization, the end of period payoff is FTm = EHTm − L0 þ B where FTm denotes the mth possible value at time T, EHTm is the expected asset price, L0 is the principal payment, and B is the reward. The optimal value in each node is the maximum between the two possibilities: the early termination payoff and the expected value in the next period. Fnm = maxð − C; e − rδt ð − rL0 þ pFnmþþ11 þ ð1 − pÞFnmþ 1 ÞÞ; 

n = 0; 1;…; T − 1

H is the asset price that makes indifferent the subject between defaulting  or continuing. The value of H can be found by solving the problem backwards.

Strategic Default with Social Interactions

49

APPENDIX B: INSTRUCTIONS Welcome! This is an economics experiment. If you pay close attention to these instructions, you can earn a significant sum of money. It will be paid to you in cash at the end of the last period. Please remain silent and do not look at other participants’ screens. If you have any questions, or need assistance of any kind, please raise your hand and we will come to you. If you disrupt the experiment by talking, laughing, etc., you may be asked to leave without compensation. We expect and appreciate your cooperation today. The Basic Idea Each period you will receive an asset that has been acquired with a loan. During the period, you will make interest payments on the loan and watch the asset value change randomly with time. You can decide to stop paying at any point, in which case the asset value no longer matters to you. If you keep the asset until the end of the period, you will receive its final value less the loan amount. Assets Over Time The session today is divided in a number of periods, each lasting 36 seconds. During the period, you will observe fluctuations in the asset value. At each point, the asset value can go UP or DOWN with a given probability. The increment value and the probability of going up/down are written on the white board. The asset value is displayed as the black line on your screen (see Fig. 1). Liabilities Over Time In the beginning of each period, you acquire the asset with a loan. The characteristics of the loan are such that you only pay interest payments until the end of the period. At the end of period, you also pay the principal. Your screen also shows the sum of future payments as a blue line. When the black line is above the blue line, it indicates that the asset value currently exceeds the future payments (liabilities). When it is below the blue line, those liabilities currently exceed the asset value. Stopping You can decide to stop paying the loan at any time by pressing the space bar. When you press, you have to pay a quit fee and you save the sum of

50

JEAN PAUL RABANAL

future interest payments and principal, but you no longer own the asset. If you want to see what the value of stopping is now, you can look at the text at the center top of the screen. It shows the current gain if you stop paying
the liabilities saved minus the quit fee and the asset value. The quit fee is shown as a red bar under the blue line, as in Fig. 1. Anytime that you stop by pressing the space bar you will observe a vertical bar that appears at the time you decide to stop. Payoffs Your payoff at the end of the period depends on whether you stopped paying. If you didn’t stop paying, you get Base pay + final asset value
all interest payments
loan amount. If you did stop paying you get Base pay
interest payments already made
quit fee. In some periods you may receive a bonus, an addition to your base pay, if you never stop the interest payments. This bonus is shown as a bar on top of the blue line. This bonus may change during the period, depending on the decisions of other players in your group. The bonus shrinks whenever someone in your group stops paying. Your screen shows the payoff you get on the top right corner and it also represent in the green bar that appears either when you decide to stop or at the end of the period. If the bonus depends on the decision of other players, you will see the fraction of players that stop paying below the “Gain Stop Paying.” Earnings You will be paid at the end of the experiment for all points earned over all periods. You can see your total points on the top left of the screen. The conversion rate from payoff points to U.S. dollars is written on the whiteboard. Frequently Asked Questions Q1. Is this some kind of psychology experiment with an agenda you haven’t told us? Answer. No. It is an economics experiment. If we do anything deceptive or don’t pay you cash as described then you can complain to the campus Human Subjects Committee and we will be in serious trouble. These instructions are meant to clarify how you earn money, and our interest is in seeing how people make decisions. Q2. Is the gain of stopping equal to the payoff I get? Answer. No. The gain of stop paying tells us what is the benefit to stop paying right now against the alternative of paying until the end of the period. The payoff you receive considers the interest payments that you have already paid and the quit fee if you decide to stop (Figs. B.1 and B.2).

Strategic Default with Social Interactions

Fig. B.1. .

Fig. B.2.

Screen Plot
Stopping Early.

Screen Plot
No Stopping.

51

52

JEAN PAUL RABANAL

APPENDIX C: ASSET PRICE SUBSAMPLES Table C.1.

Stopping Decision.

Low Dep.

Low Indep.

High Dep.

High Ind.

34 22 2,380

36 23 2,198

34 24 2,380

14 37 968

Sample: min. price < threshold Subjects 34 % Default 25 Observations 634

36 41 537

34 39 578

14 79 234

Sample: min. price < 90% of threshold Subjects 34 % Default 26 Observations 416

36 45 321

34 39 496

14 79 170

Sample: min. price < 80% of threshold Subjects 34 % Default 28 Observations 194

36 47 219

34 39 402

14 88 110

Sample: min. price < 70% of threshold Subjects 34 % Default 29 Observations 90

33 51 131

34 41 308

14 90 98

Sample: min. price < 60% of threshold Subjects 34 % Default 32 Observations 34

26 53 51

34 42 240

14 94 62

Sample: All Subjects % Default Observations

Note: Obs. is the total number of loans. Threshold refers to the predicted asset value that triggers default.

CHAPTER 4 DIVISIBLE-GOOD UNIFORM PRICE AUCTIONS: THE ROLE OF ALLOCATION RULES AND COMMUNICATION AMONG BIDDERS Martin Sefton and Ping Zhang ABSTRACT Purpose
We compare allocation rules in uniform price divisible-good auctions. Theoretically, a “standard allocation rule (STANDARD)” and a “uniform allocation rule (UNIFORM)” admit different types of low-price equilibria, which are eliminated by a “hybrid allocation rule (HYBRID).” We use a controlled laboratory experiment to compare the empirical performances of these allocation rules. Design/methodology/approach
We conduct three-bidder uniform price divisible-good auctions varying the different allocation rules (standard, uniform, or hybrid) and whether or not explicit communication between bidders is allowed. For the case where explicit communication is allowed we also study six-bidder auctions.

Experiments in Financial Economics Research in Experimental Economics, Volume 16, 53
86 Copyright r 2013 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0193-2306/doi:10.1108/S0193-2306(2013)0000016004

53

54

MARTIN SEFTON AND PING ZHANG

Findings
We find that prices are similar across allocation rules. Under all three allocation rules, prices are competitive when bidders cannot explicitly communicate. With explicit communication, prices are collusive, and we observe collusive prices even when collusive agreements are broken. Collusive agreements are particularly fragile when the gain from a unilateral deviation is larger, and an implication of this is that collusive agreements are more robust under STANDARD. Research limitations/implications
We do not find conclusive evidence of differences in performance among allocation rules. However, there is suggestive evidence that STANDARD may be more vulnerable to collusion. Originality/value
Divisible-good uniform price auctions are used in financial markets, but it is not possible to use naturally occurring data to test how alternatives to the standard format would perform. Using laboratory methods we provide an initial test of alternative allocation rules. Keywords: Multiunit auctions; divisible-good auctions; uniform price auctions; allocation rules

INTRODUCTION Uniform price auctions are widely used for selling a variety of assets, such as Treasury bills, spectrum, initial public offerings, and pollution permits. A potential drawback of this auction format is the existence of low-price equilibria which can result in large underpricing and severely reduced seller revenue (Wilson, 1979). Recently, Kremer and Nyborg (2004a, 2004b) point out that the extent of equilibrium underpricing is sensitive to the allocation rule, and in particular how excess demand is rationed. In Kremer and Nyborg (2004a) they show that the standard allocation rule (STANDARD), whereby bids placed above the market price are fully fulfilled and the rest of units are rationed among bids placed at the market price, may inhibit competition. In contrast, they show that a simple uniform allocation rule (UNIFORM), where rationing applies to all winning bids (bids placed either above or at the market price), encourages a Bertrand-like competition among bidders, and thus eliminates the low-price equilibria. However, when bidders are capacity-constrained UNIFORM admits another set of low-price equilibria in which all bidders bid for their

The Role of Allocation Rules and Communication Among Bidders

55

capacity at a low price. In this case Kremer and Nyborg (2004b) suggest that a hybrid allocation rule (HYBRID), a weighted average of the UNIFORM and STANDARD, fosters price competition and eliminates underpricing. The Kremer and Nyborg (2004b) theoretical result has clear practical implications, however it is difficult to test using field data as hybrid mechanisms are not used in practice. Thus, in this chapter we compare these three alternative allocation rules in a laboratory experiment based on their theoretical setup. In our first set of experiments we compare the three allocation rules in a three-bidder setting with limited communication among bidders. We find little difference in prices across allocation rules. For all three rules we observe aggressive bidding which results in high prices. There is little evidence of low-price equilibria or, more generally, of low-price outcomes being attained. In our second set of experiments we studied a setting that may be more conducive to coordination on low-price equilibria, namely one in which we introduced richer possibilities for communication by allowing bidders to send messages through a chat box before submitting bids. In this set of experiments coordinated strategies predominate and the market price drops dramatically under all three allocation rules. Analysis of individual behavior shows that this is not necessarily due to the play of low-price equilibrium strategies, but rather we see that bidders are often able to coordinate on nonequilibrium agreements that are apparently difficult to achieve and maintain without chat opportunities. These nonequilibrium agreements typically involve bidders submitting flat demand curves. We also see that some agreements are broken, and that larger the potential gain from a unilateral deviation, more fragile the agreements. However, a structural feature of uniform price auctions is that low-price outcomes may be maintained even if some bidders renege on an agreement. In fact, we observe collusive prices even in groups where collusive agreements were broken. In our third set of experiments we retained chat opportunities but increased the number of bidders from three to six. We find that bidders are less cooperative than in the three-bidder case, in the sense that nonequilibrium agreements are honored less frequently. Though many subjects keep to a nonequilibrium agreement to bid low, some “cheat” and raise their bids, although this has only a weak effect on price. We do, however, see some differences in behavior between small and large groups. In larger groups, bidders no longer rely on flat demand curves, but instead coordinate on steep demand curves that are less vulnerable to deviation. In particular,

56

MARTIN SEFTON AND PING ZHANG

under STANDARD, low-price equilibria involving steep demand curves are observed even with six bidders. The experimental literature on multiunit uniform price auctions goes back to Smith (1967). Much of this literature compares the standard uniform price auction with discriminatory auctions and focuses on private value settings. Goswami, Noe, and Rebello (1996) study the effect of nonbinding pre-play communication in a common value setting similar to ours, and our results add to their evidence on how communication affects auction performance in this important setting. As far as we are aware, ours is the first experimental study of UNIFORM or HYBRID. The remainder of the chapter is organized as follows. In the next section we briefly review the theoretical background on which our experiment is based. In the third section we describe our experimental design and procedures. In the fourth section we report the results, and in the fifth section we conclude.

THEORETICAL BACKGROUND Equilibrium Underpricing and Auction Rules Low-price equilibria in uniform price auctions were identified by Wilson (1979) in the context of share auctions. The basic idea is simple. In standard uniform price auctions bidders submit demand functions, and the market price is determined as the highest price such that aggregate demand is at least equal to supply. Demand for units at prices above the market price is fully fulfilled, and demand at the market price is rationed pro rata among marginal bids. By submitting inelastic demand functions that just clear the market, any attempt by a bidder to increase his or her allocation will cause a large increase in price. When the aggregate demand function is steep enough, any deviation is not profitable and a low price can be sustained in equilibrium. Further theoretical work has demonstrated the generality of Wilson’s results by showing the existence of a continuum of low-price equilibria under different information structures, risk attitudes, supply uncertainty, and continuous/discrete demand functions (see Ausubel & Cramton, 2004; Back & Zender, 1993; Klemperer & Meyer, 1989; Wang & Zender, 2002). Kremer and Nyborg (2004a) show that low-price equilibria can be eliminated by using UNIFORM so that demand at and above the market price is rationed pro rata. By bidding for the same amount at a slightly higher

The Role of Allocation Rules and Communication Among Bidders

57

price than the market price, a bidder can increase his or her allocation at the cost of a small increase in price. Thus, this rule induces price competition at the margin and eliminates low-price equilibria. Kremer and Nyborg (2004b) expand the analysis and point out a potential problem with UNIFORM which arises if bidders are capacityconstrained.1 In this case there is another set of equilibria that result in a low price. These are “single-bid equilibria” where bidders bid for their capacity at a common price. If a bidder’s capacity is less than the total number of units for sale then he or she cannot increase either the market price or the allocation by increasing his or her bid. Thus, there is no incentive to deviate and the common price is an equilibrium price. Kremer and Nyborg (2004b) suggest that a rule that incorporates properties of both allocation rules, for example, a weighted average of the two rules, encourages competition both above and at the margin and thus eliminates both types of low-price equilibria. The example below illustrates all of this. The example closely follows the basic model of Kremer and Nyborg (2004b), which in turn is consistent with the basic share auction model of Wilson (1979) and Back and Zender (1993) where a perfectly divisible-good with a common value is sold to a group of symmetric buyers. Kremer and Nyborg adapted the model to a discrete structure in both quantity and price dimensions and, for simplicity, considered the case where all bidders are risk neutral, and completely informed of the common value before submitting bids.

An Example There are 25 units for sale. There are three bidders and each bidder can bid for up to 24 units. Units have a common value of $10 per unit. Bidders can place a bid at any integer price between 0 and 10 for any integer quantity between 0 and 24. Each bidder can submit multiple bids, as long as the total number of units bid for does not exceed 24. The market price is the highest price where aggregate demand is greater than or equal to 25. If the aggregate demand at a price of 0 is less than 25 units the market price is 0 and each bidder obtains the number of units he or she bids for. Otherwise the allocation may differ under different allocation rules. Standard Allocation Rule Under STANDARD, when demand exceeds supply, bids placed above the market price are fully fulfilled. Then the remaining units are allocated

58

MARTIN SEFTON AND PING ZHANG

proportionately among the bid(s) placed at the market price. Suppose all three bidders bid for 8 units at a price of $10 and 16 units at a price of $0. Then the market price is 0. (At prices exceeding 0 aggregate demand is 24 and supply is 25.) Each bidder bids for 8 units at prices above the market price, and so immediately receives 8 units. One unit is left to be allocated proportionately, so each bidder also receives a further 1/3 unit. In fact, these bids constitute an equilibrium. Each bidder earns $83.33. The most profitable unilateral deviation would be for a bidder to bid for 8 units at a price of $10 and 16 units at a price of $1. Then the market price would be $1 and this bidder would be allocated 9 units. The bidder’s earnings would be $81, and so this deviation is not profitable. We refer to this type of equilibrium as a “Tacit Collusion Equilibrium” (TCE); similar strategies can support TCE prices at any price below $10.2 Notice that all bidders bidding for 24 units at $10 is also an equilibrium, and in this, each bidder obtains one-third of the units for sale at a price of $10. Submitting flat demand functions at any other price cannot support an equilibrium, as bidders would have an incentive to submit a flat demand function at a higher price. Uniform Allocation Rule Under UNIFORM, the units for sale are allocated proportionately among all winning bids, that is, bids placed are either above or at the market price. Suppose, as before, all three bidders bid for 8 units at a price of $10 and 16 units at a price of $0. Then the market price is 0, and the aggregate demand is 72 units at this price. Thus each bidder is awarded 25 × 24/72 units, or 8.33 units. Each bidder earns $83.33. This is not a Nash equilibrium under UNIFORM. By bidding for 8 units at $10 and 16 units at $1 a bidder could raise his or her allocation from 25 × 24/72 units to 25 × 24/40 units while increasing the price from $0 to $1: this would increase the bidder’s earnings to $135. Similarly other strategies that form a TCE under STANDARD are not equilibria under UNIFORM. Although some TCE strategies remain equilibria under UNIFORM, the lowest price that can be supported by such strategies is $8.3 However, because bidders are capacity-constrained, there exists another type of equilibrium. Suppose that every bidder bids for 24 at price 0. Each bidder receives 8.33 units at a price of $0. It is easy to see that no bidder could raise the market price by deviating as the maximum quantity he or she can bid for is 24 and there are 25 units for sale. Whatever a bidder bids, aggregate demand at prices above 0 would still be less than 25. Also, because bids placed above and at the market price are treated equally, as

The Role of Allocation Rules and Communication Among Bidders

59

long as the bidder bids for 24 units he or she would still obtain the same allocation. Similarly, if all bidders bid for 24 units at a common price, this constitutes a Nash equilibrium, and so any market price can be sustained as a Nash equilibrium. We refer to this type of equilibrium as a “Single Bid Equilibrium” (SBE). Hybrid Allocation Rule Kremer and Nyborg (2004b) further suggest HYBRID, which is a weighted average of the STANDARD and UNIFORM. In our experiment we use the simple average of the other two allocation rules. This allocation rule incorporates advantages of both the other two allocation rules and as a result, it reduces the scope for both types of lowprice equilibria. Like STANDARD, HYBRID discriminates in favor of inframarginal demand, and this eliminates single-bid underpricing equilibria; while as in UNIFORM, a bidder can significantly increase his or her allocation with a negligible increase in price, and this reduces the scope for TCE. As Kremer and Nyborg (2004b) note, HYBRID “creates incentives for bidders to be aggressive above as well as on the margin” (p. 865). For example, bidding for 24 units at a price of $10 is an equilibrium strategy, but there are no other SBE, and the lowest price that can be sustained by a TCE is $7.

Summary The example above is used in our experiment. In our experiment we also conduct auctions with six bidders. It is straightforward to obtain equilibria for different numbers of bidders. Table 1 gives the minimum equilibrium prices that can be sustained by TCE or SBE under each allocation rule. The maximum equilibrium price under all the three allocation rules is $10. Any price between the minimum and the maximum prices can be sustained in equilibrium.4 The large set of equilibria makes the performance of uniform price auctions crucially depend on equilibrium selection. Although the theoretical literature has focused on collusive, or payoff dominant, equilibria, a large experimental literature has shown that subjects often fail to coordinate on payoff dominant equilibria (e.g., see Devetag & Ortmann, 2007, for a review of coordination games). There have been several private value auction experiments with two bidders and two units for sale where the market price is the highest losing price. The payoff-dominant equilibrium requires subjects bid their private values for the first unit and 0 for the second

60

MARTIN SEFTON AND PING ZHANG

Table 1.

Minimum Equilibrium Prices Under Alternative Allocation Rules. 3 Bidders Minimum Minimum TCE price SBE price

STANDARD UNIFORM HYBRID

0 8 7

10 0 10

6 Bidders Minimum equilibrium price

Minimum Minimum TCE price SBE price

0 0 7

4 9 9

10 0 10

Minimum equilibrium price 4 0 9

unit, but the common outcome of such experiments is that bidders bid at or above the value for the first unit, and bid lower than the value but higher than 0 for the second unit (see Engelmann & Grimm, 2009; Kagel & Levin, 2001; Porter & Vragov, 2006). In general, collusive bidding is rarely observed in experimental uniform price sealed bid auctions.5 The evidence suggests collusion is only achievable if there are two bidders (see Kwasnica & Sherstyuk, 2007; Sherstyuk, 2008), subjects have a coordination device (Brown, Plott, & Sullivan, 2009; Li & Plott, 2009), anonymity is abandoned (Fu¨llbrunn & Neugebauer, 2007), auctions are in open format (Burtraw et al., 2009; Goeree, Offerman, & Sloof, 2013), or bidders are allowed to communicate prior to bidding (Goswami et al., 1996). Pre-play communication has been shown to be an effective facilitating device in a wide range of settings. It increases coordination on payoff dominant equilibria in experimental coordination games (see Devetag & Ortmann, 2007) and sustains low prices in various experimental auctions (e.g., Burtraw et al., 2009; Kwasnica, 2000; Isaac & Walker, 1985). As we describe in the next section, we implement two types of pre-play communication in our experiment.

EXPERIMENT Experimental Design Our experiment is based on the example described in the previous section. In particular, as in Kremer and Nyborg (2004b), we focus on a one-shot game by having subjects bid in only one auction for which they are paid their earnings. Although the equilibrium analysis of the one-shot game is

The Role of Allocation Rules and Communication Among Bidders

61

considerably less complex than the analysis of a repeated game, there remains a multiplicity of equilibria under all three allocation rules, and so it is unclear how subjects might coordinate on equilibrium. In order to provide better opportunity for equilibration, we preceded the auction by a series of “practice” auctions, for which subjects were not paid. We conducted three NC3 (“No Chat, 3 bidders”) treatments, one for each allocation rule. In each NC3 treatment, subjects were matched into groups of three and remained in this fixed group for 20 practice periods before a final “real” period in which subjects were paid their earnings. As well as giving subjects experience with the experimental rules and environment, these practice periods allow subjects to communicate with other bidders via their decisions in previous periods.6 Note that in addition to providing limited communication opportunities, this design incorporates several other collusion facilitating factors: a small number of bidders, symmetry among bidders, complete information, and high expected gains from collusion (Sherstyuk, 1999). As we report later, despite these factors, collusive behavior was rarely observed. In our C3 (“Chat, 3 bidders”) treatments, we introduced more explicit communication possibilities by allowing subjects to send and receive messages through a chat box prior to bidding. Finally, in our C6 (“Chat, 6 bidders”) treatments, we examined the three allocation rules with groups of six bidders, using the same explicit communication possibilities as in the C3 treatments. As collusion was rarely observed in our NC3 treatments we did not run further NC treatments with larger groups as there would be no reason to expect these to be less competitive than the NC3 groups.7

Procedures The experiment comprised 21 computerized sessions conducted at the University of Nottingham, using a total of 342 subjects recruited from a university-wide pool of undergraduate students. Subjects were recruited through the online recruiting system ORSEE (Greiner, 2004) and the experiment was programmed and conducted with the software z-Tree (Fischbacher, 2007). Two sessions were conducted for each NC3 and C3 treatment and three sessions were conducted for each C6 treatment. This resulted in 10 groups of three bidders per treatment (NC3 and C3) and 9 groups of six bidders (C6). This design is summarized in Table 2. At the beginning of each session, subjects were given a written set of instructions (reproduced in Appendix A) that the experimenter read aloud.

62

MARTIN SEFTON AND PING ZHANG

Table 2.

Summary of Experimental Design.

Practice Periods

Chat Allowed?

Bidders Per Group

Number of Groups

Number of Subjects

NC3 treatments STANDARD UNIFORM HYBRID

20 20 20

No No No

3 3 3

10 10 10

30 30 30

C3 treatments STANDARD UNIFORM HYBRID

11 11 11

Yes Yes Yes

3 3 3

10 10 10

30 30 30

C6 treatments STANDARD UNIFORM HYBRID

11 11 11

Yes Yes Yes

6 6 6

9 9 9

54 54 54

Total

342

Subjects then were allowed to ask questions by raising their hands and speaking to the experimenter in private. At the beginning of the first period, subjects were assigned to groups of either three or six (depending on treatments) and stayed in these groups for the entire session. Subjects were not informed of the identities of the other group members in any period, and no information passed across groups. In the treatments without chat subjects were not allowed to communicate with one another during the session, except via the decisions they entered on their terminals. The decision-making phase of the session consisted of 20 practice periods followed by one “real” period. At the beginning of the treatments with chat, subjects were assigned a subject ID, A, B, or C, which identified messages in the chat stage. After four practice periods without explicit communication, bidders then played seven more practice periods which consisted of a four-minute chat stage followed by the bidding stage. Thus, in these treatments there were 11 practice periods before the real period.8 In each chat stage bidders could send and receive messages via a chat box. They were prohibited from revealing personal information or making threatening, insulting, or offensive comments, but otherwise the chat was unrestricted.9 For each group a period consisted of an auction with 25 units for sale, where the value of each unit to a bidder was 10 points. There was no time

The Role of Allocation Rules and Communication Among Bidders

63

limit for submitting bids, and bidders could place bids at any integer price up to 10 points, for any integer number of units up to 24 units. Bidders could submit multiple bids as long as the total number of units bid for did not exceed 24. After all subjects had submitted their bids, the computer calculated the price and the allocation. Each subject was then informed of the bids submitted by the bidder’s group, the market price, bidder’s allocation, and bidder’s point earnings in the period.10 The period then ended and the next period began. Before the last period subjects were reminded that this was the last period and that their earnings would be based on points earned in this period. Sessions lasted 90 minutes on average and each subject was paid a show up fee of £5 plus £0.20 per point for points earned in the last period. Thus, in the last period each group was bidding for 25 units worth £2 each. Subjects earned £13.10 on average, ranging from a minimum of £5 to a maximum of £48.20.11

RESULTS Prices and Allocations We begin by focusing on market prices in the payment period.12 Fig. 1 presents histograms of prices by treatment. With the limited communication possibilities of our NC3 treatments, prices cluster at the high end of the price range and are not significantly different across allocation rules.13 When groups can chat, prices are lower. In our C3 treatments the average price is 0.60, and we find no significant differences across allocation rules (p = 0.639); in our C6 treatments the average price is 2.03, again not varying significantly across allocation rules (p = 0.839). Introducing explicit communication opportunities significantly lowers price: prices are significantly different in our C3 versus NC3 treatments (p < 0.001 for each allocation rule). The effect of the number of bidders (C3 vs. C6) is much smaller, although significant in the case of UNIFORM and HYBRID (STANDARD p = 0.363; UNIFORM p = 0.029; HYBRID p = 0.047 respectively). In principle, subjects might have used practice periods to coordinate on low-price equilibria in the NC3 treatments, but there is little evidence of this. In contrast, the effectiveness of explicit communication is clear from looking at how market prices vary across periods (see Fig. 2).

64

MARTIN SEFTON AND PING ZHANG

10

No Chat 3 Bidders

6 4 0

2

Frequency

8

STANDARD ave. price: 8.4 stdv.: 1.43 UNIFORM ave. price: 7.2 stdv.: 2.7 HYBRID ave. price: 7.9 stdv.: 2.64

0

1

2

3

4

5 Price

6

7

8

9

10

9

10

9

10

10

Chat 3 Bidders

6 4 0

2

Frequency

8

STANDARD ave. price: 1.6 stdv.: 3.2 UNIFORM ave. price: 0.1 stdv.: 0.32 HYBRID ave. price: 0.2 stdv.: 0.63

0

1

2

3

4

5 Price

6

7

8

10

Chat 6 Bidders

6 4 2 0

Frequency

8

STANDARD ave. price: 2.33 stdv.: 2.78 UNIFORM ave. price: 2.56 stdv.: 3.61 HYBRID ave. price: 1.22 stdv.: 1.64

0

Fig. 1.

1

2

3

4

5 Price

6

7

8

Histograms of Prices in Payment Period by Treatment.

65

The Role of Allocation Rules and Communication Among Bidders

0

2

Price 4 6

8

10

No Chat 3 Bidders

0

4

8

12

16

20

Period STANDARD

UNIFORM

HYBRID

0

2

Price 4 6

8

10

Chat 3 Bidders

0

4

8

12

16

20

Period STANDARD

UNIFORM

HYBRID

0

2

Price 4 6

8

10

Chat 6 Bidders

0

4

8

12

16

20

Period STANDARD

Fig. 2.

UNIFORM

Average Prices Across Periods.

