Economics of corruption and crime: an interdisciplinary approach to behavioral ethics / by Eugen Dimant, M.Sc. in Business Sciences; M.Sc. in International Economics

ECONOMICS OF CORRUPTION AND CRIME: AN INTERDISCIPLINARY APPROACH TO BEHAVIORAL ETHICS A DISSERTATION SUBMITTED TO THE FACULTY OF BUSINESS ADMINISTRATION AND ECONOMICS UNIVERSITY OF PADERBORN IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF ECONOMIC SCIENCES - DOCTOR RERUM POLITICARUM - GRADE: SUMMA CUM LAUDE BY EUGEN DIMANT DATE OF BIR TH : 10 SEP TE MBER 1986 PLACE OF BIR T H: CHIȘINĂU, MO LD OV A M.SC. BUSINE SS SCIE N CE S : WITH HIG H DI ST IN C TION & VA L ED I CT OR I A N , 2014 M.SC. INTER NA T IO NAL EC ON O MI C S: WIT H HIG H DIS T INCTION & VA L ED I C T OR I AN , 2012 DECEMBER 2015 ii DEDICATED TO MY PARENTS GALINA AND LEONID DIMANT AND MY BELOVED ONES. iii ACKNOWLEDGMENTS This dissertation would have not been possible without the persistent and substantial support, encouragement, and assistance of people from both the academic and social sphere, including my doctoral supervisor and distinguished committee members as well as my family and numerous friends. I begin by expressing my deepest gratitude and appreciation to my doctoral supervisor Professor Burkhard Hehenkamp, whose encouragement, constructive feedback, and support made this dissertation possible. He fully supported me with helpful academic and personal advice, and helped me persevere through my projects. His interest in both experimental economics and behavioral ethics was contagious, and inspired my line of research. He encouraged my development by providing me ample freedom and a conducive environment. For these, and many other reasons, I am very indebted to everything Burkhard has done for me. I would also like to thank Professor Tim Krieger who has been an inspiration since my Bachelors. He sparked my interest in economics, has been the supervisor of both my Bachelor’s and Master’s thesis, and a great colleague and collaborator on numerous of my research projects ever since. Tim is also the reason why I decided to stay in academia. Beyond establishing a professional relationship, Tim always had a sympathetic ear for all sorts of personal and academic questions along the way, for which I am very thankful. In addition, I would like to thank the remaining members on my committee, namely Professor René Fahr, Professor Luigi Mittone, and Professor Ulrich Schmidt for having triggered my interest in experimental economics, as well as being great colleagues and inspirations to my work. I am also very grateful for the chance to develop my skill set and learn from the best during my visiting researcher positions at various US institutions. In particular, I want to express my deepest appreciation to Professor Daniel Houser, Professor Lawrence Lessig, Professor Cristina Bicchieri, and Professor Gary Bolton for hosting me at the Interdisciplinary Center for Economic Science (ICES) at the George Mason University, the Behavioral Ethics Lab (BeLab) at the University of Pennsylvania, the Edmond J. Safra Center at Harvard University, iv and the Center and Laboratory for Behavioral Operations and Economics (LBOE) at the University of Texas at Dallas, respectively. Finally, yet importantly, I would like to thank my parents who have always been extremely supportive of all of my endeavors and encouraged me to follow my passions. Thank you for playing the utmost important role in my life and making me who I am. Sincere thanks are also due to my colleagues and beloved friends who gave advice and support and suffered through hours of my muttering about research. In particular, I would like to highlight Margarita Staschewski’s contribution, as she had to endure various research-related conversations and stories over the last couple of years, as well as for repeatedly proofreading my research. All of you have played an integral role in my personal and scholarly development. This dissertation has been an equally arduous and rewarding journey, and I am grateful to each and every one who was part of this enchanting venture. v CONTENTS Acknowledgments......................................................................................................................................... iii Contents.......................................................................................................................................................... v Introduction .................................................................................................................................................... 8 Chapter 1: Survey ................................................................................................................................... 12 The Nature of Corruption: An Interdisciplinary Perspective .................................................. 12 1.1 Introduction ............................................................................................................... 13 1.2 History of Corruption and Corruption Research......................................................... 15 1.3 Facets and Determinants of Corruption ..................................................................... 17 1.3.1 Internal World – Rational Choice & Behavioral Perspective...................................... 19 1.3.2 Meso World—Sociological & Criminological Factors ................................................. 22 1.3.3 External World—Economic, Legal, Political, Historical and Geographical Factors 25 1.3.4 Interdisciplinary Perspective and Empirical Findings ............................................... 28 1.4 Conclusion ................................................................................................................. 29 Chapter 2: Empirics................................................................................................................................ 32 2.1 The Effect of Corruption on Migration, 1985-2000 ...................................................... 32 2.1.1 Introduction ............................................................................................................ 33 2.1.2 Data and Methodology ............................................................................................ 34 2.1.3 Empirical results ...................................................................................................... 37 2.1.4 Conclusion............................................................................................................... 40 2.2 A Crook is a Crook … But is He Still a Crook Abroad? On the Effect of Immigration on Destination-Country Corruption ................................................................................. 42 2.2.1 Introduction ............................................................................................................ 43 2.2.2 Theoretical Considerations ..................................................................................... 45 2.2.3 Empirical Analysis ................................................................................................... 49 2.2.4 Data ......................................................................................................................... 51 2.2.5 Empirical Results ..................................................................................................... 57 2.2.6 Conclusion............................................................................................................... 71 Chapter 3: Experiments........................................................................................................................ 73 3.1 On Peer Effects: Behavioral Contagion of (Un)Ethical Behavior and the Role of Social Identity ........................................................................................................................ 73 3.1.1 Introduction ............................................................................................................ 74 3.1.2 Background and Course of Investigation ................................................................. 80 3.1.3 The Conceptual Framework of Behavioral Adaptation ............................................ 85 vi 3.1.3.1 The Mechanism of Social Interaction and Challenges of Measuring its Effects ... 85 3.1.3.2 Why Understanding Behavioral Adaptation Matters .............................................. 90 3.1.3.3 What We Know: On Peers, Behavioral Adaptation, and Neighborhood Effects ... 92 3.1.4 Drivers of Behavioral Contagion: An Interdisciplinary Perspective .......................... 98 3.1.4.1 Concepts in Economics .............................................................................................. 99 3.1.4.2 Concepts in (Social) Psychology ............................................................................. 106 3.1.5 The Role of Social Identity and Observed (Un)Ethicality in Behavioral Contagion 111 3.1.5.1 Conceptual Introduction .......................................................................................... 111 3.1.5.2 Theoretical Model ..................................................................................................... 115 3.1.6 The Experiment ..................................................................................................... 120 3.1.6.1 Experimental Design and Procedure ...................................................................... 120 3.1.6.2 Hypotheses ................................................................................................................ 124 3.1.6.3 Results and Discussion ............................................................................................. 125 3.1.7 Lessons Learned: Policy Implications .................................................................... 135 3.1.8 Conclusion and Outlook ........................................................................................ 138 3.1.9 Appendix ............................................................................................................... 140 A: Data Overview, Robustness Checks, and Additional Results ............................... 140 B: Social Identity Statements ....................................................................................... 148 C: Experimental Instructions........................................................................................ 149 D: Screenshots of Decision Screens............................................................................. 152 3.2 Tax Evasion Revised: Surprising Experimental Evidence on the Role of Principal Witness Regulations and Differences in Gender Attitudes ........................................ 156 3.2.1 Introduction .......................................................................................................... 157 3.2.2 Experimental Design ............................................................................................. 158 3.2.3 Predictions and Hypothesis ................................................................................... 160 3.2.4 Results and Discussion .......................................................................................... 161 3.2.5 Conclusion............................................................................................................. 166 3.2.6 Appendix ............................................................................................................... 167 Bibliography ............................................................................................................................................... 168 vii Confirmation of co-authorship: “With my signature, I confirm that Eugen Dimant has contributed at least the equivalent effort of 1/n to the paper that I have co-authored (with n denoting the number of authors on the respective paper).” Chapter Project Name Total Number of Authors Name of CoAuthor(s) Signature of Co-Authors Published in a Peer-Reviewed Journal1 1 The Nature of Corruption: An Interdisciplinary Perspective 2 Thorben Schulte  2.1 The Effect of Corruption on Migration, 1985-2000 3 Tim Krieger  Daniel Meierrieks 2.2 A Crook is a Crook … But is He Still a Crook Abroad? On the Effect of Immigration on Destination-Country Corruption 3 Tim Krieger  Margarete Redlin 3.1 On Peer Effects: Behavioral Contagion of (Un)Ethical Behavior and the Role of Social Identity 1 - - 3.2 Tax Evasion Revised: Surprising Experimental Evidence on the Role of Principal Witness Regulations and Difference in Gender Attitudes 3 Johannes Buckenmaier Luigi Mittone 1 At the time of this dissertation’s submission. 8 INTRODUCTION Due to its adverse nature, the study of illicit behavior has taken a center stage in current economic research. Illicit behavior can take various forms, such as corruption, fraud, and tax evasion, but they are all generally detrimental to both society and economy, and hamper our social lives. Understanding the drivers of illicit behavior and how they interrelate with institutional factors not only draws a clearer picture of how and why individuals behave in illegitimate ways, but also how they facilitate the creation and implementation of more efficient and effective policy measures. These motivations are at the heart of this dissertation. The collection of scientific works in this dissertation build upon an interdisciplinary approach. By applying survey techniques, empirical investigations of observational data, as well as experimental methodology, the research presented here sheds light on topics related to the economics of crime and corruption. I employ the encompassing term behavioral ethics to stress the fact that my research gives priority to assessing the drivers and consequences of individual decision-making in (un)ethical settings. I follow Bazerman and Gino (2012, p. 85) in defining the term behavioral ethics as “the study of systematic and predictable ways in which individuals make ethical decisions and judge the ethical decisions of others when these decisions are at odds with intuition and the benefits of the broader society.” I expand this perspective by enriching the discussion on decision-making within the unethical sphere, and contribute to a broader understanding of (un)ethical behavior. A total of five papers are arranged according to their topic in three distinct chapters. The structure of this dissertation follows a macro-to-micro approach, which is conducive to leading a comprehensive discussion on illicit behavior. The order in which my research projects are discussed in this dissertation transitions from a high-level approach to a concise behavioral investigation at the individual level. To set the stage, I first discuss the status-quo of existing theories and empirical research on the drivers of illicit behavior with a particular focus, but not limited to, corruption, before investigating the interrelation between illicit behavior and institutional environments employing empirical techniques of observational data analyzing the bilateral interdependence between corruption and migration. I complete 9 the picture with a narrow microanalysis on the individual drivers of illicit behavior with a particular focus on behavioral spillovers. In more detail, Chapter 1 deals with the general overview and discussion of the economics of corruption from an interdisciplinary perspective (Dimant & Schulte, forthcoming). Corruption, as the predominant part of organized crime, has fierce impacts on economic and societal development, with estimates suggesting the direct cost from corruption to exceed $1 trillion on a global scale (Transparency International, 2011a). As Gire (1999, p. 1) explains it, “corruption is one of the most dangerous social ills of any society. This is because corruption, like a deadly virus, attacks the vital structures that make for society’s progressive functioning, thus putting its very existence into serious peril.” By definition, corruption represents a hidden action distending under the surface of our daily life, rendering reliable estimates of its magnitude and pervasiveness nearly impossible. Nonetheless, and to the best of our knowledge, corruption has soaked through entire parts of society and the economy, both in subliminal and pervasive forms (cf. Rose-Ackerman & Soreide (2011)). It also has become a more publicly discussed topic due to increased media coverage, recently driven by the FIFA corruption scandals and exposed cases of performance enhancing drugs in professional sports (Dimant & Deutscher (2015)). Consequently, in order to understand the bigger picture of corruption it is important to break the underlying mechanisms down into antecedents and effects of corruption at the micro, meso, and macro level. This represents this chapter’s paper, arguing that only the consideration of rational and behavioral aspects, as well as sociological, criminological, and institutional factors and their interaction paints us a comprehensive picture of why, how, and to what extent, individuals engage in illicit behavior in general and corrupt behavior in particular. In Chapter 2, the empirical analyses of observational data shed light on the impact of corrupt institutional environments on migration decisions as well as on the reverse link that is the impact of immigration on destination country’s corruption levels. In the paper by Dimant, Krieger & Meierrieks (2013) discussed in Chapter 2.1, we shed light on the role of corruption in triggering emigration flows from corruption-ridden countries. We examine the influence of corruption on migration for 111 countries between 1985 and 2000. Robust evidence indicates that corruption is among the push factors of migration, especially fueling 16 Still, for a long time, corruption was mainly a research topic in the fields of political, sociological, historical and criminal law research. In the 1960s and 1970s, general approaches to assessing the mechanism of corruption created an ambiguous picture of its overall effects. Due to a lack of reliable data and methodological issues, economic research remained largely silent (Myrdal, 2011). At that time, conflicting interests between politicians and researchers were preventing corruption research from advancing. For example, trying to receive a visa for a possibly corruption-ridden country was almost impossible at that time if the trip’s purpose—a corruption study—was mentioned (Nye, 1967). On top of that, research on corruption had suffered from disagreement on a formal definition and the context dependency of an act, which may fall under the definition of corruption in one country but not in another. One of the first oft-recited definitions was coined by Nye (1967, p. 419):“Corruption is behavior which deviates from the formal duties of a public role because of private-regarding (personal, close family, private clique) pecuniary or status gains; or violates rules against the exercise of certain types of privateregarding influence.” One drawback of this definition is the inherent ambiguity, because “all illegal acts are not necessarily corrupt and all corrupt acts are not necessarily illegal.” (Peters & Welch, 2011, p. 155) In certain societies, particular actions may already be considered a form of corrupt misconduct, whereas in other societies these acts may well be part of their “formal duties” and “just politics.” (Peters & Welch, 2011) Starting in the late 1980s and early 1990s, sound theoretical approaches facilitated the scholarly efforts to study the mechanism of the economics of corruption. Especially in light of the economic acceleration of Asian countries at that time, research was still unsettled on whether corruption exhibits only adverse effects on societies and economics—sanding the wheels—or might create positive effects—greasing the wheels—under certain circumstances through the reduction of inefficient red tape (Dreher & Gassebner (2013), Vial & Hanoteau (2010)). Today, this argument was settled by sound research, indicating that corruption above all is detrimental to the society. These results are now broadly accepted. Through the use of more sophisticated methodological approaches and more reliable data, current research has settled on the fact that the general and long-term detrimental effects of corruption outweigh the context-specific and short-termed positive effects (Aidt (2009), 17 Méon & Sekkat (2005)). The broader availability of huge datasets was key for this development. For example, the PRS Group introduced the “International Country Risk Guide” in 1984 and Transparency International established the Corruption Perception Index as one of the most acknowledged measurements in 1995. In the 1990s and after the end of the Cold War, the first global anti-corruption movements occurred along with the democratization process of many developing countries. Ever since, the media has become increasingly involved in a critical assessment of corruption, drawing the public’s attention to its consequences (Lambropoulou, et al., 2005). In what follows, the mechanism, the antecedents, and the effects of corruption will be discussed from an interdisciplinary perspective on the micro, meso, and macro level. 1.3 Facets and Determinants of Corruption The next section centers on the interdisciplinary nature of corruption research. In our attempt to blend different theories from various areas, we introduce a structural framework that allows us to discuss corruption stepwise, from what we refer to as the inner-to-outerworld approach. For this reason, we start with the analysis of corrupt behavior in the internal world, which comprises a critical discussion of the rational choice theory and behavioral theories. Building on this, we then add an additional level of discussion at the meso level, where we shed light on both sociological and criminological factors. Ultimately, we discuss corrupt behavior from the perspective of the external world, which includes, among others, economic, legal, and political aspects. We believe that such an approach encompasses the breath of scientific discussion on the topic of corruption and does sufficient justice to the different theories and approaches that contribute to a better understanding of what shapes corrupt behavior. For reasons of convenience, we provide a graphical illustration to guide the reader through the next section’s discussion of factors that explain corrupt behavior. 18 Figure 1.1: Interdisciplinary Perspective External World Bureaucratic Environment Political Environment History Other Institutional Characteristics Legal Environment Economic Environment Geography Meso World Values Social Norms Strain Theory Education Culture Internal World Rational Choice Theory Behavioral Perspective Individual Corrupt Behavior Induces a Retroactive Effect Decision - making - process eventually leads to Differential Association Theory 19 1.3.1 Internal World – Rational Choice & Behavioral Perspective The internal world represents a micro perspective that highlights the individual’s intrinsic willingness to actively engage in acts of corruption. This aspect comprises purely rational behavior and behavior beyond this clear-cut rationale. Here, light will be shed on aspects that exclusively target the individual perspective. This represents a precise methodological difference in comparison to the aggregate levels that will be analyzed in subsequent chapters. We deem it important to include these different perspectives to allow for a wellrounded discussion of the antecedents and effects of corruption. For this purpose, we will start with a pure actor-based perspective and then gradually move towards an aggregate perspective. Considering rational choice, this particular approach in the context of crime has its roots in the seminal contribution of Gary S. Becker, analyzing the disposition to deviant behavior based on cost-benefit calculations (Becker, 1968). Encompassing economic theories on crime causation have evolved ever since. The rational choice, whether or not to succumb to corrupt behavior, is based on a decision process in which individuals try to maximize their utility. This is done by weighing expected benefits against expected costs of deviant behavior, including opportunity costs and the risk of being caught or punished. One can use this general approach to understand a subset of criminal behavior, namely corruption, by shedding light on the decision making process of both the briber and the bribee. Although opportunity costs and risk calculation will certainly differ for each of the parties involved, the basic decision process is similar. (1) Opportunity costs due to time allocation: Whenever time is spent on criminal engagement, less time is available for legal activities. The opportunity costs therefore represent the amount of income, which is given up to attend to the alternative action. (2) Risk calculation: The consideration of the risk of being caught or punished. Certain actions are less likely to be observed and prosecuted and thus drive the individual risk assessment. Both factors also represent viable ways to deter corrupt behavior, for example, through applying more severe punishments and increasing the probability of detection. Research indicates that both increasing the certainty and the severity of punishment are viable measures to deter criminal behavior, with the former being backed up by more, consistent, 20 empirical evidence than the latter (Nagin, 2013). Feess et al. report that increasing the magnitude of punishment—for example, up to a death penalty like in China—might even bring about perverse effects (Feess, et al., 2014). It is reasonable to assume that under such circumstances, judges would tend to be more careful in sentencing, since the condemnation would be associated with high costs for both the defendant and the judge given the risk of a potentially wrong decision. Consequently, irrespective of the corrupt acts detected, percentage of actual convictions might drop, rendering increased sanction detrimental or useless at the best. From a criminal’s perspective, in a situation in which deviant behavior becomes more lucrative due to a ceteris paribus decrease in expected costs, such a leeway might induce even more deviant behavior. After all, facing both a drop in convictions and a rising estimated number of unreported cases may tempt the government to impose even harder sanctions, leading to a vicious circle (Steinrücken, 2004). Yet, more often than not, individual behavior goes beyond clear-cut rational decisionmaking but is bounded in terms of to what extent decisions are thoroughly elaborated (Gigerenzer & Selten, 2002). As described before, the pure rational choice approach leaves no room for moral quarrels that may influence the calculus, although real life experience proves morals highly relevant. Yet, morals differ not only from society to society but also on an individual level and even from one situation to another—especially if factors such as emotions are considered. Essentially, a combination of all these aspects is needed to reach a well-elaborated internal view. Thus, in recent years, the behavioral approach, which enriches the rational perspective with the inclusion of psychological aspects and biases, has been incorporated into models trying to better explain deviant behavior in general and corrupt behavior in particular. It has been argued that even a rational decision-maker might end up engaging in seemingly irrational behavior that is guided by more than just a rational calculus, but rather is a function of the underlying environment. This stream of literature has extended the decision space of the so-called “homo oeconomicus” by incorporating factors such as reciprocity, emotions, social image and the like to draw a more realistic picture of human behavior (Barberis, 2011). Clearly, the growing body of approaches represents an addition rather than substitution of the rational choice approach. 21 Arguably, pure rational choice concerns are incapable of explaining the de facto extent of existing corruption. Lambsdorff argues that the rational choice theory brings about two seemingly conflicting outcomes, one with and one without existing corruption. On the one hand, one should observe corruption more frequently as it is the case since—at least in the absence of norms, values, and the like—criminal behavior is solely driven by rational calculus (Lambsdorff, 2012). On the other hand, since bribery is not a subgame perfect Nash equilibrium, its actual occurrence might already be surprising. In one-shot bribery settings, as is usually the case, reputation does not play any role, suggesting that the bribee has no incentive to reciprocate the behavior of the briber. Consequently, the briber anticipates the bribee’s deviant behavior—e.g. pocketing the money without providing the respective service—and, as a result, he should not pay any bribes in the first place. Even in repetitive settings, the exchange will terminate eventually, leading to what is called an endgame effect, suggesting that the bribee will deviate from the reciprocal arrangement at some point. This entails that by using backward induction, the briber will refrain from paying bribes in the first place as well. Accounting for these seemingly conflicting outcomes, current research suggests that one’s decision-making process is vastly guided by the social environment and one’s peer’s behavior (Evans et al. (1992), Glaeser et al. (1996)). Among other things, theoretical and experimental research suggests that the effect of behavioral contagion is mediated by the social proximity to the peers (Akerlof (1997), Dimant (2015)). A person’s traits and behavior are predominantly based on social interaction (LaRossa & Reitzes, 1993); people are not born with them, but rather learned and adapted through the course of social interaction. These patterns and values can vary and develop as time moves on and they can be considered to be under constant exogenous influence. What is more, existing evidence points at the importance of social norms and values, but also the impact of reputation in repeated game environments, in explaining corrupt behavior (Gächter & Falk (2002), Milinski et al. (2002)). “Reputation is a powerful force for strengthening and enlarging moral.” (Haidt, 2007, p. 998) In sum, the many factors comprising the internal world can be seen as the essential pillars in explaining corrupt behavior. Research indicates, however, that the decision to behave in a corrupt manner is not driven solely by internal factors. Instead, it is the interplay with the 22 social environment that impacts or overrides the internal world. The social nature of humans promotes the consideration of peer group affiliation and reputation, deeming it unlikely that behavior in general and unethical one in particular is purely self-driven. We now turn to the discussion of meso and macro factors that add to the understanding of corrupt decision-making and build upon the internal world. 1.3.2 Meso World—Sociological & Criminological Factors The meso world focuses on social interaction. It is plausible to assume that, beyond the intrinsic willingness, different components like typical values, rules, and norms within a given society have a strong impact on a person’s decision on whether or not to act corruptly. There are many sociological factors and criminological aspects as well as theories that can influence the level of corrupt behavior. Sociological Factors The general culture within a given country can have a significant impact on individual decisions to engage in corrupt behavior. Husted examines the effect of different cultural aspects and describes “a cultural profile of a corrupt country as one in which there is high uncertainty avoidance, high masculinity, and high power distance.” (Husted, 1999, p. 354) Other studies come to a similar conclusion. For example, Volkema and Getz (2001) analyzed power distance and uncertainty avoidance, again showing a significant positive correlation between these cultural factors and the level of corruption. Recent studies also support these results. The two dimensions of national culture (power distance and individualism) moderate the relationship between human development and corruption (Sims, et al., 2012). This is also true if norms and values are carried over from different cultures through migration. For example, Dimant et al. find some indication for such a footprint effect. In continuing to conduct business as usual, the destination countries experience deterioration in institutional quality and an increase in corruption levels in the short run. But they also find that migrants eventually assimilate to the new environment in the medium run (Dimant, et al., 2015). Aside from the cultural aspects, research also points at the relevance of education in mediating the inclination towards corrupt behavior. Education typically intensifies in the 23 process of economic development within a given country and contributes to lower levels of corruption (Treisman, 2000). A study conducted in Nepal indicates that education is one of the primary determinants of corrupt behavior. Higher education is strongly correlated with the likeliness to condemn corrupt behavior and the reluctance to accept even small bribes (Truex, 2011). Research also indicates that the composition of gender in leading positions mediates the extent of corruption (Sung & Chu (2003), Sung (2003), Sung (2012)). For example, Dollar et al. (2001) find that a greater number of women involved in parliament is typically associated with lower levels of corruption. Similar results are common in cross-country evaluations (Swamy, et al., 2001). Typically, women tend to obey society rules and are less likely to take serious risks and therefore less often commit to corruption (Esarey & Chirillo (2013), Frank et al. (2011)). Criminological Factors From a criminological perspective, corruption is at the center of general crime and it facilitates the pervasiveness of the crime (Husman & Walle (2010), Shelley (2014)). The criminological view on deviant behavior is interdisciplinary in itself. In particular, there is a strong interdependence between the sociological factors and criminology, because aspects like culture and education have an effect on general crime rates and therefore on the level of corruption. The incorporation of rational decision-making also represents an evident link to the internal world (Glueck & Glueck, 2014). Sutherland and Cressey (2014) brought forward the differential association theory, concluding that criminal behavior is commonly learned and adopted in interaction with other people. Aspects such as social class, race and unstable homes are not only factors favoring the commitment to criminal activity but they also increase the probability that people will socialize with persons of similar character. This theory is widely supported by empirical research that focuses on social learning for both criminal and conforming behavior (Akers (2014), Cohen (2014)). At the same time, social learning is not only restricted to small neighborhoods or certain areas, but also does entail an aggregate perspective on the societal level. The strain theory, first established by Merton in 1938 in a time when the 24 most widely accepted hypothesis was that criminal behavior can be primarily attributed to biological disposition, highlights the relevance of social structures and social pressure in the occurrence of criminal behavior (Merton, 1938). Whenever individuals feel they are being treated unfairly by the society—e.g. restricted access to good schooling—, they encounter a stressful situation, which in turn taxes one’s self-control (Hirschi & Gottfredson, 1990). This theory suggests that under these circumstances, people may tend to reverse the goals set by society and create their own goals conflicting with existing norms and values. They are likely to believe that the means justify the ends, which is conducive to their decision to engage in criminal activities (Cohen (2014), Messner & Rosenfeld (2014)). The basic strain theory, however, has been altered over time, eventually leading to a more generalized theory. Individuals even in a stable personal environment—e.g. well paid and secure job—are potentially willing to put everything at risk and choose to engage in criminal behavior. Such behavior might stem from a biased self-perception. Although well-educated white-collar individuals should be able to fully take stock of the consequences of their corrupt behavior, Benson argues that such offenders often do not view themselves as criminals but rather as good employees, justifying their acts solely on the basis of trying to enforce the company’s success (Benson, 2014). This theory seems to hold especially for employees in higher positions with ample responsibilities when they see the chance to, for example, secure other people’s jobs by acting corruptly (Fleming & Zyglidopoulos, 2009). Such a biased selfperception might be the result of both hypocrisy and a different understanding of what is right and wrong. As research indicates, such an understanding of, for example, what is considered a bribe or a gift is context dependent, which varies substantially across countries (Millington, et al. (2005), Steidlmeier (1999)). However, aside from varying perceptions in different countries, the rationalization process is present in every society and it is a key determinant for white-collar crime and corruption in particular. The ability to rationalize unethical behavior pushes out feelings of guilt and shame, rendering corrupt behavior justifiable if there are enough good reasons (Søreide, 2014). In line with the social learning theory introduced earlier, such work environments can be deemed highly negative. If the supervisors act corruptly without any feelings of guilt, this behavior may affect the other 25 employees’ decision-making process. Consequently, further analysis is essential with respect to the extremely high damages involved in white-collar crimes. Prosecution and quantification of such crimes turn out to be extremely tough (Lambsdorff & Schulze, 2015), and even though numerous cases with extensive damage are known, the actual ramifications remain devious. Furthermore, higher levels of corruption combined with weak institutional structures, soak through society and eventually lead to rising general crime rates, creating a hostile environment and breeding ground for even more corruption (Claros, 2013). This article now turns to the external world by shedding light at factors at the macro level that influence the extent of corruption. 1.3.3 EXTERNAL World—Economic, Legal, Political, Historical and Geographical Factors The external world includes all other elements representing extrinsic opportunities that directly or indirectly have an influence on corruption. Among others, these are economic, legal, political, historical and geographical factors. Economic Factors Existing research points at a broad range of economic factors relevant to the extent of corruption. For example, the overall quality of the government in a given country is a wellstudied determinant. “Poor governance may affect economic performance through their impact on tax revenue, public spending, and fiscal deficit.” (Tanzi, 1999, p. 10) Inefficient bureaucracy fuels corruption because it provides a fertile ground for “speed money.” Such a mechanism is designed to circumvent impeding regulatory bodies, which represent the major ingredient of the greasing the wheels hypothesis described in Section 1.2. In the context of firm entry in highly regulated countries, Dreher and Gassebner (2013) analyzed more than forty countries for several years, concluding that the greasing the wheels hypothesis holds even today. The more inefficient regulations are, the longer the delays for companies being able to start their business. In consequence, such inefficiency, coupled with the risk of losing money and business, trigger their decision to make use of speed 32 CHAPTER 2: EMPIRICS 2.1 THE EFFECT OF CORRUPTION ON MIGRATION, 1985-2000ꟸ Eugen Dimant*, Tim Krieger**, Daniel Meierrieks *** ABSTRACT: We examine the influence of corruption on migration for 111 countries between 1985 and 2000. Robust evidence indicates that corruption is among the push factors of migration, especially fueling skilled migration. We argue that corruption tends to diminish the returns to education, which is particularly relevant to the better educated. KEYWORDS: Corruption; migration; skilled migration; push factors of migration JEL CLASSIFICATION: D73; F22; O15 ꟸ Published in: Applied Economics Letters (2013), 20(13), pp. 1270-1274. * University of Paderborn. Corresponding author. ** University of Freiburg and CESifo. *** University of Freiburg. 33 2.1.1 Introduction Previous empirical research suggests that, besides socio-economic and demographic factors (e.g., underdevelopment and demographic pressures), politico-institutional conditions (e.g., political instability) are among the push factors of international migration (cf., e.g., Hatton and Williamson (2003); Mayda (2010); Dreher et al. (2011); Docquier and Rapoport (2012)). We argue that a related important push factor, which has so far been mostly neglected in both theoretical and empirical research, is corruption. Corruption is associated with a number of unfavorable outcomes. For one, corruption tends to negatively affect a country’s (short-run) level of economic activity.1 Economies plagued by high levels of corruption grow more slowly, e.g., as corruption contributes to an inefficient allocation of resources (Jain (2001); Campos et al. (2010)). For another, corruption may also worsen a country’s (structural) socio-economic situation. For instance, Gupta et al. (2002) find that high levels of corruption promote income inequality and the spread of poverty. Furthermore, corruption may also lead to suboptimal patterns of social mobility when it matters more strongly to upward mobility than actual merits. In sum, the prevalence of corruption is likely to worsen individual working and living conditions for the majority of citizens. It may therefore also matter to the calculus of a potential migrant. Individual education is a particularly important factor influencing migration decisions. Here, corruption tends to lower the returns to education, e.g., by contributing to slow economic growth and unemployment, widespread inequality and the lack of social advancement. Given the irreversibility of human capital investment, corruption may make it more attractive to migrate to recoup one’s individual education investment. Here, we expect the highly skilled to be particularly responsive to the prevalence of corruption, given their high level of human capital investment and subsequent need for particularly high (i.e., cost-effective) skill premiums.2 Based on these lines of reasoning, we hypothesize that corruption is among the push factors of migration and particularly relevant to skilled migration. 1 Note, however, that this assessment does not rule out that corruption may actually yield positive economic effects under specific circumstances (Dreher and Gassebner, 2013). 2 In addition to that, the increase of income inequality and poverty caused by corruption (Gupta, et al., 2002) may foster the political demand for redistribution. As the better skilled are typically the typical net payers of (progressive) income taxes, this may further fuel skilled emigration. 34 2.1.2 Data and Methodology To empirically examine this hypothesis, we compile data on (skilled and average) migration, corruption and further controls for 111 countries between 1985 and 2000. The summary statistics and the operationalization of the controls are reported in Table 1.3 Data on our dependent variables, the migration rates, are drawn from Defoort (2008) who provides estimates of the rates of skilled and average migration to six main receiving countries (Australia, Canada, France, Germany, the U.K. and the U.S.). Here, the skilled migration rate refers to the ratio of the number of skilled emigrants (who exhibit a post-secondary certificate) to the total number of skilled natives aged 25 or older, while the average migration rate is defined as the ratio of the total number of emigrants to the total number of natives aged 25 or older (Defoort, 2008). 3 The migration data is available only for three points in time (1990, 1995 and 2000). Therefore, we use five-year averages of the explanatory variables (for the 1986-1990, 1991-1995 and 1996-2000 periods) to estimate their influence on migration. 35 Table 2.2.1: Summary Statistics and Data Operationalization and Sources Variable Observations Mean SD Minimum Maximum Operationalization Source Skilled Migration 333 0.143 0.179 0.001 0.910 Average Migration 333 0.035 0.061 0.001 0.419 Corruption 333 2.669 1.294 0 5.983 Per Capita Income 333 8.625 1.179 5.576 10.908 Real, PPP-adjusted per capita income, logged (a) Population Size 333 9.380 1.487 6.028 14.038 Population size in thousands, logged (a) Regime Type 333 2.169 7.072 -10 10 Revised Combined Polity Score, ranging from -10 (autocracy) to +10 (democracy (b) Political Instability 333 0.300 0.758 0 3.766 Number of battle deaths in civil wars (defined as a conflict with at least 25 battle deaths per year), logged+1 (c) Youth Burden 333 0.260 0.030 0.174 0.333 Number of people between the ages of 15 and 29 as share of total population (d) Quality of Bureaucracy 333 1.890 1.200 0 4 Index of institutional strength and quality of a country’s bureaucracy (e) Trade Openness 333 4.026 0.644 0.486 5.811 Sum of exports and imports as a share of real GDP, logged (a) Former Colony 333 0.495 0.501 0 1 Time-invariant dummy variable (1=country has colonial links to one of the six major receiving countries) (f) Distance 333 7.943 1.066 4.456 9.093 Time invariant variable that measures the distance between country’s capital and the nearest capital of one of the six major receiving countries, logged (f) Notes : Source refers to (a) PENN World Tables (https://pwt.sas.upenn.edu/), (b) Polity4 Dataset (http://www.systemicpeace.org/poli ty/polity4.htm), (c) PRIO Battle Deaths Data (http://www.prio.no/Data/Armed-Conflict/Battle-Deaths/), (d) United Nations Population Division (http://esa.un.org/unpd/wpp/Excel-Data/population.htm), (e) ICRG (2009), (f) CEPII GeoDist Dataset (http://www.cepii.fr/CEPII/en/bdd_modele/presentation.asp?id=6). 36 Corruption data are drawn from the International Country Risk Guide (PRS Group, 2008).4 Surveying experts, the ICRG issues a corruption index that we use as our main explanatory variable. Here, corruption refers to financial corruption associated with conducting business (e.g., bribes) as well as other forms of political corruption such as excessive patronage, nepotism and close ties between politics and business. Note that we rescaled the ICRG corruption index, so that higher values correspond to higher corruption levels. To avoid detecting only spurious effects of corruption on migration, we also consider a number of confounding controls that may simultaneously affect corruption and migration. For instance, we control for the effect of economic development, as richer countries are both less susceptible to corruption and less likely to have high migration rates. Following the empirical literature on the determinants of corruption (Serra, 2006) and migration (Docquier & Rapoport, 2012), in our baseline specification we control for the effect of per capita income, population size, regime type and political instability. As robustness checks, we amend this model with additional controls for demographic pressure (youth burden), institutional quality (the quality of a country’s bureaucracy), trade openness and certain country-specific traits that may affect migration costs (colonial ties and distance between sending and destination countries).5 Initial tests indicate the presence of autocorrelation, heteroskedasticity and cross-sectional dependence, as it is common for panel data with country-year observations. We therefore run a series of pooled OLS and fixed-effects regressions with Driscoll-Kraay standard errors that are robust to these data characteristics (Driscoll & Kraay, 1998).6 4 We use the ICRG data because it is available since 1984, making a panel estimation approach to the corruption-migration nexus possible. Other corruption measures are available only for shorter time periods. Jain (Jain, 2001, p. 77) notes that the various corruption measures are usually highly correlated. 