HYBRID

66

MARTIN SEFTON AND PING ZHANG

When no chat option is available, prices increase in early periods and then remain high for the rest of the session under all three allocation rules. When communication via a chat box becomes available in period five prices drop dramatically. Even in HYBRID treatments, where the lowest equilibrium price is 7 (C3) or 8 (C6), bidders are able to attain lower prices when they send chat messages. Result 1: Without explicit communication opportunities prices are competitive under all three allocation rules. Introducing explicit communication opportunities leads to a substantial and significant decrease in price under all three allocation rules. Retaining explicit communication but increasing the number of bidders results in a small increase in the market price, which is statistically significant in UNIFORM and HYBRID. ’

Next we consider allocations. In panel (a) of Table 3 we report the fraction of groups that attained a perfectly symmetric allocation. The picture across the three allocation rules is similar. Without chat only 1 of 30 groups attained a perfectly symmetric allocation, whereas 18 of 30 (C3) and 10/27 (C6) attained a perfectly symmetric allocation with chat. Thus, chat messages facilitate equal allocations. Closer inspection of the data from STANDARD showed that although many groups in the C3 and C6 treatments coordinated on a perfectly symmetric allocation, those that failed to do so achieved highly asymmetric allocations. Out of 19 groups in the C3 and C6 treatments, 11 failed to coordinate on a symmetric allocation. Across these 11 groups the average standard deviation in allocation was 8.38, notably higher than the variability exhibited in the NC3 treatment. Similar inspection of UNIFORM data shows that groups that fail to coordinate on a perfectly symmetric allocation nevertheless exhibit less variability than in the NC3 treatment (the average is 5.0625), and the pattern in HYBRID is in between STANDARD and UNIFORM (average variability among groups that failed to coordinate on a perfectly symmetric allocation is 6.41). The variability in allocations, as measured by the standard Table 3.

Within-Group Variability in Allocations. (a) Fraction of Groups with Symmetric Allocation

STANDARD UNIFORM HYBRID

(b) Average Within-Group Standard Deviation of Allocation

NC3

C3

C6

NC3

C3

C6

0/10 1/10 0/10

5/10 7/10 6/10

3/9 4/9 3/9

5.84 6.37 5.50

4.37 1.44 2.45

5.39 2.90 4.40

The Role of Allocation Rules and Communication Among Bidders

67

deviation of the allocation within each group and averaged over groups, is reported in panel (b) of Table 3. In UNIFORM and HYBRID, introducing chat results in a significant decrease in variability (UNIFORM: NC3 vs. C3 p = 0.006; HYBRID: NC3 vs. C3 p = 0.039). However, in STANDARD we find no significant treatment effects (p = 0.253). Thus, although chat promotes collusive prices, under STANDARD, it appears to result in one of two extreme effects: either all bidders share the benefit of the low price equally, or else there is a high degree of asymmetry in the distribution of benefits. Result 2: Explicit communication facilitates coordination on symmetric allocations under all allocation rules. This results in lower within-group variability in allocations under UNIFORM and HYBRID. Under STANDARD, explicit communication results in a higher frequency of highly asymmetric allocations as well as a higher frequency of perfectly symmetric allocations, and so within-group variability in allocations, averaged across groups, is similar across the NC3, C3, and C6 treatments. ’

Bidder Behavior: No Chat Treatments In the no-chat treatments, bidding is very heterogeneous and no clear strategy predominates under any allocation rule. In only 3 of 600 practice periods did bids form an equilibrium. Given this, it is perhaps not surprising that groups failed to coordinate on equilibrium in the payment period. Only 2 of the 30 payment period auctions resulted in an equilibrium (and these at prices of 8 and 10). Thus, there is little evidence of coordination on equilibrium, let alone a low-price equilibrium. There is some evidence of bidders shading their bids in practice periods and bidding for more units at higher prices in the payment period. Consider bidders’ aggregate demand curves, as shown in Fig. 3. Average aggregate demand curves are shown for the first 10 practice periods, the last 10 practice periods, and the payment period. The demand curves for the earlier practice periods and later practice periods are similar in all treatments, but the demand curves for the payment periods clearly shift outward in all treatments. Most subjects bid more aggressively in the payment period than in the last practice period, by either bidding at a higher price, bidding for more units, or both.14 In principle, subjects could have used practice periods to coordinate on low-price equilibria, but in fact they did not. Coordination may have been difficult because different bidders were choosing different types of strategies such as flat demand curves or steep demand curves, and without the help of explicit communication it may be difficult to interpret signals sent from

68

MARTIN SEFTON AND PING ZHANG STANDARD

10

UNIFORM

HYBRID

Price

8 6 4 2 0 0

20

40

60

80

0

Period 1-10

Fig. 3.

20

40

60

Period 11-20

80

0

20

40

60

80 Units

Period 21

Average Aggregate Demand Curves in NC3 Treatments.

other bidders who are following a different strategy. Moreover, some bidders may have been deliberately attempting to mislead rival bidders, bidding low in the practice periods, whereas planning to bid higher in the payment period. Bidder Behavior: Chat Treatments We next turn to the chat treatments. Bidder behavior in periods 1
4, where chat is not available, is not very different from the no-chat treatments. As soon as chat is introduced, most groups develop coordinated strategies. In particular, chat promotes symmetric bids under all allocation rules, as shown in Fig. 4, which shows the proportions of groups where all bidders placed identical bids. In almost all groups, bidders reached an explicit agreement about what bids to place and kept to the agreement. Some coordinated on the agreement until the end of practice periods, whereas other groups developed different agreements, sometimes using more sophisticated strategies that made profitable deviation impossible. By the last practice period, most groups (43 out of 57) have settled on symmetric strategies. These strategies vary across groups, but invariably lead to a price of 0.15 However, there was a clear change in the bidding behavior in the payment period: across all allocation rules 26 of 57 groups submitted symmetric strategies. Pooling across allocation rules, most strategy combinations can be classified as belonging to one of three categories. In a Tacit Collusion strategy,

69

The Role of Allocation Rules and Communication Among Bidders

.6 .4 0

.2

Proportion

.8

1

No Chat 3 Bidders

0

4

8

16

12

20

Period STANDARD

HYBRID

UNIFORM

.6 .4 0

.2

Proportion

.8

1

Chat 3 Bidders

0

4

8

12

Period STANDARD

HYBRID

UNIFORM

.6 .4 0

.2

Proportion

.8

1

Chat 6 Bidders

0

4

8

12

Period STANDARD

Fig. 4.

UNIFORM

HYBRID

Proportions of Groups where All Bidders Plaed Identical Bids.

70

MARTIN SEFTON AND PING ZHANG

combination bidders submit strategies which would form a TCE under STANDARD.16 In a Single Bid strategy, combination bidders submit strategies corresponding to a SBE strategy in UNIFORM: all bidders bid for 24 units at a single price. In a Low Demand strategy combination, each bidder bids for a total of 8 units (or 4 in a group of 6) at a price lower than 10, keeping the aggregate group demand below 25 and leading to a market price of 0. Table 4 reports the number of groups in each category in the last practice period and the payment period. For completeness we also provide the breakdown of strategy combinations by allocation rule. Notice that the number of groups submitting Tacit Collusion strategy combinations in the payment period is similar to that in the last practice period. However, compared with the last practice period, there are far fewer groups submitting Single Bid or Low Demand strategy combinations. Almost all of the groups that successfully coordinated on a Tacit Collusion, Single Bid, or Low Demand strategy combination in the last practice period used the chat stage to agree on doing, in the words of more Table 4.

Strategy Combinations in the Chat Treatments. Number of Groups Submitting Strategy Combinations

Last practice period C3 STANDARD UNIFORM HYBRID C6 STANDARD UNIFORM HYBRID Total Payment period C3 STANDARD UNIFORM HYBRID C6 STANDARD UNIFORM HYBRID Total

Total

Tacit collusion

Single bid

Low demand

Other

2 2 1

4 4 3

1 2 3

3 2 3

10 10 10

3 4 3 15

1 1 2 15

0 2 3 11

5 2 1 16

9 9 9 57

3 1 1

3 2 2

0 2 2

4 5 5

10 10 10

3 3 2 13

0 0 1 8

0 1 0 5

6 5 6 31

9 9 9 57

71

The Role of Allocation Rules and Communication Among Bidders

than one subject, “the same again” in the payment period. However, 17 of the 41 groups broke the agreement. Whether a collusive agreement in the last practice period carries over to the payment period depends on the incentives for unilateral deviation. We calculate the “unilateral deviation gain” in a group as the highest increase in point earnings that a bidder could have attained by unilaterally changing his or her strategy in the last practice period.17 Then, we categorize each group as “successfully colluding” if they submit the same bids in the payment period as in the last practice period. Fig. 5 shows, for each allocation rule, the numbers of groups with various unilateral deviation gains (in 10 point bin widths), broken down by groups that successfully colluded (left panel) and groups that did not (right panel). Also shown are histograms of unilateral deviation gains for all allocation rules combined. Again, we report separate histograms for groups that did or did not successfully collude (in each panel the fractions sum to one). Pooling across all allocation rules, 24 groups are classified as successfully colluding, and the average unilateral deviation gain for these groups is 54, while 33 are classified as not successfully colluding, and the average unilateral deviation gain for these groups is 91. A rank-sum test for differences in unilateral deviation gains between groups that did or did not successfully

Deviation

Successful Collusion

3 2

number of groups

0

1

4 3 2

0

50 100 150 200 250 0 50 100 150 200 250 Unilateral deviation gain -- STANDARD Deviation

Successful Collusion

Deviation

0

.2 0

1

2

3

fraction of groups

.3

4

Successful Collusion

50 100 150 200 250 0 50 100 150 200 250 Unilateral deviation gain -- UNIFORM

.1

number of groups

1 0 0

number of groups

Deviation

4

Successful Collusion

0

50 100 150 200 250 0 50 100 150 200 250 Unilateral deviation gain -- HYBRID

Fig. 5.

0 50 100 150 200 250 0 50 100 150 200 250 Unilateral deviation gain -- ALL ALLOCATION RULES

Unilateral Deviation Gain and Successful Collusion.

72

MARTIN SEFTON AND PING ZHANG

collude indicates a significant difference (p-value = 0.031). It is not always the case that low-price equilibria attained in the last practice period are repeated in the payment period, and sometimes nonequilibrium outcomes in the last practice period are repeated in the payment period. However, the general pattern is that agreements are more stable the lower is the incentive to deviate from them. This may reflect that subjects incur costs from working out ways to increase earnings and small deviation gains do not justify this decision cost. Alternatively, subjects may incur psychological costs from breaking promises or misleading other group members and small deviation gains do not justify incurring these costs. Result 3: Explicit communication opportunities enable bidders to coordinate on strategies that generate low-price outcomes in practice periods. The extent to which this coordination translates to the payment period depends on the type of strategy and incentives for unilateral deviation. Strategies that offer lower unilateral gains from deviation are more stable than strategies offering higher gains. ’

Comparison of Strategy Combinations and Allocation Rules The extent to which strategy combinations are vulnerable to deviation is related to the type of strategy, the allocation rule, and the number of bidders. The analysis of unilateral deviation gain suggests that the greater the gain from a unilateral deviation, the more fragile the agreements, and so one might expect Tacit Collusion strategies to be stable in STANDARD, where they form an equilibrium, and less so in other allocation rules. In fact this is the case: under STANDARD, all five of the groups using a Tacit Collusion strategy combination in the last practice period also submitted a Tacit Collusion strategy combination in the payment period, whereas under the other allocation rules ten groups used Tacit Collusion strategy combinations in the last practice round but four of them changed in the payment period.18 The finding that Tacit Collusion strategy combinations, once achieved, are stable under STANDARD shows that STANDARD is susceptible to collusive equilibria, at least when bidders can explicitly communicate. Interestingly, we find this to be as true for six-bidder groups as for threebidder groups, reinforcing a point made by Wilson (1979) that sellers may not benefit from an increase in the number of bidders under this allocation rule. One might also expect Single Bid strategies to be more stable in UNIFORM, where they form an equilibrium, than under other allocation

The Role of Allocation Rules and Communication Among Bidders

73

rules. This, however, is not observed in the data: under UNIFORM two of five groups stayed with Single Bid strategies after using them in the last practice period. This compares with six of ten groups under the other allocation rules. Result 4: Tacit Collusion strategies featuring steep demand curves are more stable than Single Bid strategies featuring flat demand curves. In particular under STANDARD agreements to play Tacit Collusion strategy combinations in STANDARD are always kept, whereas under UNIFORM agreements to play Single Bid strategy combinations usually fail. ’

One explanation for why Single Bid strategies are more fragile than Tacit Collusion strategies may be the following. In STANDARD, if a bidder unilaterally deviates from a TCE by bidding more aggressively, he or she will suffer a reduced payoff. In UNIFORM, if a bidder unilaterally deviates from a SBE by raising his or her bid, he or she will not change the price or allocation, and so the bidder’s payoff will not change. Thus a unilateral deviation from SBE in UNIFORM does not make a bidder better off or worse off. Moreover, if a bidder worries that others may raise their bids, the best response on bidder’s part will be to raise his or her bid as well rather than following the Single Bid strategy. This may also explain why Single Bid strategies are particularly rare in our C6 treatments. As shown in Table 4, of 30 groups in our C3 treatments 11 submit Single Bid strategies in the last practice period and 7 submit Single Bid strategies in the payment period. By comparison, only 4 of 27 groups in the C6 treatments submitted Single Bid strategies in the last practice period and only 1 group submitted Single Bid strategies in the payment period. It may be that subjects are less confident that all bidders will follow a Single Bid strategy in a six-bidder group, and this makes playing a Single Bid strategy less attractive. Result 5: Tacit Collusion strategies featuring steep demand curves are more frequently used in six-bidder than three-bidder groups. In contrast, Single Bid strategies featuring flat demand curves are used less frequently in six-bidder than three-bidder groups. ’

DISCUSSION AND CONCLUSIONS We find little evidence of revenue differences due to alternative allocation rules for uniform price auctions. In our no chat treatments, subjects play repeatedly in practice periods before playing a single auction for payment, and so in principle can

74

MARTIN SEFTON AND PING ZHANG

communicate via decisions in practice periods in order to coordinate on low-price equilibria. However, we find bidding is competitive under all three allocation rules. Introducing explicit communication opportunities by allowing bidders to send messages via a chat box before placing bids leads to a substantial and significant decrease in the market price: low prices are observed under all three allocation rules. These low prices are generally not achieved through one of the low-price equilibria identified by theory. This is perhaps why the HYBRID, which is designed to eliminate such low-price equilibria, has such a weak effect. The effectiveness of explicit communication is reminiscent of the result that communication plays a more important role than repeated play in fostering successful collusive agreements in first price sealed bid auctions (Kwasnica, 1998). The effectiveness of explicit communication for facilitating collusion is also in line with the results of Goswami et al. (1996), who study a similar setting to ours using multiple pay periods. They also find that nonbinding pre-play communication increases bidders’ tendency to play collusive strategies, though not necessarily equilibrium strategies, in uniform price auctions under STANDARD, and low-price outcomes are sustained over multiple pay periods.19 A closer look at bidding behavior in the chat treatments reveals that different groups attained low prices in different ways. In three-bidder groups, a particularly frequent strategy combination consists of single bid strategies whereby bidders submit flat demand functions at a common price. In sixbidder groups such strategies are observed less frequently, and instead we more often observe Tacit Collusion strategies whereby bidders submit steep demand functions. We also observe that although many groups agreed to submit the same bids in the payment period as in the last practice period, these agreements were not always kept. In general, the success of an agreement depended on the incentives for unilateral deviation and the type of strategy. Strategies that offer low unilateral gains from deviation are more stable than strategies offering higher gains, and single-bid strategies are vulnerable even under UNIFORM where they form an equilibrium. In contrast, groups that followed Tacit Collusion strategies in practice periods tended to rely on these in the payment period as well. Although the stability of different types of strategy depends on the allocation rule, this does not translate into price differences. The reason is that many bidders keep to nonequilibrium agreements, and in groups where there is a deviation it is often only a single bidder who deviates and this results in the deviator receiving a higher allocation but not paying a higher price. In six-bidder groups nonequilibrium agreements are more vulnerable, but at the same time bidders protect themselves by coordinating on

The Role of Allocation Rules and Communication Among Bidders

75

different strategy combinations. In particular steep demand curves are frequently submitted. These are more complex but have the advantages that they give a smaller incentive to deviate and also allow the bidder to secure some allocation even if other bidders deviate. Moreover, under STANDARD, these strategies form an equilibrium and when groups agree to use these strategies in practice periods, they subsequently stay with the equilibrium in the payment period. An implication is that STANDARD is more susceptible to collusive equilibria, at least when bidders can explicitly communicate.

NOTES 1. They note: “Such constraints could be built into the auction by design
for example in U.S. Treasury auctions, (a bidder can bid up to 35% of the quantity for sale)
or from bidders’ limitations” (Kremer & Nyborg, 2004b, p. 863). 2. Any strategies supporting an equilibrium price below 10 under this allocation rule are termed TCE strategies. These strategies use high inframarginal bids to support a market price below 10. For example, to sustain an equilibrium price of 9, each bidder might demand 8 units at a price of 10 and 16 at a price of 9. In general a TCE requires that the total number of units requested at prices exceeding the market price is 24 (otherwise a bidder could increase his or her allocation without raising the market price), and moreover these inframarginal bids must be placed at sufficiently high prices. 3. All TCE can be eliminated if the tick price is sufficiently small. 4. Although we use a discrete setup in the example, as well as in the experiment, the effect of allocation rules also applies in a continuous setup (see Kremer & Nyborg, 2004b, for details). 5. For example, see Zhang (2009), who finds no evidence that repeated play helps bidders coordinate on collusive strategies in a multiunit uniform price auction with common values and four bidders. 6. Some previous experiments have found that subjects use such communication opportunities to attain more cooperative outcomes. For example, Burton and Sefton (2004) find that communication via decisions helps subjects coordinate on an efficient, but risky, equilibrium. In experimental multiunit ascending auctions with private values, Kwasnica and Sherstyuk (2007) find that repeated play allows subjects to signal and retaliate, fostering collusion. 7. On the other hand, one might expect less collusion in C6 than C3. For example, Kwasnica and Sherstyuk (2007) found that the collusive behavior observed in groups of two bidders failed in groups of five bidders. 8. We did not provide the chat option from period one because we wanted to give subjects some experience with the experimental rules and environment before sending messages. Also, time constraints meant that we could not conduct as many practice periods as in our NC3 treatments.

76

MARTIN SEFTON AND PING ZHANG

9. Subjects were told that they would forfeit their earnings if they violated these rules. An inspection of chat data reveals no instances where subjects violated these rules. 10. Note that the information feedback did not allow subjects to observe the demands of individual bidders. An alternative procedure, that would have perhaps been more favorable to collusion, would be to give information on the demands of individual bidders, and, in the chat treatments, attach these to chat identifiers. 11. At the time of the experiment £1 ≈ $1.45. 12. 99.3% of units auctioned were sold, so there is a tight link between price, seller revenue, and average bidder profit. In all cases of treatment effects (or noneffects) concerning price, there are corresponding effects (or noneffects) in terms of seller revenue and average bidder profit. Raw data, including chat logs, are available at www.nottingham.ac.uk/economics/cedex/papers/supplement/szdata.html 13. A Kruskal-Wallis test yields a p-value of 0.647. Unless otherwise noted, we report p-values from Kruskal-Wallis tests (equivalent to a two-sided Wilcoxon rank-sum test when two treatments are being compared). 14. Overall, out of 90 subjects 47 bid more aggressively. Of the other subjects, 21 bid less aggressively and 22 either bid the same or cannot be classified because they bid more aggressively at some prices and less aggressively at others. The pattern is similar across allocation rules. 15. The proportion of groups playing symmetric strategies understates the degree to which chat promotes agreements. Some groups agreed on asymmetric strategies that generated a symmetric outcome (e.g., each bidder agreed to bid for 8 units, resulting in a price of 0, but then the bidders submitted their bids at different prices). 16. We also classify some strategy combinations as Tacit Collusion, where 24 units are purchased at a price of 0 but the additional unit is ignored: for example, a common strategy in the experiment is that all bidders submit a single bid for 8 units (or 4 units in C6) at a price of 10. These bids do not form an equilibrium since a bidder could increase his or her allocation by one unit without raising the price. However, a bidder could only increase his or her allocation by more than this by raising the price to 10. 17. For example, suppose a group uses the symmetric Tacit Collusion strategy combination of bidding for 8 units at 10 (or 4 units at 10 with 6 bidders) and the remaining units at 0. This results in a payoff of 83.33 (or 41.67 with 6 bidders) per bidder. In SC3, this is an equilibrium and so the unilateral deviation gain would be 0. In the other treatments the most profitable unilateral deviation would be to bid for the remaining units at a price of 1. This increases the deviator’s payoff by 51.67 (UC3) or 24.67 (HC3), and by 3.33 (SC6), 81.11 (UC6), or 42.33 (HC6). 18. It should be noted that the difference in proportions of groups changing their strategies conditional on adopting a Tacit Collusion strategy in the last practice period (0/5 versus 4/10) is not significant according to Fisher’s exact test (two-sided p = 0.231). Note, however that there are relatively few observations of a given strategy type in any of our treatments, and so formal statistical tests may suffer from low power. For this reason we do not report formal statistical tests in the remainder of this section, and the results should be interpreted accordingly as suggestive. 19. Sade, Schnitzlein, and Zender (2006) also conduct experiments with uniform price auctions under the standard rule, allowing communication. They only report

The Role of Allocation Rules and Communication Among Bidders

77

on “perfect collusion,” in which all bidders bid the same amount at the lowest permissible price. This corresponds to our Single Bid strategy combination at a price of 0. In their experiment four out of thirteen groups follow this strategy, while in our STANDARD C3 and C6 treatments three out of nineteen groups follow this strategy.

ACKNOWLEDGMENTS We thank the coeditors and an anonymous referee, as well as conference participants at the Economic Science Association 2009 International Meeting, Washington D.C., and the 2009 International Conference on Game Theory, Stony Brook University, New York, for useful comments. We also thank the ESRC for financial support (PTA-026-27-1408).

REFERENCES Ausubel, L., & Cramton, P. (2004). Auctioning many divisible goods. Journal of the European Economic Association, 2, 480
493. Back, K., & Zender, J. (1993). Auctions of divisible goods: On the rationale for the treasury experiment. Review of Financial Studies, 6, 733
764. Brown, A., Plott, C., & Sullivan, H. (2009). Collusion facilitating and collusion breaking power of simultaneous ascending and descending price auctions. Economic Inquiry, 47, 395
424. Burton, A., & Sefton, M. (2004). Risk, pre-play communication and equilibrium. Games and Economic Behavior, 46, 23
40. Burtraw, D., Goeree, J., Holt, C., Myers, E., Palmer, K., & Shobe, W. (2009). Collusion in auctions for permission permits: An experimental study. Journal of Policy and Management, 28, 672
691. Devetag, G., & Ortmann, A. (2007). When and why? A critical survey on coordination failure in the laboratory. Experimental Economics, 10, 331
344. Engelmann, D., & Grimm, V. (2009). Bidding behavior in multi-unit auctions: An experimental investigation. Economic Journal, 119, 855
882. Fischbacher, U. (2007). z-Tree: Zurich toolbox for ready-made economic experiments. Experimental Economics, 10, 171
178. Fu¨llbrunn, S., & Neugebauer, T. (2007). Anonymity deters collusion in hard-close auctions: Experimental evidence. New Zealand Review of Economics, 43, 131
148. Goeree, J., Offerman, T., & Sloof, R. (2013). Demand reduction and preemptive bidding in license auctions. Experimental Economics, 16, 52
87. Goswami, G., Noe, T., & Rebello, M. (1996). Collusion in uniform-price auctions: Experimental evidence and implication for treasury auctions. Review of Financial Studies, 9, 757
785.

78

MARTIN SEFTON AND PING ZHANG

Greiner, B. (2004). An online recruitment system for economic experiments. In K. Kremer & V. Macho (Eds.), Forschung und Wissenschaftliches Rechnen (pp. 79
93). GWDG Bericht 63. Go¨ttingen: Gesellschaft. fu¨r Wissenschaftliche Datenverarbeitung. Isaac, R., & Walker, J. (1985). Information and conspiracy in sealed bid auctions. Journal of Economic Behavior and Organization, 6, 139
159. Kagel, J., & Levin, D. (2001). Behavior in multi-unit demand auctions: Experiments with uniform price and dynamic vickrey auctions. Econometrica, 69, 413
454. Klemperer, P., & Meyer, M. (1989). Supply function equilibria in oligopoly under uncertainty. Econometrica, 57, 1243
1277. Kremer, I., & Nyborg, K. (2004a). Divisible good auctions: The role of allocation rules. Rand Journal of Economics, 35, 147
159. Kremer, I., & Nyborg, K. (2004b). Underpricing and market power in uniform price auctions. Review of Financial Studies, 17, 849
877. Kwasnica, A. (1998). Collusion in multiple object simultaneous auctions: Theory and experiments. Caltech Social Science Working Paper 1010, California Institute of Technology, Pasadena. Kwasnica, A. (2000). The choice of cooperative strategies in sealed bid auctions. Journal of Economic Behavior and Organization, 42, 323
346. Kwasnica, A., & Sherstyuk, K. (2007). Collusion and Equilibrium Selection in auctions. Economic Journal, 117, 120
145. Li, J., & Plott, C. (2009). Tacit collusion in auctions and conditions for its facilitation and prevention: Equilibrium selection in laboratory experimental markets. Economic Inquiry, 47, 425
448. Porter, D., & Vragov, R. (2006). An experimental examination of demand reduction in multiunit versions of the uniform-price, vickrey, and English auctions. Managerial and Decision Economics, 27, 445
458. Sade, O., Schnitzlein, C., & Zender, J. (2006). Competition and cooperation in divisible good auctions: An experimental examination. Review of Financial Studies, 19, 195
235. Sherstyuk, K. (1999). Collusion without conspiracy: An experimental study of one-sided auctions. Experimental Economics, 2, 59
75. Sherstyuk, K. (2008). Some results on anti-competitive behavior in multi-unit ascending price auctions. In C. Plott & V. Smith (Eds.), Handbook of experimental economics results. Amsterdam: Elsevier. Smith, V. (1967). Experimental studies of discrimination versus competition in sealed-bid auction markets. Journal of Business Research, 40, 56
84. Wang, J., & Zender, J. (2002). Auctioning divisible goods. Economic Theory, 19, 673
705. Wilson, R. (1979). Auctions of shares. Quarterly Journal of Economics, 93, 675
689. Zhang, P. (2009). Uniform price auctions and fixed price offerings in IPOs: An experimental comparison. Experimental Economics, 12, 202
219.

The Role of Allocation Rules and Communication Among Bidders

79

APPENDIX The instructions below are for three-bidder treatments. Instructions for six-bidder treatments are similar. Text that differs across treatments with different allocation rules appears in square brackets and indicated with S (for STANDARD), U (for UNIFORM), or H (for HYBRID). Text that differs in treatments with or without the chat stage appears in braces and starts with C (for treatments with the chat stage) or NC3 (for treatments without the chat stage).