5 Our findings are also robust to the inclusion of further controls for religious fractionalization, oil production, government size, further geographic and historic country characteristics (landlocked location, common language) and education (years of schooling). 6 We also experimented with instrumental variable (IV) estimations, as reverse causation may be an issue. However, pooled and fixed-effects IV-estimations (where corruption is instrumented by the quality of judicial institutions and the degree of democratic participation) do not indicate that corruption is endogenous to migration. Also, Durbin-Wu-Hausman tests suggests that any endogeneity among the regressors does not bias our estimates. 37 2.1.3 Empirical results The pooled OLS estimates are reported in Table 2. We find that corruption has a positive and statistically significant effect on both skilled and average migration. However, the marginal effect of corruption on skilled migration tends to be approximately three to four times higher than its effect on average migration. This finding provides first support for our hypothesis that corruption as among the push factors of migration and especially matters to the migration decisions of the highly skilled. The fixed-effects estimates—which truly consider the panel structure of our dataset—are reported in Table 3.7 We find that corruption only has a positive, statistically significant and specification-robust effect on skilled migration, but has no significant impact on average migration. This result further strengthens our previous finding that the decision of the highly skilled to emigrate is strongly affected by the disincentive of corruption at home. 7 Note that all constant influencing factors (colonial ties, distance etc.) are now subsumed under the fixed effects. 38 Table 2.2.2: Corruption and Migration (Pooled OLS Estimates) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Skilled Migration (1) - (5) Average Migration (6) - (10) Corruption 0.025 0.022 0.036 0.024 0.035 0.009 0.005 0.009 0.008 0.012 (0.006)** (0.004)** (0.005)** (0.006)** (0.006)** (0.001)*** (0.001)** (0.001)** (0.001)*** (0.001)*** Per Capita - 0.046 - 0.043 - 0.057 - 0.048 - 0.045 0.003 0.005 0.002 0.002 0.002 Income (0.002)*** (0.003)*** (0.007)** (0.003)*** (0.002)*** (0.001)** (0.001)** (0.002) (0.001) (0.001) Population Size - 0.044 - 0.044 - 0.047 - 0.040 - 0.041 - 0.014 - 0.013 - 0.014 - 0.012 - 0.013 (0.001)*** (0.001)*** (0.002)*** (0.001)*** (0.001)*** (0.001)*** (0.001)*** (0.002)** (0.001)** (0.001)*** Regime Type 0.006 0.007 0.006 0.007 0.007 0.003 0.003 0.003 0.003 0.003 (0.001)** (0.001)** (0.001)** (0.001)** (0.001)*** (0.001)*** (0.001)*** (0.001)*** (0.001)*** (0.001)*** Political 0.002 0.002 0.007 0.005 0.004 - 0.003 - 0.003 - 0.003 - 0.002 - 0.003 Instability (0.005) (0.005) (0.005) (0.005) (0.007) (0.003) (0.003) (0.002) (0.003) (0.004) Youth Burden 0.389 0.403 (0.361) (0.041)** Quality of - 0.027 - 0.002 Bureaucracy (0.010) (0.005) Trade 0.020 0.009 Openness (0.001)*** (0.001)*** Former Colony 0.068 0.019 (0.001)*** (0.001)*** Distance - 0.026 - 0.010 (0.003)** (0.002)** R 2 0.252 0.254 0.262 0.255 0.294 0.200 0.225 0.200 0.206 0.237 N*T 333 333 333 333 333 333 333 333 333 333 Notes : Constant not reported. Driscoll - Kraay standard errors in parentheses. **p<0.05, ***p<0.01. 39 Table 2.2.3: Corruption and Migration (Fixed-Effects Model Estimates) (1) (2) (3) (4) (5) (6) (7) (8) Skilled Migration (1) - (4) Average Migration (5) - (8) Corruption 0.005 0.005 0.004 0.005 - 0.003 - 0.003 - 0.002 - 0.003 (0.001)*** (0.001)*** (0.001)** (0.001)** (0.001) (0.001) (0.001) (0.001) Per Capita - 0.048 - 0.043 - 0.045 - 0.051 0.002 - 0.001 0.001 0.001 Income (0.010)** (0.010)** (0.010)** (0.008)** (0.003) (0.003) (0.002) (0.003) Population Size 0.001 - 0.001 0.002 - 0.002 0.012 0.013 0.012 0.011 (0.009) (0.009) (0.010) (0.011) (0.002)** (0.002)** (0.002)** (0.002)** Regime Type 0.002 0.002 0.002 0.002 0.001 0.001 0.001 0.001 (0.001)* (0.001)* (0.001)* (0.001)* (0.001)** (0.001)** (0.001)** (0.001)** Political 0.008 0.008 0.007 0.008 - 0.002 - 0.002 - 0.002 - 0.002 Instability (0.001)** (0.001)** (0.001)** (0.001)*** (0.001)* (0.001)** (0.001)* (0.001)* Youth Burden 0.190 - 0.087 (0.028)** (0.018)** Quality of 0.003 - 0.001 Bureaucracy (0.001) (0.005) Trade 0.006 0.003 Openness (0.006) (0.001)** Within - R 2 0.174 0.182 0.176 0.176 0.145 0.145 0.132 0.130 N*T 333 333 333 333 333 333 333 333 Notes : Constant not reported. Driscoll - Kraay standard errors in parentheses. *p<0.1, **p<0.05, ***p<0.01. 40 Finally, note that the results for the control variables are largely in line with previous research. For instance, both the pooled OLS and fixed-effects estimates suggest that skilled migration is less common in richer countries, as previously reported in Docquier and Rapoport (2012). As another example, the positive effect of political instability on skilled migration in the fixed-effects regressions is in line with Dreher et al. (2011). 2.1.4 Conclusion We examine the impact of corruption on migration for a panel of 111 countries between 1985 and 2000. Our empirical results indicate that corruption especially drives skilled migration, while its effect on average migration is less pronounced and not statistically robust. Our main finding is consistent with the hypothesis that corruption lowers the returns to education and consequently matters most to the calculus of (prospective) highly skilled migrants. Corruption control may therefore be an important policy tool to rein the brain drain, particularly when this brain drain is associated with predominantly poor development outcomes. 41 Appendix Albania Dom. Republic Italy Papua N. Guinea Tunisia Algeria Ecuador Jamaica Paraguay Turkey Angola Egypt Japan Peru Uganda Argentina El Salvador Jordan Philippines U. Arab Emirates Australia Ethiopia Kenya Poland United Kingdom Austria Finland Kuwait Portugal United States Bahrain France Liberia Qatar Uruguay Bangladesh Gabon Libya Romania Venezuela Belgium Gambia Madagascar Saudi Arabia Vietnam Bolivia Germany Malawi Senegal Zambia Botswana Ghana Malaysia Sierra Leone Zimbabwe Brazil Greece Mali Singapore Bulgaria Guatemala Mexico Somalia Burkina Faso Guinea Mongolia South Africa Cameroon Guinea-Bissau Morocco South Korea Canada Guyana Mozambique Spain Chile Haiti Netherlands Sri Lanka China Honduras New Zealand Sudan Colombia Hungary Nicaragua Sweden Congo (Republic) India Niger Switzerland Costa Rica Indonesia Nigeria Syria Cote d’Ivoire Iran Norway Tanzania Cuba Iraq Oman Thailand Cyprus Ireland Pakistan Togo Denmark Israel Panama Trinidad 48 corruption are deeply entrenched within the people’s mindset. Bilodeau (2014) finds that the immigrants’ relationship with their destination country’s politics is substantially affected by the political environment in their home country, thus sustainably imprinting their personal attitudes. Along these lines, Helliwell (2014) also find support for the footprint effect of trust levels, which is of high relevance in the corruption context (cf. Bjørnskov (2011); Rothstein and Eek (2009)). Their results suggest that migrants from low-trust environments carry over their trust-attitudes to their destination countries in a much more pronounced way than migrants from high trust environments, indicating an asymmetric interrelation between migration and stickiness of norms (see also Uslaner (2008)). Consequently, value assimilation becomes unlikely in the short run, and corrupt behavior remains persistent. Third, as Varese (2011) notes, successful criminal behavior in a new and unknown environment does not only require a criminal mind, but also an opportunity. It might take some time after entering the destination country to comprehensively adapt to the new environment, and to find ways and means for successful corruption. If immigrants show persistent corruption attitudes, the full effect of immigrants’ corrupt behavior may become visible in the target country only after some period of time. This leads us to our next hypothesis. Hypothesis H2: The effect on the destination country’s level of corruption, related to immigration from a more corrupt sending country to a less corrupt destination country, is positive. However, it might take some time before the effect becomes noticeable. In the following section, we will test our hypotheses to investigate which effects dominate. Beforehand, a caveat is in order. Endogeneity is a widely acknowledged issue in empirical corruption research.7 As is generally true for empirical panel data research, a correlation exposes a general coherence rather than rendering a clear causal relationship. In our case, corruption could potentially be both the antecedent and the effect of other factors. As already indicated above, Cooray and Schneider (2014) and Dimant et al. (2013) show reverse causality between migration and corruption, finding that excessive corruption decisively impacts migration flows. Hence, immigration might very well leave a corrupt footprint in the 7 For example, the literature indicates that the relationship between corruption and economic growth also holds in the reverse direction (cf. Dreher and Gassebner (2013); supporting the ‘greasingthe-wheels’ hypothesis, and Méon and Sekkat (2005), supporting the ‘sanding-the-wheels’ hypothesis). 49 destination country because of (self-) selection effects. If, for instance, an honest outsider decides to leave a corrupt country it is unlikely that he/she will (voluntarily) choose an equally corrupt destination. That is, the level of corruption in the destination country might be relevant in shaping migration flows. Evidently, it is important to control for endogeneity as the results might potentially suffer from a reverse-causality bias. Our approach of how to address this problem will be presented in the following section. 2.2.3 Empirical Analysis The Empirical Model Based on the previous theoretical considerations, the discussion in this and the following section aims at testing the hypotheses developed in Section 2.2.2. Our starting point is a panel model of the form it corruption = α + ф it q m ig r a tio n  + β 1 it X   + i  + it  where it corruption is the level of corruption in country i and year t, it q m ig r a tio n  is the total migration stock with a time lag q, 1 it X   is a conditioning set of lagged control variables and the disturbance term is composed of the individual effect ηi and the stochastic disturbance it  which is assumed to be generated by a white noise process. This specification allows testing the general effect of migration on corruption according to hypothesis H1. Since we assume the migration variable to have a time-shift effect on corruption, we let the independent variable of interest enter the model with a time lag q, which may take values from one to five if, for example, the maximum lag is five years. This lag structure allows us to differentiate between immediate and delayed effects. Additionally, lagging the independent variable of interest dampens the problem of a possible endogenous relationship between corruption and migration by eliminating the correlation between the explanatory variables and the error term. We report the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) to allow for a comparison of the model fit of the alternative lag 50 selections.8 Assuming control variables also do not have an immediate effect in the same period, all other controls enter the model with a lag of t-1. We provide results for fixed-effects panel regressions that allow us to account for country-specific effects. Furthermore, we explore the effect of immigration from highly corrupt countries on the corruption level of the target countries according to hypothesis H2. This is tested by regressing the total migration stock from countries that exhibit a corruption level which is higher than the total average over all 207 countries of migration origin on the corruption level of the destination country, so that we can test if a higher migration stock from more than proportionally corrupt countries drives the corruption level in the destination country. Dealing with Potential Endogeneity We account for potential endogeneity by applying a Difference-GMM estimation in order to exclude results that might be driven by the underlying econometric approach, and thus do not allow for statistically reliable inference. The dynamic GMM approach developed by Arellano and Bond (1991) appears as an appropriate approach, as it allows calculating consistent and efficient estimates by using lagged levels dated in period t-2 and earlier as instruments. The corresponding moment condition can be checked using the Sargan statistic that tests the validity of the instruments. In following Arellano and Bond (1991), we provide results for Difference-GMM estimations. In general, the results are in line with the estimations presented before. An alternative estimation proposed by Arellano and Bover (1995) and Blundell and Bond (1998) is the SystemGMM approach, which performs well with highly persistent data under mild assumptions. However, there is an important point to be made about using System-GMM. Given that System-GMM uses more instruments than the Difference-GMM, it may not be appropriate to use System-GMM with a dataset that consists of a small number of countries. In this case, this method is likely to exhibit a finite sample bias as the number of instruments increases exponentially with the number of periods used. As argued by Roodman (2009), such an overfitting of endogenous variables is likely to lead to false positive results. In addition, the assumption of lagged control variables being exogenous to the error term is non-trivial. For this 8 Plümper et al. (2005) illustrate that in fixed-effects models the lag structure of the independent variable has a large impact on the coefficient and the level of significance. They argue that there is no generally accepted indicator for the determination of the length of the lag, however, there are several candidates like the t-statistic, the R2, the AIC and the BIC that facilitate the choice. 51 reason, we resort to the Difference-GMM approach, as the ratio of countries and time periods used in our panel is well balanced, thus ruling out a potential small sample bias (cf. Alonso-Borrego and Arellano (1999)).9 2.2.4 Data Dependent Variable We use the cross-national corruption rating from the International Country Risk Guide (ICRG) (PRS Group, 2008). It relies on the subjective assessment of country experts typically operating within international non-governmental organizations. As a component of the political rights index, it is concerned with actual or potential corruption in the form of excessive patronage, nepotism, job reservations, ‘favor-for-favors’, secret party funding and suspiciously close ties between politics and business.10 Originally, the value of the index ranges from 0 to 6, with 0 indicating a high level of corruption and 6 representing a low level. We transpose the scale to simplify the interpretation of the results so that higher values of the index indicate a higher extent of corruption. The main advantage of this index is that it is available annually for a large sample of countries beginning in the early 1980s, and so enables us to analyze the corruption–migration nexus within a panel framework.11 The summary statistics can be found in Table S1. Main Independent Variable of Interest Our main independent variable is immigration (migration). We use the OECD International Migration Database which provides annual series on migration flows and stocks into OECD countries from 207 countries of origin for the period 1975–2011 (OECD, 2014a). The major advantage of this dataset is that it provides bilateral data and so allows distinguishing between countries of destination and countries of origin, allowing us not only to analyze the general effect of migration on corruption but also to group source countries according to 9 To check the robustness of the model specification, we also run all regressions using the System-GMM. The results support our main findings. However, the rule of thumb – to keep the number of instruments less than or equal to the number of groups – cannot be met. Even if only the second lag is used as an instrument for the System-GMM the number of instruments exceeds the number of countries. The results are available on request. 10 http://www.prsgroup.com/ICRG_Methodology.aspx. 11 Other common corruption measures like the Corruption Perceptions Index (CPI) constructed by Transparency International or the Control of Corruption Rating published by the World Bank are available only from 1995 and 1996, respectively. Svensson (2005) and Treisman (2007) show that all three measures are highly correlated. 52 their level of corruption. We weigh migration by the respective destination country’s population in thousands in order to account for the inherent population size heterogeneity across the OECD countries. Since different countries use different definitions of immigration12 and different sources for their migration statistics, the OECD database offers both data on immigrants by nationality and on immigrants by country of birth. Especially in the case of the migration stock variable, the differences in the definition play an important role and must be considered. The ‘country of birth’ approach takes into account the foreign-born population, for example, the first generation of immigrants, including immigrants that have obtained citizenship. The ‘nationality’ approach includes second and higher generations of foreigners, but does not cover naturalized citizens. Thus, the nature of the countries’ legislation on citizenship and naturalization plays a role (Pedersen et al. (2008)). We use the immigrants stock by ‘nationality’ variable for three reasons. First, this variable is available for more country-time observations than the immigration stock by ‘country of birth’, thus allowing for more meaningful estimations. Second, we act on the assumption that naturalized citizens should be put on an equal footing with the domestic population as it is reasonable to assume that the naturalized citizens’ magnitude of assimilation is well advanced. Third (and closely connected to the previous argument), our hypothesis H2 takes the assimilation process into consideration assuming that the full effect of immigration on destination country corruption occurs only after some time. The ‘nationality’ approach takes up this time dimension more naturally. Control Variables To avoid spurious relationships between the dependent variable and the independent variable of interest, we employ a set of control variables commonly identified as potential determinants of corruption. In our baseline specification we control, first, for the impact of economic development measured by (logged) real per capita income (GDP p.c.) extracted from the Penn World Table (PWT) (Heston, et al., 2012). It is a commonly used variable to explain corruption. The theoretical argument stresses that economic development fosters 12 Countries like Australia, Canada, the Netherlands, New Zealand, Poland, the Slovak Republic and the United States define an ‘immigrant’ by country of origin or country of birth, while some countries like Austria, the Czech Republic, Denmark, Finland, Greece, Iceland, Italy, Norway and Sweden define an immigrant by citizenship and finally some countries like Belgium, France, Hungary, Germany, Japan, Luxembourg, Portugal, Spain, Switzerland and the United Kingdom rely on selfreported nationality (Pedersen, et al., 2008). 53 higher institutional quality, which in turn will provide fewer breeding grounds for corruption via implementation of more sophisticated anticorruption measures. A higher chance of identification and punishment of corruption will increase the expected costs, and crowd out incentives to engage in deviant behavior (cf. Serra (2006)). Along these lines, several empirical studies find a robust negative correlation between economic development and perceived corruption, suggesting that poorer countries exhibit higher corruption rates (cf. La Porta et al. (1999); Serra (2006); Treisman (2007)). However, panel studies based on fixedeffect estimation by Braun and Di Tella (2004) find that an increase in a country’s wealth measured by GDP per capita also increases corruption. A potential explanation for a positive nexus between growth and corruption is provided by Kindleberger (2000). He argues that moral standards vanish in times of economic booms due to a more pronounced manifestation of greed, eventually undermining the individual’s disposition to obey the law. Overall, we follow the empirically settled mainstream argument and expect that more developed countries (in terms of GDP per capita) experience lower rates of corruption. We also account for the effect of population size (PWT) on corruption. From the theoretical perspective, Knack and Azfar (2003) suggest that larger polities may benefit from economies of scale in establishing political and administrative structures, so that a large country size might be negatively correlated with corruption. On the other hand, small countries may benefit from higher manageability, and more efficiency and transparency in administrative management, leading to a positive correlation between population size and corruption. Empirical evidence shows mixed results. For one, Knack and Azfar (2003) show that there is indeed no clear relationship between country size and corruption and that existing results suffer from selection bias. On the other hand, a cross-country study by Tavares (2003) shows a negative impact of population on corruption, while Root (1999) finds that a larger population is significantly associated with more corruption indicating that smaller countries are less corrupt than larger countries. We follow the majority of existing evidence and assume that population size and extent of corruption go hand in hand, due to a higher number of potential bribers and bribees and issues of effective monitoring, which are likely to be more extensive with a growing population size. Ali and Isse (2003) argue that a large government sector (government size) may create opportunities for corruption. The larger the size and scope of the bureaucracy, the more likely 54 it is to find corrupt behavior. On the contrary, Goel and Nelson (2010) indicate that government size might be inversely related to a country’s corruption level. Not a large public sector per se determines the magnitude of corrupt activity, but larger governments might in fact devote a higher share of public spending to operative law enforcement aimed at deterring deviant behavior (cf. Fisman and Gatti (2002); Goel and Nelson (2010)).13 Although not explicitly tested for a subset of OECD countries, we follow the majority of existing empirical evidence and expect a large government sector to have a negative impact on a country’s corruption level. The variable is measured by government expenditure divided by GDP and extracted form the World Development Indicators (WorldBank, 2013). Furthermore, we control for democracy measured by the Polity2 index (Polity IV Project, 2012), which is found to be highly relevant in existing theoretical and empirical research on corruption. In general, both strands of research indicate that more democratic countries tend to be less corrupt (e.g., Braun and Di Tella (2004); Knack and Azfar (2003); Kunicová and Rose-Ackerman (2005); Shen and Williamson (2005)). From a theoretical perspective, Shen and Williamson (2005) contend that states with democratic governments are likely to have more sophisticated policies and legal institutions that are more likely to be independent of the elites’ impairment. Seldadyo and de Haan (2006) argue that political liberty imposes transparency and provides checks and balances within the political system and so tends to reduce corruption. Kunicová and Rose-Ackerman (2005) suggest that electoral rules and political structures can influence the level of corruption. Political participation, political competition and constraints on the chief executive make it easier to monitor the political system and limit political corruption.14 Overall, since both theory and empirics resonate with each other, we would expect a negative impact of democracy on corruption. In addition, existing research acknowledges the important link between economic freedom (ICRG) and corruption. From a theoretical perspective, one can argue that, especially in modern economies, many restrictions on economic freedom– in particular restrictions of 13 However, it is worth noting that parts of the existing literature also point at a different relationship between government sector and corruption. Corrupt governments may impose detrimental effects on public goods delivery, weaken the tax morale and the bureaucratic quality whose functional interaction, ceteris paribus, likely leads to a smaller government sector (cf. Frey and Torgler (2007); Hall and Jones (1999); Johnson et al. (1997); Tanzi (2013)). 14 Treisman (2007) indicates that the relationship between democracy and corruption might be more complex, suggesting that democratization increases corruption in the short run and reduces it as democracy deepens. However, the composition of our data does not allow us to examine long-run effects of controls such as democratization. Thus, we resort to a short-run examination of the control’s impact on corruption. 55 capital and financial markets – provide opportunities for corruption (cf. Graeff and Mehlkop (2003)). This notion is strongly supported by the empirical literature. Goel and Nelson (2005) find a strong negative relationship between economic freedom and corruption, where the relationship depends on a country’s level of development. Paldam (2002) presents similar results suggesting that countries with high regulation and little economic freedom have a larger potential for rent seeking, resulting in higher levels of corruption. Supportive results of a negative relationship between economic freedom and corruption are also found by Ali and Isse (2003), and Kunicová and Rose-Ackerman (2005). We expect that more economic freedom and fewer restrictions imposed on trade are inversely correlated with a country’s corruption level. We measure economic freedom by the investment profile variable of the ICRG, arguing that a high investment risk accompanies lower economic freedom.15 Finally, religion may also matter for explaining corruption. Religion is believed to play a decisive role in affecting corruption levels through its inherent heterogeneity in putting emphasis on moral values, honesty and being in thrall to authority. Consequently, religious structures that are more hierarchical are believed to be more conducive to the inception and development of corrupt structures (Paldam, 2001). Empirical research finds that countries with a predominantly protestant population tend to have lower corruption levels, while more hierarchically structured religions (such as Catholicism, Eastern Orthodoxy and Islam) tend to increase corruption (La Porta, et al., 1999). We follow Blomberg and Hess (2008) in using religious tensions as a control in order to get an impression whether a dominant role of a specific religious group, and the suppression of religious freedom, has an effect on the level of corruption. The argument is that a dominant religion in a country creates differential access to power, leading to a situation in which less powerful religious groups resort to corruption for leveling the political and economic landscape. A set of variables does not enter our baseline model, in particular economic growth, trade openness, internal and external conflicts and regime stability. Rather, they are used to assess the robustness of our findings. The first of these variables is economic growth (PWT) (in addition to the level of development). Ali and Isse (2003) argue that if countries with lower corruption levels grew faster, this positive experience ought to give way to a stricter fight 15 As an alternative, we also employed the ‘Economic Freedom’ index provided by the Fraser Institute (Gwartney & Lawson, 2008). The results (based on a significantly smaller dataset) support our main findings and are available on request. 56 against corruption in the future. That is, economic growth should be negatively correlated with future corruption. However, the empirical evidence on this argument is mixed. While Leite and Weidmann (1999) find that GDP growth has a dampening effect on country level corruption, Berdiev et al. (2013) find the opposite effect. However, for the subset of OECD countries (in which we are interested in) their results remain insignificant. Furthermore, other studies find no significant effect at all (cf. Ali and Isse (2003); Brunetti et al. (1997); Mauro (1995)). Consequently, due to the focus on the same subset of countries, we expect our results to be in line with Berdiev et al. (2013) for their subset of OECD countries and expect no significant effect in either direction of GDP growth on corruption levels. We furthermore assess the impact of trade openness – measured by exports and imports as a share of GDP (PWT) – as an indicator of competition.16 Leite and Weidmann (1999) suggest that openness to foreign trade, which is equivalent to a relatively strong economic competition, is a primary factor for experiencing relatively low levels of corruption. This argument is backed up by empirical research. Sandholtz and Koetzle (2000) find that economic integration decreases corruption activity, albeit not directly.17 In particular, the existing research sheds light on the interrelation between openness of financial markets and corruption levels. Although not entirely congruent, for the most part the existing literature points at the idea that restrictions bring about individual effort to bypass regulations with the use of deviant behavior, such as bribing public officials (cf. Dreher and Siemers (2009); Edwards (1999)). We thus expect an inverse relationship between trade openness and corruption, which is along the lines of the previous discussion on the impact of economic freedom on corruption. We also account for a potential effect of internal and external conflicts (ICRG) on corruption. Conflicts – in terms of domestic and transnational terrorism or civil war – may have a destabilizing effect on the economy which is what we expect to show up in our analysis. For instance, Dreher et al. (2010) and Meierrieks and Gries (2013) show that terrorism affects the 16 Alternatively, we proxy trade openness by the ratio of import to GDP (Herzfeld & Weiss, 2003). Here, a low import share implies high import restrictions. Consequently, the presence of such restrictions offers an opportunity to bribe (Seldadyo & De Haan, 2006). 17 However, Knack and Azfar (2003) argue that trade share and import share of GDP are strongly related to country size. Smaller countries tend to have a higher trade share, so not controlling for population the coefficient on openness is likely to reflect selection bias. 57 economy negatively and contributes to political instability. This in turn may create a breeding ground and may also provide opportunities for corruption. Regardless of the regime type, regime stability (Polity IV Project, 2012) is another political variable that may affect corruption. As suggested by Treisman (2007), it may take decades for democratic institutions to translate into low perceived corruption so that not the current regime type but the regime stability affects the corruption level. This is supported by an extreme-bounds analysis by Serra (2006), who provides evidence that actual democracy is weakly interrelated with corruption, whereas political stability measured by uninterrupted democracy results in reducing corruption. It is reasonable to assume that the political vacuums inherent to unstable regimes enable fraudsters to more easily find means to precipitate successful acts of corruption. Consequently, and in compliance with the previously lead discussion on the interrelation between democracy and corruption, we expect countries with stable regimes to be less prone to corruption. 2.2.5 Empirical Results In this section, we report our empirical results using different econometric approaches to ensure robustness and to account for possible endogeneity issues. Table 1 presents results for the baseline model with an alternative lag length for both fixed effects (FE) and Difference-GMM estimations, while Tobit results are generally presented in the supporting information.18 18 It should be noted that the results of the GMM estimates differ from those of the FE and the Tobit estimates in some cases. This has at least two reasons. For one, the Difference-GMM includes the lag of the dependent variable as an additional regressor, resulting in a reduction in the number of observations of about 10%. For another, based on the rule of thumb, which declares the number of instruments to be smaller than the number of cross sections, only one lag is used for instrumentation. However, in this case the GMM estimator is not necessarily efficient since it does not make use of all available moment restrictions. 64 Table 2.1.2: Continued from Previous Page VIF 1.28 1.98 1.77 1.58 1.34 2.19 1.91 1.58 Adjusted R2 0.310 0.379 0.396 0.411 0.443 0.483 0.489 0.493 AIC 696.9321 672.1169 661.6321 648.9388 420.2725 394.3588 392.0972 387.4652 BIC 725.3623 700.7084 698.3926 677.5303 447.2379 421.3242 426.7670 414.4306 Observations 429 439 439 439 348 348 348 348 Note: * p < 0.10, ** p < 0.05, *** p < 0.01; robust standard errors in parentheses; migration stock is weighted by population. 65 Finally, the Difference-GMM estimations yield a significant and positive effect of regime stability, indicating that, in the short run, countries that are wealthier and possess a more stable regime structure are more prone to corruption. While the overall direction is the same, the effect’s magnitude is more conservative than the coefficients derived from the FE approach and only shows up significantly in the short run. We present the Difference-GMM results in Table 3.21 After we could not identify a consistent and significant effect of general immigration on corruption supporting H1, we now turn to the question raised in hypothesis H2, whether corruption migrates and how long it may take to infiltrate the destination country. More specifically, we explore whether immigrants from highly corrupt countries carry over their behavior, so that immigration from countries with a level of corruption that is higher than the average leads to an increase in corruption in the destination country. The results of this exercise are shown in Table 4. Again, we present the FE and Difference-GMM estimations jointly. The results are based on a calculation of the total average of corruption levels over all countries for each year from which we then derive the most corrupt countries at the top 50% level.22 Overall, the results indicate that immigration from highly corrupt countries boosts the corruption level in the host country, thus supporting our hypothesis H2. According to the FE estimation, we find a significant and positive effect of selected migration on host countries’ corruption levels. The coefficient rises to a value of 0.0099 (for a lag of three periods), which means that an increase in the migration stock of one hundred migrants per one thousand citizens affects corruption significantly, increasing the corruption value by 0.99 points (out of 7). This is a raise of 14.1% of the maximum scale. As it has previously been the case, the results of the Difference-GMM estimations are more conservative, thus representing a lower bound result with a raise of up to 4.4% of the maximum scale for the same increase in the migrants-to-citizens ratio. Conversely, the results of the alternative Tobit regressions point to an upper bound result, indicating a raise of up to 18.1% of the maximum scale. In general, 21 We also calculate a Tobit version of the regressions with and without alternative control variables. The results are in line with the FE estimation and are presented in Table S3. 22 Our findings do not change when we consider migration from even more corrupt countries at the top 40% (30%, 20% and 10%) level. 66 the Difference-GMM results are more conservative and turn out to be significant less often compared to the FE and Tobit estimations. We trace this back to the limited amount of cross sections, which is a problem inherent to our focus on OECD countries.23 Future research might potentially overcome this drawback by extending the research scope beyond OECD countries. Noticeably, while we initially observe an escalating effect of selective migration on corruption levels, the results are indicative of an assimilation process over time. 23 Further possible explanations were offered in the beginning of this section. 67 Table 2.1.3: Migration and Corruption - Difference-GMM Regression with Alternative Controls corruption (1) (2) (3) (4) (5) (6) (7) (8) corruption t-1 0.4531*** (0.1466) 0.4341*** (0.1587) 0.4909*** (0.1740) 0.5372** (0.2237) 0.3699* (0.1919) 0.3327* (0.1975) 0.3743* (0.2000) 0.6012*** (0.0921) migration t-1 0.0006 (0.0038) -0.0026 (0.0061) -0.0006 (0.0047) -0.0040 (0.0056) migration t-5 0.0075 (0.0053) 0.0050 (0.0037) 0.0064* (0.0038) 0.0030 (0.0032) GDP p.c. t-1 -0.0115 (0.5432) -1.0297*** (0.3747) -0.1208 (0.2843) -0.9525 (0.7889) population t-1 0.0000 (0.0000) 0.0000 (0.0000) 0.0000 (0.0000) 0.0000 (0.0000) 0.0000 (0.0000) 0.0000 (0.0000) 0.0000 (0.0000) -0.0000 (0.0000) gov size t-1 -0.0115 (0.0337) 0.0006 (0.0288) -0.0102 (0.0328) -0.0189 (0.0337) -0.0096 (0.0259) 0.0087 (0.0288) 0.0042 (0.0260) -0.0257 (0.0278) democracy t-1 -0.0131 (0.0251) -0.0182 (0.0237) 0.0068 (0.0313) 0.0207 (0.0950) -0.0259 (0.1338) 0.0672 (0.0821) econ freedom t-1 0.0669*** (0.0254) 0.0604** (0.0242) 0.0567** (0.0222) 0.0406** (0.0162) 0.0464*** (0.0127) 0.0382*** (0.0131) 0.0423*** (0.0154) 0.0192 (0.0137) religious tension t-1 0.0418 (0.0861) 0.0520 (0.0740) 0.0792 (0.0865) 0.0707 (0.0913) 0.0674 (0.0583) 0.0735 (0.0612) 0.0555 (0.0708) 0.1102* (0.0609) GDP p.c. growth -0.5576 (0.8686) -0.7175 (0.7638) trade openness t-1 0.0033 (0.0043) 0.0021 (0.0030) internal conflict t-1 -0.0129 (0.0380) -0.0896* (0.0458) external conflict t-1 0.0311 (0.0384) 0.0822* (0.0430) regime stability t-1 0.0457** (0.0201) 0.0397 (0.0250) 68 Table 2.1.3: Continued from Previous Page Sargan (p-Value) AR (2) (p-value) 0.5787 0.2176 0.6018 0.2107 0.7435 0.2153 0.6894 0.1494 0.7741 0.5403 0.7274 0.5610 0.7305 0.6099 0.7374 0.4782 Instruments 30 31 33 31 27 27 29 27 Observations 396 406 406 406 319 319 319 319 Note: * p < 0.10, ** p < 0.05, *** p < 0.01; robust standard errors in parentheses; GMM results based on the two-step Difference-GMM estimator, second lag of the dependent variable used as GMM-style instrument; AR (2) refers to the Arelano Bond test for autoregressive correlation (order 2); Sargan refers to the Sargan test of over identification restrictions; migration stock is weighted by population. 69 Table 2.1.4: Migration from Corrupt Countries and Corruption - Fixed Effects and GMM Regression corruption Fixed Effects GMM (1.1) (1.2) (1.3) (1.4) (1.5) (2.1) (2.2) (2.3) (2.4) (2.5) corruption t-1 0.5278*** (0.1875) 0.4694** (0.1974) 0.5085*** (0.1926) 0.3542** (0.1716) 0.3792* (0.2219) migration t-1 0.0087** (0.0037) -0.0016 (0.0014) migration t-2 0.0095*** (0.0033) 0.0026* (0.0014) migration t-3 0.0099*** (0.0030) 0.0031** (0.0015) migration t-4 0.0073*** (0.0023) -0.0004 (0.0008) migration t-5 0.0040* (0.0020) -0.0007 (0.0011) GDP p.c. t-1 1.0742* (0.5622) 1.0156 (0.5998) 1.0565 (0.6270) 1.2442** (0.5403) 1.6372*** (0.4360) -0.3686 (0.5195) -0.3201 (0.3661) -0.4154 (0.4664) -0.1656 (0.3887) 0.1622 (0.5569) population t-1 0.0001 (0.0001) 0.0001 (0.0001) 0.0001 (0.0001) 0.0001 (0.0001) 0.0001 (0.0001) 0.0001 (0.0001) 0.0001 (0.0001) 0.0001 (0.0001) 0.0001 (0.0001) 0.0001 (0.0001) gov size t-1 0.0522 (0.0388) 0.0734** (0.0338) 0.0855*** (0.0290) 0.0997*** (0.0273) 0.0989*** (0.0250) -0.0109 (0.0332) -0.0078 (0.0265) -0.0171 (0.0353) -0.0013 (0.0272) 0.0050 (0.0358) democracy t-1 -0.1738*** (0.0295) -0.1251** (0.0512) 0.0682 (0.1657) 0.1052 (0.1667) 0.1565 (0.1573) 0.0332 (0.0887) 0.0210 (0.2159) -0.0095 (0.1567) 0.0446 (0.1538) -0.0003 (0.1068) econ freedom t-1 0.0741** (0.0327) 0.0721** (0.0348) 0.0672* (0.0363) 0.0616* (0.0313) 0.0425* (0.0229) 0.0505** (0.0212) 0.0382* (0.0230) 0.0410** (0.0160) 0.0393** (0.0171) 0.0433** (0.0192) religious tensiont-1 0.2573* (0.1401) 0.2462* (0.1341) 0.2584* (0.1303) 0.2035 (0.1209) 0.1564 (0.1077) -0.0273 (0.0807) -0.0234 (0.0752) -0.0104 (0.0795) -0.0405 (0.0418) -0.0564 (0.0480) 70 Table 2.1.4: Continued from Previous Page VIF Adjusted R2 1.24 0.382 1.26 0.381 1.27 0.423 1.26 0.459 1.27 0.494 AIC 625.1091 579.3207 501.4671 432.3486 367.1703 BIC 653.2051 607.0477 528.7663 459.2535 393.6571 Sargan (p-value) 0.7940 0.9690 0.8070 0.8624 0.6690 AR (2) (p-value) 0.3989 0.3811 0.6287 0.6532 0.8664 Instruments 31 30 29 28 27 Observations 409 388 365 345 325 379 358 338 318 298 Note: * p < 0.10, ** p < 0.05, *** p < 0.01; robust standard errors in parentheses; GMM results based on the two-step Difference-GMM estimator, second lag of the dependent variable used as GMM-style instrument; AR (2) refers to the Arelano Bond test for autoregressive correlation (order 2); Sargan refers to the Sargan test of over identification restrictions; migration stock is weighted by population. 71 These observations are in line with the previously discussed arguments presented by Chiswick (1978), which are supportive of the idea that the migrants’ assimilation happens at different speeds.24 Moreover, the results of the control variables are broadly in line with our previous findings presented in Tables 1 and 2. 2.2.6 Conclusion In this paper, we shed light on the impact of migration on corruption in the destination country. Capitalizing on a comprehensive dataset consisting of annual series on migration flows and stocks into OECD countries from 207 sending countries for the period 1984–2008, we explored different channels through which corruption might migrate. Initially, the implications might go into various directions as different effects are in place at the same time. On one side, the existing literature suggests that migration could be the result of a positive selection. For example, highly skilled people might leave their home countries as they expect their individual living conditions to improve. On the contrary, however, poor socioeconomic conditions typically constitute push factors of migration, not only for a small positive selection of honest people but also for the corruptible average individual. Independent of the econometric methodology applied, we consistently find that (i) general migration has an insignificant effect on the destination country’s corruption level, and (ii) that immigration from corruption-ridden countries boosts corruption in the destination country. This holds even after controlling for potential endogeneity by means of a Difference-GMM estimation. Hence, the fear by international legislators (as expressed in recent agreements by the G20 group) that immigration may cause a problematic inflow of corruption appears justified. Policy-makers will, therefore, have to take precautions to avoid this problem. However, it is not immediately obvious what the optimal response will be. One possibility could be to restrict immigration by only selecting immigrants originating from non-corrupt countries. Alternatively, very careful checks ad personam could be conducted. The downside of this pol24 The results of the Tobit regression are presented in Table S4 in the online appendix. The results are coherent and survive when using the alternative set of control variables. 72 icy is that the remaining inflow of migrants could be rather small, which might not be optimal, given that most OECD countries face a severe ageing problem and are in need of immigration to keep their social security systems sustainable. An arguably better strategy could be to immunize the domestic population against a corrupt attitude brought into the country by some immigrants. This would be in line with Varese’s (2011) argument which we may rephrase as follows: successful corruption needs both a corrupt mind and an opportunity. 73 CHAPTER 3: EXPERIMENTS 3.1 ON PEER EFFECTS: BEHAVIORAL CONTAGION OF (UN)ETHICAL BEHAVIOR AND THE ROLE OF SOCIAL IDENTITY Eugen Dimant* Version: December 2015 ABSTRACT: Social interactions and the resulting peer effects loom large in both economic and social contexts. This is particularly true for the spillover of (un)ethical behavior in explaining how behavior and norms spread across individual people, neighborhoods, or even cultures. Although we understand and observe the outcomes of such contagion effects, little is known about the drivers and the underlying mechanisms, especially with respect to the role of social identity with one’s peers and the (un)ethicality of behavior one is exposed to. We use a variant of a give-or-take dictator game to shed light on these aspects in a controlled laboratory setting. Our experiment contributes to the existing literature in two ways: first, using a novel approach of inducing social identification with one’s peers in the lab, our design allows us to analyze the spillover-effects of (un)ethical behavior under varied levels of social identification. Second, we study whether contagion of ethical behavior differs from contagion of unethical behavior. Our results suggest that a) unethical behavior is more contagious, and b) social identification with one’s peers and not the (un)ethicality of observed behavior is the main driver of behavioral contagion. Our findings are particularly important from a policy perspective both in order to foster pro-social and mitigate deviant behavior. KEYWORDS: Conformity, Contagion, Peer Effects, Social Identity, Unethical Behavior JEL: D03; D73; D81 * University of Paderborn and Harvard University. This work has greatly benefited from conversations with Edward Glaeser, Daniel Houser, Lawrence Katz, David Laibson, Ulrich Schmidt, and Wendelin Schnedler on the early version of the experimental design. I am particularly thankful for valuable input from Max Bazerman, Cristina Bicchieri, Gary Bolton, Elena Katok, Judd Kessler, and Robert Kurzban during my visiting research positions at the Harvard University, the University of Pennsylvania, and the University of Texas at Dallas, respectively. I also want to thank René Fahr, Uri Gneezy, Burkhard Hehenkamp, Rosemarie Nagel, Arno Riedl, and Tim Salmon as well as the participants at numerous workshops and conferences for interesting discussions and helpful suggestions. Financial support by the German Research Foundation (DFG) through the SFB 901 at the University of Paderborn is gratefully acknowledged. 