Instructions Welcome! You are about to take part in an experiment in the economics of decision making. You will be paid in private and in cash at the end of the experiment. You will be paid a participation fee of five pounds, plus an additional amount that will depend on your decisions. How this additional amount is determined is explained in these instructions, so please follow the instructions carefully. It is important that you do not talk to any of the other participants until the experiment is over. If you have a question at any time, raise your hand and a monitor will come to your desk to answer it.

Description of the Experiment The experiment will consist of {C: 12} {NC3: 21} periods. In each period you will be in a group of three bidders. At the beginning of the experiment, the computer will randomly form groups of three bidders from the participants in the room. A bidder ID, which is either A, B, or C, will be allocated to you randomly. You will be in a group with the same two other participants throughout the experiment. No participant will ever learn the identities of the other members of his group in any period. In each period you can earn points. The first {C: eleven} {NC3: twenty} periods will be practice periods in order to help you understand the experimental rules and the experimental environment. The points you earn in these periods will NOT count toward your final earnings. Your point earnings in period {C: twelve} {NC3: twenty-one} will be converted to British Pounds at the exchange rate of 1 point = 20 pence to determine your

80

MARTIN SEFTON AND PING ZHANG

additional earnings. You will be paid the participation fee and your additional earnings at the end of the experiment. {C: From period five there will be up to four minutes for communication at the beginning of each period. You will be able to communicate with the other two bidders through a message box on your computer screen. You are free to discuss any aspect of the experiment that you wish, except that: • You must not reveal any personal information (e.g., your name, where you are seated in the lab, where you live, your email address, phone number, etc.). • No threatening, insulting, or offensive comments are allowed. If you violate these rules your payment will be forfeited.}

Description of a Period In each period there are 25 units for sale in each group. You buy units by submitting bids. A bid is a price-quantity pair indicating how many units you bid for at which price. You can submit multiple bids, as long as the total number of units you bid for does not exceed 24 units. Thus, the most any bidder can bid for is less than the total number of units for sale. After all bidders have submitted their bids, the computer calculates a market price and allocates units among bidders. For each unit allocated to you, you receive 10 points and pay the market price. The market price rule and the allocation rule will be explained shortly. At the beginning of each period you will see a bidding box on your computer screen (see Fig. A.1). To place a bid you enter a price and the number of units that you bid for at that price in the corresponding boxes, then click on the “Place this Bid” button. A price must be an integer between 0 and 10. The number of units must be an integer between 0 and 24. After you click on the “Place this Bid” button, your bid will appear in the table on the right of your screen (see Fig. A.2). Repeat the above procedure if you want to place multiple bids. Your bids will be sorted by price from high to low in the table. The total number of units that you have bid for is shown above the table (see Fig. A.2). At any time before the final submission of bids, you can withdraw a bid that you have placed. To do this click on the bids that you wish to withdraw from the table (the bids then will be highlighted), then click on the

The Role of Allocation Rules and Communication Among Bidders

81

“Withdraw a Bid” button. If you want to withdraw all bids that you have placed, simply click on the “Clear All Bid(s)” button. When you are ready, click on the “Final Submission” button to submit your bids. After all members in your group have finally submitted their bids, the computer calculates the market price and allocates units among your group. You will then be informed of the bids submitted by your group, the market price, your allocation, and your profit for the period. The period then ends and the next period begins.

Market Price and Allocation Rules If the total number of units your group bids for is less than 25, the market price will be 0 and each group member will obtain the number of units he or she bid for. [S: If the total number of units your group bids for is greater than or equal to 25, the market price will be the highest price at which all 25 units can be sold. Bids placed above the market price will be fully fulfilled. The remaining units will be allocated proportionately among the bid(s) placed at the market price.] [U: If the total number of units your group bids for is greater than or equal to 25, the market price will be the highest price at which all 25 units can be sold. The 25 units will be allocated proportionately among all winning bids, that is, bids placed at or above the market price.] [H: If the total number of units your group bids for is greater than or equal to 25, the market price will be the highest price at which all 25 units can be sold. Two methods will then be used to allocate the units. According to the first method, the 25 units will be allocated proportionately among all winning bids, that is, bids placed at or above the market price. According to the second method, bids placed above the market price will be fully fulfilled, and the remaining units will be allocated proportionately among the bid(s) placed at the market price. Your allocation in each period will be the average of the allocation based on the two methods.]

82

MARTIN SEFTON AND PING ZHANG

Fig. A.1.

Fig. A.2.

83

The Role of Allocation Rules and Communication Among Bidders

We use three examples to demonstrate how the market price and allocation rules work. The examples are based on a table that is similar to the table that will be displayed on your screen at the end of each period. However, in the examples prices range up to 30, while in the experiment prices range between 0 and 10. Also, the bids in the experiment will be determined by the participants’ choices. In all examples the column headed “Units Group Bids for” is the sum of the columns headed “Units You Bid for” and “Units Other Group Members Bid for.” The column headed “Accumulated Units Group Bids for” shows the number of units bid for by your group at and above the corresponding price. The line corresponding to the market price will appear in blue on your screen. Example 1: Price

Units You Bid for

Units Other Group Members Bid for

Units Group Bids for

Accumulated Units Group Bids for

6 0 5 0

0 10 3 0

6 10 8 0

6 16 24 24

30 26 23 0

In Example 1, the total number of units that your group bids for is 24, which is less than 25. Thus the market price is 0 and you obtain the total number of units you bid for, which is 11 units. Example 2: Price 30 26 23 0

Units You Bid for

Units Other Group Members Bid for

Units Group Bids for

Accumulated Units Group Bids for

6 0 5 0

0 10 4 0

6 10 9 0

6 16 25 25

Compared with Example 1, the only difference in Example 2 is that at price 23 the other group members bid for 4 units. The highest price at which all 25 units can be sold is 23. Thus the market price is 23. [S: Bids placed above the market price are fully fulfilled. Thus, you are allocated 6 units and the other group members are allocated 10 units. This leaves 9 units to be allocated proportionately among the bids

84

MARTIN SEFTON AND PING ZHANG

placed at the price of 23. Since the group bids for 9 units at a price of 23, each bid is fully fulfilled. Thus you are allocated an additional 5 units. In total you are allocated 11 units.] [U: The units are allocated to all winning bids, that is, bids placed at or above the market price. There are 25 units available and the group bids for 25 units at or above the market price. Hence each unit bid for at or above the market price is fully fulfilled. Thus you obtain 11 units.] [H: Two methods are used to allocate the units. According to the first method, the units are allocated to all winning bids, that is, bids placed at or above the market price. There are 25 units available and the group bids for 25 units at or above the market price. Hence each unit bid for at or above the market price is fully fulfilled. Thus your allocation following the first method is 11 units. According to the second method, bids placed above the market price are fully fulfilled. Thus you are allocated 6 units and the other group members are allocated 10 units. This leaves 9 units to be allocated proportionately among the bids placed at the price of 23. Since the group bids for 9 units at a price of 23, each bid is fully fulfilled. Thus you are allocated an additional 5 units. Your allocation following the second method is 11 units in total. Your final allocation is the average of the allocation based on the two methods, that is, (11 + 11)/2 = 11 units.] Example 3: Price 30 26 23 0

Units You Bid for

Units Other Group Members Bid for

Units Group Bids for

Accumulated Units Group Bids for

6 0 5 0

0 16 4 0

6 16 9 0

6 22 31 31

Compared with Example 2, the only difference in Example 3 is that at price 26 the group bids for 16 units. The highest price at which all 25 units can be sold is 23. Thus the market price is 23.

The Role of Allocation Rules and Communication Among Bidders

85

[S: Bids placed above the market price are fully fulfilled. Thus you are allocated 6 units and the other group members are allocated 16 units. This leaves 3 units to be allocated proportionately among the bids placed at the price of 23. Since the group bids for 9 units at a price of 23, a bidder receives 3/9 of a unit for each unit he or she bids for at this price. Thus, you are allocated an additional 5 × 3/9 = 1.667 units. In total you are allocated 7.667 units.] [U: The units are allocated proportionately among all winning bids, that is, bids placed at or above the market price. There are 25 units available and the group bids for 31 units at or above the market price. Hence a bidder receives 25/31 of a unit for each unit he or she bid for at or above the market price. Thus you are allocated 11 × 25/31 = 8.871 units.] [H: Two methods are used to allocate the units. According to the first method, the units will be allocated proportionately among all winning bids, that is, bids placed at or above the market price. There are 25 units available and the group bids for 31 units at or above the market price. Hence a bidder receives 25/31 of a unit for each unit he or she bid for at or above the market price. Thus your allocation following the first method is 11 × 25/31 = 8.871 units. According to the second method, bids placed above the market price are fully fulfilled. Thus you are allocated 6 units and the other group members are allocated 16 units. This leaves 3 units to be allocated proportionately among the bids placed at the price of 23. Since the group bids for 9 units at a price of 23, a bidder receives 3/9 of a unit for each unit he or she bid for at this price. Thus you are allocated an additional 5 × 3/9 = 1.667 units. Your allocation following the second method is 7.667 units in total. Your final allocation is the average of the allocation based on the two methods, that is, (8.871 + 7.667)/2 = 8.269 units.]

Your Earnings In each period you receive 10 points for each unit allocated to you and pay the market price for each unit allocated to you. Thus, your profit in points in a period is: Your profit = ð10 − market priceÞ × number of units allocated to you

86

MARTIN SEFTON AND PING ZHANG

Remember, the first 11 periods are just for practice. The points you earn in these periods will not affect your final cash earnings. Your profits from the last period, period {C: 12} {NC3: 21}, will determine your cash earnings. Your profit in period {C: 12} {NC3: 21} will be converted into British Pounds at a rate of 20 pence per point.

Beginning the Experiment We are now ready to begin the experiment. Please look at your screen and follow the prompts. If you have a question at any time please raise your hand.

CHAPTER 5 THE DIVIDEND PUZZLE: A LABORATORY INVESTIGATION Sascha Fu¨llbrunn and Ernan Haruvy ABSTRACT Purpose
We investigate the implications of the misalignment between manager and shareholder interests and the effects of initial ownership stakes and reinvestment of unpaid dividends on managerial self-dealing. Methodology
We collect and analyze data from controlled laboratory experiments with an experimental setting which captures the role of ownership in managerial considerations. Findings
We see the emergence of both investor-aligned outcomes and managerial self-dealing outcomes. We find that increasing managers’ initial endowment of shares makes it harder for managers to coordinate on an outcome and lowers return on investment. Moreover, allowing managers to reinvest unpaid dividends results in a transfer of wealth to management. Research limitations
The results and the conclusions are drawn upon data from the particular setting we investigate. Generalizing them beyond the specific setting should be done with caution.

Experiments in Financial Economics Research in Experimental Economics, Volume 16, 87
110 Copyright r 2013 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0193-2306/doi:10.1108/S0193-2306(2013)0000016005

87

88

SASCHA FU¨LLBRUNN AND ERNAN HARUVY

Practical implications
Higher managerial ownership stake means that managers have a greater incentive to reward shareholders, but we find that it may also imply a more difficult coordination problem between managers
sometimes to the detriment of shareholders. Originality
This study is the first to consider the direct relationship between managers’ portfolios and voting decisions regarding dividends and investment levels. Keywords: Dividend policy; corporate finance; experiment

INTRODUCTION Why do firms pay dividends? If a firm has excess cash, this cash could be used to pay off debt, to invest in future productive capacity, or it could simply be kept for unanticipated future needs. As long as the firm acts to maximize shareholder value, any undistributed cash should be immediately factored into the price per share, and shareholders should be at least as well off in comparison to getting the cash payment. The general answer is that managers might not use the cash in the shareholders’ best interest. This conflict of interest is called the agency problem between shareholders and management (e.g., Rozeff, 1982). The literature highlights the opposing interest of management and shareholders in terms of dividend payout. On the one hand, shareholders seek more dividends to reduce agency costs, also known as the free cash flow problem (Jensen, 1986). On the other hand, managers may seek to maximize their own payoffs which may depend on firm performance. Empirically, the evidence is mixed. In some cases, management or insider ownership is negatively correlated with dividend payout (Truong & Heaney, 2007)
consistent with greater ability of managers to reinvest earnings. In other cases, management ownership can be shown to be positively correlated with dividends, which may be due to better alignment of managers and shareholder preferences. The premise of the present work is that dividend payouts and investment decisions are the result of interaction between managers and investors. The theoretical and empirical literatures indicate that dividends may be useful in addressing the agency problem
the misalignment between stakeholders. Stakeholders may include shareholders, managers, board members, lenders,

The Dividend Puzzle: A Laboratory Investigation

89

institutional investors, private investors, old investors, and new investors (Easterbrook, 1984). While the literature has been somewhat noncommittal on what that misalignment might be, the literature has been able to link different stakeholder compositions with different outcomes. These different stakeholders have complex and diverse sets of considerations. The misalignment problem has been generally recognized in the finance literature and formed the bases of the literature on managerial compensation (Hall & Liebman, 1998; Hall & Murphy, 2000). Specifically, the goal of that literature has been to better align managerial incentives with those of shareholders and thus reduce agency costs and increase efficiency. We add an extra layer of interaction by positing that it is not enough to align incentives between a key decision maker and shareholders, but rather a sufficient number of decision makers have to be aligned with shareholders and these decision makers have to form beliefs about other decision makers’ intents regarding future dividend payouts. We use the term “managers” to refer to decision makers and this is done for ease of exposition with no intent to imply anything about corporate decisionmaking structure. This is a lab investigation and therefore we make simplifying abstractions about management, while isolating the key variable we mean to investigate
namely dividend payouts and investment decisions. We find that laboratory subjects in the role of managers behave largely in their own self-interest and thus their individual choices in a given period depend on the alignment between managerial incentives and shareholder interests. However, that self-interest is not in itself sufficient to achieve coordination on a particular outcome. It appears that the managers find it difficult to coordinate among themselves when there are multiple equilibrium outcomes. The coordination is a dynamic problem which critically depends on initial allocation of shares to decision makers. When decision makers are endowed with sufficient shares to align their interests with those of shareholders, they are less able to send signals to other managers regarding their intentions, thus making coordination more difficult. This is paradoxical because in a way it suggests that initial endowments that align managers’ incentives with investors’ interests may have negative repercussions in the future because they hamper the ability of managers to signal intent through market activities. This chapter is structured by first discussing the literature leading to the current problem, then the experimental design, results, and analysis. Implications for compensation are discussed in the conclusions.

90

SASCHA FU¨LLBRUNN AND ERNAN HARUVY

LITERATURE REVIEW The costs attributed to the misalignment between managers’ and shareholders’ interests are collectively referred to as agency costs. Agency costs also include the costs of putting in place a monitoring system and thus incurring monitoring costs to keep management in check. Rozeff (1982, p. 250) suggests that a “wealth-maximizing firm adopts an optimal monitoring package which acts to reduce agency costs.” Jensen (1986) argues that a firm is likely to share more of its profit with investors when it faces lower monitoring costs. Some argue that dividend policy is a manifestation of the extent to which agency conflicts between shareholders and managers exist within the firm (Easterbrook, 1984). One documented class of misalignment-related activities involves expansion activities which do not benefit investors. Jensen (1986) cites food industry and broadcasting industry mergers as a clear evidence of such a phenomenon. Jensen (1986) termed this inefficient allocation of firm resources the free cash flow hypothesis and argued that a high cash balance induces managerial inefficiencies. According to Jensen, paying dividends and financing projects through debt prevents firms from wasting resources on low-return projects. The free-cash flow theory implies “managers of firms with unused borrowing power and large free cash flow are more likely to undertake low-benefit or even value-destroying” activities. A related misalignment is that management may wish to increase managerial pay and perks and this desire could eat into profits. Excessive bonuses are often a subject of shareholder disgruntlement as well as company perks like private jets. This is known as management self-dealing (as opposed to inefficiencies, e.g., Fu¨llbrunn & Haruvy, 2011; Oprea, 2008), implying that managers pursue activities that benefit them directly at the expense of shareholders. In our investigation, the role of the “managers” is an abstraction for a wide array of decision makers. In the empirical literature, there are three major classes of decision makers: the management class, institutional owners, and external directors. We discuss the evidence pertaining to each class. The management class includes top management, officers, executives, and board members. The evidence in the literature suggests that greater inside ownership is correlated with lower dividend payment (Jensen, Solberg, & Zorn, 1992; Rozeff, 1982; Short, Zhang, & Keasey, 2002; Truong & Heaney, 2007). This is entirely consistent with the agency explanation for dividends. Since managers have better information and therefore lower agency costs, there is less advantage to dividend payment as a means

The Dividend Puzzle: A Laboratory Investigation

91

of reducing agency costs. It is important to emphasize that the agency costs are reduced for the manager-owners with regards to firm decisions. The agency costs may actually increase for external investors who may find themselves at a disadvantage regarding surplus division between managerowners and external investors. Institutional owners may have board representation or otherwise have influence on the direction of the company and its management. Institutional ownership can affect dividends in one of two ways. On the one hand, institutions can squeeze management for greater dividends. On the other hand, institutional owners reduce the agency problem and are therefore similar to management in that regard. Since the class of institutional owners represents a broad category of institutions which in turn represent different types of investors with different interests, the evidence on the effect of institutional ownership on dividend payouts is mixed. Short et al. (2002) find that institutional ownership is helpful in increasing dividend payout. Truong and Heaney (2007) find the opposite effect. External directors may serve to reduce the agency cost, thus lowering the need for dividends (Bathala & Rao, 1995). Or they may serve to represent the interests of those stakeholders who wish for higher dividends (e.g., Belden, Fister, & Knapp, 2005; Kaplan & Reishus, 1990; Schellenger, Wood, & Tashakori, 1989). Thus, the evidence for this class is mixed as well. In summary, the power, ownership, and concentration of different decision makers in the firm clearly affect decisions and outcomes which in turn impact all stakeholders in the firm. The extant empirical evidence points to loss of efficiency being the result of a misalignment between decision makers and investors, which leads to an agency problem. The idea that asset market activities and the resulting portfolios of various market actors can affect the outcomes and efficiency of subsequent coordination games has been previously proposed and investigated in the laboratory by Kogan, Kwasnica, and Weber (2011). In that study, experimental asset markets preceded a coordination game with Pareto ranked outcomes. As in the current study, Kogan et al. (2011) find a negative effect of asset market trading on subsequent coordination efficiency and documented that portfolio positions of some market actors were responsible for this loss of efficiency. As in the current experiments, Kogan et al. (2011) allowed for a small (or large in some treatments) insider group (akin to managers in our setting) to determine the input levels and thus the production outcome of the coordination (firm’s input decision) game. Kogan et al. (2011) thus captured the idea that a small group of managers’ portfolio

SASCHA FU¨LLBRUNN AND ERNAN HARUVY

92

decisions could be responsible for a company’s activity level and output and that this affects all stakeholders and results in possible efficiency loss. Their study captures some of the same elements and concepts of the study reported here. While their study and the present study were conceived of and conducted independently, there are certainly overlaps in the general conceptual underpinnings. Nevertheless, our settings have direct financial parallels to dividend decisions and investment decisions that we think nicely complement the findings of Kogan et al. (2011). We particularly focus on the direct relationship between managers’ portfolios and voting decisions regarding dividends and investment levels.

EXPERIMENTAL STUDY Study 1 The purpose of Study 1 is to study the joint managerial decision regarding dividend payout as an outcome of a coordination problem. We assume opposing interests between managers and shareholders. This assumption is captured in our setting as a simple division of surplus between management and shareholders, which is jointly decided upon by the management team in a majority vote. Following a period of trading, three managers vote on whether to award dividends to shareholders. Shareholders consist of managers and external investors who own shares but have no voting power. The majority vote of the managers determines the payoffs to managers, who may or not be shareholders, and externals investors. A decision to pay dividend results in a dividend of 12 francs (the experimental currency) to each share. A majority vote to withhold dividends (a vote to “Not pay”) results in a payoff of 16 francs to each manager and no dividends for shares (Table 1). This represents a 50% efficiency loss in accordance with the free cash flow principle. provides a fast overview on the voting situation. Table 1. Majority Vote Pay Dividend Not pay

Payoffs from Voting in Study 1.

Payoff to Outside Investors 12 francs per share in the current period 0 francs per share in the current period

Payoff to Managers 12 francs per share in the current period 16 francs per manager in the current period

93

The Dividend Puzzle: A Laboratory Investigation

In the absence of managerial ownership stake, the surplus division decision on the part of managers would be simple
self-interested managers would vote to withhold dividends. However, managers may have some ownership stake. The dilemma facing manager-owners is as follows: on the one hand, if the manager owns a sufficient number of shares, he would like to reward shareholders with a dividend payment. On the other hand, the manager must form beliefs about the intentions of other managers to own sufficient shares toward the same goal. If the manager does not believe that other managers intend to own sufficient shares, then regardless of his/ her own ownership position, the manager should sell his own shares and vote against shareholder interests. Each treatment consists of four subsequent markets with five periods each. Shares and cash for each participant are reset in Period 1 of each market, although earnings are cumulative. Each period includes a double auction stage which lasts for 90 seconds and a voting stage. We run two treatments, named treatment LOW and treatment HIGH. The two treatments differ only on initial endowments of shares to managers and external investors. Treatment LOW has managers starting with zero shares and investors starting with all eight shares in Market 1. The managers’ endowment of shares increases over markets until in the last market, managers begin with five shares. Treatment HIGH has managers starting with five shares and that endowment decreases to zero shares by Market 4. A summary of the endowments in the two treatments is shown in Table 2. Each cell has three numbers, separated by commas, representing the number of shares each of the three managers and three investors begins with in Period 1 of a given market. The cash endowment was adjusted according to the share endowment, where participants starting with zero shares were endowed with 270 francs, participants starting with one share were endowed with 210 francs, participants with two shares were Table 2. Treatment Participants Market 1 Market 2 Market 3 Market 4

Endowments of Shares and Cash under the Two Endowment Treatments. Treatment LOW

Treatment HIGH

Managers 1, 2, 3

Investors 1, 2, 3

Managers 1, 2, 3

Investors 1, 2, 3

0,0,0 1,1,1 1,2,2 2,2,1

3,3,2 1,2,2 1,1,1 1,1,1

2,2,1 1,2,2 1,1,1 0,0,0

1,1,1 1,1,1 1,2,2 3,3,2

94

SASCHA FU¨LLBRUNN AND ERNAN HARUVY

endowed with 150 francs, and participants with three shares were endowed with 90 francs.

Study 2 The task in Study 1 involved a division of a pie between managers and external investors. In Study 2, we relax the fixed pie assumption and consider a different type of misalignment between managers and owners. As in Study 1, each treatment consists of four subsequent markets with five periods of two stages each. There are six subjects divided equally into the fixed roles of managers and investors
three managers and three investors. There are five periods in a market and four markets in total. Each market is a double auction that is open for 90 seconds. The study involved two treatments with endowments exactly as in Study 1. In the scenario of Study 2, the managers can invest current earnings in a productive activity which translates to higher future income. In addition to investing, managers may also self-deal or award dividends, as in Study 1. The investment activity is in the managers’ self-interest if the managers intend to own shares in the future. This makes for a difficult coordination decision for managers because they have to form longer term expectations about the intent of other managers for their future ownership position. Paying dividends is a dominated alternative relative to investment in terms of earnings for investors. However, dividends could serve to effectively reduce the agency problem. Without dividends, investors cannot tell whether managers are investing the earnings or awarding the earnings to themselves. This study involves managers voting for a choice among three alternatives: (1) Pay Dividend, (2) Invest, and (3) Not Pay and Not Invest. A majority vote to Pay Dividends pays 12 francs for each share in the current period. A majority vote to Invest pays 15 francs for each share, but this payment is made only at the end of the market and only to the future owners of these shares. A majority vote to Not Pay and Not Invest pays 16 francs to each manager in the current period and pays nothing to the external investors. Externals investors are only informed whether the managers vote to Pay Dividends or not. They are not informed about the investment decision. The payoff table is shown below (Table 3). To get a better sense of the decision facing managers, note that a consistent majority vote to self-deal (Not Pay and Not Invest) earns a total surplus of 5 periods × 3 managers × 16 francs = 240 francs, and this surplus

95

The Dividend Puzzle: A Laboratory Investigation

Table 3. Majority Vote Pay Dividend Invest Not Pay and Not Invest

Payoffs from Voting in Study 2.

Payoff to Outside Investors 12 francs per share in the current period 15 francs per share in the final period 0 francs per share in the current period

Payoff to Managers 12 francs per share in the current period 15 francs per share in the final period 16 francs per manager in the current period

is divided between managers only, with nothing for external investors, yielding 80 francs per manager. A majority vote to Invest provides 5 investments × 8 shares × 15 francs = 600 francs, which is to be divided in the final period according to share ownership, yielding 75 francs per share. A majority vote to pay a dividend provides 5 periods × 8 shares × 12 francs = 480 francs, which to be divided according to share ownership, yielding 60 francs per share. Thus, we see that the decision to pay dividends is strictly dominated for managers in terms of earnings, but it is informative to outside investors. The decision of which action to vote for depends on the manager’s ownership stake.

THEORETICAL DEVELOPMENT Study 1 A risk-neutral manager would vote to Pay Dividends if the number of shares he owns is at least two. Given the high endowment setting and common expectations, managers would sell their shares (from here on, s denotes the number of shares owned by a manager; t denotes the number of periods in a market) if the future payoff from self-dealing 16(5 − t) plus the price received (p) exceeds profits from dividends, that is, if 16(5 − t) +sp > 12(5 − t)s or if the price exceeds (5 − t)(12s − 16)/s. However, since investors know that buying a high number of shares from managers would lead to a zero dividend policy they would not do so. In the low endowment setting, with common expectations, managers would buy shares if the future payoff from dividend payments 12(5 − t)s minus the price paid exceeds profits from self-dealing, that is, if 12(5 − t)s − sp > 16(5 − t) or if the price falls below (5 − t)(12s − 16)/s. But this argument depends on the

96

SASCHA FU¨LLBRUNN AND ERNAN HARUVY

inventory share of the other managers since managers would need to coordinate on a single strategy.

Study 2 A risk-neutral manager would vote to Invest if the number of shares he owns is at least two. A vote to Pay Dividends is always dominated. Given the high endowment situation and common expectations, managers would sell their shares if 16(5 − t) plus the price received exceeds the return from the final return 15t~ þ 15ð5 − tÞ, where t~ equals the number of decisions to invest. However, since the investors know that buying a sufficient number of shares from the managers would lead to zero investments in the future they would not buy and keep their money and/or receive dividends given they have shares. They would not even sell their shares, since so far it is not clear whether the managers voted to Invest or Not Invest. Hence, trades should be merely observed. Given the low endowment situation and common expectations, managers have low incentives to buy shares, because without investment decisions no final return will be paid. But, all of this thinking depends on the share inventory of the other managers since they need to coordinate on one strategy. We begin by noting that regardless of expectations, self-interested managers should vote to award or withhold dividends in a manner consistent with their selfish best interest. While we cannot rule out managers voting against their interest in an attempt to signal some future action, there is no obvious signaling strategy. It is also possible that pivotal voting managers will deviate in order to dislodge the group from one outcome and we will examine this shortly. We state the individual rationality conjecture as H1. H1. Rational individual myopic best response. In Study 1, a self-interested manager votes to Pay Dividends if he has two or more shares and votes to Not Pay Dividends (self-deal) if he has one or fewer shares. In Study 2, a manager votes to Invest if he has two or more shares and to Not Pay and Not Invest (self-deal) if he has one or fewer shares. Table 4 summarizes the differences between the two studies in terms of ownership conditions required for managers to favor each of the available outcomes. We next look at the ability of the group, through a majority vote, to coordinate on the myopic best response for the majority of the group’s members. Naturally, one would expect that individual best response would

97

The Dividend Puzzle: A Laboratory Investigation

Table 4.