80 chapter 3.1.6. We close with a concise discussion on potential policy recommendations in chapter 3.1.7 and a conclusion and an outlook in chapter 3.1.8. 3.1.2 Background and Course of Investigation A basic principle of classical economic theory suggests that individuals form rational expectations based on available information and act on them accordingly. However, even the great John Maynard Keynes (1936) expressed his concern about the rationality of individuals to realize efficient investment decisions in the long run already 80 years ago. Instead, Keynes expected individuals to follow the herd, thus stressing the importance of peers for many economic decisions. Since then, a contrasting strain of literature emerged that accounts for the relevance of behavioral traits on the individual decision-making process. Understanding the underlying mechanism of peer effects is key to comprehending its impact on economic decisions and outcomes. For the bigger part, existing research on peer effects mainly resorts to field experiments or purely observational studies that are generally inferior to controlled lab experiments in terms of, among others, a clean identification of the relevant channels, endogeneity, and reflection problems (see Manski (2000), Falk & Fischbacher (2002)). Only recently, there has been a push to study peer effects in the lab, allowing us to gain a deeper and often a more reliable understanding of the underlying mechanism Angrist (2014). Social interactions in general and the potentially resulting peer effects in particular play an instrumental role from both the societal and economic perspective. Existing literature indicates that standard economic forces alone cannot encompass many of the outcomes that we observe in real life. Examples are, among others, the escalation of crime rates or the massive surge in female labor participation rates in World War II (cf. Mulligan (1998), Levitt (1999). See also recent findings on paternity leave by Dahl, Løken & Mogstad (2014)). Instead, social interactions are found to offer explanations helpful to understanding the causes of rapid shifts in economic fundamentals. Such ripple effects are likely the result of social interactions, thus raising the awareness about the importance of understanding the underlying mechanism of peer effects (Glaeser & Scheinkman, 2004). Although explicit research on peer effects and the resulting behavioral spillovers (in the literature sometimes referred to as behavioral or social contagion) has its origins in the late 81 19th century, the underlying concept has been observed long before.2 Reportedly, an abstruse-seeming stream of suicides happened after reading Goethe’s The Sorrows of the Young Werther two hundred years ago. “My friends […] thought that they must transform poetry into reality, imitate a novel like this in real life and, in any case, shoot themselves; and what occurred at first among a few took place later among the general public […]” (Goethe, quoted in Rose (1929, p. 29)). The widespread imitation of this behavior gave rise to fear among the population and governments, ultimately leading to a ban of the book in Italy, Leipzig, and Copenhagen (Phillips, 1974). The outbreak of the Tanganyika laughter epidemic of 1962 in Uganda is another infamous example of behavioral contagion. There, a mass hysteria infected almost 100 pupils with contagious laughter, forcing several schools to close down for days (Rankin & Philip, 1963). Initially, the concept of social contagion has been introduced in the form of a social phenomenon – as opposed to a biological one – explaining why and how certain forms of behavior soak through society (for early work see Baldwin (1894), Tarde (1903)). Since the 1950s, empirical research on this topic has been on the rise with evidence suggesting that the mere exposure to and contact with individuals or culture is sufficient to trigger behavioral contagion. Conditional on a sufficiently salient trigger, behavioral contagion leads to behavioral adaptation towards observed behavior. In this paper, we aim at expanding the existing knowledge on the drivers of saliency, in particular with a focus on both the ethicality of the observed behavior and the degree of social distance or proximity with the observed individual. A long tradition in social science highlights the importance of social identity in understanding individual behavior within the framework of social interactions (cf. Bogardus (1928)). A salient state of social identity is found to trigger favoritism towards those of stronger social kinship. The term social identity is eclectic and several of its facets have been studied in existing economic research. While the term encompasses a broad range of conceptual elements, 2 Research on behavioral contagion is fragmented and different disciplines have introduced own notions and definitions referring to the same or closely related concept. Existing research interchangeably uses different terms to describe such situations, among others: conformity, behavioral contagion, imitation, or behavioral adaptation. Due to its more generic nature and the context of our research, we will mainly resort to the term behavioral contagion or behavioral adaptation. For the sake of comprehensibility, we will abstain from clearly defining and delimiting those concepts for now. 82 from shared preferences and experiences to shared cultural and religious beliefs, in this paper we follow the primal understanding of this term as introduced by Tajfel and Turner (1979). More precise, we refer to the existence of social identity if a person derives self-esteem from belonging to the peer group and has a preference for exhibiting similar behavioral patterns (for a similar approach, see Chen & Li (2009)). Using this definition in combination with a simple implementation of social identity, as applied in our experiment via the observation of preference similarity, is conducive to both deriving lower bound results for the role of social identity in facilitating (un)ethical behavior and easier reproducibility of our results. This line of research is important from a policy perspective in generating effective measures to trigger both more pro-social and less anti-social behavior, consequently reducing the otherwise resulting economic and social inefficiencies. With notable exceptions, existing research has been struggling to overcome a number of challenges to study clean peer effects in the lab, especially in contexts where social identity plays a mediating role. Among these, inducing or at least proxying the natural occurring variation of social identity has proven to be difficult. Following the tradition of the minimalgroup paradigm (Tajfel & Turner, 1986), psychological research has introduced a number of ways to proxy social identity in the lab, such as having participants interact with participants that were assigned the same color avatar (Tajfel, 1982), a similar name (Pelham, et al., 2005), same birthday (Cialdini & DeNichols, 1989), or imagined closeness (Gunia, et al., 2009). So far, it has proven to be challenging to study peer effects under controlled settings in general, let alone within a more sophisticated social environment such as varied levels of social identity. Existing economic research has resorted to using a dual approach, studying peer effects in the lab and in the field, with both approaches having their limitations. Applying a novel methodological approach to induce varying levels of social identity in the lab allows us to combine the best features of both worlds to study peer effects beyond what had been possible so far. 83 To our knowledge, the first economic contribution examining this question in a controlled setting by varying social identity among individuals is the study by Bohnet & Frey (1999).3 They used a rough proxy for social distance by varying the degrees of identification and found evidence that social distance is decisive in predicting the extent of other regarding behavior in a dictator game setting.4 Other examples of economic approaches that bridge anonymity and induce social identity, among others, vary the wording of the experimental instructions (Hoffman, et al., 1996), use face-to-face interaction (Bohnet & Frey, 1999) or show pictures to one party only (Eckel & Petrie, 2011), reveal names (Charness & Gneezy, 2008), reveal preferences such as those for paintings (Chen & Li, 2009), or recruit friends and family members (Brandts & Solà, 2010). Such studies typically yield the robust finding that stronger social identity triggers favoritism. Two natural problems arise with the concepts used in economics so far. Firstly, using face-to-face communication or allowing participants to interact with friends or family members introduces serious biases and crowds-out the revelation of true preferences (Roth, 1995). Secondly, and more relevant to the point of our experiment, the degree of social identification measured in the lab by these concepts can hardly be varied and is rather binary. Our approach of using dating-website questions as a matching device allows us to induce and exogenously vary different levels of social proximity in the lab in order to study their role in the spillover of (un)ethical behavior. It is this paper’s aim to shed light on three questions stated above and to contribute to a better understanding of the general mechanism of peer-effects. For this reason, we propose a novel approach to proxy different levels of social proximity among peers in a laboratory setting. Such an approach allows us to exogenously vary social characteristics and study their role in behavioral contagion. We mimic social proximity by the use of questions taken from a major American dating website to capture individual preferences and interests and use the matching scores of overlapping answers among lab participants as an exogenous 3 Although motivated by a previous study of Hoffman, McCabe & Smith (1996), the work of Bohnet & Frey (1999) is the first to directly vary (a proxy for) social distance among peers. Instead, Hoffman, McCabe & Smith (1996) varied the language used in the distributed instructions, arguing that “subjects bring their ongoing repeated game experience and reputations from the world into the laboratory, and […] dictator instructions […] may imply that the objective is to share the money with someone, who, though anonymous, is socially relatively near to the decision maker” (Hoffman, et al., 1996, p. 655). 4 Research usually refers to this concept as social identity, which encompasses both social distance and its inverse, social proximity. Throughout this paper, we will mainly refer to social proximity. Another approach to mimic social proximity used in economic and psychological experiments has been to ask participants to bring along their friends or relatives and study their interaction in a controlled environment. Among other things, we will discuss potential drawbacks of these and related approaches in chapter 3.1.5 in more detail. 84 matching device across treatments. This allows us to study decision-making going beyond simple ingroup - outgroup comparisons. Rather, our approach provides us with an extensive array of possibilities to match participants according to their shared similarities. To the best of our knowledge, we are the first to use such an approach. Thus, we not only complement existing field studies, but also broaden the scope and utilization of lab experiments in explaining behavior and behavioral changes in peer settings, especially within the unethical domain. For this purpose, we extend the currently existing approaches by a social component that sufficiently considers the relevance of social distance and proximity to one’s peers in affecting behavioral decisions. This attempt will be at the heart of this paper, leading to a proposed theoretical extension of Akerlof’s (1997) and Glaeser & Scheinkman’s (2004) seminal work on social interaction, social distance, and conformity, where the extent of social distance is a function of geographic location. We, however, emphasize the role of social distance as a function of an actual overlap in personality based on e.g. personality traits or interests. Another substantial contribution of our research is the direct comparison of behavioral contagion of ethical and unethical behavior. In general, existing research has focused on shedding light separately on behavioral spillovers of either ethical behavior (cf. Thöni & Gächter (2015)) or unethical behavior (cf. Gino et al. (2009)). Considering the differing settings and games used in existing experiments to study behavioral contagion, current research does not help to understand whether and to which extent behavioral contagion in either direction differs from each other. Instead, experiments that add to the understanding of the potentially different mechanisms should place participants in a uniform environment and allow for the spillover of ethical and unethical behavior. Our experimental set-up allows us to study these questions. Participants play a one-shot dictator game in which they decide how much money to donate to or take away from a charity, which resembles a variation of a dictator game implemented by List (2007) and Bardsley (2008)). Here, the (un)ethicality of (taking away) donating money to the charity is stressed by explaining the consequences of their behavior clearly to the participants. That is, all the money that is (taken away) donated to the charity will (not) be given forward to the charity, thus individual behavior will (harm) benefit the charity. Hence, participants face a riskless 85 but in terms of its (un)ethicality precisely defined situation in which they have to decide whether or not to personally benefit at the expense of a charity of their choice. After reaching their initial decision, participants are given the opportunity to learn about other participants’ initial decisions followed by the option to revise their own initial decision. Several economic and psychological theories are able to explain behavioral contagion even under full anonymity and without observability of one’s own initial and potential revision decision, as implemented in our experiment. Among these are concepts relating to social decisions and social distance (Akerlof (1997), Glaeser & Scheinkman (2004)), imitation of preferences (Sliwka, 2007), social learning (Bandura, 1971), norms (Cialdini et al. (1990) and Bicchieri (2006)), self-expansion (Aron & Aron, 1986) or even guilt (Kandel & Lazear, 1992). Many of these concepts are not strictly distinct, in both their assumptions and predictions. Thus, in this paper we will not attempt to resolve which approach explains behavioral contagion best but rather focus on shedding light on the drivers of behavioral contagion and its interrelation with the social identity dimension and the extent of (un)ethicality of observed behavior. 3.1.3 The Conceptual Framework of Behavioral Adaptation 3.1.3.1 The Mechanism of Social Interaction and Challenges of Measuring its Effects A growing body of literature suggests that social interactions are principal not only to humans, e.g. in making social or economic decisions, but also to animals, e.g. in finding the right strategy or place to maximize one’s hunting success (Laland, 2002). Evidently, social interactions trigger different reactions and outcomes, which are referred to as social effects. Different streams of literature refer to such interactions in different ways: bandwagons, conformity, epidemics, herd behavior, imitation, neighborhood effects, peer influences, social learning, or social norms (see Hyman (1942), Merton (1957), Granovetter (1979), Jones (1984), Manski (2000)). In addition, Pingle & Day (1996) subsume these and other types of 86 behavior, including following an authority, habit, thoughtless impulse, and hunch, as economizing behavior. Some call it simply peer effects.5 Although these notions refer to different mechanisms, their outcome is often (but not always) similar; that is, the individual’s adaptation to observed behavior. However, some mechanisms encompass stronger interaction with the peers than others do (e.g. the peer’s ability to observe my reaction and exact change in behavior) and involve more or less deliberation.6 As we will discuss in this chapter in more detail, existing research indicates that the existence and persistence of social effects are context specific and may even lead to, among others, higher consumption of alcohol and drugs, cheating, and smoking. The same is also found to be true for pro-social and cooperative behavior. Thus, before turning to the specific effects of social interaction, we shall make the effort to understand the underlying mechanisms first. The concept of social interaction lies at the heart of social psychology and sociology, for example in order to explain the formation of tastes (Weber, 1978). The importance of social interaction in explaining social phenomena has a long tradition and is often ascribed to Sutherland’s Differential Association Theory (Sutherland, 1939). Akerlof (1997) stresses the view that the theory of social interaction is key to understanding why individuals do not succumb to isolated and purely self-maximizing decision-making. Instead, this concept gives rise to conceive an individual as someone who constantly interrelates with the underlying social environment and produces and deals with the resulting externalities. A principle consequence of extending the rational model based on Becker’s early work by such a social dimension is that the type of the resulting individual is more sophisticated and resembles more closely to the intuition of sociologists than the classical economists. 5 Often, the term “peer effects” is used to refer to all of these mechanisms without specifying the exact channel through which behavioral adaptation arises. Although it is important to shed light on the different channels through which social interaction potentially transitions into behavioral adaptation, the focus of this study lies in understanding the role of exogenous factors such as social proximity that have the potential to influence the intensity of behavioral adaptation. We shall not attempt to settle the argument of which approach explains the mechanism of our experiment best. Instead, we will discuss some of the more prominent concepts and mechanisms in the fields of economics and social psychology that resemble the mechanism of behavioral adaptation the way it is implemented in our experiment in chapter 3.1.4. For our purpose, however, we will assume peer effects to be in play whenever an individual ′ behavior changes after having been exposed to behavior of individual , irrespective of the direction of the behavioral change. In turn, however, we assume behavioral contagion to be in place whenever individual ′ behavior changes in the direction of individual  behavior. 6 It is worth noting that it is far beyond this paper’s scope to shed light on all of these concepts. Instead, we will pick out and discuss those concepts of which we believe are of bigger importance to what we analyze within our experiment. 87 Manski (1993) distinguishes between two types of social interactions.7 One is endogenous interaction in the form of, for example, information exchange among criminals or social norms. The behavior of the relevant peer group mediates the likelihood that the individual will engage in the same kind of behavior. In addition, exogenous interaction emphasizes that the propensity of an individual to behave in a certain way is also mediated by exogenous characteristics of the group such as their attitude toward crime or social and economic status.8 In social interactions, externalities abound. The key mechanism of social interaction implies that one’s personal net benefit is a function of the behavior exhibited by one’s relevant social group or contact person (Glaeser & Scheinkman, 2004).9 Inherent to such interactions are strategic complementarities in the form of circular cascades where “even if changes in fundamentals create only a small change in the level of activity for each individual, each individual’s small change will then raise the benefits for everyone else pursuing the activity” (Glaeser & Scheinkman, 2004, p. 84). In principle, small changes in fundamentals may cause large shifts in outcomes, which are sometimes referred to as the butterfly effect, a term hailing from the chaos theory (Lorenz, 1963). Social interactions are also highly relevant in understanding the spread of criminal behavior. Social interaction plays a decisive role in the formation of gangs and the recruitment of young criminals (see Reiss (1988) and Jankowski (1991)). This is particularly true for the criminals’ decision to engage in illicit behavior jointly (Reiss (1980)). “Social interactions seem to create a sense of invulnerability and a willingness to violate social norms and take 7 Glaeser & Scheinkman (2004) provide a more distinct categorization of the mechanisms that generate social interaction: physical, learning, stigma, and taste-related interactions. For the purpose of this experiment, we will extend this categorization in order to better capture the mechanism of behavioral changes we are interested in. We will discuss these points in more detail in chapter 3.1.6. 8 While these two aspects represent interactions that are shaped by the underlying social environment, Manski (1993) also introduces correlated effects that explain similar behavior as the result of facing similar institutional environments (see also Manski (2000)). This third aspect is not considered any further since correlated effects are not social effects and are thus neither created by social interactions nor create social multipliers (see Glaeser, Sacerdote & Scheinkman (1996), (2003)). Using a novel approach in our experiment, a clean variation of both endogenous and exogenous interaction allows us to draw causal inferences that are more precise than what has previously been possible. We will return to this important point in our design section of chapter 3.1.6. 9 The term ‘relevant social group or contact person’ is vague in existing research. A social group or contact person is selfreported and defined from one’s individual point of view and refers to one or more individuals whose behavior either has a direct or indirect impact on one’s well-being, e.g. through resource externalities or other-regarding preferences. While it is important to understand the different channels through which behavioral spillovers occur, the focus of our study is to shed light on the drivers instead. See chapter 3.1.4 for a discussion. 88 risks, as long as one is in the company of like-minded individuals” (Glaeser, et al., 1996, p. 511). Social interaction also strongly affects stigmatization, which is important from an evolutionary perspective. As the number of criminals rise, illicit behavior becomes more common and thus potentially more accepted; and as criminality converges towards ‘normality’, the (social) rents of illicit behavior increase due to increased attractiveness, social acceptability and a crowding-out of legal activities where earnings from legal activities are stolen by criminals (see Rasmussen (1996) and Murphy, Shleifer & Vishny (1993)). Researchers face substantial difficulties measuring social interactions and its resulting effects in a clean way. In a real-world instance, assigning changes in individual conformity to distinct mechanisms are subject to identification problems (Angrist, 2014). As argued before, this problem arises because individual behavior is affected by both endogenous (e.g. the group’s behavior) and exogenous effects (e.g. group characteristics) and, in addition, uniform behavior can be the result of similar unobserved characteristics (Manski, 1993). Previous empirical research involved regressing a person’s actions on the action of his peers. However, Manski (1993) points at three fundamental problems concerning this methodology: first, drawing causal inference is difficult when endogeneity is a problem, in particular when the individual’s and its peer’s behavior is interactive and influences each other circularly. Second, omitted variables increase the likelihood of spurious correlations between actions. Third, in reality, sorting and self-selection into particular neighborhoods renders it difficult to understand what actually drives behavior. Arguably, empirical research in particular faces these challenges because one only observes the behavior of individuals who selfselected themselves into, for example, moving to a better neighborhood, but not of those who decided to turn down the opportunity (Glaeser & Scheinkman, 2004).10 Of particular interest to our research are social interactions leading to the spread of unethical behavior. Existing research points to a strong presence of positive covariates across individuals’ decision to engage in criminal behavior. In particular, Glaeser, Sacerdote & 10 Prominent examples and forceful ways of addressing these issues include, among others, the work of Case & Katz (1991), Katz, Kling & Liebman (2001), Angrist & Lang (2004), Kling, Ludwig & Katz (2005), Kling, Liebman & Katz (2007), Ludwig & Kling (2007), Damm & Dustmann (2014), and Chetty, Hendren & Katz (2015). We will discuss these and other empirical contributions examining peer effects and behavioral adaptation in the field in more detail in this chapter. 89 Scheinkman (1996) state that only covariance across criminal decisions of individuals explains existing variance in crime rates, which is far beyond any theoretical prediction of crime rates. In order to mitigate any term-related confusion on the side of the interested reader, we are in need of a term that allows us to capture a particular type of behavior that resembles a subset of social interaction and its resulting social effects. More specifically, we are not interested in the drivers of all kinds of behavioral changes resulting from observation, but only in behavioral changes that lead to a convergence of behavior. We are thus proposing the impartial term behavioral adaptation to capture such behavior.11 The term behavioral adaptation refers only to a subset of social interaction because social interaction may lead to all kinds of social effects where the resulting behavior may or may not converge towards what has been observed. In its basic form, the existence of social effects does not tell us much about the specific behavioral reaction, if any, that follows from this exposure. In principle, the result of social interaction could lead to either behavioral alienation or adaptation. For the purpose of this study, we will focus on the latter. In turn, behavioral adaptation refers to those situations only in which the resulting behavior is coherent to the behavior one has been exposed to. Our study sheds light on the key drivers triggering an individual’s response to become more like others, in one way or another.12 Existing macroand micro-level data inadequately catches the underlying mechanisms and the causal relationships relevant to our project. That is, answering the questions of how behavioral adaptation varies with different levels of social proximity and whether this interplay is different for adaptation towards ethical versus unethical behavior. As has been argued by Angrist (2014), with rare exceptions like Kling et al.’s (2007) Moving-to-Opportunity research, most field studies on peer-effects suffer in one way or the other from endogeneity, 11 Behavioral adaptation has also been studied in the field of evolutionary game theory as well as in the theory of learning in games (see Selten (1978), Roth & Erev (1995), Schlag (1998), and Apesteguia, Huck & Oechssler (2007), to only name a few). However, it is beyond the scope of this paper to touch upon all streams of literature that have shed light on the general process of behavioral adaptation. Instead, we focus on the studies most relevant to our experiment and extend our apologies to colleagues whose research remains unnamed in this paper. 12 Studying the drivers of behavioral alienation is a potential venue for future research, as this line of research has yet to catch attention from economists. Akerlof (1997) refers to this behavior as the result of status-seeking efforts. Beyond this, however, behavioral alienation can be the result of e.g. one’s desire to not be identified with a particular social group. 96 their findings indicate that social influence is transmitted through both channels and point in the same direction. Some studies, however, find little evidence for neighborhood or peer effects. In a highly regarded study, Evans, Oates & Schwab (1992) show that after controlling for selection bias, any measurable peer effect on teenage pregnancy and school dropout rates disappear. While being careful in not claiming that no peer effects exist at all, they rather point critically to methodological issues measuring peer effects in a clean way. Angrist & Lang (2004) use data from the Metropolitan Council for Opportunity (Metco) desegregation program in which mostly black students are sent to more affluent suburbs. Their findings indicate that there are little, if at all, positive spillovers on students. Similarly, Burke & Sass (2013) find little evidence of classroom peer effects on student achievement for Florida public school students. Likewise, Ludwig & Kling (2007) find little evidence for the contagion of crime hypothesis using MTO data. Instead, their findings indicate that crime rates are merely driven by neighborhood racial segregation. In conclusion, the existence of peer effects is up for scholarly debate, which is mainly driven by methodological challenges and data problems. However, in following Angrist (2014), the previously discussed MTO program yields the most promising setting to study clean peer effects in the field. Previous research that utilized MTO data yields, among other things, strong gender asymmetries in terms of the evolution of criminal behavior. The authors suggest that these findings can be attributed to differences in which males and females respond to their environment and its influences. Ultimately, this leads to differences in magnitude and speed at which (illicit) behavior is picked up. Evidence from the Lab In what follows, we shall not attempt to provide an exhausting overview of the comprehensive literature dealing with peer effects in general. Instead, we will focus on a range of influential studies using lab experiments to shed light on mechanisms relating to behavioral spillovers that are more in line with our paper’s focus. Later in the paper, we will refer to these studies in more detail where necessary. Over the last decade, a comprehensive stream of literature studying spillover-effects and behavioral adaptation in the lab has emerged that complement the ongoing important work 97 in the field. As discussed previously, the methodological shift was strongly driven by challenges relating to identifying these effects in a clean way using observational data. Several researchers claim that although the most recent generation of studies measuring such effects with observational data has succeeded to make important steps towards tacking the challenges outlined before, controlled lab experiments are still the gold standard in reducing noise and potential confounds (Angrist, 2014). “However, even if the setting offers an almost perfect opportunity to identify peer effects in many of these studies, the impossibility of controlling for all local or personal confounding factors and for endogenous sorting makes the identification strategy not fully convincing" (Falk & Ichino, 2006, p. 40). Early laboratory research studying peer effects and social identification jointly has been pioneered by Hoffman, McCabe & Smith (1996) and Bohnet & Frey (1999). These studies made use of variation in the instruction’s wording or enhanced face-to-face communication to study the role of social identification in giving decision, equivocally finding support for its relevance (see also Charness & Gneezy (2008)). We will return to these studies in more detail in chapter 3.1.5. Other studies have looked into peer effects in productivity decisions. In a highly regarded study, Falk & Ichino (2006) found robust evidence for the existence of peer effects in a productivity task. Their results indicate that low-productivity workers are particularly susceptible to peer effects, which results in an over-proportional raise in productivity. Following the work of Mas & Moretti (2009), subsequent studies tried to disentangle the naturally occurring channels of simultaneously observing peers and being observed by peers. For the most part, these studies found the latter channel to be more effective than the former in boosting productivity (cf. Georganas, Tonin & Vlassopolous (2013); for exceptions see Veldhuizen, Oosterbeek & Sonnemans (2014)). Along the lines of studying behavior in the workplace, Gächter et al. (2012) set up an experiment that investigates reciprocal behavior under observability of other people’s actions. They find that the individual’s extent to comply with norms of reciprocity is significantly driven by both pay and effort comparison information. Zafar (2011) experimentally examines charitable giving in a social context. He finds that by systematically revealing information, both the learning about descriptive norms (through observing what others do) and the image-related concerns (through revealing own behavior to the reference group) drive 98 individual contribution levels. In a more delinquent context, Falk & Fischbacher (2002) investigate peer effects in the form of conditional stealing behavior. In particular, they investigate whether an individual’s inclination to steal is dependent on other peer’s stealing behavior. Their main findings suggest that, on the aggregate level, people make stealing decisions conditional on the behavior of their peers. In economics, a limited number experimental research has also pointed at the contagion of both selfish behavior and dishonesty. Bicchieriy and Xiao (2009) study a dictator game with varying information on other participant’s selfish or fair behavior, finding that fairness in actions is contagious. More to the point of our research, Innes & Mitra (2013) use a variant of Gneezy’s (2005) deception game to study whether dishonesty breeds dishonesty. Their findings suggest that the beliefs about other’s dishonesty is indeed contagious, potentially driven by the wiggle-room created by such social cues and thus representing a justification device for one’s personal dishonest behavior. 3.1.4 Drivers of Behavioral Contagion: An Interdisciplinary Perspective Prior to delving into the subject in more detail, some effort will be made to disentangle many of those existing concepts explaining why and under which circumstances people demonstrate a change in behavior as a function of their peer’s behavior. In both economics and psychology, several concepts have been developed over the past few decades referring to similar and often the same reasoning to change one’s own behavior.17 The complexity of the self and the dependency of one’s own behavior on what one observes of peers has received a lot of scholarly attention (cf. Baumeister (1987), Kahneman & Tversky (2009)). It is thus important to shed light on the underlying concepts driving behavioral adaptation. In this context, we will differentiate between concepts that originated in the (I) economic literature and in the (II) literature of (social) psychology. The economic concepts include (I.1) social decisions and social distance (Akerlof (1997), Glaeser & Scheinkman (2004)), (I.2) image related concerns (Bernheim, 1994), (I.3) taste for conformity (Bernheim, 1994), and (I.4) imitation (Alós-Ferrer & Schlag (2009), Sliwka (2007)). In addition, we also discuss some (social) 17 See discussion in chapter 3.1.3. 99 psychological concepts including (II.1) social learning (Bandura (1971)), (II.2) norms (Cialdini et al. (1990), Bicchieri (2006)), and (II.3) psychological closeness and vicarious dishonesty (Aron & Aron (1986), Goldstein & Cialdini (2007)). One point is worth clarifying. In this paper, we shall not attempt to settle the argument which scholarly approach explains behavior best. Rather, for our purpose, we will use the previously defined unifying terms behavioral adaptation or behavioral contagion throughout the paper in order to refer to one person’s decision to change initial behavior as a function of observed behavior from at least one other person. Using this umbrella term will allow us to focus on what is relevant without getting lost in conceptual debates. We deem it important to help understand the role social proximity plays in the change of behavior and to what extent proximity mediates the spillover of ethical and unethical behavior. It will be the next sub-chapters’ aim to dissect the different approaches in more detail and to explain the reasons why this is the case. At times, the predictions and the empirical outcomes are identical, but the underlying forces causing such outcomes are diverse. In what follows, we will discuss the different theories outlined above and their implications with respect to changes in behavior. At the end of this chapter, we will relate these theories to our experiment and provide a systematic breakdown with respect to the fit of each particular theory to explain behavior in our design.18 3.1.4.1 Concepts in Economics Social Decisions and Social Distance Seminal contributions by Akerlof (1997) and Glaeser & Scheinkman (2004) were among the first to provide an economic framework highlighting the relevance of social distance in affecting social interaction and behavioral adaptation. Beyond rationalizing one’s own behavior in an isolated environment, research indicates that behavior is a function of both pure own-maximizing and other-regarding concerns. 18 It is worth noting, however, that our discussion here will be purely descriptively. We will contrast the explanatory power of these concepts within the frame of our experiment in chapter 3.1.6. 100 For a while, traditional economics has neglected such interdependence and rather put emphasis on individualism. In reality, however, decisions are rarely brought about in total isolation, but are rather the result of an interplay with one’s (social) environment. Arguably, individuals care about both status concerns (in absolute and relative comparisons) and other’s well-being (Fehr & Schmidt (1999), Bolton & Ockenfels (2000), Charness & Rabin (2002)). Because social interactions typically render externalities with the potential of slowing down conversion towards socially (un-)desirable equilibria, it is important to understand the underlying mechanism. In his attempt to explain the connection between social interaction and behavioral adaptation (conformity, in particular), Akerlof (1997) made use of the Newtonian theory of gravity. Akerlof’s approach centers on explaining conformist behavior as a function of distance in the social space. Such difference in one’s social space is characterized by the difference in behavior between oneself and a person or group of relevance. For his purpose, Akerlof applies the concepts of a gravity model allowing him to argue that conformity leads to benefits (e.g. higher individual utility or gains from trade) that are negatively correlated with distance in social space.19 Akerlof puts a reduced form concept forward in which one’s utility declines as the individual’s behavior deviates from the behavior of others. Assuming the existence of representative agents, his model predictions yield a set of equilibria in which the ultimate behavior of all individuals is the same, thus clearly characterizing the behavior of every party.20 19 Because geographic location is one determinant of social interaction, this approach is a generalization of what sociologists have coined “social geography”, which has been inspired by Krugman’s (1991) work on economic geography. 20 Although ex-post behavior is uniform across individuals in equilibrium, such an approach still models conformism because of the ex-ante desire to be and behave like others. Akerlof also proposes an extension with individual heterogeneity that allows for the formation of social sub-groups with own norms and values. The extension models a mechanism in which (randomly distributed) past social location for each individual is inherited. In combination with static expectations about “the positions to be occupied by others in social space […] such a model can portray stable groups in low level equilibrium traps because individual’s incentives to choose x to conform with those whose inherited social locations are close may overwhelm their incentives to choose x for intrinsic reasons” (Akerlof, 1997, p. 1010). In our experiment, we will study this mechanism by holding the other’s behavior fixed and thus allowing for deliberate behavioral adaptation free of unstable, higher-order, or even nonexistent beliefs about the other person’s next move. We will return to this argument in our design section. 101 Ultimately, in order to maximize one’s intrinsic utility, the individual will converge towards the observed behavior and thus reduces the existing social distance.21 With reduced distance, social interactions become more favorable through facilitation of e.g. mutually beneficial trade. Image Related Concerns In understanding individual behavior, one fundamental question arises: “How much of what we do is the result of our genetic code (nature); and how much of it is a function of our environment, including the actions of those around us and our own past actions (nurture)?” (Cabral, 2005, p. 15). The threat of reputation loss and being punished in the case of wrongdoing is an integral component of individual decision making in general. By weighing costs against benefits, the impact of possible reputation loss might deter individuals to engage in illicit behavior of any kind. Individuals are striving for social inclusion and recognition. They want to belong to one or more social groups and engage in social interaction. Striving for social identity is an inherent characteristic, thus decisively driving the individual’s yearning for maintaining an adequate reputation. In this sense, losing face as the result of deviant behavior may serve as an avoidance, as this might result in exclusion from the group (cf. Bernheim (1994)). Along these lines, Akerlof (1980) argues that the deviation from social standards might lead to loss of social reputation and consequently to ostracism. However, it is worth noting that deviant behavior not necessarily implies bad behavior per se, but has to be understood as a deviation from the underlying social norm in either direction. Reputation can also work as a means to preserve a coherent self-image that allows keeping an internal consistency (cf. Baumeister (1998)). The concept of the self has multiple facets and impact behavior in different ways, as level of self-regard and self-esteem, the extent and content of self-identity as well as the structure of the self-concept exhibit motivational implications (cf. Wells 21 A point not made clear in this approach is whether such desire to conform is also dependent on the signaling value of one’s behavioral adaptation. The model argues that the desire to conform and reduce social distance is entirely driven by intrinsic concerns. However, the inability to signal conformity to one’s peers runs counter to the initial concept of conformism (introduced by, for example, Ash (1958) and Bernheim (1994). Thus, this concept rather resembles the ideas of imitation of preferences (Sliwka, 2007) or self-expansion (Aron & Aron, 1986) that will be addressed in the next subchapters. 102 (1978)). By that, people try to avoid a negative update of their self-image through their actions, which otherwise would result in an internal conflict. Such a mechanism possesses the power to prevent individuals from engaging in delinquent behavior in the first place. This mechanism is more distinct when the group is salient, consequently giving rise to more extensive alignment with the group’s behavior. What is more, own actions could also be subject to social signaling, which highlights the perception of oneself by others and thus potentially triggering conform behavior (cf. Grossman (2010)). It is important to clearly distinguish between these two motives and mechanisms when analyzing one’s individual drivers for conform behavior.22 However, for image related concerns to be effective, a setting of repeated interactions has to be in place. In a one-shot setting, the threat of a reputation loss is negligible as the players won’t face each other for a second time. If recurring interaction is not taking place, any deviant behavior can neither be traced back to the individual nor will be colored negatively with regard to future collaboration. Existing research supports this perspective. For example, there is comprehensive experimental literature pointing to the fact that contributions and compliance with underlying norms rise in multiple-shot interactions. In particular, this is true when transparency (e.g. ability to observe other participants’ contribution decisions) or punishment (e.g. for non-compliance with social norms) is possible (cf. Andreoni (1995), Fehr & Gächter (2000), Cameron et al. (2009), Chaudhuri (2011)). Consequently, in order for reputation to be effective (even in absence of a punishment-mechanism), a setting of repeated interaction has to be in place and actions have to be observable to other people. Taste for Conformity Research and observations of real life situations indicate that underlying social factors decisively influence individual behavior. Motivation for a particular behavior is driven by the inherent desire to be valued and to win prestige, esteem, popularity, and acceptance (Ellingsen & Johannesson, 2008). Arguably, both the introversive coherence and outward appreciation matters and can be achieved through conformity in behavior. The basic idea here is that any infinitesimal deviation from effective norms might be punished by the social 22 These aspects will be taken into consideration in the experimental design. We will return to this point later. 