The Correspondence between Manager’s Ownership Stake and Favored Outcome.

Favored Outcome Pay Dividend Not Pay and Not Invest (self-deal) Invest

Study 1

Study 2

# Shares ≥ 2 # Shares ≤ 1 Not available

Dominated # Shares ≤ 1 # Shares ≥ 2

translate directly to group best response. However, this is where a breakdown can occur. Individual group members who hold the pivotal vote may find it desirable to deviate from their myopic individual best response in order to dislodge the group from one outcome to the next. Thus, we examine group best response as a separate hypothesis. H2. Rational group voting behavior. The group voting outcome corresponds to the majority’s self-interested myopic best response. The next hypothesis is potentially the most critical so far. It suggests that the outcome in this setting is a result of an implicit coordination between managers. A path-dependent coordination, by which managers condition their actions on the trading of managers in the most recent period, is a strong evidence for coordination. Path-dependent actions imply that managers who observe other managers increasing their ownership stake would expect those other managers to vote for dividends, therefore making buying shares a more attractive proposition. H3. Path-dependent coordination. Managers’ buying/selling decisions are positively correlated to the other managers’ buying/selling in the most recent period. In the following hypotheses, we suggest that the investor’s willingness to pay (WTP) depends on recent decisions and market outcomes. They reflect the responses of the investor to the previous period’s result. H4. Investor WTP is responsive to past dividend decisions. The insider’s WTP should reflect the expectation about future dividends. However, only past dividend payments may signal the manager’s WTP dividends in the future. H5. Investor WTP is responsive to past share prices. Higher prices in previous periods reflect trust of the shareholders in the manager’s WTP dividends in the future. Thus, a correlation between recent prices and the current WTP should be found in the data.

98

SASCHA FU¨LLBRUNN AND ERNAN HARUVY

H6. Investor WTP is responsive to managers’ most recent ownership stake. Given the self-interest of the managers, only if at least two managers have more than one share a dividend payment can be expected. Thus, the investor’s WTP should reflect the manager’s incentive to pay dividends. The remaining hypotheses allow discussing the impact of the initial endowment on the ability to coordinate on a certain dividend policy. H7. The return from investment is influenced by initial endowment. Only if managers have sufficient ownership stake they vote to Invest and, thus, hold shares until maturity. Since a low initial share ownership increases the incentives to self-deal
at least at the beginning
the initial endowment should have an effect on the return from investment. H8. Initial endowments affect final dividend outcomes. Given common expectations, managers in the low share endowment treatment should vote to Not Pay and to get rid of all their shares for a positive price. In this case they end up with zero shares. In the high endowment treatment, managers should vote to Pay Dividends and therefore end up with a positive number or shares.

RESULTS The experiments were conducted at the University of Texas at Dallas with students from the School of Management who had never participated in an asset market experiment before. The experiment was computerized and programmed using zTree software (Fischbacher, 2007). In total, we collected data from 56 subjects arranged in 7 independent groups. Participants earned on average $25 for a one-and-a-half-hour session. Instructions were read aloud and participants were trained in the market facility before the experiment started.

MANAGERIAL BEHAVIOR Support for individual rationality (H1): In Study 1, 93% of the managers’ choices corresponded with the individual best reply to one’s current period

The Dividend Puzzle: A Laboratory Investigation

99

holdings. We find that 35 out of 42 managers selected the best reply in at least 90% of the cases (18/20) and this is significant by the binomial test (one-sided, p < 0.00001).1 According to the Fisher’s Exact test, there is a significant correlation between the decision to pay dividend and the incentive to do so (p < 0.001). In Study 2, which is far more complex due to the greater reliance on long-term forward-looking expectations, 70% of the managers’ choices corresponded with their myopic individual best response. According to the Fisher’s Exact test, there is a significant correlation between the vote to Invest and the myopic incentive to do so (p < 0.001). However, we find that none of the managers selected the myopic individual best response decision in at least 90% of the cases. This difference in Study 2 from Study 1 is not surprising given the greater role of long-term intentions and expectations. Support for group rationality (H2): In study 1, voting groups ended up with outcomes consistent with the majority’s myopic best response in 92% of the cases. According to the Fisher’s Exact test, there is a significant correlation between the decision to pay dividends and the incentive to do so (p < 0.001). We find that in 12 out of 14 sessions, the majority vote was in the majority’s self-interest in at least 90% (18/20) of the cases (one-sided p = 0.006).2 In Study 2, voting groups selected the majority’s myopic best response in 68% of the cases. According to the Fisher’s Exact test, there is a significant correlation between the decision to Invest and the incentive to do so (p < 0.001). However, in none of the sessions was the majority vote in the majority’s myopic best-interest in more than 90% of the cases. Support for path-dependent coordination (H3): The regression in Table 5 shows the relationship between the number of a participant’s purchases Table 5.

Random Effects Regression, Dependent variable: Number of purchases in period t.

Market


0.012 (0.032)

Market period Purchases of others lag Study 2 dummy Intercept R2


0.031 (0.032) 0.086** (0.023) 0.052 (0.079) 0.168 (0.145) 0.016

Standard errors in parentheses. **p < 0.05, *p < 0.1.

100

SASCHA FU¨LLBRUNN AND ERNAN HARUVY

in period t (negative numbers represent a sale) and the number of all other participants’ purchases in the previous period (Purchases of others lag). Market and Market Period capture the time trend. The dependent variable is the number of purchases in period t. The coefficient on purchases in period t − 1 by others is 0.086. It is significantly positive at p < 0.001.

INVESTOR BEHAVIOR We find mixed support for the hypotheses about investors’ WTP(H4
H6). The dependent variable in the random effects regression shown in Table 6 is investor WTP, that is, the median bid in the respective period. Market and Market Period capture the time trend. The variable Pay dividend lag takes on the value 1 when the voting in the former period resulted in dividend payment and is 0 otherwise. The coefficient for this variable is not significant at the 10% level under any specification. Thus, we find no statistical support for H4. That is, investors do not appear sensitive to managers’ dividend decisions in evaluating the value of shares. This may be an indication of bounded rationality on the part of investors or it may be an indication that expectations may not be driven by current managerial actions. Median price lag is the median price in the previous period. We see from the regression results that investors are highly sensitive to the previous period’s median price in determining valuations in the current period. In other words, price exhibits strong inertia. Thus, we find strong support for H5. The variable Should pay lag is 1 when it is the majority of the manager group’s best interest is to vote for dividend payment; that is, at least two managers own at least two shares each. We find no statistical significance for the coefficient estimated for this variable. Thus, we have no statistical support for H6. Investors do not appear sensitive to managers’ ownership stake. The results show that experience with markets has no effect on the WTP in either study. Time within markets has a negative effect on the WTP in Study 1. This is due to the fact that the number of potential dividend payments decrease over time. However, this effect cannot be observed in Study 2. This may be due to the fact that investment decision allow for a final payment after the last period.

Standard errors in parentheses. **p < 0.05, *p < 0.1.

Observations Number of Sessions r2 overall

Constant

Shouldpay_lag

Medianprice_lag

Paydividend_lag

47.12** (4.221) 280 14 0.0135

0.209 (0.590) −1.405** (0.466)

Market

Market period

1

Study

(1)

45.41** (6.363) 120 6 0.00374

0.845 (0.660) 0.006 (0.522)

2

(2)

(4.423)

(1.630)

43.75** (6.472) 96 6 0.00130

0.862 (0.739) 0.430 (0.739) −1.416

−0.442 (0.622) −2.410** (0.619) 0.0384

52.76** (4.760) 224 14 0.0260

2

(4)

1

(3)

25.20** (4.015) 224 14 0.580

24.04** (5.983) 96 6 0.540

0.322** (0.0692)

0.490**

0.751 (0.707) 0.455 (0.707)

2

(6)

(0.0492)

−0.448 (0.566) −2.046** (0.567)

1

(5)

52.27** (4.752) 224 14 0.0272

43.67** (6.627) 96 6 0.00930

0.256 (2.014)

1.202

0.883 (0.735) 0.387 (0.750)

2

(8)

(1.602)

−0.435 (0.618) −2.406** (0.618)

1

(7)

Table 6. Random Effects Regression, Dependent Variable: Investors’ WTP (median bid in Period t).

The Dividend Puzzle: A Laboratory Investigation 101

SASCHA FU¨LLBRUNN AND ERNAN HARUVY

102

THE EFFECT OF INITIAL ENDOWMENT The initial number of shares allocated to managers should have an influence on the voting for two reasons: (1) it affects the managers’ best-interest given the endowment and (2) it affects the potential signaling by managers through their transactions. Thus, one might speculate that the initial endowment would affect both the return to the managers through their ability to coordinate decisions, and the convergence outcome of the market. We find evidence for the first conjecture but not for the second. Support for effect of initial endowments on return on investment (H7): In Table 7, the dependent variable in the random effects regressions is the return on investment, that is, the final payout for each share after the final period. Market captures the time trend and the other variable is the sum of the manager’s initial stock. The results show that the initial share endowment has an effect on the return: a higher endowment of shares decreases the return from shares. This is also confirmed with the spearman rank coefficient which is 0.3685 (with a p-value of 0.080). Figure 1 shows no differences across starting conditions. Support for the effect of initial endowment on dividend outcome (H8). Regression analysis finds no statistically significant effect of initial endowment on the manager’s final number of shares, the final voting outcome, or final prices. Table 8 shows the final share holdings of all managers in a market and the frequency of self-dealing outcomes3 at the end of the market. Managers own on average 4.57 shares at the end of the market in the low endowment setting (LOW) of Setup 1. In the high endowment setting of setup 1 (HIGH), managers own on average 4.25 shares at the end of the market. In the low endowment setting of setup 2 (LOW), managers hold on average 5.3 shares at the end of the market. In the high endowment setting of setup 2 (HIGH), managers own 4.75 shares at the end of the market. We find no significant difference across treatments. We therefore pool the data in the tests that follow. Using a test of proportion across Table 7.

Linear Regression. Dependent Variable: Investment Payout in the Final Market Period.

Market Sum of managers’ initial shares Intercept R2 Std. errors in parentheses, ** p < 0.05, *p < 0.10.

1.250 (3.315)
3.843** (1.811) 58.74** (12.54) 0.115

103

The Dividend Puzzle: A Laboratory Investigation 70 60

Taler

50 40 30 20 10 0 000

111

122/221

000

LOW

111

122/221

HIGH

Fig. 1. Return on Investment in the Final Market Period. The bars show the average return on investment in the final market period, separated by starting LOW and starting HIGH, and by the initial market endowment with no shares for the managers (000), one for each manager (111) or 2 for one manager and 1 for two managers (122/221).

Table 8.

Managers’ Final Share Ownership. Number of Managers Shares in Period 5

LOW (1.5/100%) HIGH (5/0%)

Frequency of Self-Deal Situations

Setup 1

Setup 2

Setup 1

Setup 2

4.57 (1.5) 4.25 (5)

5.3 (1.5) 4.75

54% 71%

50% 58%

LOW: Managers’ initial share endowment is either 0,0,0 or 1,1,1, HIGH: Managers’ initial share endowment is either 1,2,2 or 2,2,1, Self Deal Situation: if at least two managers have one or fewer shares, average of initial conditions in brackets.

endowment classes we find the frequency of self-dealing outcomes in HIGH settings to be higher than in LOW settings. This means that managers on average buy shares in LOW settings (+3.3 shares) and sell shares in HIGH settings (
0.6 shares). We find the managers’ final number of shares to be higher in comparison to their initial endowments in LOW settings (p < 0.001) and lower in HIGH settings (p = 0.080). Without any transactions, the number of markets with self-dealing voting outcomes should equal 0/40 in the LOW settings and 40/40 in the HIGH settings. However, in LOW settings, the number increases from 0 in Periods 1 to 19 in the final period and decreases from 40 to 13 in HIGH. Hence, in LOW settings,

104

SASCHA FU¨LLBRUNN AND ERNAN HARUVY

managers buy shares in order to bring themselves into paying dividends while in HIGH settings this pattern is reversed.

CONCLUSIONS The premise of the present work is that dividend outcomes are the result of decisions by managers. These decisions themselves are the result of a dynamic process that may or may not lead to eventual coordination. The two equilibrium outcomes in the basic setup are one where managers dump their shares and proceed to extract surplus from the firm by self-dealing and the other where the majority of managers accumulate shares and proceed to award dividends to all shareholders. We show that both outcomes may emerge and we demonstrate the pathdependence of these outcomes on the history of the market. We thus demonstrate the importance of initial conditions on final outcomes. We also show that while the vast majority of individual decision makers behave in their own best interests, group voting outcomes were far more challenging for managers to coordinate on. Investors’ actions could potentially affect the outcomes. For example, investors could respond to dividend outcomes by selling shares, which would transfer all shares to management. Or investors could condition their valuation of shares on managers’ past purchase and sale activities. We found no evidence for this in our experiment. We investigated two settings. In the first
the basic setting
dividends are welfare improving, both in terms of payoff to external investors and in terms of societal welfare. In the second setting
the investment setting
investment decisions are incorporated and dividends come at the expense of a potentially beneficial investment. In the investment setting, paying dividends is a dominated decision, although it is one that potentially reduces the free cash flow problem. We found that in this setting, coordination proved far more challenging and groups voted in manners inconsistent with the best interest of the majority of voters far more often than in the basic setting. In both settings, increasing the managers’ initial endowment of shares
presumably in order to align managers’ interests with those of investors’
ended up having the opposite effect. Managers found it harder to coordinate on an outcome and this lowered return on investment. This held for both the benchmark and investment settings and so it appears to be a

The Dividend Puzzle: A Laboratory Investigation

105

robust finding. Intuitively, when managers are endowed with a lot of shares, they unload some shares (perhaps due to risk aversion) and this inevitably sends a signal to other managers regarding the outcome to coordinate on. We indeed find overwhelming evidence that managers condition the changes in their ownership position in a given period on the purchase and sale activities of other managers in the previous period. Vice versa, when managers are endowed with few shares, any purchases they make will move the path in the opposite direction. Outside the laboratory, it is difficult to find very clear examples for this pattern because the most publicized examples of managers unloading their shares involve large shareholders of the firm, and this could actually have the effect of diffusing more shares into the upper echelon of management. Nevertheless, the overall pattern documented in the literature, as discussed in the introduction, is that managerial ownership translates to lower dividends. While we find no direct effect on final period dividends, we do find that overall return on investment falls with higher managerial ownership in both the benchmark setting, where self-dealing is observed, and in the investment setting, where the investment outcome is observed. Thus, our overall recommendation would be for investors seeking to align management interests’ with their own should be careful in allocating shares to management, making sure that the allocation is both well distributed among managers and not too excessive in order to avoid incentives to unload shares which would lead to the less desirable equilibrium for investors.

NOTES 1. We test whether the proportion of managers that have right responses in at least 90% of the decisions is higher than would be expected by chance. 2. We test whether the proportion of groups that vote in the majority’s best response in at least 90% of the decisions is higher than would be expected by chance. 3. We create a dummy variable that equals 1 if at least two managers have more than two shares and zero otherwise (self-deal situation).

ACKNOWLEDGMENTS We are grateful to Tibor Neugebauer for making this research possible. We gratefully acknowledge financial support from the Fonds National de la Recherche Luxembourg (PDR 09 044) und die Universite´ Du Luxembourg (Project IDIAM).

106

SASCHA FU¨LLBRUNN AND ERNAN HARUVY

REFERENCES Bathala, C. T., & Rao, R. P. (1995). The determinants of board composition: An agency theory perspective. Managerial and Decision Economics, 16, 59
69. Belden, S., Fister, T., & Knapp, B. (2005). Dividends and directors: Do outsiders reduce agency costs? Business and Society Review, 110(2), 171
180. Easterbrook, F. H. (1984). Two agency-cost explanations of dividends. American Economic Review, 74(4), 650
659. Fischbacher, U. (2007). z-Tree
Zurich toolbox for readymade economic experiments. Experimental Economics, 10, 171
178. Fu¨llbrunn, S. & Haruvy, E. (2011). The takeover game. Luxembourg School of Finance Research Working Paper Series, No. 11-05, Luxembourg. Hall, B. J., & Murphy, K. J. (2000). Optimal exercise prices for executive stock options. American Economic Review, 90(2), 209
214. Hall, B. J., & Liebman, J. B. (1998). Are CEOs really paid like bureaucrats? The Quarterly Journal of Economics, 113(3), 653
691. Jensen, M. C. (1986). Agency costs of free cash flow, corporate finance, and takeovers. The American Economic Review, 76(2), 323
329. Jensen, G. R., Solberg, D. P., & Zorn, T. S. (1992). Simultaneous determination of insider ownership, debt, and dividend policies. Journal of Financial and Quantitative analysis, 27(02), 247
263. Kaplan, S. N., & Reishus, D. (1990). Outside directorships and corporate performance. Journal of Financial Economics, 27(2), 389
410. Kogan, S., Kwasnica, A. M., & Weber, R. A. (2011). Coordination in the presence of asset markets. The American Economic Review, 101(2), 927
947. Oprea, R. (2008). Free cash flow and takeover threats: An experimental study. Southern Economic Journal, 75(2), 351
366. Rozeff, M. (1982). Growth, beta and agency costs as determinants of dividend payout ratios. Journal of financial Research, 5(3), 249
259. Schellenger, M. H., Wood, D. D., & Tashakori, A. (1989). Board of director composition, shareholder wealth, and dividend policy. Journal of Management, 15(3), 457
467. Short, H., Zhang, H., & Keasey, K. (2002). The link between dividend policy and institutional ownership. Journal of Corporate Finance, 8(2), 105
122. Truong, T., & Heaney, R. (2007). Largest shareholder and dividend policy around the world. The Quarterly Review of Economics and Finance, 47(5), 667
687.

The Dividend Puzzle: A Laboratory Investigation

107

APPENDIX In the following, find the instructions for the experiment in Study 2. The Study 1 instructions are the same apart from the investment decision.

Instructions 1. General Instructions This is an experiment in the economics of market decision-making. If you follow the instructions and make good decisions, you might earn a considerable amount of money, which will be paid to you in cash at the end of the experiment. The experiment will consist of a sequence of trading periods in which you will have the opportunity to buy and sell shares. Money in this experiment is expressed in francs (1 franc = 1 cent). 2. How to Use The Computerized Market The goods that can be bought and sold in the market are called Shares. On the top panel of your computer screen you can see the Money you have available to buy shares and the number of shares you currently have. Selling a share
You may sell shares that you own. When you sell a share, your Shares decrease by 1 share and your Money increases by the sale price. You may sell a share by pressing the “SELL” button or by making an offer to sell and having someone select your offer. You may sell shares as long as you have more than 0 shares. To make an offer to sell a share
For that use the text area entitled “Enter Ask Price.” In that text area you can enter the price at which you are offering to sell a share, and then select “Submit Ask Price.” Please submit an Ask Price now. Note the Ask Prices you submit have to be lower than the current lowest Ask Price. You will notice that a number of ask prices appear in a table called “Ask Price.” The lowest Ask Price will always be on the top and highlighted. Buying a share
When you buy a share, your Shares increase by 1 share and your Money decreases by the purchase price. You may buy a share by pressing the “BUY” button or by making an offer to buy and having someone select your offer. You may buy shares as long as your Money is greater than the price and you have fewer than 4 shares. To offer to buy a share
For that use the text area entitled “Enter Bid Price.” In that text area you can enter the price at which you are offering

108

SASCHA FU¨LLBRUNN AND ERNAN HARUVY

to buy a share, and then select “Submit Bid Price.” Please submit a Bid Price now. Note the Bid Prices you submit have to be higher than the current highest Bid Price. You will notice that a number of Bid Prices now appear in a table called “Bid Price.” The highest Bid Price will always be on the top and highlighted. By now, there are both Bid Prices and Ask Prices listed in their respective tables. You may press the BUY or SELL buttons to accept the best offer from the corresponding table. Please do so now. Notice that all recent prices are provided in the table “Purchase Price.” You will now have a practice period. Your actions in the practice period do not count toward your earnings and do not influence your position later in the experiment. The goal of the practice period is only to master the use of the interface. Please be sure that you have successfully submitted Bid Prices and Ask Prices. Also be sure that you have accepted both Bid and Ask Prices. You are free to ask questions, by raising your hand, during the practice period. 3. Specific Instructions for This Experiment One market will consist of five trading periods. In each period, there will be a market open for 90 seconds, in which you may buy and sell shares. Shares are assets with a life of five periods, and your inventory of shares carries over from one trading period to the next. You may receive dividends for each share in your inventory at the end of each period. You may also receive an additional Return per Share at the end of the market (after Period 5). Each share produces revenue of 12 francs after every period. There are 8 shares in total in the market and, thus, in each period 8 × 12 = 96 francs are to be distributed among the participants. You are either of type A (3 traders) or of type B (3 traders) during the entire experiment. If you are of type B you make an additional decision at the end of every period in that you vote to either Pay Dividend, or Not Pay and Invest, or Not Pay and Not Invest. 4. How Dividends and Returns Are Distributed Among Participants Type B traders need to make two decisions: First, whether to pay dividend or to not pay dividends and, second, whether to invest or to not invest if the first decision was not to pay. If at least 2 of 3 type B traders vote Pay Dividend, a dividend is paid immediately and each share pays 12 francs. If at least 2 of 3 type B traders vote Not Pay and Not Invest, no dividend is paid and each share earns zero in dividends. However, trader B

The Dividend Puzzle: A Laboratory Investigation

109

equally split half the unpaid revenue (i.e., half of the total revenue of 96 divided by 3). In that case, only type B traders receive 16 francs each immediately. Traders of type A receive 0 francs per share. If at least 2 of 3 type B traders vote Not Pay and Invest, a deferred payment of 15 francs is given to each Share that any trader hold in the inventory after Period 5. These deferred payments accumulate to the Return per Share, that is, for every period the majority voted Not Pay/Invest the Return per Share at the end of the market increases by 15. Thus, if the majority vote is Not Pay/Invest in all periods, the Return equals 5 × 15 = 75 francs; if the majority vote is Not Pay/Invest in 4 periods the Return equals 4 × 15 = 60; if the majority vote is Not Pay/Invest in 3 periods, the Return equals 3 × 15 = 45; if the majority vote is Not Pay/Invest in 2 periods the Return equals 2 × 15 = 30; if the majority vote is Not Pay/Invest in 1 period the Return equals 15; and if the majority vote is never Not Pay/Invest the Return equals 0. The table summarizes type B’s payoff opportunities after each period. Note again, the Return is only paid if the shares are in the inventory after Period 5. Payoff Table for Type B # of shares you own

Type B Majority vote Pay Dividend

Type B Majority vote Not Pay and Not Invest

0 1 2 3 4

0 12 24 36 48 Immediate Payment

16 16 16 16 16

Type B Majority vote Not Pay and Invest 0 15 30 45 60 Payment for any Share in Inventory after Period 5

Note all traders receive information whether the majority vote of type B traders was Pay Dividend or Not Pay. But the investment decision is only known to the type B traders. Given a Pay Dividend vote in every period, the expected cumulative payment for each share held until the end of the market is equal to the number of periods remaining times the dividend payment. For example, in the first period, 5 dividend payments remain and the expected cumulative payment equals 5 × 12 = 60. In Period 2, the expected cumulative payment equals 4 × 12 = 48. This becomes 36 in Period 3, 24 in Period 4, and 12 in Period 5.

110

SASCHA FU¨LLBRUNN AND ERNAN HARUVY

At the end of the period, dividends or profits from splitting the revenue are added to your cash balance. At the end of the market the Return per Share is added to your cash balance. Further on, each type B trader gets a fixed ID (=1, 2, 3). Each Ask Price or Bid Price will be labeled with this ID. Type A traders have no ID. All traders get information about the number of shares that type B traders have. 5. Endowment and Earnings You participate in 4 subsequent markets. You role as type A or type B will stay the same in all markets. The endowments of cash and shares for the markets are stated right before the market starts and can change from market to market. In any market, your earnings will be the Money that you hold at the end of Period 5, after the last dividend payment or the profits from splitting the revenue and the Return per Share have been paid, that is: + − + +

money you have at the beginning of the market, money received from sales of shares, money spent on purchases of shares, money from immediate payments you receive (dividends/profits from splitting the revenue), and Money from deferred payments after Period 5 (Return per Share).

Summing up all the Money of all markets plus a participation fee is your final earnings.

CHAPTER 6 AN EXPERIMENTAL ANALYSIS OF MYOPIC LOSS AVERSION Tomoki Kitamura and Munenori Nakasato ABSTRACT Purpose
Previous studies showed mixed results as to the cause of myopic loss aversion (MLA). This paper reexamines the main driver of MLA, considering two factors from previous studies and an additional factor. Design/methodology/approach
Experimentally investigate whether flexibility of investment, frequency of information feedback, or timing of decision cause MLA. Findings
Timing of decision and flexibility of investment explain most differences in subject behavior. Frequency of information feedback makes only a marginal contribution. Originality/value of the paper
The differences in subject behavior can be interpreted by a shift in their reference points depending on the difference in flexibility of investment, frequency of information feedback, or timing of decision. Keywords: myopic loss aversion; reference points; flexibility of investment; information feedback JEL classifications: G10; D80

Experiments in Financial Economics Research in Experimental Economics, Volume 16, 111
143 Copyright r 2013 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0193-2306/doi:10.1108/S0193-2306(2013)0000016006

111

112

TOMOKI KITAMURA AND MUNENORI NAKASATO

INTRODUCTION We experimentally investigate what factors cause myopic loss aversion (MLA). Previous studies showed mixed results as to whether the main driver of MLA is flexibility of investment or frequency of information feedback. In this chapter, using a multi-period stock model, we reexamine these factors and also introduce one additional factor: timing of decision. We interpret our results in terms of the investor’s shift of reference points being determined by these three factors. The concept of MLA was introduced by Benartzi and Thaler (1995), and it predicts that investors who evaluate investments frequently will tend to reduce investments in risky securities. MLA begins with loss aversion and mental accounting. Loss aversion is an investor’s tendency to weigh losses more heavily than gains and occurs when investors apply a utility (value) function to evaluate investment results according to the gains and losses determined by a particular reference point. This function is concave (risk-averse) in the gain region and convex, (risk-inclined) with a steeper slope, in the loss region (Kahneman & Tversky, 1979). Empirical studies by Kahneman, Knetsch, and Thaler (1990) and Tversky and Kahneman (1992) estimated that losses weigh about twice as much as gains. Mental accounting is a behavioral concept wherein individuals tend to divide their current and future assets into separate portions (Thaler, 1980). For example, investors who have long-term investment horizons, such as those involving preparation for retirement, may repeat asset evaluations within a shorter period, such as every year. MLA is a possible explanation of the equity premium puzzle (Mehra & Prescott, 1985). Although Benartzi and Thaler (1995) did not show empirical evidence of MLA, Gneezy and Potters (1997) experimentally analyzed whether MLA could explain relatively low levels of investment in risky investments. They compared behaviors in stylized high-frequency and low-frequency investment treatments and found evidence that supported the concept of MLA: investment in the high-frequency treatment was significantly lower. Thaler, Tversky, Kahneman, and Schwartz (1997) conducted a similar experiment and found further support for MLA. Haigh and List (2005) replicated the study by Gneezy and Potters (1997) on professional traders and found even stronger effects of MLA. Gneezy, Kapteyn, and Potters (2003) demonstrated that market prices for a risky security are significantly higher when the investment evaluation period is short. The above studies confirm that investment behavior differs according to investment evaluation frequency.

An Experimental Analysis of Myopic Loss Aversion

113

In previous studies, the frequency of investment evaluation comprised two parts: frequency of information feedback and flexibility of decision making. Bellemare, Krause, Kroger, and Zhang (2005) added a treatment to the Gneezy and Potters (1997) design, which separated the roles of investment flexibility and frequency of information feedback. They found evidence that information feedback is the main factor explaining the treatment effect reported by Gneezy and Potters (1997). However, Langer and Weber (2008) conducted a similar experiment and found the opposite result: investment flexibility is the main driver of low investment in risky securities. The experimental design of Langer and Weber’s (2008) study differed from those of Bellemare et al. (2005) and Gneezy and Potters (1997) in one important aspect: returns on investments in each period were added to the endowments that investors could invest in subsequent periods. Therefore, investors theoretically decided on investments by considering subsequent periods. Langer and Weber (2008) called this a multiplicative model. In contrast, gains or losses in Bellemare et al. (2005) and Gneezy and Potters (1997) did not affect future endowments that investors could invest in subsequent periods, and was termed by Langer and Weber (2008) as an additive design. Langer and Weber (2008) implicitly argued that this important design difference might explain why their conclusions differed from those of Bellemare et al. (2005). However, these mixed results necessitate further investigation. In Langer and Weber’s (2008) experimental design, subjects invested in stock over multiple periods (30 periods). In this chapter, we use a modified (actually, a simplified) version of Langer and Weber’s (2008) experimental design, employing a two-period binomial model to represent stock price fluctuations. In order to reexamine the cause of MLA, however, we consider the frequency of investment from three angles: frequency of information feedback, flexibility of decision making, and timing of decision. We also interpret the differences in investment behavior in our experiment in terms of changing investor reference points. This stock model offers the advantage of indicating whether investor reference points move as stock prices change. We use this model repeatedly in a manner similar to an additive design. Our overall results are comparable to the findings of Langer and Weber (2008): timing of decision and flexibility of investment explain most of the differences in investor behavior, while frequency of information feedback makes only a marginal contribution. The differences in investment behavior can be interpreted by a shift in investor reference points depending on the differences in timing of decision, flexibility of investment, or frequency of information feedback. Investor behavior was consistent,

114

TOMOKI KITAMURA AND MUNENORI NAKASATO

with him/her having completely shifted his/her reference points on the current stock price in the most flexible and high frequency of information feedback situation. This shift seems to be the main reason for the reduction of investment in stock. On the contrary, investors did not adjust their reference points and did not reduce investments in stock during the least flexible and low frequency of information feedback situation. They seem to adapt their reference points asymmetrically, for gains or losses, between those two situations. The chapter proceeds as follows. We present the experimental design and possible behaviors of investors, followed by a description of the results, a robustness check, and the conclusions.

EXPERIMENTAL DESIGN Stock Price and General Experimental Procedure In this chapter, exogenous stock price is represented in a two-step binomial model.1 There are three time periods in this model: time 0, time 1, and time 2. Time 0 is the initial time period, and time 2, the final time period. The stock price is 10 dollars at time 0. It either rises to 25 dollars or falls to 5 dollars at time 1. When the stock price is 25 dollars at time 1, it either rises to 40 dollars or falls to 20 dollars at time 2. When the stock price is 5 dollars at time 1, it either rises to 20 dollars or falls to 0 dollars at time 2. In other words, when the stock price appreciates, it increases by 15 dollars, and when it depreciates, it decreases by 5 dollars.2 The probability of rising is 1/3, and that of falling is 2/3. Most of the subjects for our experiments were undergraduate students at Aoyama Gakuin University in Tokyo, Japan, while the remainder were undergraduate students at Keio University or University of Tokyo. Subjects were recruited through Aoyama Gakuin University’s part-time job recruiting system.3 They were provided printed copies of the experimental instructions, which were also read aloud. A sample of the experimental instructions translated into English can be found in Appendix A. For the experiment, subjects were given 100 dollars of “cash” (within the experiment) at time 0, and they had to decide how many units of stock to buy, within a range of integer values between 0 and 10. Funds not used to buy stock were carried over unchanged to the subsequent period without any interest accrual. Borrowing cash and short-selling stock were

An Experimental Analysis of Myopic Loss Aversion

115

prohibited. Subjects were able to change the number of stocks that they had bought during the intermediate period (time 1) according to treatments that are explained below. After the final stock price was realized at time 2, the total wealth of each subject was calculated and recorded on a computer. Subjects were also asked to manually record the number of stocks held by them and their total wealth at time 2 on their recording sheets (an example can be found in Appendix B). We call the sequence of operations from time 0 to time 2 one “trial.” Subjects were involved in several practice trials after the instructions were communicated, and they subsequently took a brief quiz about recognizing stock price movements and calculating final wealth and payment (see Appendix C), which aimed to check their understanding of their expected actions during the experiment. They were then given the correct answers to the quiz and had several minutes to think about their investment strategies. We did not inform the subjects regarding the number of trials they would be participating in. Most experiments comprised 9 trials; however, some experiments comprised 10 or 11 trials. This was done to negate potential end effects. If subjects had been given information on the exact number of trials involved, they may have taken greater risks toward the end of the trials in order to compensate for their losses in earlier trials, or they may have reduced risks in order to stabilize earlier gains.4 After finishing the last trial, we randomly chose one trial to determine the payment for each subject. We paid ten times the final wealth in period 2 for the chosen trials in Japanese yen (JPY), with an additional 1,000 JPY for participating in the experiment.5 We used this random incentive system to offset the potential that subjects who had accumulated relatively high levels of wealth would reduce their investments more than usual in order to keep their wealth or that subjects who had accumulated relatively low levels of wealth would increase their investments to recover losses. We believe that our random incentive system successfully reduced such behaviors. The total time involved for conducting each experiment, including the explanation for the instructions, quizzes, trials, and payment, was approximately 90 minutes. Overall, we conducted 12 experiments from November 26, 2009 to March 4, 2011. The total number of subjects involved was 162.

TREATMENTS We use four main treatments (FH-IH, FHa-IH, FL-IH, and FL-IL) to test whether a subject’s investment behavior differs according to the flexibility

116

TOMOKI KITAMURA AND MUNENORI NAKASATO

of investment, frequency of information feedback, or timing of decision. Panel A of Fig. 1 shows differences in the main treatments. The circles in Fig. 1 mean that a stock price is observable for subjects in a particular state. The squares mean that subjects can decide the number of stocks in a particular state. Treatment FH-IH involved a high-flexibility (FH) and high-frequency information feedback (IH) treatment. Subjects could decide on the number of stocks to buy at both time 0 and time 1. They could observe the stock prices at time 0, time 1, and time 2 and could decide on the number of stocks that they wanted to hold at time 0 and time 1 based on the stock prices at those times. Treatment FHa-IH also involved a high-flexibility and high-frequency information feedback treatment; however, the decision timing was different from that in FH-IH. In this treatment, subjects had to decide all investments at time 0. Namely, subjects could observe the stock price at time 0, time 1, and time 2 and could decide on the number of stocks that they wanted to hold at both time 0 and time 1, which is similar to treatment FH-IH. However, at time 0, they had to predecide the number of stocks that they would have at time 1, before they had the chance to observe the stock price at time 1. Treatment FL-IH involved a lowflexibility (FL) and high-frequency information feedback treatment. Subjects could observe the stock price at time 0, time 1, and time 2, as in treatments FH-IH and FHa-IH; however, they could only decide on the number of stocks to hold at time 0 and could not change the number of stocks until time 2 was over. Subjects were compelled to maintain the number of stocks that they had decided on at time 0. Treatment FL-IL involved a low-flexibility and low-frequency information feedback (IL) treatment. Subjects could observe stock prices only at time 0 and time 2. They could decide on the number of stocks to hold only at time 0, and could not change this number until time 2 had elapsed. They were forced to fix the number of stocks, as in treatment FL-IH, but only learned of their aggregated losses or gains at time 2. The difference between treatments FH-IH and FHa-IH was in the timing of the decisions. In treatment FHa-IH, subjects had to decide on the number of stocks for both time 0 and time 1 in advance at time 0. The difference between treatments FHa-IH and FL-IH was in whether subjects were permitted to change the number of their stocks at time 1. In treatment FL-IH, subjects could not change the number of stocks and had to maintain their initial positions. The difference between treatments FL-IH and FL-IL lay in whether the subjects were aware about their stocks’ price paths. In treatment FL-IL, subjects could not distinguish between the

117

An Experimental Analysis of Myopic Loss Aversion (A)

FH-IH Time 0 Time 1 Time 2 40 25

20

10

H

20

5

0 FHa-IH Time 0 Time 1 Time 2 40 25 H

Flexibility

Timing of decision

20

10

20

5

0 FL-IH Time 0 Time 1 Time 2 40 25

20

10

L

FL-IL FL IL Time 0 Time 1 Time 2 40 20

10

20

5

0

0 L

H Information feedback (B)

FL-IH(Price=9) Time 0 Time 1 Time 2 40 25 10 9

FL-IH(Price=39) Time 0 Time 1 Time 2 40 39

20 20 0

10 5

FL-IH(Sell) Time 0 Time 1 Time 2 40 25

20 20 0

10 Sell

5

20 20 0

Fig. 1. Treatment. (A) Panel A: Four main treatment. (B) Panel B: Extra treatments. Note: The circles in the figures mean that the stock prices are observable at a particular time, whereas the squares mean that subjects can decide the number of stocks at a particular time or in a particular state. In treatment FL-IH (sell), subjects were initially endowed with 10 units of stock and no cash, and they could decide to sell units of their stock.

118

TOMOKI KITAMURA AND MUNENORI NAKASATO

upstate (stock price of 25 dollars) and downstate (stock price of 5 dollars) at time 1. In other words, they did not know how they had arrived at their total gains or losses at time 2. Fig. 2 is a screenshot from a computer used in treatment FH-IH. The other treatment screens varied as necessitated by the treatment differences already described. The meaning of the terms on the screen and how their positions and total assets were calculated were explained carefully in the experimental instructions. The gains, losses, and final wealth at time 2 were automatically calculated when subjects changed the number of stocks that they wanted to invest in. They were permitted to use this software freely Trial 1

Time 1

Remaining seconds 60

Time 0

Time 1

Time 2

Time 0 Send your decision

Time 1 Send your decision

Unable to send

Able to send

+$15 (Prob=1/3)

Stock Price Stock Cash Total Assets

+$15 (Prob=1/3)

Stock price units stock cash Total Assets

Stock price u n its stock cash Total Assets

$80 $130

-$5 (Prob=2/3)

Stock Price Stock

20 $40

Cash

$80

T o ta l A s s e ts

$120

$20 $80 $100

0 1 2

+$15 (Prob=1/3)

-$5 (Prob=2/3)

S to c k P ric e Stock Cash Total Assets

Stock price Y o u r d e c is io n

$160

25 2 $50

0 1 2

10 2

40 $80 $80

0

u n its

$10

cash

$80

0 1 2

$120

5 2

stock Total Assets

20 $40 $80

$90

-$5 (Prob=2/3)

Stock Price Stock Cash Total Assets

0 $0 $80 $80

Results Time 0 S to c k p ric e T o ta l a s s e ts

10 $100

Time 1 S to c k p ric e T o ta l a s s e ts

Time 2 S to c k p ric e T o ta l a s s e ts

Fig. 2. Subjects’ Computer Screens (for Treatment FH-IH). Note: When subjects changed the number of their stocks, their positions and total assets were automatically recalculated for each time period. They were able to use the software at any time during the experiment in order to examine the relationship between their number of stocks and investment results.

An Experimental Analysis of Myopic Loss Aversion

119

during the experiment and could examine the relationship between the number of stocks and final wealth at time 2. In addition, the subjects were explained that they could fix or change the number of stocks involved in each different trial. Once the experimenter had announced that the trial had begun, the subjects had 1 minute to decide on the number of stocks that they wanted and to click the “send your decision” button. If the subjects did not click the button, 0 units of stock were assigned.6 The decisions that they sent to the experimenter were displayed in the “your decision” part of the screen. At the end of 1 minute, the realizations of stocks and results of investments were displayed on the lower part of the computer screen.7

EXPECTED SUBJECT BEHAVIOR In this section, we examine the expected behavior of the subjects, based on Kahneman and Tversky’s (1979) prospect theory.8 We assume that subjects derive utility from losses and gains, and not from wealth, which is captured by the following value function.
α if x > 0 x vðxÞ = − λ⋅ð − xÞα if x ≤ 0 The coefficient λ > 1 represents loss averseness, whereas 0 < α < 1 represents the curvature of the function. We also assume that subjects decide the number of stocks by maximizing the expected value function. Here, we examine the following three types of expected value functions. Although these are not all the possible expected value functions, they are the most plausible ones. Thereafter, we compare our experimental results with our expected value functions to analyze the behavior of subjects. We let θ0 be number of stocks at time 0, θu be the number of stocks at the upstate of time 1, θd be that at the downstate of time 1, and g1 and g2 be the gain or loss at time 1 and time 2, respectively. These gains and losses are the results of subjects’ decisions for θ0 , θu , and θd . The expected value function type V1 is characterized by a reference point that is the initial stock price, and only one period of future stock price change is considered. This means that subjects evaluate gain and loss myopically, ignoring the two times ahead. V1 ≡ E[vðg1 Þ]

120

TOMOKI KITAMURA AND MUNENORI NAKASATO

Type V2 is characterized by reference points, which move perfectly along with changes in stock price. The reference point at time 0 is set at 10 dollars. The reference point at the upstate and the downstate of time 1 is set at 25 dollars and 5 dollars, respectively. This means that the gains and losses are first evaluated separately for each time, and their values are then accumulated. V2 ≡ E[vðg1 Þ þ vðg2 Þ] Type V3 is characterized by a reference point that does not move and remains at the initial stock price of 10 dollars. This means that the gains and losses for each time are first accumulated and then evaluated. V3 ≡ E[vðg1 þ g2 Þ] Subjects identify the optimal number of stocks by maximizing their expected value functions. For example, the optimal number of stocks for investors who employed V1 is found by solving the following integer programming. 1 2 max [vð15θ0 Þ] þ [vð − 5θ0 Þ] θ0 3 3 s:t: 0 ≤ θ0 ≤ 10; which is the budget constraint. Appendix D shows that the optimal number of stocks for those who employ V1 is as seen below.
10 if 3α − 2λ > 0 θ0 = ð1Þ 0 if 3α − 2λ ≤ 0 The optimal number of stocks for investors who use V2 can be found by solving the following integer programming. max

θ0 ;θu ;θd

1 2 [vð15θ0 Þ þ vð15θu Þ] þ [vð15θ0 Þ þ vð − 5θu Þ] 9 9 2 4 þ [vð − 5θ0 Þ þ vð15θd Þ] þ [vð − 5θ0 Þ þ vð − 5θd Þ] 9 9

s:t: 0 ≤ θ0 ≤ 10 0 ≤ θu ≤ ð15⋅θ0 þ 100Þ=25 0 ≤ θd ≤ ð − 5⋅θ0 þ 100Þ=5

An Experimental Analysis of Myopic Loss Aversion

121

The above constraints are a result of budget constraints. Appendix D shows that the optimal number of stocks for those who employ V2 is as follows.
10 if 3α − 2λ > 0 θ0 = θu = θd = ð2Þ 0 if 3α − 2λ ≤ 0 These two results indicate that the optimal number of stocks for myopic investors is the same as those for long-term investors, wherein investor reference points move perfectly along with stock price changes. Since θ0 is the same for V1 and V2 , we consider only V2 . The optimal number of stocks for investors who use V3 can be found by solving the following integer programming. 1 2 [vð15θ0 þ 15θu Þ] þ [vð15θ0 − 5θu Þ] θ0 ;θu ;θd 9 9 2 4 þ [vð − 5θ0 þ 15θd Þ] þ [vð − 5θ0 − 5θd Þ] 9 9 max

s:t: 0 ≤ θ0 ≤ 10 0 ≤ θu ≤ ð15⋅θ0 þ 100Þ=25 0 ≤ θd ≤ ð − 5⋅θ0 þ 100Þ=5 Although the optimal number of stocks cannot be solved analytically, it can be solved numerically by considering all possible integer combinations of the numbers of stocks, given α and λ. If the flexibility of investments is low, which means that subjects are not able to change the number of stocks at time 1 and have to fix the number of stocks from time 0 to time 2, the above budget constraints are replaced by 0 ≤ θ0 = θu = θd ≤ 10. Fig. 3 shows the boundaries at which the optimal number of stocks for each expected value function is 0 or positive at time 0 with the given parameters: 1:01 ≤ λ ≤ 2:00 and 0:20 ≤ α ≤ 0:99 (the increment in both cases is 0.01). V3 (low flexibility) in Fig. 3 represents the boundary lines for V3 with low flexibility, which means that investors have to fix the number of stock over all time periods. The optimal number of stocks at time 0 is 0 for the regions below each line. On the contrary, the optimal number of stocks is at least 1 unit (θ0 ≥ 1) in the region above each line. As we can see in Fig. 3, V3 has the largest area where the optimal number of stocks is positive. Therefore, subjects who employ V3 should invest more at time 0 than subjects who employ other expected value functions. In

122

TOMOKI KITAMURA AND MUNENORI NAKASATO V3(low flexibility)

1.0 0.9

V3

0.8

α

0.7 0.6 V1 = V2 = V2(low flexibility)

0.5 0.4 0.3 0.2 1.00 1.10 1.20 1.30 1.40 1.50 1.60 1.70 1.80 1.90 2.00 λ V2 V3 V3(low flexibility)

Fig. 3. Boundaries for Each Expected Value Function. Note: The region to the lower right of each line represents the optimal number of stocks, given that λ and α are at 0 units at time 0. The region on the upper left-hand side of each line represents the optimal number of positive stocks at time 0. The boundaries for V1 , V2 , and V2 (low flexibility) are the same.

contrast, type V2 has the smallest positive area, which indicates that subjects who employ this function should invest the least at time 0. Subjects are expected to use different expected value functions, according to flexibility of investment or frequency of information feedback, which were distinguished in the treatments. In treatment FL-IL, subjects had to decide on the number of stocks that they wanted to hold only at time 0, and they maintained this number until time 2, with no information feedback at time 1. Therefore, they were expected to use only V3 (low flexibility). In treatment FL-IH, there was information feedback at time 1; however, the flexibility was low. Therefore, subjects could use V3 (low flexibility) or V2 . In treatments FHa-IH and FH-IH, since there were no limitations on the frequency of information and investment flexibility, subjects could employ any expected value functions. According to the existing literature, subjects in treatment FH-IH should be expected to exhibit the lowest average number of stocks among all treatments, whereas subjects in treatment FL-IL should exhibit the highest average number of stocks. However, in our experimental setting, treatment

An Experimental Analysis of Myopic Loss Aversion

123

FH-IH subjects may exhibit the highest average investment among all treatments at time 0 if they use V3 .

EXPERIMENTAL RESULTS Panel A of Appendix E shows histograms by treatment, whereas Panel B shows changes in the average number of stocks for each trial by treatment. The average number of stocks tends to increase in most studies, such as Bellemare et al. (2005) and Langer and Weber (2008); however, there is no such tendency in our experiments. Table 1 shows the average number of stocks (reflecting subjects’ decisions to invest) at time 0. The upper half of the table represents the average number of stocks and its standard deviation in each treatment, and the lower half shows the Mann
Whitney z-statistics and their p-values, testing the null hypothesis that the average number of stocks is the same for two particular treatments against the alternative hypothesis that the average number of stocks for the same two treatments differs. Initially, we compare treatments FH-IH and FL-IL. The average number of stocks in treatment FH-IH is 4.253 and that in treatment FL-IL is 6.506. The value of the Mann
Whitney z-statistic is
6.931, which shows that these averages are significantly different for p < 0.01 ((1) FH-IH vs. FL-IL). This indicates that high flexibility of investment and high frequency of information feedback jointly reduce the average number of stocks, and the result supports MLA in a manner similar to that found in other studies, such as Gneezy and Potters (1997). Then, we compared treatments FH-IH and FHa-IH. These treatments differ in terms of decision timing. Subjects in treatment FHa-IH had to decide on their time 1 investment at time 0. The average number of stocks in treatment FHa-IH is 5.258. The difference between these treatments is statistically significant at the 1% level ((2) FH-IH vs. FHa-IH). The average number of stocks in treatment FL-IH is 6.176, which was significantly higher than that in treatment FHa-IH ((3) FHa-IH vs. FL-IH). The difference between these treatments is in terms of the flexibility in decision making. This result indicates that higher flexibility of investment reduces investments in risky securities. The average number of stocks in treatment FL-IL is 6.506 and that in treatment FL-IH is 6.176. The treatments differ only in terms of frequency of information feedback. Their averages are not statistically different ((5) FLIH vs. FL-IL), which is in line with the results of Langer and Weber (2008) and contrary to those of Bellemare et al. (2005).

124

TOMOKI KITAMURA AND MUNENORI NAKASATO

Table 1. Average Number of Stocks at Time 0. FH-IH Average number of stocks at time 0 Avg 4.253 (SD) (3.586) N 261 #subjects 29

FHa-IH

FL-IH

FL-IL

5.258 (3.530) 414 46

6.176 (3.446) 306 34

6.506 (3.315) 243 27

Mann-Whitney z-value

z-value p-value

(1)

(2)

(3)

(4)

(5)

(6)

FH-IH vs. FL-IL

FH-IH vs. FHa-IH

FHa-IH vs. FL-IH

FH-IH vs. FL-IH

FL-IH vs. FL-IL

FHa-IH vs. FL-IL

−6.931 (0.000)

−3.604 (0.000)

−3.425 (0.001)

−6.245 (0.000)

−1.160 (0.246)

−4.405 (0.000)

FH-IH

H

(2)***

Timing of decision (1)***

Flexibility H

(4)***

FHa-IH (3)*** (6)***

L

FL-IH

(5)

H

FL-IL L

Information feedback Note: Column (1) “FL-IL vs. FH-IH” represents the test of the null hypothesis that investment units would be the same in treatments FL-IL and FH-IH. The other columns follow the same logic.

Although an analysis of Table 1 provides evidence that the flexibility of investments is the main factor in MLA, we made little attempts to control the panel nature of our data. Table 2 shows the results of a random-effect TOBIT regression, with the left censored at 0 and the right censored at 10. The dependent variable is the number of stocks at time 0. Column (1) shows the regression coefficients and their standard errors, using data from treatments FH-IH and FL-IL. The independent treatment variable

An Experimental Analysis of Myopic Loss Aversion

125

FH-IH (dummy) is an indicator variable, which equals 1 if subjects belong to treatment FH-IH or 0 otherwise (the other treatment independent variables used below are similar). The coefficient of FH-IH (dummy) is negative and significant at 1%. We confirm that high frequency of investment timing, high flexibility of investment, and high frequency of information feedback jointly reduce investment in stock. Column (2) is the regression result using data from treatments FH-IH and FHa-IH. The coefficient of FH-IH (dummy) is not significant when we consider a panel structure. Column (3) shows the result from using data from treatments FHa-IH and FL-IH. The coefficient of FHa-IH (dummy) is also not significant. However, in column (4), which shows the regression result using data from treatments FH-IH and FL-IH, the coefficient of FH-IH (dummy) is negative and significant at 5%. This indicates that higher flexibility of investments reduces investments in risk securities. Column (5) presents the result using data from treatments FL-IH and FL-IL. The difference between these two treatments is in the frequency of information feedback. Here, the coefficient of FL-IH (dummy) is not significant, which is the same as the result shown in Table 1. Column (7) displays regression results using all the data. The coefficient of FH-IH (dummy) is negative and significant at 1%, and the coefficient of FHa-IH (dummy) is negative and significant at 10%; however, the coefficient of FL-IH (dummy) is not significant. This result indicates a tendency, similar to that shown in columns (1)
(6). Column (8) depicts the results using all the data. The independent variables are INFOH (dummy) and FLEXH (dummy). INFOH (dummy) is an indicator variable that equals 1 if the frequency of information feedback is high (which means that the treatment involved is FH-IH, FHa-IH, or FL-IH) or 0 otherwise. FLEXH (dummy) is an indicator variable that equals 1 if the flexibility of investments is high (which means that treatment is FH-IH or FHa-IH) or 0 otherwise. Although the coefficient of FLEXH (dummy) is negative and significant at 5%, the coefficient of INFOH (dummy) is not significant. This confirms that flexibility of investment is the main contributor to MLA. Table 3 shows the subjects’ average number of stocks for treatments FH-IH and FHa-IH at time 1 (the intermediate time) when the stock price is 25 dollars (upstate), at time 1 when the stock price is 5 dollars (downstate), and at time 0. At time 1, in treatments FH-IH and FHa-IH, subjects were able to change their number of stocks. The average number of stocks for treatment FH-IH at time 0 is 4.253 and that for the up-state at time 1 is 4.172. The difference between these two averages is not statistically significant. The average number of stocks for the downstate at time 1 is 5.379.

Table 2.