103 group (Bernheim (1994); see also Akerlof (1980) and Tirole (1996) for related work on the role of reputation in individual decision making and behavioral alignment). In his seminal contribution, Bernheim (1994) formalized the concept of taste conformity. He argues that the individual’s preference for status is in line with psychological, evolutionary, and behavioral considerations. In particular, natural selection favors concerns for status as this goes hand in hand with greater opportunities for reproduction. From a behavioral point of view, such concerns act as a reinforcing device to form preferences for higher esteem because esteemed individuals are more likely to receive better treatment. Although partly representing a departure from the traditional formulation of preferences, this approach does not necessarily require the abandonment of consistent, self-interested optimization processes. By assuming that status depends on “public perceptions about an individual’s preferences over actions […] esteem is determined by expectations about future actions and that tastes and proclivities are the best predictors of future actions” (Bernheim, 1994, p. 843). Unsurprisingly, research indicates that people share similar preferences and opinions within the same group, which can arguably be both the antecedent and the effect of social interaction and norm alignment.23 Aligning with existing norms is a way to signal one’s (wishful) belonging to a certain social group. As argued by Hogg & Tindale (2002), enacting in-group-prototypical behavior is a way to validate one’s own group membership, not only to the social group but also to themselves. By that, aligned behavior might function as a signal to both the external (e.g. the peers) and the internal world (e.g. themselves) in order to solidify the sense of belonging. If the individual cares about status, one has to set the right signals to the peers in order to create esteem. This esteem will be provided in case the individual behaves in a way that is expected from him. In this vein, Bernheim (1994, p. 844) derives the following proposition, highlighting the decision process leading to behavioral adaptation: “When popularity is sufficiently important relative to intrinsic utility (defined as the utility derived directly from consumption), many individuals conform to a single, 23 Manski (1993) refers to this as the previously mentioned reflection problem, rendering it extremely difficult to study clean peer-effects in the field. We will return to this point shortly in the experiment’s design section of chapter 3.1.6. 104 homogeneous standard of behavior, despite heterogeneous underlying preferences. They are willing to suppress their individuality and conform to the social norm because they recognize that even small departures from the norm will seriously impair their popularity.” Arguably, the extent at which people are willing to align with their peer’s behavior is driven by the degree of social identification to the social group. Social identity theory provides helpful guidance to distinguish between different magnitudes of personal identification with the peer’s behavior (cf. Tajfel & Turner (1979), Tajfel (1982)). According to this approach, people categorize themselves and other persons into different social groups. Membership in a group is defined as the social identity and individuals strive to enhance their position and self-esteem through actions. These actions encompass the alignment with existing group norms. “[…] the social identity analysis of categorization processes suggests that group cohesion or solidarity is not only attraction among group members, but also attitudinal and behavioral consensus, ethnocentrism, in-group favoritism and intergroup differentiation, and so forth – the entire range of effects of categorization-based depersonalization” (Hogg & Tindale, 2002, p. 65). A comprehensive stream of literature suggests that such identification is driven by in-group-out-group concerns, finding that in-group favoritism is found across different contexts and social settings (cf. Tajfel (1982), Hoffman, McCabe & Smith (1996), Bohnet & Frey (1999), Eckel & Petrie (2011), Charness & Gneezy (2008)). From this one can derive the assumption that the stronger the desire to be part of the in-group, the stronger one’s intrinsic willingness to comply with prevalent rules and engage in conformity by adapting social norms. Such a desire is driven by, for example, the inherent relevance of the social image component (cf. Andreoni & Bernheim (2009)). The stronger the social identity with a reference group, the stronger the effect of social comparison on individual conformity. Overall, taste for conformity is a function of existing social identity and the extent of social comparison to one’s reference group. Deviating from the group’s behavior (even in extreme cases where the group behavior is clearly wrong, cf. Asch (1958)) and thus being perceived as different might create discomfort in form of substantial disutility. 105 Imitation In this subchapter, we will discuss two streams of literature suggesting that behavioral adaptation is the result of either pure behavioral imitation (e.g. due to observing behavior that leads to a superior outcome) or the adjustment of preferences (e.g. contingent on the behavior of the peers). With no claim to completeness, this chapter’s focus will lie on the selected sample of contributions discussing the topics of behavioral imitation and adjustment of preference as a mechanism for behavioral adaptation. Pure Behavioral Imitation Imitation is a common behavioral trait of both animals and humans (Laland, 2002). While it is likely that imitation is driven by evolutionary and cultural facets, its pervasiveness can only be explained if it sufficiently often leads to a desirable outcome. Imitation has a number of desirable features because it allows one, among other things, to free ride on superior information (in the spirit of social learning, but see also Sinclair (1990)), to save mental resources (in the spirit of bounded rationality, but see also Conlisk (1980)), or as a means to be regarded similar to others (in the spirit of the social esteem argument, but see also Cho & Kreps (1987)). Along these lines, Alós-Ferrer & Schlag (2009) abstract from the standard Bayesian beliefbased approach and introduce the concept of imitation as a belief-free behavioral rule. The authors make the case that imitation is triggered by the observation of a superior outcome. In face of different imitation concepts, the degree of sophistication involved in assessing the extent of the observed agent’s better performance clearly dominates a simple ‘follow whoever performs better than me’ rule. The reason behind this argumentation is that “the reluctance to switch when the observed choices are only slightly better than the own might not be due to switching costs, but rather to the fact that the payoff-sensitivity of the imitation rule allows the population to learn the best option” (Alos-Ferrer & Schlag, 2009, p. 273). In essence, this approach highlights the idea that behavioral adaptation can be triggered by behavioral imitation that results from observing (sufficiently) superior outcomes. Adjustment of Preferences Another strain of literature models behavioral adaptation in a way that is hard to swallow by traditional economists: preferences are not stable per se but are subject to interrelations 112 may induce a carry-over effect of behavior (Goldstein & Cialdini, 2007), whereas the perspective-taking theory assumes that the imagined social closeness is already sufficient to facilitate behavioral adaptation (Galinsky, et al., 2005). As has been discussed previously, psychological research has introduced a number of ways to proxy social identity in the lab. Due to methodological differences, existing economic research has resorted primarily to using rough proxies to introduce social identification in one form or another. Exemplarily, this has been achieved by varying the wording of the experimental instructions (Hoffman, et al., 1996), using face-to-face interaction (Bohnet & Frey, 1999) or showing pictures (Eckel & Petrie, 2011), revealing names (Charness & Gneezy, 2008), revealing preferences such as those for paintings (Chen & Li, 2009), and recruiting friends and family members (Brandts & Solà, 2010). In line with most psychological research, these studies find that social identity triggers in-group favoritism. However, two natural problems arise with these approaches to introduce social identification in the lab: firstly, the introduction of potential biases caused by face-to-face communication within the lab or letting participants interact with their kin (Roth, 1995). Exemplarily, it has been shown that individuals discriminate using social cues such as looks or skin color (Eckel & Petrie, 2011). Additional research has investigated other-regarding behavior as a function of social distance. Using the dictator game setting, Bohnet & Frey (1999) find that other-regarding behavior is more pronounced under stronger identification. They incepted stronger identification by the use of identification prior to the dictator game. Again, their approach is only a rough proxy for social proximity and does not control for potential biases stemming from, among others, unobserved attraction effects. Secondly, and more along the lines of our experiment, applying these methods to introduce social identification does not usually allow for a variation of social proximity beyond a simple binary outcome of no social identification versus some social identification. In our experiment, we propose a method that allows us to naturally vary the degree of social identification in the lab without having to deal with the issues raised above. We will return to this in our design section. Overall, it may be stated that different streams of existing research point at the significance of social identity in explaining the magnitude of behavioral adaptation thus making it worthwhile to examine this topic in more detail. 113 The ethicality of observed behavior might also be decisive to the magnitude of behavioral adaptation. As we will argue throughout this paper and will be explained more formally in this chapter, we expect that observations of unethical behavior render behavioral adaptation more likely and more extensive as compared to observations of ethical behavior. Existing empirical literature on the slippery-slope effect (cf. Gino & Bazerman (2009), Welsh et al. (2015)) as well as on the “broken windows effect” (cf. Beckenkamp et al. (2014), Lefebvre, Pestieau, Riedl & Villeval (2015)) support this point of view.28 Arguably, the magnitude of behavioral adaptation is asymmetric in one direction or another, indicating that the likelihoods for adapting behavior are different when observing ethical or unethical behavior. Below, we will provide some theoretical arguments for this claim before turning to a more formal approach. Individuals engage in ethical and unethical behavior for all sorts of reasons and sometimes turn evil extremely quickly under particular circumstances (see the infamous Milgram experiment (Milgram, 1963) and Zimbardo’s Stanford prison experiment (Zimbardo, 1971)). Different perspectives are helpful in understanding one’s rationalization to engage in (un)ethical behavior, such as self-rationalization, obedience to norms, or active signaling to one’s peers. Individuals tend to rationalize their behavior with themselves while trying to avoid cognitive strain stemming from, among others, negative self-image updating (cf. Wells (1978), Baumeister (1998)). Here, individuals do not want to think of themselves as a bad person and thus tend to engage in self-deception and moral rationalization and create positive disillusions that sustain a positive self-image (cf. Taylor (1989), Mazar, Amir & Ariely (2008)). “Self-deception allows one to behave self-interestedly while, at the same time, falsely believing that one’s moral principles were upheld. The end result of this internal con game is that the ethical aspects of the decision “fade” into the background, the moral implications obscured” (Tenbrunsel & Messick, 2004, p. 223). Arguably, by avoiding thinking of immoral acts as such, mental processes leading to a denial of facts facilitate a person’s inclination towards unethical behavior. 28 Although not explicitly along the lines of (un)ethical behavior, a similar argument can be made based on the findings within the extensive literature on conditional cooperation and punishment (see Chaudhuri (2010) for a selective review). 114 From the perspective of norm theory, the exposition of unethical behavior comes with a relatively higher risk of being condemned by and excluded from the social group (Bicchieri, 2006). The degree to which one can behave unethically (if at all) among one’s social peers is typically grounded in uncertainty. Consequently, a person’s decision to engage in either ethical or unethical behavior is not only simple cost-gain maximization under risk but also involves accounting for the inherent uncertainty. The uncertainty involves, among others, the deviation from what is deemed appropriate within the social group, reputation loss, and the potential exclusion from the group. From a psychological perspective, engaging in unethical behavior involves higher costs, as one has to forcefully convince oneself that gaining personally at the expense of harming others is acceptable. Processes leading up to this reasoning involve, among others, self-deception, justification, and hypocrisy. Typically, only the acting individual profits from unethical behavior in (non-) monetary terms, such as stealing money, while another person is losing out. In turn, however, normally both individuals gain from ethical behavior such as donating money: the receiving individual gains in monetary terms and the acting individual, while losing out in monetary terms, is (over-)compensated from a behavioral perspective by either the positive feelings of warm glow of giving and altruism (Becker (1974), Andreoni (1989) & (1990)) or social pressure (Akerlof & Kranton, 2000). Plausibly, to convince oneself to behave unethically involves more (psychological) effort than to behave ethically. From a signaling perspective, observing someone who is socially closer to oneself as compared to observing a stranger sends a more salient form of social learning about appropriate (or at least tolerated) behavior within the social group. This assumption is also along the lines of Schelling’s (1968) prominently stated “the more we know, the more we care” and a number of experimental results in various settings (see Eckel & Grossman (1996), Bohnet & Frey (1999), Charness & Gneezy (2008), and Gino & Gallinsky (2012)). The resulting resolution of uncertainty allows the individual to overrule their own concerns about the inappropriateness of unethical behavior. This stream of literature suggests that observing unethical behavior might trigger stronger contagion as compared to observing ethical behavior. For these and more reasons, we expect behavioral adaptation to be asymmetrically biased towards unethical behavior. 115 3.1.5.2 Theoretical Model We extend the model introduced by Akerlof (1997) and expand the underlying concepts in a way that is conducive to understanding the role of social identity and the (un)ethicality of observed behavior in behavioral contagion.29 Previous work put emphasis on modeling social interactions and behavioral adaptation as a function of distance in the social space, e.g. resulting from distance in behavior and geographic location to one’s peers. We, however, shed light on the relevance of social distance and proximity resulting from overlapping and common interests and preferences in driving behavioral adaptation. Intuitively, one could reasonably assume that the behavior of one’s family and friends exhibits a more salient signal and thus is taken more strongly into consideration than observing random strangers. What is more, we follow the previous discussion and assume additionally an asymmetry in behavioral contagion depending on the ethicality of the observed behavior. This is a substantial extension to Akerlof (1997), where an asymmetric adaptation in behavior cannot be illustrated. The purpose of this section is to outline a simple reduced-form model that allows one to draw predictions about individual contribution decisions in the various contexts that are reflected by the experimental design. These predictions can then be tested empirically. The principal goal of this exercise is to show how behavioral contagion is mediated by social identity and the (un)ethicality of observed behavior and how these factors affect the adaptation gap, that is the difference between one’s own and observed behavior, after contagion has taken place. We will introduce the symmetric model first before turning to the extension. Symmetric Contagion Prior to introducing the formal model, we begin with the intuition of what the model should capture. The underlying idea is that individual behavior is a function of the peer’s actions 29 In particular, we focus on Akerlof’s (1997) quadratic utility version of the conformity model as this generates unique equilibria from which we can derive testable hypotheses for our experiment, which we also deem to be more in line with the general story of this paper. While Akerlof’s general model puts more emphasis on one’s own decisions, the quadratic utility model has an a priori assumption that puts one’s own and peer behavior on an equal footing with respect to how behavior affects individual utility. Assuming that individuals are on a continuum between extremely selfish and extremely altruistic, the quadratic utility approach is conclusive. It should be noted that this assumption is not crucial to our model’s predictions as we mainly focus on the relevance of two factors, i.e. social identity and the (un)ethicality of observed behavior, in affecting one’s own decisions. Thus, the equal-weight assumption is not decisive in predicting the direction in which individual behavior changes as a function of those two factors. 116 and thus encompasses more than one’s self-referentiality. In a situation of social interaction, an individual is expected to face a trade-off, weighing one’s own preferences against the peer’s revealed preferences. To stay in Manski’s (1993) terms, social interaction and the understanding that one’s own actions are reflected by the peers and thus have an impact on other (closely related) people shapes the individual’s willingness to engage in behavioral adaptation. The magnitude to which one is willing to revise one’s own initial behavior and adapt is first and foremost a function of the social proximity to the peer. “As a consequence, the impact of my choices on my interactions with other members of my social network may be the primary determinant of my decision, with the ordinary determinants of choice the direct additions and subtractions from utility due to the choice) of only secondary importance” (Akerlof, 1997, pp. 1006-1007). It is thus reasonable to assume that the individual’s utility is subject to a relative evaluation of one’s own behavior and the behavior of the peers. Similar to Akerlof (1997), the underlying characteristic of our model is the feature that individual utility is declining with increasing distance between one’s own behavior and the peer’s behavior. While the aim of this section is to outline a model that explains changes in general behavior, we will use the resulting predictions to generate hypotheses in the context of (un)ethical behavior. In our design, we resort to a two-stage dictator game in which each individual is paired with a charity of his choice. Both, the individual and the charity start with an initial endowment ==∈ ℝ of equal size. At each stage ∈[1,2], each individual  faces the choice, , of either (a) donate (part or all of) one’s own money to the charity, (b) retain the equal split, or (c) take away (part or all of) the charity’s money and add it to one’s own income. We will refer to (a) and (c) as ethical or unethical behavior, respectively. Naturally, the individual’s decision is of the form  ∈ [−I, +I ]. The only difference between both stages is the information set that the individual possesses about his peer’s behavior. That is, after completion of stage 1, the individual observes a random individual’s behavior from stage 1. At stage 2 (that is, after the observation), the individual is given the opportunity to revise his initial decision, if desired. Let .∈(0,1) depict the social proximity of an individual at time t. Importantly, let an individual’s inherent attitude towards (un)ethical behavior be described by . That is,  represents the individual ’s preference to give or take a particular monetary amount within the 117 boundaries of one’s income in a given situation, thus being defined as  ∈ [−I, +I ]. What is more, let  represent individual ′ prior (stage 1) or actual observation (stage 2) of individual ′ (un)ethical behavior. (1) =  = 1 (   )  ( ) From this it follows that each player maximizes own utility at Stage 1 of the following form: (2)   =−1−∗−−∗− At stage 1, individual  has no information about individual ’s decision, as all participants carry out their decisions simultaneously. Consequently,  resorts to forming beliefs about the behavior of . As depicted above, individuals face a trade-off decision at stage 1, in which deviation from the individual inherent characteristic  has to be weighed against deviating from one’s own beliefs about the peer’s behavior . Because no information about the peers was given at stage 1, the resulting decision is a simple maximization problem of the form: (3)  = 2∗−−2∗−1∗− . ⇔∗==−+ yielding the comparative static: (4)  =− We can infer that a change of  in  depends on the social proximity (that is either a prior in = or an updated belief in =) in the following way: (5)  = > 0  >(   ) = 0  = (   ) < 0  < (   ) 118 In order to study the adaptation gap at stage 2, that is the difference in behavior between individual  and  after the second stage, one has to hold ’s behavior from stage 1 constant while giving individual  the ability to revise his initial decision.30 Proposition: Equation (2) provides a solution to the maximization problem and reduces the adaptation gap to: (6) −=−+−=∗1−+(−1) Observe that unlike in Akerlof’s (1997) general conformity model, this approach generates a unique equilibrium prediction due to restrictions put on the social proximity parameter, that is 0 < .< and the linear reaction function. Essentially, by assuming that behavioral adaptation is symmetric in either direction this indicates that the gap is driven by the social proximity to the peer that one is observing: the closer individual are in terms of proximity the smaller is the expected gap and the more similar is their behavior. An alternative interpretation is that the peer effect is stronger and leads to a more extensive behavioral adaptation the higher their social proximity is. Asymmetric Contagion The main purpose of the extended model is to allow for an asymmetric adaptation to unethical and ethical behavior. It is plausible to assume that not only observing but also starting to act on unethical behavior requires a different mindset and triggers other cognitive processes than it is the case for ethical behavior. Here, the ability to self-justify behavior depends on the (un)ethicality of the (observed) act to the extent that it varies the boundaries of the moral wiggle room (Dana, et al., 2007). Empirical support is provided by the slipperyslope effect (cf. Gino & Bazerman (2009), Welsh et al. (2015)) as well as results on the “broken windows effect” (cf. Beckenkamp et al. (2014), Lefebvre, Pestieau, Riedl & Villeval (2015)), which suggest that unethical behavior is likely to be more contagious than unethical behavior. We introduce , which represents a factor biasing behavioral adaptation towards the observation of unethical behavior. We assume that the degree of the adaptation bias depends on 30 Which is exactly what we will do in our experiment. See chapter 3.1.6. 