Results of the Random Effect TOBIT Regression. (1)

Data used

(2)

FH-IH

(3)

(4)

FH-IH FHa-IH

(5)

FL-IH

FH-IH(dummy)

−3.722

(1.271)

FL-IH

FH-IH

FH-IH

FHa-IH

FHa-IH

FHa-IH

FL-IH

FL-IH

FL-IL

FL-IL

FL-IH FL-IL

−1.550

−3.015

(1.097)

FL-IL

−3.641

**

(1.239)

(8)

***

(1.285)

−1.396

FHa-IH(dummy)

(7)

FH-IH FHa-IH

FL-IL ***

(6)

−2.069*

(1.090)

(1.145)

FL-IH(dummy)

−2.084* (1.167)

−0.662

−0.646

(1.297)

(1.243)

INFOH(dummy)

−0.649

FLEXH(dummy)

−2.043**

(1.254)

(1.001) Cons. sigma_u sigma_e left-censored right-censored N χ2 p-value Log likelihood

8.565*** (1.060) 4.483*** (0.553) 3.709*** (0.176) 80 134 504 59.98*** 0.000 −1004

6.503*** (0.802) 4.414*** (0.448) 3.550*** (0.140) 115 144 675 25.51*** 0.002 −1398

7.453*** (0.918) 4.605*** (0.457) 3.377*** (0.128) 83 196 720 8.80 0.456 −1453

7.577*** (0.969) 4.675*** (0.525) 3.567*** (0.158) 86 150 567 34.85*** 0.000 −1128

7.917*** (1.084) 4.779*** (0.569) 3.500*** (0.158) 48 186 549 34.68*** 0.000 −1063

8.351*** (1.012) 4.471*** (0.477) 3.549*** (0.142) 77 180 657 16.99** 0.049 −1340

8.298*** (0.985) 4.563*** (0.353) 3.571*** (0.106) 163 330 1224 46.00*** 0.000 −2472

8.305*** (0.992) 4.604*** (0.356) 3.571*** (0.106) 163 330 1224 44.06*** 0.000 −2472

FH-IH

H (2)

Timing of decision (1)***

Flexibility H

(4)**

FHa-IH (3) (6)*

L

FL-IH

(5)

H

FL-IL L

Information feedback Note: The dependent variable is the number of stocks at time 0. FH-IH (dummy) is a treatment indicator variable that equals 1 if the treatment is FH-IH and 0 otherwise. The other treatment indicators follow the same logic. INFOH (dummy) is an indicator variable that equals 1 if the treatment is FH-IH, FHa-IH, or FL-IH and 0 otherwise. FLEXH (dummy) is an indicator variable that equals 1 if the treatment is FH-IH or FHa-IH and 0 otherwise. Each trial dummy variable is included, but not shown, in this table. “sigma_e” and “sigma_u” indicate the overall and panel-level variance, respectively. ***, **, and * represent significance at the 1%, 5%, and 10% levels. The results are computed by STATA SE (12.1).

127

An Experimental Analysis of Myopic Loss Aversion

Table 3. Average Number of Stocks at Time 1. Time

State

FH-IH

FHa-IH

time 0

(Price = 10)

Avg SD N

4.253 (3.586) 261

5.258 (3.530) 414

time 1

Up -State (Price = 25)

Avg SD N diff. z-value p-value

4.172 (3.353) 145 −0.080 −0.058 (0.954)

4.415 (3.263) 414 −0.843 −3.437 (0.001)

Down-State (Price = 5)

Avg SD N diff. z-value p-value

5.379 (5.446) 116 1.127 0.939 (0.348)

4.795 (4.778) 414 −0.464 −3.423 (0.001)

Note: Of the total 261 trials performed for treatment FH-IH, 145 and 116 relate to the upstate and downstate, respectively. These numbers are different from what would be statistically expected with probabilities of 1/3 and 2/3, because all subjects in one experiment day experienced the same ups and downs in stock fluctuations. “diff.” represents the difference of averages between time 1 and time 0. “z-value” and “p-value” represent the z-value and p-value of the Mann
Whitney test.

The difference between the average number of stocks at time 0 and for the downstate at time 1 is not statistically significant. On the contrary, for treatment FHa-IH, the average number of stocks for the up-state at time 1 is 4.415, which is significantly lower than that at time 0 at the 1% level. The average number of stocks for the downstate at time 1 is 4.975, which is also statistically significantly different from that at time 0. Overall, treatment FH-IH subjects did not change their average number of stocks, whereas treatment FHa-IH subjects reduced their average number of stocks at time 1 from that at time 0. We now interpret all of the above results in terms of the change in reference points. An investigation of our results, based on the expected value functions examined in the previous section, reveals several interesting points. FL-IL subjects were not permitted to change their number of stocks at time 1 and had no information feedback at time 1. These subjects received accumulated investment results only at time 2. Therefore, the only possible expected value function that these subjects could employ, among

128

TOMOKI KITAMURA AND MUNENORI NAKASATO

the ones we considered in the previous section, was V3 , for which the reference point was fixed at the initial stock price. We use treatment FL-IL as a baseline for investigating investment behavior in the other treatments. In treatment FL-IH, the average number of stocks is not statistically different from that involved in treatment FL-IL. This suggests that treatment FL-IH subjects also fixed their reference points at the initial stock price. In treatment FH-IH, subjects had decision-making flexibility and received information feedback at high frequency. Our results show that the average number of stocks in treatment FH-IH at time 0 is the lowest among all treatments, which implies that treatment FH-IH subjects did not solely use V3 and might have employed V1 or V2 . In addition, Table 3 shows that these subjects did not change the number of their stocks at time 1, which is consistent with the use of V1 or V2 . These results indicate that treatment FH-IH subjects perfectly moved their reference points along with stock price changes. The average number of stocks in treatment FHa-IH was intermediate between those in treatments FH-IH and FL-IL. This suggests that treatment FHa-IH subjects did not solely use V1 ; V2 , or V3 , thereby indicating that the reference points of treatment FHa-IH subjects neither completely moved in line with stock prices, nor were fixed at the initial stock price. Instead, these subjects shifted their reference points partially, or asymmetrically, which is similar to the results of Arkes, Hirshleifer, Jiang, and Lim (2008).9

FURTHER INVESTIGATION INTO FREQUENCY OF INFORMATION FEEDBACK Bellemare et al. (2005) demonstrated that only information feedback induces investment behavior that is in line with MLA. Our results indicate that, although flexibility is important, information feedback has no impact on investment behavior that is in line with MLA. However, our results could have been influenced by our particular stock price process or the manner of subjects’ decision making in our experimental settings. With this in mind, we check the robustness of our results by conducting three extra treatments. In one, we consider a stock price that was in a different upstate or downstate from that of our main treatments at time 1, but which was the same at time 0 and time 2. In treatment FL-IH, subjects could not change the number of their stocks at time 1. According to our results, changing the intermediate stock price in treatment FL-IH has no impact on the

129

An Experimental Analysis of Myopic Loss Aversion

average number of stocks held at time 0. However, if information feedback is an important factor, the average number of stocks at time 0 under the new stock price tree would have been different from that in the original treatment FL-IH. Panel B of Fig. 1 shows our three extra treatments. In treatment FL-IH (price = 9), the downstate stock price at time 1 was 9 dollars instead of 5 dollars, whereas the other conditions were the same as in the original treatment FL-IH. In treatment FL-IH (price = 39), the upstate stock price at time 1 was 39 dollars instead of 25 dollars, whereas the other conditions were the same as in the original treatment FL-IH. These two extra treatments tested whether the stock price at the intermediate time influenced subject behavior. If subjects’ investment behavior remained unchanged, it would confirm that there was no information feedback effect. In the last extra treatment, FL-IH (sell), we tested for the effects of a different initial endowment on treatment FL-IH. In the original treatment FL-IH, subjects were initially endowed with 100 dollars and no units of stock, and then, they were asked how many stocks they wanted to buy. In contrast, in the FL-IH (sell) treatment, subjects were initially endowed with 10 units of stock and no cash, and they had to decide how many units of stock they wanted to sell. If the subjects did nothing, 10 stocks would remain in their accounts. With this treatment, we note the initial endowment effect of

Table 4.

Average Number of Stocks at Time 0 in Extra Treatments. FL-IH (Price = 9)

Average Number of Stocks at Time 0 Average 6.540 (SD) (3.212) N 63 Number of subjects 7

FL-IH (Price = 39)

FL-IH (Sell)

FL-IH

6.900 (2.427) 90 10

5.457 (4.150) 81 9

6.176 (3.446) 306 34

Mann
Whitney z-value

z-value p-value

FL-IH (Price = 9) vs. FL-IH

FL-IH (Price = 39) vs. FL-IH

FL-IH (Sell) vs. FL-IH

−0.663 (0.508)

−1.610 (0.107)

1.174 (0.240)

Note: “FL-IH (price = 9) vs. FL-IH” represents the test of the null hypothesis that the average number of stocks between treatments FL-IH (price = 9) and FL-IH would not differ. The other columns are similar.

130

TOMOKI KITAMURA AND MUNENORI NAKASATO

whether subjects’ behaviors regarding information feedback at stock purchase differed at stock sale. Table 4 shows the average number of stocks in the three extra treatments and in treatment FL-IH. The average number of stocks in treatments FL-IH (price = 9), FL-IH (price = 39), FL-IH (sell), and the original FL-IH was 6.540, 6.900, 5.457, and 6.176, respectively. The differences between the extra treatments and the original treatment FL-IH are not statistically significant. This confirms that the frequency of information feedback has no impact on investment behavior.

CONCLUSION We experimentally reexamined what factors cause MLA. We investigated three factors under multi-period settings: timing of decision, flexibility of investment, and frequency of information feedback. Our overall results are comparable to the findings of Langer and Weber (2008): timing of decision and flexibility of investment explain most of the differences in investor behavior, and the frequency of information feedback makes only a marginal contribution in explaining these differences.10 The differences in investment behavior can be interpreted by a shift in investor reference points depending on the differences in timing of decision, flexibility of investment, or frequency of information feedback. Investor behavior was consistent, with him/her having completely shifted his/her reference points on the current stock price in the most flexible and high frequency of information feedback situation. This shift seems to be the main reason for MLA, namely, the reduction of investment in stock. On the contrary, investors did not adjust their reference points and did not reduce investments in stock during the least flexible and low frequency of information feedback situation. They seemed to adapt their reference points asymmetrically for gains or losses between those two situations. Our results suggest how real household assets should be managed. Investors now frequently trade stocks through the internet. The fact that they have frequent opportunities to change their portfolios according to the fluctuations of stock prices means that they tend to move their reference points to the current stock price and evaluate investments myopically. This causes them to invest less in stock and may lead to lower long-term investment performance. Our results indicate that when the flexibility of investment is high, as in the current internet trading era, planning

An Experimental Analysis of Myopic Loss Aversion

131

investments in advance and gaining adequate advice on investments, such as life planning for retirement, is of great importance.

NOTES 1. See Fig. 2 for stock price assumptions. 2. The reason for setting the same gain or loss at time 0 and time 1, instead of setting the same rate of return for both times, was as follows. As shown in Appendix D, the optimal number of stocks for one-period (myopic) and two-period (long-term) investors who moved their reference points to the current stock price was the same. Accordingly, we could disregard investors’ time horizons and concentrate our analysis on the movement of reference points. 3. This system serves the students of Aoyama Gakuin University; however, we also permitted subjects from other universities, who responded to our recruitment advertisements on the internet, to participate in the experiments. 4. End effects have been evidenced in other experiments, such as that by Langer and Weber (2008), wherein subjects’ payments were decided according to their final wealth in the experiment. Subjects reduced investment in risky assets for the last three periods. Langer and Weber explained that this was due to the conventional wisdom that one should lower risk exposure at the end of the planning horizon. On the contrary, under the random incentive system, the subject is unaware of the trial chosen for the payment. Hence, the subject tends to behave homogeneously with respect to payment among trials. 5. In some (extra) treatments, we paid 15 times or 20 times the final wealth. The final number was determined based on the average wages offered for part-time day jobs and the time taken for the experiment. 6. We were not able to distinguish between subjects who were assigned 0 units due to automatic choice and the subjects who did not invest intentionally. However, we investigate the effect of automatic choice in the fourth section of this chapter. 7. The subjects of treatments FHa-IH and FL-IH had to make all their decisions at time 0. However, these subjects were also able to consider their gain or loss at times 1 and 2 by using their computer at any time during the experiment. 8. The expected payoff for subjects who invested 10 units of stock for both time 0 and time 1 (“full investment strategy”) was 133.3. There was no other strategy that had a higher expected payoff than the full investment strategy. If investors are risk neutral, they would employ the full investment strategy regardless of treatment. Because the expected return and risk of stock differs according to the state, riskaverse investors may employ some other strategies. For example, one can employ the “contrarian strategy” or the “trend follower strategy.” The former includes no investment at time 0, full investment when the stock goes down at time 1, or no investment when the stock goes up at time 1. The latter includes full investment at time 0, no investment when the stock goes down at time 1, or full investment when the stock goes up at time 1. However, the full investment strategy dominates the other strategies in terms of the expected return and risk.

132

TOMOKI KITAMURA AND MUNENORI NAKASATO

9. We explain our result in terms of the movement of reference points; however, it can also be explained by the change of decision weight π, as done by Tversky and Kahneman (1992), depending on the differences in flexibility of investment or frequency of information feedback. 10. Our result appears to be different from that of Bellemare et al. (2005); however, this can be explained in terms of the movement of reference points. Bellemare et al. (2005) used an additive experimental design, which means that all investment opportunities are identical. In their treatment involving high information feedback and high flexibility of investment (H), and high information feedback and low flexibility of investment (M), subjects were compelled to move their reference points perfectly, because stock prices were set at the same price at the beginning of each round. In such an environment, subjects find it difficult to accumulate investment results for multiple periods, tend to evaluate investments separately, and reduce investment. On the contrary, in a treatment involving low information feedback and low flexibility of investment (L), subjects have to decide on three lotteries at the same time, and only the accumulated results of the lotteries are disclosed. In this situation, subjects find it hard to move their reference points and do not reduce investment.

ACKNOWLEDGMENT We are grateful to Takao Kusakawa, Yasuhiro Yonezawa, Kunio Nakashima, Shinichi Hirota, Ryoko Wada, Kenju Akai, Keiko Aoki, and anonymous referee for helpful comments and suggestions. We also appreciate valuable comments from participants at the Association of Behavioral Economics and Finance 2008 Tokyo Meeting, the Economic Science Association 2009 Tucson Meeting, the Economic Science Association 2010 Melbourne Meeting, and the Nippon Finance Association 2010 Meeting. We would like to thank Andrew Palaski for language support. We would like to acknowledge the financial support provided by the Japanese Ministry of Education, Culture, Sports, Science and Technology and NLI Research Institute, Tokyo Japan. Neither institution has any role in the study design, the collection, analysis and interpretation of data, the writing or in the decision to submit for publication of this chapter.

REFERENCES Arkes, H. R., Hirshleifer, A. D., Jiang, D., & Lim, S. S. (2008). Reference point adaptation: Tests in the domain of security trading. Organizational Behavior and Human Decision Processes, 105, 67
81. Bellemare, C., Krause, M., Kroger, S., & Zhang, C. (2005). Myopic loss aversion: Information feedback vs. investment flexibility. Economics Letters, 87, 319
324.

An Experimental Analysis of Myopic Loss Aversion

133

Benartzi, S., & Thaler, R. (1995). Myopic loss aversion and the equity premium puzzle. Quarterly Journal of Economics, 110, 73
92. Gneezy, U., & Potters, J. (1997). An experiment on risk taking and evaluation periods. Quarterly Journal of Economics, 112, 631
645. Gneezy, U., Kapteyn, A., & Potters, J. (2003). Evaluation period and asset prices in a market experiment. Journal of Finance, 58, 821
837. Haigh, M. S., & List, J. A. (2005). Do professional traders exhibit myopic loss aversion? An experimental analysis. Journal of Finance, 60, 523
534. Kahneman, D., & Tversky, A. (1979). Prospect theory: An analysis of decision under risk. Econometrica, 47, 263
291. Kahneman, D., Knetsch, J. L., & Thaler, R. H. (1990). Experimental tests of the endowment effect and the Coase Theorem. Journal of Political Economy, 98, 1325
1348. Langer, T., & Weber, M. (2008). Does commitment or feedback influence myopic loss aversion? An experimental analysis. Journal of Economic Behavior & Organization, 67, 810
819. Mehra, R., & Prescott, E. C. (1985). The equity premium: A puzzle. Journal of Monetary Economics, 15, 145
161. Thaler, R. H. (1980). Towards a positive theory of consumer choice. Journal of Economic Behavior & Organization, 1, 39
60. Thaler, R. H., Tversky, A., Kahneman, D., & Schwartz, A. (1997). The effect of myopia and loss aversion on risk taking: An experimental test. Quarterly Journal of Economics, 112, 648
661. Tversky, A., & Kahneman, D. (1992). Advances in prospect theory: Cumulative representation of uncertainty. Journal of Risk and Uncertainty, 5, 297
323.

134

TOMOKI KITAMURA AND MUNENORI NAKASATO

APPENDIX A: EXPERIMENTAL INSTRUCTIONS (The following instructions were provided to the subjects for all main treatments.)

Thank You for Participating in this Experiment These experimental instructions should be returned to the experimenter after the experiment is finished. Please do not discuss the details of this experiment with any other individual. This experiment is for academic research, and we ask you to take it seriously. Please read the following instructions carefully. They explain how your payment will be determined. Each of you has a subject number, which is shown on the computer in front of you. Please fill in your subject number at the top of this instruction sheet. Do not talk or try to communicate in any manner with other participants during the experiment. If you have any questions, please raise your hand silently. The experimenter will then answer your questions privately.

Your Role and Stock Prices Your role in this experiment is to be an investor in stocks. In this experiment, there are three time periods: time 0, time 1, and time 2. You will start by receiving 100 dollars “cash” (within the experiment) and will have to decide how much you would like to invest in stocks. Please look at Fig. 3, which is attached. The stock price is 10 dollars at time 0. At time 1, it will either increase by 15 dollars to 25 dollars (the probability of this is 1/3) or decrease by 5 dollars to become 5 dollars (the probability of this is 2/3). At time 2, if the stock price was 25 dollars at time 1, it will either increase by 15 dollars to become 40 dollars (the probability of this is 1/3) or decrease by 5 dollars to become 20 dollars (the probability of this is 2/3). At time 2, if the stock price is 5 dollars at time 1, it will either increase by 15 dollars to become 20 dollars (the probability of this is 1/3) or decrease by 5 dollars to become 0 dollars (the probability of this is 2/3). At time 0, the total probability that the stock price will become 40 dollars is 1/9, the total probability that it will become 20 dollars is 4/9, and the total probability that it will become 0 dollars is 4/9. The stock price fluctuations are determined randomly by a computer.

An Experimental Analysis of Myopic Loss Aversion

135

How to Calculate Your Total Assets At time 0, you can buy 1 to 10 units of stock or choose not to invest in any stock. The amount that you buy is automatically deducted from your cash balance. You can buy stock as long as you have cash. For example, if you buy 2 units of stock, with a value of 20 dollars, you will have 80 dollars left. If the stock price increases by 15 dollars to become 25 dollars at time 1, the value of your 2 units of stock will become 50 dollars. You will not get any interest on cash, so your cash will still be worth 80 dollars. Your total assets will then be 130 dollars. If you continue to hold your 2 units of stock, and the price per stock at time 2 appreciates by 15 dollars to become 40 dollars per unit, your stock value will become 80 dollars, whereas your cash will remain worth 80 dollars. Your total assets will then be 160 dollars. If at time 2, the stock price depreciates by 5 dollars to become 20 dollars per unit, your stock value will become 40 dollars, but your cash will still be worth 80 dollars; therefore, your total assets will be 120 dollars. If the stock price depreciates by 5 dollars at time 1 to become worth 5 dollars per unit, your 2 units will be worth 10 dollars, whereas your cash will still be 80 dollars, meaning that your total assets will be 90 dollars. If the stock price then increases to 20 dollars per unit at time 2, your total assets will be worth 120 dollars. If the price per unit decreases to 0 dollars at time 2, your total assets will be worth 80 dollars only. Your investment result is based on your final total assets after time 2. You can change the number of stocks by changing the list box on the software. Your total assets are automatically recalculated when you change your number of stocks. You can freely examine the relationship between the number of stocks and your total assets as the experiment proceeds.

How to Decide on Your Investments (The following instructions apply to treatment FH-IH only.) Please look at Fig. 3, which is attached. When the experimenter says, “Time 0 starts now,” you must choose your number of stocks from the list box and click the “send your decision” button within 1 minute, before the time expires. You can click the “send your decision” button when the blue “able to send” indicator is displayed. You will see the yellow “your decision has been sent” indicator once your decision has been sent. If you see the red “unable to send” indicator, you are not permitted to send your decision for that particular period. If you never click the “send your decision”

136

TOMOKI KITAMURA AND MUNENORI NAKASATO

button, your number of stocks will remain unchanged at 0 units. After your decision has been sent to the experimenter, you will see how the stock price has changed. Next, the experimenter will say, “Time 1 starts now.” You then have to decide your time 1 stock numbers and click the “send your decision” button located in the “time 1” area within 1 minute. When the stock price appreciates at time 1, please use the list box in the upper portion. When the stock price depreciates at time 1, please use the lower portion. You must choose the total number of stocks from the list boxes for period 1, which is different from the number of shares that you want to add or subtract from your period 0 total. (The following instructions apply to treatment FHa-IH only.) Please look at the attached figure (not shown in this chapter). When the experimenter says, “The trial starts now,” select the number of stocks that you initially want from the list box located in the “time 0” area and the number of stocks that you want for time 1 from the list boxes located in the “time 1” area. You have 1 minute to click the “send your decision” button before the time expires. Note: On the computer screen for treatment FHa-IH, there is only one “send your decision” button, which is located in the “time 0” portion of the screen. Both investments for time 0 and time 1 are sent at the same time. (The following instructions apply to treatments FL-IH and FL-IL only.) Please look at the attached figure (not shown in this chapter). When the experimenter says, “The trial starts now,” choose the number of stocks that you want from the list box and click the “send your decision” button within 1 minute, before the time expires. You will not be able to change your number of stocks at time 1, and your number of stocks at time 1 will be the same as that at time 0. Note: On the computer screen for treatment FL-IH, subjects can see the stock price and total assets at time 1, whereas on the computer screen for treatment FL-IL, subjects cannot see the stock price or total assets at time 1. For both treatments, there is only one “send your decision” button, which is located at the “time 0” portion of the screen. (The following instructions apply to all treatments.) When the blue “able to send” indicator appears, you will be able to click the “send your decision” button. Once your decision has been sent, you will see the yellow “your decision has been sent” indicator. When you see the red “unable to send” indicator, you are not permitted to send your decision for the particular period. After your decision has been sent to the

An Experimental Analysis of Myopic Loss Aversion

137

experimenter, you will see the final stock price and your final total assets on your screen. If you do not click the “send your decision” button, your number of stocks will be counted as 0.

Your Payment The process described above represents one “trial.” Several trials will be conducted; however, you will not be told exactly how many. For each trial, you will start by obtaining 100 dollars, and your final total assets at the end of time 2 will be recorded on the computer. When a new trial begins, your total assets from the last trial will not be carried over, and you will start again with 100 dollars. Several practice trials will occur before the actual trials begin. Your payment will be determined based on only one actual trial. After finishing all trials, the computer will randomly decide your payment trial. Your total assets at the end of time 2 of the selected trial will be multiplied by 10, and the resulting number will be given to you in yen as payment. For example, if your total assets for one trial are 200 dollars, for another are 50 dollars, and for a third are 90 dollars, and the computer decides that your payment will be based on the first trial, your payment will be 2,000 JPY. In addition to the experimental reward, you will receive 1,000 JPY for attending. Your payment amount will remain confidential and the payment will be made in cash at the end of the experiment.

138

TOMOKI KITAMURA AND MUNENORI NAKASATO

APPENDIX B: RECORDING SHEET Subject Number Trials

The number of stocks you invest in Time 0

Time 1 25 dollars

Total Assets Time 2

5 dollars

Note: For each trial, subjects were asked to fill in the number of stocks they invested in at time 0, at time 1 when the stock price was 25 dollars, and at time 1 when the stock price was 5 dollars, as well as their final total assets at time 2. This recording sheet was used for treatments FH-IH and FHa-IH. The recording sheet for treatments FL-IH and FL-IL did not include columns for time 1.

An Experimental Analysis of Myopic Loss Aversion

139

APPENDIX C: BRIEF QUIZ AFTER THE PRACTICE TRIALS Q1: How much cash do you have at the beginning of time 0?

dollars

Q2: What is the stock price at time 0?

dollars

Q3: By how much does the stock price increase when it appreciates at time 1? What is the probability that it will appreciate? dollars, % Q4: By how much does the stock price decrease when it depreciates at time 1? What is the probability that it will depreciate? dollars, % Q5: How much cash do you have if you buy six shares of stock at time 0? dollars Q6: What is the stock price when it appreciates at time 1? Q7: Can you trade stock at time 1?

dollars Yes/No

Q8: The stock price goes up at time 1 and goes down at time 2. What is the final stock price? dollars Q9: Imagine that you buy six shares at time 0 and continue to hold the six shares at time 1. If the stock price is the same as in Q8, what will your final total assets be for this trial? dollars Q10: If your payment was determined by the above trial, how much would your experimental reward be? dollars

140

TOMOKI KITAMURA AND MUNENORI NAKASATO

APPENDIX D: OPTIMAL SOLUTIONS FOR INVESTORS USING V1 AND V2 The type V1 expected value function in our stock process is as follows. V1 ðθ0 Þ =

1 2 ð15θ0 Þα − λð5θ0 Þα 3 3

ðD:1Þ

where 0 < α < 1 and λ > 1. The budget constraint is 0 ≤ θ0 ≤ 10. The first order derivative of Eq. (D.1) with respect to θ0 is as follows. ∂V1 ðθ0 Þ 15 10 1 αð15θ0 Þα − 1 − αλð5θ0 Þα − 1 = α5α θα − 1 ð3α − 2λÞ = ∂θ0 3 3 3

ðD:2Þ

The sign of Eq. (D.2) depends on that of 3α − 2λ. Eq. (D.2) increases monotonically if 3α − 2λ > 0 and decreases monotonically if 3α − 2λ ≤ 0. Therefore, Eq. (D.1) is maximized when the budget constraint is binding. The optimal solution for Eq. (D.1) is as seen below. θ0 =




10 0

if if

3α − 2λ > 0 3α − 2λ ≤ 0

The expected value function V2 in our stock process is as seen below. V2 ðθ0 ; θu ; θd Þ =

1 2 [ð15θ0 Þα þ ð15θu Þα ] þ [ð15θ0 Þα − λð5θu Þα ] 9 9 2 4 α þ [ − λð5θ0 Þ þ ð15θd Þα ] þ [ − λð5θ0 Þα − λð5θd Þα ] 9 9

ðD:3Þ

The budget constraints are 0 ≤ θ0 ≤ 10, 0 ≤ θu ≤ ð15θ0 þ 100Þ=25, and 0 ≤ θd ≤ ð − 5θ0 þ 100Þ=5. The maximization problem for Eq. (D.3) can be solved using backward induction. The first order derivative of Eq. (D.3) with respect to θu is as follows. ∂V2 ðθ0 ; θu ; θd Þ 15 10 1 αð15θu Þα − 1 − αλð5θu Þα − 1 = α5α θαu − 1 ð3α − 2λÞ ðD:4Þ = ∂θu 9 9 9 The sign of Eq. (D.4) depends on that of 3α − 2λ. Eq. (D.4) increases monotonically if 3α − 2λ > 0 and decreases monotonically if 3α − 2λ ≤ 0.

141

An Experimental Analysis of Myopic Loss Aversion

Therefore, when the stock price is 25 dollars in period 1, Eq. (D.3) is maximized when the following constraint for θu is binding. θu =




ð15θ0 þ 100Þ=25 0

if if

3α − 2λ > 0 3α − 2λ ≤ 0

ðD:5Þ

Similarly, the first order derivative of Eq. (D.3) with respect to θd is as seen below. ∂V2 ðθ0 ; θu ; θd Þ 30 20 2 αð15θd Þα − 1 − αλð5θd Þα − 1 = α5α θαd − 1 ð3α − 2λÞ = ∂θd 9 9 9 Eq. (D.3) is maximized when the following constraint for θd is binding. θd =




ð − 5θ0 þ 100Þ=5 0

if if

3α − 2λ > 0 3α − 2λ ≤ 0

ðD:6Þ

Substituting Eqs. (D.5) and (D.6) in Eq. (D.3), we obtain the following equation. V2 ðθ0 Þ =

 1 α ð3 − 2λÞ 3⋅ð5θ0 Þα þ ð3θ0 þ 20Þα þ 2⋅5α ð − θ0 þ 20Þα g 9

ðD:7Þ

The first order derivative of Eq. (D.6) with respect to θ0 is as seen below. f ðθ0 Þ≡

 dV2 ðθ0 Þ 1 α = ð3 − 2λÞ 3⋅5⋅αð5θ0 Þα − 1 þ 3αð3θ0 þ 20Þα − 1 dθ0 9  − 2⋅5α αð − θ0 þ 20Þα − 1

The second order derivative of Eq. (D.6) with respect to θ0 is as follows. sðθ0 Þ≡

 d2 V2 ðθ0 Þ 1 α = ð3 − 2λÞ 3⋅52 αðα − 1Þð5θ0 Þα − 2 9 dθ0 2 þ 32 αðα − 1Þð3θ0 þ 20Þα − 2 þ 2⋅5α αðα − 1Þð − θ0 þ 20Þα − 2 g

If 3α − 2λ > 0, sðθ0 Þ < 0 (∵α − 1 < 0). We also have f ð10Þ = ð3 − 2λÞ8αð50Þα − 1 =9 > 0. Then, f ðθ0 Þ > 0 for ∀θ0 ≤ 10. Therefore, Eq. (D.7) α

142

TOMOKI KITAMURA AND MUNENORI NAKASATO

is maximized at θ0 = 10. On the contrary, if 3α − 2λ ≤ 0, sðθ0 Þ ≥ 0. We also have f ð10Þ ≤ 0. Then, f ðθ0 Þ ≤ 0 for ∀θ0 ≤ 10. Therefore, Eq. (D.7) is maximized at θ0 = 0. In conclusion, the optimal solutions for Eq. (D.3) are as seen below. θ0 = θu = θd =




10 if 0 if

3α − 2λ > 0 3α − 2λ ≤ 0

Since these optimal solutions are integers, they are also the optimal solutions for original integer programming.

143

An Experimental Analysis of Myopic Loss Aversion

APPENDIX E: HISTOGRAMS AND CHANGE IN THE AVERAGE NUMBER OF STOCKS FOR EACH TRIAL BY TREATMENT Panel A: Histograms FHa-IH

FL-IH

FL-IL

0 1 0

.5

Density

.5

1

FH-IH

0

5

10 0 Number of stocks

5

10

Graphs by treatment

Panel B: Change in the average number of stocks for each trial by treatment 10

Stock Units

8

6

FH-IH FHa-IH FL-IH FL-IL

4

2

0 Tr1

Tr2

Tr3

Tr4

Tr5

Tr6

Tr7

Tr8

Tr9

CHAPTER 7 DOES EVERYONE ACCEPT A FREE LUNCH? DECISION-MAKING UNDER (ALMOST) ZERO-COST BORROWING Michael Insler, James Compton and Pamela Schmitt ABSTRACT Purpose
This chapter examines the debt aversion of a group of college students who have the opportunity to take out a sizable, low-interest, non-credit dependent loan. If the loan is simply invested in low-risk assets, it would effectively yield a free lunch in net interest earnings. Methodology
The research uses survey data to examine demographic, socio-economic, personality traits, and other characteristics of those willing and unwilling to accept the loan offer, as well as their intentions of early repayment. Findings
Individuals willing to accept the loan tend to have prior debt, longer planning horizons, come from middle-income families, and may have higher cognitive ability. Anticipated early repayment of the loan is more likely among those with prior investments, no prior debt, from

Experiments in Financial Economics Research in Experimental Economics, Volume 16, 145
170 Copyright r 2013 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0193-2306/doi:10.1108/S0193-2306(2013)0000016007

145

146

MICHAEL INSLER ET AL.

STEM majors, with upper income parents, and those who expect to buy a home soon. Research limitations/implications
We find no consistent relationships between debt aversion and intellectual ability or gender, but this finding may be hampered by our small sample of female loan-rejecters. Our limited sample size also precludes examining interactions between the dimensions of personality types. Originality
We suggest consideration of policies to encourage “smart” borrowing, focusing on the financially disadvantaged, particularly for education loans. This study examines a uniquely occurring natural experiment regarding the opportunity to accept a non-credit dependent loan. Our results describe the behavior of young adults, an infrequently studied yet important segment of the population, especially in the context of borrowing behavior. Keywords: Survey methods; behavioral economics; personal finance; investment decisions JEL classification: C83; D03; D14; G11

INTRODUCTION In their junior year, students at the United States Naval Academy (USNA) may accept a sizable loan (up to $36,000) at interest rates at or below market rates of return on even very conservative investments
that is, they are effectively offered a ‘free lunch’ in potential net interest.1 The federal government does not subsidize these loans; the Navy Federal Credit Union (NFCU) and the United Services Automobile Association (USAA) privately offer them in order to establish a relationship with future officers who may then bank with them for many years. Since all graduates enter the armed services, their future employment and salary schedules are stable and nearly identical, making the decision to accept the loan very low risk (all must serve at least five years as junior officers in the United States Navy or Marine Corps). This chapter examines the characteristics of those willing and unwilling to take the loan, characterizing the latter as debt averse. For those who accept the loan, we also study a weaker indicator of debt aversion: whether they anticipate repaying it early. We explore how these two decisions relate to individual and family demographics,

Does Everyone Accept a Free Lunch?

147

socio-economic characteristics, personality traits (as measured by the Myers-Briggs Type Indicator, MBTI), cognitive ability (as measured by the Cognitive Reflection Test, CRT), and intellectual ability (as measured by Scholastic Assessment Test, SAT scores and grade point average, GPA). We view this opportunity as a uniquely occurring natural experiment, as all students receive the same loan offer regardless of credit scores and personal income; in fact, students do not even apply for the loan
lenders directly advertise it to them. Despite the clear benefits associated with the loan opportunity, we observe that 10 percent of the students do not take it; we conjecture that this is primarily due to debt aversion. It is important to understand debt aversion if we hope to design policies and loan opportunities for groups that may have difficulty obtaining credit due to socio-economic characteristics. Recent research has noted that even though credit has loosened since the recession of 2008
2009, young adults have continued to avoid taking on debt (Fry, 2013). While this may be partially due to demographic factors
young adults are marrying later, staying in school longer, and paying down education loan debt instead of taking out mortgages
evidence also suggests an uptick in a “general aversion” to debt (Fry, 2013). It is important to understand such a motive, as it has important implications for the fledgling housing market recovery, future education prospects and costs, and smaller ticket purchases such as vehicles and electronics. Unlike previous studies of high-interest credit card debt or education loans, we isolate debt aversion from all credit-history-related factors as well as motives to invest in human capital. Johnes (1994), Gayle (1996), Hayhoe (2002), and Lyons (2004) focus on credit card debt, while Olson and Rosenfeld (1984), Johnes (1994), Gayle (1996), Eckel, Johnson, and Montmarquette (2005), and Eckel, Johnson, Montmarquette, and Rojas (2007) examine education loans. In our study, the loan is similar to a credit card in that it can be used for any purpose, but it is also similar to an education loan, as it has an extremely low interest rate. Loan recipients may smooth their consumption (Hayhoe, 2002; Lyons, 2004); in fact, anticipating the loan opportunity at the beginning of their junior year, students may take on additional debt in their first two years to invest (Compton, Insler, & Schmitt, 2013) or to absorb personal or family-related financial shocks (Wilson, Findlay, Meehan, Wellford, & Schurter, 2010). Our study utilizes survey data from a sample that is relatively homogenous
college students at USNA
which helps to isolate the characteristics associated with debt aversion (captured by the decision to take out

148

MICHAEL INSLER ET AL.

the loan and the anticipation of early repayment) in a real, high-stakes setting. Despite the similarities of the subjects, we find substantial variation in the choice to take out the loan and to opt for early repayment; approximately 10 percent of students reject the loan and 56 percent of loan-takers anticipate early repayment, and we find that these differences are correlated with a wide range of factors.2 The individual-specific data contains demographic information, including gender, GPA, SAT scores, major, and MBTI; cognitive ability, as measured by Frederick’s (2005) CRT3; self-reported socio-economic characteristics, such as family education levels and income; self-reported information on sources of financial advice; and future expectations about starting a family and purchasing a home. Previous literature has explored how demographics and socio-economic characteristics relate to debt aversion via credit cards and education loans. Johnes (1994) finds that women are less likely to take a subsidized education loan. Xiao, Noring, and Anderson (1995) see no connection between credit card debt and major or gender. Gayle (1996) finds that married students with credit card debt are more likely to take a student loan and that family income factors are not significant predictors of loan acceptance. Constanides, Donaldson, and Mehra (2002) also found that young adults typically do not hold debt. Lyons (2004) finds that financially at-risk students are more likely to misuse credit cards; many such students include minorities and females. Callender and Jackson (2005) determine that prospective students in the United Kingdom from lower income families are more debt averse. Controlling for credit score, credit history, income, employment status, and homeownership, Ravina (2012) finds that personal characteristics significantly affect the likelihood of receiving a loan. Specifically, physically attractive borrowers, whites, and females are more likely to receive a loan at more favorable terms. Due to the nature of our subjects’ loan offer, we are able to test and verify many of these findings in a more controlled environment. Fewer papers examine loan opportunities in true experimental settings. For example, Wilson et al. (2010) have subjects participate in a payday loan. Similar to our survey, these loans require no formal credit check. However, in contrast to ours, their annual percentage rate exceeds 300 percent. In the laboratory, Wilson et al. find that payday loans help recipients deal with financial shocks and aid financial survival, but the researchers do not consider demographic information. Eckel et al. (2005) use experimental subjects from the working poor in Canada to examine

Does Everyone Accept a Free Lunch?

149

high stakes4 decisions on whether subjects choose a smaller sum of money immediately or a larger sum later (ranging from one day to one year). They examine discount rates and focus on the differences between short-horizon and long-horizon investment decisions. They find that age and gender impact patience and that those with low incomes are less patient, while women and older people tend to be more patient. Using the same dataset, Eckel et al. (2007) examine the relationship of education grants and loans to human capital investments. They focus on how individual characteristics (demographics and socio-economics), risk and time preferences, and debt aversion impact individuals’ decisions to take out loans and subsequently enroll in school or training programs. They find that individuals are less likely to accept a loan if the price of the loan increases, if age and wealth levels increase, and if subjects are risk averse. Our study integrates compelling aspects of both experimental and nonexperimental investigations. We collect a wide variety of individual-specific information and study how it relates to subjects’ decisions regarding a purely exogenous loan offer. Linear and non-linear regression methods reveal that, holding all observable factors constant, debt aversion is not consistently associated with gender or intellectual ability. Those willing to accept the loan often have prior debt. Socio-economic background also matters; those with greater family income are more willing to take the loan (except for the wealthiest group). We also uncover some characteristics that explain a weaker brand of debt aversion: respondents’ reported intention to repay the loan early. Women tend to prefer early repayment of the loan, as do individuals with low GPA, STEM5 majors (relative to economics majors), students from the wealthiest families, loan-takers with no prior debt, and “thinkers” (T), based on MBTI score.6 Although we did not conduct a standard risk assessment procedure, such as in Holt and Laury (2002),7 we use Frederick (2005)’s CRT as a measure of patience rather than risk preference. Dohmen, Falk, Huffman, and Sunde (2007) find that risk aversion coincides with lower cognitive ability. We consider CRT as a proxy for risk aversion, but do not find it to be significantly correlated to loan-related decisions. However, signs of CRT’s coefficient estimates associate it with higher probability of taking the loan and lower probability of an intention to repay the loan early, coinciding with the work of Dohmen et al. (2007). The chapter proceeds as follows. The second section discusses the data. The third section presents the methods and results. The last section concludes.

150

MICHAEL INSLER ET AL.

THE DATA The survey was designed for two main purposes. The first, reported here, was to examine a unique opportunity (similar to a high stakes experiment) in which college students have the choice to take out a large, virtually riskfree loan that is not credit-dependent. The opportunity is particularly unique in the context of the previously discovered connections between debt aversion and age, marital status, and personal income
all of which are automatically constant in our sample
establishing a much more controlled natural experiment. The second purpose, reported in a companion paper (Compton et al., 2013), is to examine subjects’ loan allocation decisions.

Loan Details The Naval Academy students entering their junior year may accept a “Career Starter Loan.” The NFCU and the USAA offered this loan, for the class of 2013, in slightly different forms (Table 1 compares the terms of the banks’ loan offers). The loans are not government subsidized; the NFCU and the USAA provide them privately, as they hope to build and maintain banking relationships with newly minted Naval and Marine Corps officers and their families. After accepting the loan, recipients must direct-deposit their military paychecks at their lender’s bank, which precludes taking both loans and ensures (for the bank) at least a minimal financial relationship. Repayment begins three months after graduation. If a student or officer separates from the Naval Academy or the Navy/Marine Corps before repaying the loan, the loan rate reverts to the prevailing signature loan rate at the time. Unlike in Ravina (2012), the ability to (and likelihood of) default on the Table 1. Lender Availability Principal amount Rate Monthly repayment Total repayment Time window

Career Starter Loan Details. NFCU

USAA

August of junior year $32,000 1.25% APR $564.97 $33,898.20 Closes at graduation

October
November of junior year $36,000 0.75% APR $619.80 $37,636.20 Closes one year after commissioning

Does Everyone Accept a Free Lunch?

151

loan is nearly negligible because banks automatically deduct payments from recipients’ paychecks and because all recipients possess excellent job security and stable salaries.

Survey Details The survey questionnaire began with the three CRT questions (Frederick, 2005), which we aggregate into an integer from 0 to 3, counting the number of questions answered correctly. A higher CRT score tends to indicate higher cognitive ability (Dohmen et al., 2007). Students then self-assessed their financial literacy (“subjectively,” on a scale of 1
10) and provided details on their prior assets, prior debt, and which lender (USAA or NFCU) they chose. Students next answered various questions about their specific usage of the Career Starter Loan,8 planned timing of repayment, sources of financial advice, and expectations regarding possible earnings (or losses) on loan-funded investments. Students also answered an “objective” financial literacy question proposed by Pingue (2011),9 as well as a number of socio-economic questions about their family background (education levels for each parent, family income, and financial support given/received to/from parents). Finally, respondents reported future plans including their expected tenure in the military and the timing of their intentions to purchase a home and have children. In addition to the online, self-reported questionnaire, the Office of Institutional Research at USNA merged information for each respondent on gender, major, home state, verbal SAT, math SAT, GPA, MBTI, and class year. The external data helped to minimize reporting errors and survey fatigue. The office also coded the survey instrument, collected the data, and randomly distributed two $100 gift cards to incentivize survey completion. In this way, the researchers were minimally associated with data collection and were at no point involved with identified data. We administered the survey online to students in the classes of 2013 and 2014. In total, 609 students completed the survey. Our response rate was quite high (28 percent), given that the total number of students in those classes in the fall of 2012 was 2,182. To construct our final sample, we discard respondents for whom we could detect critical reporting errors or omissions: We drop four respondents who did not specify whether they took the loan and one respondent whose answers were blatantly incorrect. We drop 49 individuals who did not indicate whether they held any prior investments or debt as well as 15 individuals who did not report their choice of the

152

MICHAEL INSLER ET AL.

NFCU or the USAA loan option. We also omit two individuals who did not report their repayment timing, five who did not report their gain or loss expectations, 26 who did not respond to the financial literacy questions, and 11 respondents who did not report on transfers to or from their family.10 Lastly, we discard information from four severe outliers, with respect to their finances.11 This leaves us with 492 usable observations. Although the number of discarded observations may seem relatively large, our final sample passes a number of simple robustness checks for sample selection bias.12

Myers-Briggs Type Indicator Before starting their freshman year,13 students complete the MBTI exam. Unlike in Yang, Coble, and Hudson (2009), students do not associate MBTI exam responses with their loan acceptance decision. The MBTI captures how individuals process information and behave. It is based on Carl G. Jung’s (1923) theory of psychological types, which argues that certain psychological preferences prefer to perform certain tasks. The MBTI employs four dichotomous dimensions that identify personality types. Individuals may have a preference for either extraversion (E) or introversion (I) in orientation, a preference for sensing (S) or intuition (N) in perception, a preference for thinking (T) or feeling (F) in judgment, and a preference for judgment (J) or perception (P) in attitude.14 Financial behavior literature ties preferences for extraversion (E), intuition (N), thinking (T), and perception (P) to risk-seeking choices (Filbeck, Hatfield, & Horvath, 2005; Li & Liu, 2008; Su-li & Ke-fan, 2010). Unlike these studies, we find only a limited connection between personality type and the decision to accept the loan; those with a preference for thinking (T) are more likely to accept the loan, but they often plan to repay their loans early. As suggested by Myers et al. (1998), this could be because thinking (T) types act “by linking ideas together through logical connections” (p. 24), suggesting that such individuals may be forward looking, perhaps in hopes of establishing good credit.

RESULTS Table 2a presents descriptive statistics, grouped by loan-takers versus loan-rejecters and Table 2b presents descriptive statistics for loan-specific

Seniors in College Femalea SAT (verbal) SAT (math) GPA Major: Economicsa Major: STEMa Major: Othera MBTI: Extravert (vs. introvert)a MBTI: Sensing (vs. intuition)a MBTI: Thinking (vs. feeling)a MBTI: Judging (vs. perceiving)a Mom’s Educ.: Less than Bach. Deg.a Mom’s Educ.: Bach. Deg.a Mom’s Educ.: Graduate Deg.a Dad’s Educ.: Less than Bach. Deg.a Dad’s Educ.: Bach. Deg.a Dad’s Educ.: Graduate Deg.a Family Income: $40,000 or lessa Family Income: $40,001 to $60,000a Family Income: $60,001 to $80,000a Family Income: $80,001 to $100,000a Family Income: $100,001 to $120,000a Family Income: $120,001 or morea Family Income: No responsea Annual transfer from family ($)b Annual transfer to family ($)b Self-reported fin. literacy (1
10 scale)

a

Table 2a.

0.45 0.21 648 679 3.10 0.15 0.64 0.21 0.53 0.71 0.81 0.64 0.29 0.44 0.26 0.26 0.36 0.37 0.08 0.11 0.11 0.14 0.13 0.28 0.16 1,500 1,305 5.90

Mean

Took Loan

0.50 0.41 77 79 0.57 0.35 0.48 0.41 0.50 0.45 0.39 0.48 0.46 0.50 0.44 0.44 0.48 0.48 0.27 0.31 0.31 0.35 0.34 0.45 0.37 1,830 3,364 1.80

Std. Dev. 0.28 0.28 676 688 3.13 0.13 0.64 0.23 0.53 0.70 0.74 0.62 0.17 0.49 0.34 0.13 0.45 0.40 0.06 0.06 0.06 0.15 0.09 0.40 0.17 2,243 634 5.70

Mean 0.45 0.45 73 70 0.59 0.34 0.49 0.43 0.50 0.46 0.44 0.49 0.38 0.51 0.48 0.34 0.50 0.50 0.25 0.25 0.25 0.36 0.28 0.50 0.38 3,696 1,245 2.30

Std. Dev.

Did Not Take Loan

Descriptive Statistics of Loan-Takers and Loan-Rejecters.

Does Everyone Accept a Free Lunch? 153

0.84 0.29 0.10 0.03 0.16 0.01 0.32 0.05 0.05 0.11 0.26 0.28 12.3 0.30 4.4 0.25 6.7 445

Binary variable denotes sample proportion. Statistics are conditional on observations reporting positive values.

b

a

Fin. literacy question (interest/risk)a Held investments prior to loan offera Held debt prior to loan offera Service selection: Restricted linea Service selection: Nukea Service selection: Medical Corpsa Service selection: Naval Aviationa Service selection: Special Forcesa Service selection: Unsurea Service selection: Surface Warfarea Service selection: Marine Corpsa Expected years of service (unsure)a Expected years of serviceb Expected years to house purchase (unsure)a Expected years to house purchaseb Expected years to having kids (unsure)a Expected years to having kidsb Number of Observations:

Mean

Table 2a. Took Loan

0.37 0.45 0.30 0.17 0.37 0.12 0.47 0.22 0.22 0.32 0.44 0.45 8.2 0.46 2.4 0.43 2.5

Std. Dev.

(Continued )

0.77 0.28 0.02 0.09 0.17 0.04 0.26 0.09 0.11 0.04 0.21 0.21 13.0 0.45 4.5 0.30 7.8 47

Mean 0.43 0.45 0.15 0.28 0.38 0.20 0.44 0.28 0.31 0.20 0.41 0.41 8.8 0.50 2.5 0.46 2.4

Std. Dev.

Did Not Take Loan

154 MICHAEL INSLER ET AL.

0.87 0.10 0.46 0.44 0.51 0.51 0.13 0.41 0.06 0.62 0.06 0.31 7.6 7.0 445

Mean

Descriptive Statistics Expectations for Loan-Takers.

Binary variable denotes sample proportion. Statistics are conditional on observations reporting positive values.

b

a

USAA loan (alternative: NFCU loan) Plans to repay in 0
2 yearsa Plans to repay in 2
4 yearsa Plans to repay in 4
6 yearsa Advice: Financial advisora Advice: Friends/familya Advice: News sourcesa Advice: Personal researcha Advice: Othera Expects to gaina Expects to losea Expects to break evena Expected annual gain (%)b Expected annual loss (%)b Number of observations

a

Table 2b. 0.33 0.30 0.50 0.50 0.50 0.50 0.33 0.49 0.23 0.49 0.23 0.46 4.7 4.9

Std. Dev.

Does Everyone Accept a Free Lunch? 155

156

MICHAEL INSLER ET AL.

variables for the loan-takers only. Sample standard deviations account for the known, finite size of the population. At first glance, this simple cut of the data shows some discrepancies between the two groups amongst point estimates of gender (28 percent of loan-rejecters are women, while only 21 percent of loan-takers are women), verbal SAT, family income (particularly for the highest income group), family transfers, and debt holdings. However, given the small sample of loan-rejecters, confidence intervals for population means from each subgroup generally overlap. To better understand such patterns, we next condition on all observable individual-specific information to identify the factors that most strongly relate to loan choices. The following two subsections investigate respondents’ loan acceptance decisions and respondents’ anticipated early repayment of the loan, respectively.

Decision to Take the Loan The loan acceptance decision is a binary response, which we model via linear probability (estimated by OLS) and a probit model (estimated by maximum likelihood). Table 3 contains coefficient estimates and standard errors, which account for the finite, known population size. In all four specifications, the dependent variable is equal to one if a respondent accepted the loan and zero otherwise. The first two columns display probit and linear probability model (LPM) results, respectively, utilizing the full set of controls available in our data. The third and fourth columns present analogous models using a smaller set of covariates. The results do not substantially change as we eliminate less important control variables. The sign and significance of key variables is consistent across probit and LPM; thus we focus our discussion on the easier-to-interpret LPM results. We find that verbal SAT score matters significantly: On average, a 100 point increase in verbal SAT score is associated with a smaller probability of taking the loan by 5 percentage points. Estimates of an alternate measure of intellectual ability, GPA, are positive but insignificant. Cognitive ability, captured by CRT score, is also positive but insignificant. The female dummy variable is negative and insignificant. Academic major is not a significant predictor of the loan-acceptance decision. These results qualitatively agree with previous work (Dohmen et al., 2007; Johnes, 1994; Xiao et al., 1995), but some inferences are not strong. We may simply not have enough observations of loan-rejecters (47 obs.) to tease out weaker relationships. For CRT, at least, in comparing the two probit models, t-ratios increase notably as degrees of freedom increase, suggesting that

(0.189) 0.366 (0.247)

(reference: Bach. Deg.)

Dad’s Educ.: Less than Bach. Deg.

(reference: Bach. Deg.)

(0.0319)

0.0306

(0.0313)

(0.0270) −0.0129

(0.209) −0.0788

Mom’s Educ.: Graduate Deg.

0.0360 (0.0239) −0.0345 (0.0373) −0.000473** (0.000200) −0.000131 (0.000200) 0.0393 (0.0265) 0.000132 (0.0111) −0.0101 (0.0331) −0.00500 (0.0410) −0.00807 (0.0228) 0.00783 (0.0247) 0.0485 (0.0334) 0.0251 (0.0241) 0.0144

LPM (All Controls)

(reference: Bach. Deg.)

Mom’s Educ.: Less than Bach. Deg.

MBTI: Judging (vs. perceiving)

MBTI: Thinking (vs. feeling)

MBTI: Sensing (vs. intuition)

MBTI: Extravert (vs. introvert)

Major: Other (reference: Economics)

CRT

GPA

SAT (math)

SAT (verbal)

Female

Major: STEM (reference: Economics)

**

Probit (All Controls) 0.407 (0.155) −0.142 (0.228) −0.00347*** (0.00134) −0.00112 (0.00138) 0.315* (0.175) 0.0439 (0.0870) −0.243 (0.255) −0.00887 (0.280) −0.101 (0.157) 0.0440 (0.163) 0.344* (0.204) 0.102 (0.156) 0.187

Senior in College

Dependent Variable: “Took loan”

Table 3. Probability of Taking the Loan (N = 492).

−0.142 (0.154) −0.0636 (0.165) 0.378* (0.194) 0.137 (0.151)

0.344 (0.148) −0.190 (0.196) −0.00321*** (0.00130) −0.00161 (0.00130) 0.235 (0.157) 0.0929 (0.08170)

**

Probit (Subset)

−0.0158 (0.0234) −0.000913 (0.0254) 0.0541 (0.0334) 0.0241 (0.0242)

0.0406* (0.0222) −0.0332 (0.0331) −0.000493*** (0.000199) −0.000171 (0.000200) 0.0317 (0.0261) 0.00689 (0.0117)

LPM (Subset)

Does Everyone Accept a Free Lunch? 157

0.0659 (0.0491) 0.0490 (0.0395) 0.0599 (0.0436) 0.0541 (0.0381) 0.0697* (0.0369) 0.0426 (0.0405) −0.0705 (0.0984) −0.00474 (0.0142) 0.221* (0.132) 0.0317 (0.0198) 0.00253 (0.00846) 0.0518 (0.0353) −0.00413 (0.0263) 0.0623*** (0.0228)

0.628* (0.344) 0.469 (0.304) 0.410 (0.329) 0.569** (0.233) 0.666** (0.271) 0.362 (0.253) −0.449 (0.729) −0.0124 (0.103) 1.945** (0.822) 0.286** (0.137) 0.0298 (0.0576) 0.351* (0.199) −0.169 (0.180) 0.884*** (0.334)

Family Income: $40,000 or less (reference: >$120,000) Family Income: $40,001 to $60,000 (reference: >$120,000) Family Income: $60,001 to $80,000 (reference: >$120,000) Family Income: $80,001 to $100,000 (reference: >$120,000) Family Income: $100,001 to $120,000 (reference: >$120,000) Family Income: No response (reference: >$120,000) No transfer from family

Held debt prior to loan offer

Held investments prior to loan offer

Fin. literacy question (interest/risk)

Self-reported fin. literacy (1
10 scale)

ln(Annual transfer to family ($))

No transfer to family

ln(Annual transfer from family ($))

(0.0284)

0.393 (0.173)

0.0425

LPM (All Controls)

(reference: Bach. Deg.)

**

Probit (All Controls)

Dad’s Educ.: Graduate Deg.

Dependent Variable: “Took loan”

Table 3. (Continued )

0.405 (0.294) 0.374 (0.267) 0.257 (0.276) 0.414* (0.222) 0.558** (0.252) 0.233 (0.226) −0.681 (0.690) −0.0582 (0.0965) 1.864*** (0.762) 0.283** (0.126) −0.00309 (0.0540) 0.416** (0.198) −0.0917 (0.175) 0.887*** (0.327)

Probit (Subset)

0.0493 (0.0463) 0.0523 (0.0368) 0.0444 (0.0375) 0.0511 (0.0396) 0.0694* (0.0362) 0.0332 (0.0401) −0.0888 (0.101) −0.00786 (0.0146) 0.256** (0.128) 0.0360* (0.0188) 0.000268 (0.00827) 0.0629* (0.0366) −0.00149 (0.0263) 0.0611** (0.0217)

LPM (Subset)

158 MICHAEL INSLER ET AL.

(reference: Surface Warfare)

Expected years of service (unsure)

*

p < .10, **p < .05, ***p < .01.

Constant

Expected years to having kids

Expected years to having kids (unsure)

Expected years to house purchase

Expected years to house purchase (unsure)

Expected years of service

(0.365)

Service selection: Marine Corps 0.0318 (0.0342) −0.00171 (0.00183) −0.0528 (0.0425) 0.00385 (0.00576) −0.0743 (0.0476) −0.0134** (0.00567) 0.919*** (0.204)

0.258 (0.222) −0.0117 (0.00987) −0.323 (0.274) 0.0376 (0.0424) −0.582* (0.349) −0.0964*** (0.0365) 1.066 (1.381)

(0.0424)

(0.0670) −0.0324

(0.393) −0.424

(reference: Surface Warfare)

(0.0669) −0.107

(0.429) −0.694*

Service selection: Unsure

Service selection: Special Forces

(reference: Surface Warfare)

(0.0401) −0.0448

(0.340) −0.349

(reference: Surface Warfare)

(0.140) −0.0184

(0.567) −0.283

Service selection: Naval Aviation

Service selection: Medical Corps

(reference: Surface Warfare)

(0.0455) −0.123

(0.378) −0.816

(reference: Surface Warfare)

(0.0892) −0.0477

(0.452) −0.542

Service selection: Nuke

−0.171*

(reference: Surface Warfare)

−1.043**

Service selection: Restricted line

−0.427 (0.259) 0.00626 (0.0403) −0.564* (0.329) −0.0944*** (0.0350) 1.858 (1.328)

−0.0611 (0.0410) 0.00156 (0.00590) −0.0682 (0.0461) −0.0129** (0.00571) 0.953*** (0.202)

Does Everyone Accept a Free Lunch? 159

160

MICHAEL INSLER ET AL.

additional data collection may strengthen these findings. The only MBTI coefficient that is statistically significant is for “thinkers” (T), who we associate with having a greater probability of taking the loan. Models incorporating various interactions of MBTI categories yielded no interesting results, likely due to too-high dimensionality relative to the small number of loan-rejecters. Students from upper-middle income families ($100,000
120,000 annually) are almost 7 percentage points more likely to take the loan than students from the highest income families (over $120,000 annually). Estimates for lower income groups are positive but insignificant, in reference to the highest income group. These findings are contrary to Callendar and Johnson (2005) but in line with Eckel et al. (2007). Holding prior debt is associated with a larger chance of taking the loan by 6 percentage points (similar to Gayle, 1996), and reporting no transfer to one’s family adds 26 percentage points to the predicted probability of loan acceptance. Students who expect to have children soon are more likely to accept the loan, holding all other controls constant, compared to students with either long-term or uncertain planning horizons. It is important to note that we cannot infer causal relationships amongst variables in these specifications. On the one hand, there is not likely to be severe simultaneity bias in this setting: It is difficult to imagine how the loan-acceptance decision itself could subsequently influence demographic traits, family background variables, or measured cognition.15 But on the other hand, we do not claim that all selection is explained by the set of observable controls. It is very likely that unexplained factors, such as unseen financial advice, extenuating life circumstances, or non-classical measurement errors, may be correlated with both the loan decision and the explanatory variables. Thus, our inferences reflect correlational relationships, with an underlying conjecture that some may have causal foundations.

Anticipation of Early Repayment Table 4 presents estimations of probit and LPMs for which the outcome is a binary variable equal to one if a subject anticipated early repayment of the loan and zero otherwise.16 The models in the third and fourth columns employ a smaller set of covariates. As before, standard error estimates account for the finite, known population size. The sample includes only the 445 students who decided to take out the loan. It is important to clarify that the estimation results in this subsection pertain only to the subset of

(0.140)

(reference: Bach. Deg.)

(0.0494)

(0.0535) −0.110**

(0.152) −0.305**

Mom’s Educ.: Graduate Deg.

0.160 (0.0430) 0.126** (0.0561) 0.000505 (0.000347) −0.000123 (0.000382) −0.0721 (0.0466) −0.0207 (0.0207) 0.128* (0.0651) 0.0566 (0.0771) −0.0513 (0.0408) −0.0228 (0.0475) 0.101* (0.0520) −0.0139 (0.0432) −0.0648

***

LPM (Full)

(reference: Bach. Deg.)

Mom’s Educ.: Less than Bach. Deg.

MBTI: Judging (vs. perceiving)

MBTI: Thinking (vs. feeling)

MBTI: Sensing (vs. intuition)

MBTI: Extravert (vs. introvert)

Major: Other (reference: Economics)

CRT

GPA

SAT (math)

SAT (verbal)

Female

Major: STEM (reference: Economics)

***

Probit (Full) 0.460 (0.124) 0.352** (0.166) 0.00161 (0.000999) −0.000292 (0.00110) −0.224* (0.135) −0.0561 (0.0588) 0.334* (0.180) 0.136 (0.215) −0.142 (0.117) −0.0680 (0.133) 0.272* (0.152) −0.0373 (0.124) −0.202

Senior in College

Dependent Variable: “Repay soon” 0.376 (0.121) 0.243 (0.156) 0.00142 (0.000967) −0.000129 (0.00107) −0.214* (0.126) −0.0398 (0.0583) 0.396** (0.175) 0.259 (0.209) −0.130 (0.115) −0.0748 (0.129) 0.261* (0.150) −0.0273 (0.121)

***

Probit (Smaller) 0.134*** (0.0429) 0.0879 (0.0539) 0.000451 (0.000344) −0.0000781 (0.000381) −0.0711 (0.0441) −0.0141 (0.0209) 0.151** (0.0638) 0.0974 (0.0766) −0.0487 (0.0411) −0.0242 (0.0469) 0.0943* (0.0527) −0.00802 (0.0434)

LPM (Smaller)

Table 4. Probability That “Plans Early Repayment” of Loan (N = 445, Using Only Those That Took the Loan). Does Everyone Accept a Free Lunch? 161

Fin. literacy question (interest/risk)

Self-reported fin. literacy (1
10 scale)

ln(Annual transfer to family ($))

No transfer to family

ln(Annual transfer from family ($))

(0.141) −0.00850 (0.261) −0.725*** (0.237) −0.152 (0.226) −0.173 (0.193) −0.171 (0.188) −0.268 (0.194) 0.924* (0.487) 0.0862 (0.0703) −0.275 (0.598) −0.0167 (0.0929) −0.0340 (0.0388) 0.141 (0.170)

(0.0499) 0.00289 (0.0899) −0.246*** (0.0820) −0.0503 (0.0774) −0.0606 (0.0687) −0.0581 (0.0658) −0.0892 (0.0698) 0.326* (0.176) 0.0316 (0.0253) −0.139 (0.209) −0.0130 (0.0323) −0.00993 (0.0135) 0.0435 (0.0595)

−0.0197

−0.0649

Dad’s Educ.: Graduate Deg.

(reference: Bach. Deg.) Family Income: $40,000 or less (reference: >$120,000) Family Income: $40,001 to $60,000 (reference: >$120,000) Family Income: $60,001 to $80,000 (reference: >$120,000) Family Income: $80,001 to $100,000 (reference: >$120,000) Family Income: $101,001 to $120,000 (reference: >$120,000) Family Income: No response (reference: >$120,000) No transfer from family

0.0923 (0.0574)

0.266 (0.166)

LPM (Full)

(reference: Bach. Deg.)

Probit (Full)

Dad’s Educ.: Less than Bach. Deg.

Dependent Variable: “Repay soon”

Table 4. (Continued )

0.112 (0.254) −0.545** (0.221) 0.0100 (0.209) −0.125 (0.186) −0.153 (0.186) −0.186 (0.189) 0.825* (0.471) 0.0670 (0.0678) −0.201 (0.583) 0.0000672 (0.0903) −0.0211 (0.0382) 0.191 (0.165)

Probit (Smaller)

0.0377 (0.0874) −0.194** (0.0783) 0.00240 (0.0748) −0.0466 (0.0667) −0.0582 (0.0674) −0.0676 (0.0699) 0.292* (0.171) 0.0241 (0.0247) −0.109 (0.205) −0.00599 (0.0314) −0.00682 (0.0136) 0.0657 (0.0585)

LPM (Smaller)

162 MICHAEL INSLER ET AL.

0.151*

(0.0783) 0.222*

(0.0830) −0.0383 (0.0596) −0.00540* (0.00305) −0.135* (0.0719) −0.0118 (0.0113) −0.0141 (0.0852) −0.00464 (0.0100)

0.451* (0.251) 0.121 (0.491) 0.161 (0.221) 0.626* (0.334) 0.178 (0.317) 0.523** (0.235) −0.126 (0.171) −0.0156* (0.00885) −0.370* (0.205) −0.0322 (0.0314) −0.0469 (0.244) −0.0138 (0.0286)

Service selection: Nuke

(reference: Surface Warfare) Service selection: Medical Corps

(reference: Surface Warfare)

Service selection: Naval Aviation

(reference: Surface Warfare)

Service selection: Special Forces

(reference: Surface Warfare)

Service selection: Unsure

(reference: Surface Warfare) Service selection: Marine Corps

(reference: Surface Warfare)

Expected years of service (unsure)

Expected years to having kids

Expected years to having kids (unsure)

Expected years to house purchase

Expected years to house purchase (unsure)

Expected years of service

(0.113) 0.173**

0.0504

(0.115)

0.0517

(0.171)

(0.0864) 0.0416

(0.132)

(0.381)

(reference: Surface Warfare)

Service selection: Restricted line

Held debt prior to loan offer

0.0805 (0.0493) −0.202*** (0.0641) 0.0617

0.238* (0.140) −0.623*** (0.193) 0.140

Held investments prior to loan offer

−0.0972 (0.167) −0.0165** (0.00836) −0.358* (0.198) −0.0365 (0.0309) −0.0552 (0.238) −0.00432 (0.0273)

0.181 (0.137) −0.580*** (0.187)

−0.0320 (0.0594) −0.00570* (0.00296) −0.132* (0.0702) −0.0136 (0.0112) −0.0180 (0.0848) −0.00131 (0.00980)

0.0663 (0.0493) −0.196*** (0.0632)

Does Everyone Accept a Free Lunch? 163

−0.0289 (0.0639) −0.0329 (0.0431) 0.0185 (0.0428) −0.0116 (0.0638) 0.0308 (0.0490) −0.131 (0.0882) 0.0311 (0.0642) 0.0438 (0.145) −0.00615 (0.00555) 0.00693 (0.0169) 0.428 (0.377)

−0.0971 (0.180) −0.0897 (0.124) 0.0621 (0.123) −0.0195 (0.179) 0.0784 (0.138) −0.400 (0.256) 0.104 (0.184) 0.252 (0.483) −0.0189 (0.0155) 0.00933 (0.0568) −0.351 (1.068)

USAA loan (reference: NFCU loan)

*

p < .10, **p < .05, ***p < .01.

Constant

Expected annual loss (%
if expects loss)

Expected annual gain (%
if expects gain)

Expects to lose (reference: break-even)

Expects to gain (reference: break-even)

Advice: Other

Advice: Personal research

Advice: News sources

Advice: Friends/family

Advice: Financial advisor

LPM (Full)

Probit (Full)

Dependent Variable: “Repay soon”

Table 4. (Continued ) −0.108 (0.175) −0.0927 (0.122) 0.0406 (0.119) 0.00270 (0.180) 0.0367 (0.134) −0.349 (0.251) 0.0366 (0.183) 0.255 (0.448) −0.0144 (0.0154) 0.00732 (0.0519) −0.338 (1.018)

Probit (Smaller) −0.0349 (0.0627) −0.0349 (0.0432) 0.0122 (0.0424) 0.000166 (0.0655) 0.0160 (0.0485) −0.120 (0.0883) 0.0108 (0.0652) 0.0620 (0.142) −0.00501 (0.00559) 0.00452 (0.0160) 0.445 (0.363)

LPM (Smaller)

164 MICHAEL INSLER ET AL.

Does Everyone Accept a Free Lunch?

165

students who accept the loan. We can envision a two-stage econometric specification that sequentially models the students’ two decisions: First, they choose whether to take out the loan, and second, they report their repayment-timing intentions. Due to data limitations
predominantly low sample size of loan-rejecters
we have been unable to obtain viable results from such a model. From the second column of Table 4, we find the predicted probability of anticipating early repayment to be greater, on average, by 12.6 percentage points for women, 12.8 percentage points for STEM majors (versus economics majors), 10.1 percentage points for those with a preference for thinking (T), and 8 percentage points for those who held investments prior to the loan (although only significant in the probit model). The predicted probability is lower by 7.2 percentage points per one-point increase in student GPA (although only significant in the probit model), and it is lower by 13.5 percentage points for those uncertain about a future home purchase. Family income dummies are negative, but only one is statistically significant; there is evidence that students from lower income backgrounds are less likely to intend to repay the loan early. As stated earlier, we view the expectation of early repayment as a weaker form of debt aversion, as compared to the decision to reject the loan altogether. In this regard, the results in Table 4 are generally analogous to the results in Table 3. In Table 4, lower family income relates to standard loan repayment plans (less debt aversion), while in Table 3, lower family income predicts a greater likelihood of loan acceptance (also indicative of less debt aversion). In Table 3, cognitive ability (as measured by CRT score) is not significantly related to loan-acceptance, but its coefficient estimate’s positive sign suggests less debt aversion. In Table 4, we again do not capture a significant relationship for CRT score, but a negative slope estimate implies a potential alignment with less debt aversion. We see similarly consistent relationships
with some statistically significant and some not
between the two tables for gender, college major (STEM versus economics), prior debt holdings, and planning horizon variables (at least when comparing family planning and house-purchase planning). The only statistically significant inconsistencies between the two tables’ predictions of debt aversion come from estimates for Myers-Briggs thinking (T) types17 and intellectual ability.18 As mentioned in the Data Section under Myers-Briggs Type Indicator, we conjecture that since thinking (T) types are particularly strong at logically connecting ideas (Myers et al., 1998), such individuals may be forward looking, perhaps in hopes to establish good credit. Since the literature does not take a stand on intellectual ability, we leave this

166

MICHAEL INSLER ET AL.

factor as open to further investigation. It is also true that the results in Table 4 reflect debt aversion specifically for the subsample of students who are already revealed to be sufficiently “debt-loving” to have taken out the loan in the first place. This pre-condition may also contribute to such discrepancies. As before, we cannot infer causal relationships between regressors and loan-repayment anticipations. Although it is indeed possible that, for instance, an orientation toward thinking (T) directly affects one’s repayment plans, it is also likely that other unobserved traits correlated with personality type, such as specific family circumstances, contribute to the effect. Simultaneity bias is another likely culprit in these models; for instance, repayment plans may impact the choice of USAA versus NFCU loans because of their differences in timing and terms.

CONCLUSION Despite the challenges posed by causal inference, we establish several compelling links between debt aversion and observable characteristics. In this chapter, we exploit a natural experiment
USNA students’ Career Starter Loans
to verify and strengthen knowledge of debt aversion. Furthermore, our results directly describe the behavior of young adults, an infrequently studied yet important segment of the population, especially in the context of borrowing behavior. To summarize, we find limited evidence of debt aversion in women, but this finding may be hampered by our small sample of female loan-rejecters. Measures of intellectual ability do not have consistent relationships with debt aversion; a superior verbal SAT score predicts a lower likelihood of taking the loan, but GPA portends greater chances of early repayment. Thinking (T) types are similarly conflicted. We find that subjects from upper-middle income families, who do not contribute money to their family, who hold prior debt, and who expect children earlier are more likely to accept the loan. Anticipated early repayment of the loan is more likely among those with prior investments, no prior debt, from STEM majors, with upper income parents, and those who expect to buy a home soon. As there is not a statistically significant relationship between CRT score and the probability of taking the loan, cognitive ability and debt aversion may not be related; coefficient estimates’ signs provide weak evidence that they are negatively linked.

167

Does Everyone Accept a Free Lunch?

To summarize our main results, we have framed the loan-acceptance decision
or rather, loan-rejection
as “debt aversion,” and we build on previous work that investigates related factors. Along with studies such as Johnes (1994), Gayle (1996), Eckel et al. (2005), and Eckel et al. (2007), we find links between individuals coming from “better off” backgrounds (i.e., from wealthier roots, no prior debt, higher cognitive ability) and choosing optimally, which we view as taking the free lunch. Why are individuals with less ability or from less fortunate backgrounds more debt averse? Such students may see no credible way to commit to not transferring some of the principal to family, thus accepting the additional repayment burden that would accompany such a choice. Therefore, they would have a greater propensity to reject the loan altogether. Interestingly, financial literacy and advice variables are not connected to debt aversion in our results.19 Could educational interventions help potential borrowers make more informed decisions? We suggest consideration of policies to encourage “smart” borrowing, focusing on the financially disadvantaged, particularly for education loans.

NOTES 1. For example, the United Services Automobile Association (USAA) offers a $36,000 loan at 0.75% annual percentage rate. According to the U.S. Treasury in February 2013, the annual yield on a 5-year security was roughly 0.9%; investing $36,000 at 0.9% for five years, compounded annually, would return $37,649. But in that time, cumulative interest payments for the 0.75% USAA loan would only total about $670. Thus, students can profit nearly $1,000, virtually risk-free. Therefore, the decision to reject the loan is akin to refusing a free lunch. 2. We define students anticipating early repayment as those who report that they expect to repay the loan no later than four years after accepting it. Standard repayment plans require students to complete repayment six years after accepting the loan. 3. The CRT consists of three basic questions: (1) A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost? ___ cents; (2) If it takes 5 machines 5 minutes to make 5 widgets, how long would it take 100 machines to make 100 widgets? ____ minutes; and (3) In a lake, there is a patch of lily pads. Every day, the patch doubles in size. If it takes 48 days for the patch to cover the entire lake, how long would it take for the patch to cover half of the lake? ___ days. The intention of the CRT is to test two decision-making characteristics: time preference and risk preference. The “intuitive” answers to these three questions are generally incorrect, and people who answer them incorrectly tend to believe they are easier than those who answer correctly. Frederick (2005) finds the correlation of CRT score to SAT, ACT, and other tests to be weak.

168

MICHAEL INSLER ET AL.

He hypothesizes that, more than any of the tests the CRT is compared to, it measures the need for resisting the first “intuitive” answer and instead solving for the correct solution to a problem, and therefore it functions as a test of cognitive ability. 4. In their experiments, earnings averaged $130 and ranged from $0 to $600. 5. Refers to disciplines in the fields of science, technology, engineering, and mathematics. 6. The Data Section under Myers-Briggs Type Indicator contains details on MBTI. 7. Because students at the USNA are paid for their service while students, they cannot receive payments for participation in experiments. Because of this, we decided not to use the risk assessment tool as shown in Holt and Laury (2002). In almost every study using Holt and Laury (2002), actual payments are used as an incentive; Jacobson and Petrie (2009), however, do not pay all subjects and they find no difference if payoffs are real or hypothetical. Camerer and Hogarth (1999) find ambiguity. To avoid ambiguity, acknowledging the fact we cannot pay subjects, we focus on the CRT and patience rather than Holt and Laury (2002)’s riskaversion assessment. 8. See Compton et al. (2013). 9. The “objective” and “subjective” financial literacy questions have a moderate positive correlation of 0.27. The question from Pingue (2011) is: “Which of the following best describes the relationship between the interest rate charged to a person for a loan and that person’s risk of nonpayment of the loan?” Answer options: (a) Lower interest rates are charged on loans with a lower risk of nonpayment; (b) Higher interest rates are charged on loans with a lower risk of nonpayment; (c) Lower interest rates are charged on loans with a higher risk of nonpayment; (d) I don’t know. 10. The reader should bear in mind that the accounting of discarded observations is reported sequentially. We have recounted the omitted responses in the same order as the survey’s questions. Thus, recalling that the survey was administered online, the drops predominantly represent submissions of incomplete surveys at various stages of survey completion. 11. One student reported a transfer to his family far greater than the size of the loan; one student reported debt holdings far greater than the size of the loan; two students reported prior investments well over $100,000. 12. We observe the official data (Class year, gender, MBTI, GPA, SAT, major, home state) from the Office of Institutional Research, even for “discarded respondents” who did not submit complete surveys. We performed t-tests for equal means (and proportions where appropriate) between the truncated and un-truncated samples. Even at the 10 percent significance level, we could not reject mean-equality for class year, gender, MBTI, GPA, or SAT. For the same set of variables, we could not reject variance-equality when performing F-tests for equal variances between the two samples. We did not check for sample consistency on major or home state due to high dimensionality of these variables. 13. Freshmen participate in “plebe summer” for approximately two months prior to the start of academic year in August. During this time, students receive the MBTI and other class placement exams. 14. For more detailed information on the MBTI, see Myers, McCaulley, Quenk, and Hammer (1998).

Does Everyone Accept a Free Lunch?

169

15. One exception is the case of planning horizon variables. A $36,000 windfall in one’s junior or senior year of college may certainly influence his or her plans to purchase a home, remain in the armed services, or to have children. 16. We consider loan-takers to anticipate early repayment if they report plans to repay the loan in four years or fewer, versus the standard six-year plan. Ten percent of loan-takers report plans to repay within two years; 46 percent anticipate repayment in 2
4 years. 17. Table 3 associates thinkers (T) with the less debtaverse choice, while Table 4 associates them with the more debt-averse option. 18. In Table 3, verbal SAT scores reveal that more intellectually able subjects are, on average, more debt averse, while in Table 4, we see that individuals with greater GPA’s tend to be less debt averse. 19. Compton et al. (2013) determine that these variables matter much more as they relate to investment behavior, once students have made the decision to accept the loan.

ACKNOWLEDGEMENTS The authors would like to thank Lou Cox for her expertise in programming the survey. We would also like to thank Kurtis Swope, participants at the 2012 Southern Economic Association Meetings in New Orleans, the coeditor, and an anonymous referee for helpful comments. IRB human subject’s approval received on September 12, 2012 under HRPP Approval #USNA.2012.0004-AM03-EP7-A.

REFERENCES Callender, C., & Jackson, J. (2005). Does the fear of debt deter students from higher education? Journal of Social Policy, 34, 509
540. Camerer, C. F., & Hogarth, R. M. (1999). The effect of financial incentives in experiments: A review and capital-labor-production framework. Journal of Risk and Uncertainty, 19(1
3), 7
42. Compton, J., Insler, M., & Schmitt, P. (2013). “Open line of credit:” Under no borrowing constraints, how do young adults invest? Unpublished manuscript. Dohmen, T., Falk, A., Huffman, D., & Sunde, U. (2007). Are risk aversion and impatience related to cognitive ability?, IZA Discussion Papers, No. 2735, urn:nbn:de:101:1-20080402396, http://hdl.handle.net/10419/34592 Eckel, C., Johnson, C., & Montmarquette, C. (2005). Saving decisions of the working poor: Short- and long-term horizons, Field Experiments in Economics: Research in Experimental Economics (10, 219
260). Eckel, C., Johnson, C., Montmarquette, C., & Rojas, C. (2007). Debt aversion and the demand for loans for postsecondary education. Public Finance Review, 35(2), 232
262.

170

MICHAEL INSLER ET AL.

Filbeck, G., Hatfield, P., & Horvath, P. (2005). Risk aversion and personality type. The Journal of Behavioral Finance, 6(4), 170
180. Frederick, S. (2005). Cognitive reflection and decision making. Journal of Economic Perspectives, 19(4), 25
42. Fry, R. (2013). Young adults after the recession: Fewer homes, fewer cars, less debt. Pew Research Center, Technical Report. Gayle, V. (1996). The determinants of student loan take-up in the United Kingdom: Another gaze. Applied Economics Letters, 3, 25
27. Hayhoe, C. (2002). Comparison of affective credit attitude scores and credit use of college students at two points in time. Journal of Family and Consumer Sciences, 94, 71
77. Holt, C., & Laury, S. (2002). Risk aversion and incentive effects. American Economic Review, 95(5), 1644
1655. Jacobson, S., & Petrie, R. (2009). What do inconsistent choices over risk tell us? Journal of Risk Uncertainty, 38, 143
158. Johnes, G. (1994). The determinants of student loan take-up in the United Kingdom. Applied Economics, 34(10), 999
1005. Jung, C. G. (1923). Psychological types. Princeton, NJ: Princeton University Press. Li, S., & Liu, C. (2008). Individual differences in a switch from risk-averse preferences for gains to risk-seeking preferences for losses: Can personality variables predict the risk preferences? Journal of Risk Research, 11(5), 367
686. Lyons, A. C. (2004). A profile of financially at-risk college students. The Journal of Consumer Affairs, 38(1), 56
80. Myers, I., McCaulley, M., Quenk, N., & Hammer, A. (1998). MBTI manual: A guide to the development and use of the Myers Briggs Type Indicator. Palo Alto, CA: Consulting Psychologists Press. Pingue, L. (2011). Can you pass a high school financial literacy test? March 9, 2011. Reuters. Ravina, E. (2012). Love & loans: The effect of beauty and personal characteristics in credit markets, Columbia University Working Paper. Su-li, Z., & Ke-fan X. (2010). Research on the entrepreneurial team members’ personality traits influence on group risk decision-making. 2010 International conference on Management Science & Engineering (17th). November 24
26, 937
942. http://ieeexplore.ieee.org/xpl/ articleDetails.jsp?arnumber = 6328181 Wilson, B., Findlay, D., Meehan J. W., Wellford, C., & Schurter, K. (2010). An experimental analysis of the demand for payday loans. The B.E. Journal of Economic Analysis & Policy, 10(1) (Topics), Article 93. Xiao, J., Noring, F., & Anderson, J. (1995). College students’ attitudes toward credit cards. Journal of Consumer Studies and Home Economics, 19, 155
174.