119 the difference between one’s own and observed unethical behavior −. With  still retaining its properties from (2), the individual’s maximization is of the form: (7)   =−1−∗−−∗∗− with: (8) = max (−+ 1, 1) This definition illustrates that whenever one’s own initial behavior is more ethical than what is being observed from the peer behavioral contagion is stronger than observing more behavior that is more ethical. The strength of the difference in this contagion force depends on the distance between own and observed behavior. This implies the following maximization of the individual’s utility: (9)  = 2k∗−−2∗−1∗− . ⇔=−+ −+ 1 with the comparative static: (10)  =∗()  Under these assumptions, the resulting adaptation gap looks as follows: (11) −=|…|=󰈅−()∗()  󰈅 We can easily see that the slightly differently appearing adaptation gap retains the same properties as in the symmetric model. In conclusion, we expect behavioral adaptation to be driven by social proximity with a bias towards unethical behavior. Put differently, we expect the spillover of unethical behavior to be more pronounced, thus leading to a stronger degeneration of behavior relatively to the rise of good Samaritans. 120 3.1.6 The Experiment 3.1.6.1 Experimental Design and Procedure In order to study behavioral contagion, we are mainly interested in answering two questions with our experimental design: first, whether individuals revise their initial behavior in light of observing peer behavior. Second, whether and how a behavioral change depends on both the ethicality of the observed behavior and the social identification with the observed peers. Our design draws on a unique approach to study behavioral contagion in the lab that allows us to account for potential confounds potentially inherent into peer effects studies as argued by Manski (1993) and Angrist (2014) (see Thöni & Gächter (2015) for a similar approach). In order to be regarded as behavioral contagion, revised behavior has to be more similar to observed behavior than one’s initial behavior that one has decided upon prior to learning peer information. That is, revision of one’s initial behavior must follow the direction of observed initial peer behavior. Our basic design follows this straightforward procedure: action – observation of a peer – reaction.31 Consider a variant of a two-player dictator game in which the participant (dictator) is matched with a charity (recipient). The dictator’s action space entails taking away money from the charity, leaving the initial situation unchanged, or give money to a self-chosen charity (the basic design follows List (2007) and Bardsley (2008)). In following Eckel & Grossman (1996), we use a charity to increase the saliency of the involved decisions. The experiment is played one-shot with a possibility to revise one’s initial behavior. Between the initial decision and potential revision, individuals are given the opportunity to observe the initial behavior of another random participant. Alongside the actual behavior, treatment variations include the alteration of unveiled social proximity information of the observed participant.32 That is, in addition to learning actual behavior and the amount that was taken away or given by this participant, additional information on the participant’s social proximity to oneself is varied with the random treatment assignment. The treatment variation lies in the information given about the social proximity to the observed peers: no information on proximity (Baseline), as well as high proximity (T1) and low proximity (T2) information. 31 Note that in order to exclude any hedging concerns throughout the whole experiment, information about the specifics of the design were only provided where necessary in order to reach a deliberate decision. That is, at Stage 1 participants were neither aware of the possibility to observe peers later on nor to revise their initial decision, ensuring unbiased initial behavior. 32 Henceforth, we will use the terms social proximity and social identity interchangeably. 121 Proximity is calculated based on overlapping answers in the list of statements used in the beginning of each session and then presented to the participants in the form of belowor above-average proximity information to the observed peer.33 We capitalize on a shortened 25 items list of statements compiled from a major US American dating website to ensure the validity of the questions in successfully matching people (see Gibbs, Ellison & Heino (2006) and Hitsch, Hortaçsu & Ariely (2010) for a discussion).34 Since the business concept of dating websites is based on achieving high matching success rates, the use of such validated questions improves the success of incepting social identification between participants in the lab.35 The experimental procedure is represented by a single iteration of the following three stages: First Stage - The Action: Starting with an equal distribution of money, each individual decides whether to (i) donate own money to the charity’s account, (ii) not change the initial equal distribution, or (iii) take money from the charity and add to one’s own account. Second Stage – The Observation: Each active player observes one passive player of random who has engaged in either ethical or unethical behavior. In all three treatments, the exact information entails the monetary amount taken away from or given to the charity. Ex33 The implementation of the lowand high-proximity information followed a very straightforward calculation. For each participant of the active group, an individual proximity score to both participants of the passive group was calculated based on overlapping answers in the list of statements. From each active participant’s individual perspective, the passive participant with the higher (lower) score was labeled as the high (low) proximity peer. In fact, this calculation approach allows for the same passive person to be of high (low) proximity to one active person, while being of low (high) proximity to another active person, thus truly randomizing information. We abstained from providing explicit matching scores or percentages to retain maximum control. In addition, this allows us to alleviate the false-consensus effect, in which people systematically overestimate the degree of similarity to others. The provision of social cues of this kind allows the participants to update their beliefs reliably with respect to the actual degree of similarity. See Ellingsen & Johannesson (2008, p. 995) for a discussion. 34 We should stress the fact that we report lower-bound results. We induce social proximity in a very simple way by providing participants with either the highor low-proximity signal in the social identification treatments 1 and 2. Although this approach allows us to provide a comprehensive set of information to induce even more salient and distinct forms of social identity (i.e. by providing the exact matching score, the exact answers to the questions, letting participants put different weights on questions according to their individual importance and so on), we resort to this easy-to-use-easy-toreproduce approach. See Appendix B for the exact list of questions used in our experiment. 35 In addition, we elicit the strength of social identity to the observed peer using a variant of the self-evaluation scale of one’s social identity following Luhtanen & Crocker (1992) to verify the robustness of our social identity implementation approach. The non-parametric and regression estimations yield the same overall results both in direction and in magnitude. 128 the charity as compared to their initial behavior prior to having observed a peer’s behavior. The results are significant at the 1% level (p = 0.000, z = 4.365) and indicate a change in behavior almost three times as large when unethical behavior was observed as compared to ethical behavior. The results are highly suggestive of unethical behavior being more contagious. Thus, these results strongly support our hypothesis H1 and confirm that unethical behavior is indeed more contagious than ethical behavior independent of social identification.45 Figure 3 illustrates our findings. Figure 3: Amount Revised (%) and Observed (Un)Ethicality. The figure depicts the amount revised as percentage of one’s initial behavior. Any value above (below) zero indicates that more (less) ECU were given after the revision to the charity relative to one’s initial decision. The analysis is broken down into the (un)ethicality of observed behavior by the active participants. We are also interested in the type of behavioral contagion that is triggered by social identification. Following hypothesis H2, we assume that social identification amplifies the contagion of unethical behavior in an over-proportional way as compared to contagion of ethical 45 Unless noted otherwise in the results section, we obtain similar results in terms of evidence and significance when using absolute ECU numbers rather than percentages. 129 behavior. We thus examine the role of social identification in affecting the magnitude and direction of revision choices. Both findings are illustrated below in figure 4. Our results robustly indicate that higher social identity indeed triggers stronger behavioral contagion, particularly contagion of unethical behavior. As social identity increases, the magnitude of revised behavior increases as well, peaking at -29.8% for the high proximity condition. The differences in behavior are significant at the 1% level (p = 0.000, z = 4.759) when comparing behavior in the unknown proximity with the high proximity condition. Likewise, the results are significant at the 1% level (p = 0.000, z = -3.448) when comparing the high proximity with the low proximity condition. Here, again, the negative numbers of amount revised suggest that, in terms of magnitude of revised behavior, unethical behavior is strongly pronounced and thus more contagious than ethical behavior. Overall, we find ample support for our hypothesis H2 and thus conclude that the magnitude of a revision of one’s initial behavior is indeed strongly correlated with social identification. Consistent with our theoretical model, the results also yield strong support for the idea that the reduction in the adaptation gap is driven by social identification. As predicted by ., the stronger the social identification to one’s peer is a robust predictor (all at the 1% level) of one’s adaptation gap to the observed peer after observation. By the numbers, we obtain p = 0.000 and z = 6.104 (p = 0.000 and z = -3.441) when comparing unknown proximity versus high proximity condition (high proximity versus low proximity condition). We find that the adaptation gap is inversely correlated to the social identity. 130 Figure 4: Amount revised (%) and adaptation gap (%) by social proximity. Along similar lines, behavioral contagion as a function of both social identification and the (un)ethicality of observed behavior is illustrated in figure 5. The behavioral space is described by three alternatives and also speaks to the (non-)existence of observed peer effects:  No Contagion: after observing peer behavior, the participant either did not revise his/her initial behavior or revised it into the opposite direction of what he/she observed the peer has done.  Contagion: after observing peer’s behavior, the participant did revise his/her initial decision. The revision was directed into the direction of the observed behavior. This behavior indicates the existence of behavioral contagion caused by peer effects. The breakdown of behavioral changes by different levels of social proximity provides additional evidence for hypothesis H2. When comparing to the condition where no social proximity to one’s observed peer was induced, behavioral is more contagious for both low and high proximity situations in both the ethical and unethical domain. Again, since we report 131 lower-bound results, this is a strong indication that our method of inducing social identification works and is likely to produce even stronger results when social proximity would be introduced in a more sophisticated way. Figure 5: Behavioral change by treatment and observed (un)ethicality. This figure illustrates the fraction of participants exhibiting behavioral contagion broken down into both the observed (un)ethicality of behavior and the social identity to one’s observed peer. Unless the active participant revises his/her behavior into the direction of observed behavior the behavior is not classified as contagion. In terms of explaining the drivers of behavioral contagion, two literature streams have not yet been brought together in existing research: the role of social identification on behavioral contagion on the one hand, and the role of observed behavioral unethicality on the other. Our experimental design allows us to do exactly this for the first time and ascertain the main driver of behavioral contagion. Following our previous discussion, we assume social identification to be a stronger driver of behavioral contagion than the unethicality of observed behavior (H3). As a first, we investigate whether behavioral contagion is different under varied levels of social proximity when directly comparing contagion in the ethical versus the unethical domain using non-parametric comparisons. Our results provide strong support that behavioral contagion is asymmetric. In particular, a variation in social identification yields no significant variation in behavioral changes in the ethical domain. However, the results are strongly statistically significant when looking at behavioral changes as a function 132 of social identification in the unethical domain. Here, when comparing the high proximity condition to the no proximity (low proximity) condition, the Mann-Whitney-U statistics yield results that are highly significant at the 1% level with p = 0.000 and z = 6.025 (p = 0.000 and z = -4.005). These results highlight the importance of our contribution in this paper: peer effects are not uniformly in place, but rather strongly depend on both the (un)ethicality of observed behavior and the degree of social identification to the observed peer. Figure 6: Amount revised (%) by social proximity and observed (un)ethicality. What is more, we ran several regressions, including OLS, Logit, and Tobit, where applicable, in order to assess the robustness of our results. In sum, across different specifications our results strongly suggest that the observation of (un)ethical behavior does not trigger any particular behavior, neither ethical nor unethical. Thus, the mere observation of behavior alone is insufficient for the existence of peer effects in the (un)ethical sphere, but is rather dependent on the social identification to one’s peers. These findings are in support of our hypothesis H3. Findings are presented in Appendix A. 133 All in all, we find convincing support for all of our three hypotheses: unethical behavior is indeed more contagious than ethical behavior (H1), social identification drives the magnitude of behavioral contagion (H2), and social identification is a more reliable predictor of behavioral contagion than the observed (un)ethicality of observed behavior (H3). Here, it is important to stress a particular point with respect to the interpretation of the results. Arguably, the experiment’s framed environment created by the presence of a charity could potentially carry a norm of prosocial behavior in the lab in and of itself. That is, prior to observing one’s peers, some participants at the margin of behaving unethically might carry the normative belief that taking from a charity represents inappropriate behavior and thus initially refrain for it. If so, it would come as no surprise to observe stronger contagion of unethical as compared to ethical behavior because those who wanted to behave unethically in the first place but refrained from doing so might now find justification in their peer’s behavior. In this respect, two important remarks should be made: first, if anything, such an assumption would only explain level effects but not treatment differences because such beliefs are by experimental design uncorrelated with the treatments. Consequently, irrespective of the existence of potential norms, our design renders our main finding valid: behavioral contagion is highest where social identity is strongest.46 In addition, we elicited incentivized beliefs about what participants thought about his peer’s behavior prior to observing it. As the regressions results suggest, such beliefs yielded no explanatory power and thus play no role in neither explaining the magnitude nor the differences of behavioral contagion. From a policy perspective, our results stress that social proximity renders it difficult to change individual behavior, but it rather amplifies one’s initial (un)ethicality. We will return to this point in our policy recommendations in chapter 3.1.7. In light of the very conservative inception of social identification of providing very limited information on social identification, we deem these results to represent a lower bound thus strengthening the role of social identification within the context of behavioral contagion. Our lower-bound approach comes from the fact that participants were neither told the exact matching percentage nor the actual interests and preferences they had in common. If having been randomly assigned to one of the social identification treatments, participants only knew whether they were observing a peer with above or below average congruence. A more sophisticated way to induce 46 I would like to thank Gary Bolton, René Fahr, and Elena Katok for point this out and for related discussions. 134 social identification and match participants accordingly is likely to produce results that are more pronounced. The well-engineered mechanisms implemented by dating websites to match people and achieve high success rates are a shining example for what is possible: excluding matching partners based on personality traits that represent a no-go (e.g. smoking), putting emphasis on particular interests (e.g. sports, religion), or individual characteristics (e.g. looks, education). We deliberatively refrained from applying sophisticated measures of this kind and rather resorted to an easy-to-use-easy-to-reproduce methodological approach that could be used in future experiments in which inducing salient social identity is key. Along these lines, the presence of a potential experimenter demand effects (EDE) is worth addressing since their presence has potentially been problematic to prior peer effect studies (for a discussion see Thöni & Gächter (2015)). Because we are mainly interested in treatment differences rather than in overall levels, the experimenter demand effect is deemed less problematic as long as its existence and magnitude is orthogonal to the treatment variation (Zizzo, 2010). Nonetheless, we considered existing experimental studies to rule out experimenter demand effects to the extent possible. Not exclusively to but prominently existing in peer effect studies, forced learning (i.e., forced observation of one’s peer’s behavior) might potentially induce EDE or even lead to resentment on the side of the participants. Forced observation might trigger thoughts related to being expected to use the information to reconsider and potentially revise initial behavior. In previous general and peer effect studies in particular, this issue has normally been overlooked, mainly to avoid self-selection problems. However, when the option not to learn is withheld from participants, the obtained results are potentially confounded. We deem this challenging to the study of peer effects and should thus be discussed. Existing research indicates that individuals sometimes choose to deliberately remain ignorant about the state of nature (see Carrillo & Mariotti (2000), Dana, Weber & Kuang (2007), Conrads & Irlenbusch (2013), Bartling, Engl & Weber (2014), and Grossman (2014)). If present, such strategic ignorance might be an important component of our experiment. A potential reason not to acquire costless information is related to the avoidance to indulge in negative self-image updating or guilt aversion (cf. Wells (1978), Baumeister (1998), Charness & Dufwenberg (2006)). One can plausibly assume that such aversion is even stronger when 135 studying peer effects within an (un)ethical dimension. Thus, forcing participants to learn potentially unpleasant information might lead to biased behavior and might even increase EDE. In order to address this challenge, our design follows Conrads & Irlenbusch (2013) and Bartling, Engl & Weber (2014) and draws on a mechanism in which learning peer information is voluntary (see also Eckel & Petrie (2011)).47 Additionally, in order to rule out any reputational concerns, social learning, or reciprocity, the experiment includes an anonymously played one-shot interaction with another participant. Such an experimental design allows us to study behavioral contagion in the lab in an unbiased way. To our knowledge, this represents a novel design in studying peer effects in the lab in general and the behavioral contagion of (un)ethical behavior in particular while controlling for potential confounds that peer effect studies suffer from regularly (see Manski (2000), Falk & Fischbacher (2002), Angrist (2014)). In the light of our experimental set-up, any treatment-specific information is provided only after one’s deliberate decision to learn peer behavior. Thus, any still potentially existing form of EDE would be fully uncorrelated with the treatments and thus exhibit only a general level-effect, if any. 3.1.7 Lessons Learned: Policy Implications As argued before, understanding social interactions in general and in particular the potentially resulting peer effects is fundamental from a policy perspective. It does not only help to understand societal and economic outcomes beyond what standard economic forces can explain (i.e., the massive surge in female labor participation rates in World War II (Mulligan, 1998) or the escalation of crime rates (Levitt, 1999)). It also allows us to implement bettertargeted policy measures to tackle a battery of challenges such as reducing crime rates, improving health conditions, or increasing labor market participation. “To the extent that theory and measurement of social interactions enables us to understand these massive changes, the study of social interactions potentially has major policy relevance” (Glaeser & Scheinkman, 2004, p. 84). In this chapter, we will discuss some of the policy implications 47 However, deliberately allowing participants to remain ignorant about peer behavior bears the risk of self-selection effects. It is worth noting that in our experiment 10% of all the participants decided not to acquire peer information. Importantly, however, this choice is unconditional on the participant’s initial behavior, thus strongly suggesting the absence of any self-selection mechanism leading to potential biases in our analysis. 136 that one can infer from our results. We will follow the main theme of this paper and approach this topic from two sides: the ethical and unethical context (for a broader discussion see Irlenbusch & Villeval (2015)). Starting with the ethical perspective, voluntary redistribution of income e.g. in the form of charitable giving is an integral part of humaneness, with up to 90% of Americans donating to charities. Understanding the drivers of charitable giving, such as altruism and social pressure, has been at the heart of last decade’s research, both in the field and in the lab (cf. Levitt & List (2009), DellaVigna, List & Malmendier (2012)). Still, we seem to have an imperfect understanding of what really motivates giving (Andreoni, 2006). Beyond what we have since learned about the existence of peer effects with respect to contribution decisions, our experiment yields several new insights helpful to understanding the extent to which such behavior shows up, especially as compared to unethical behavior. We will return to this comparison shortly. In turn, unethical behavior in its various forms impairs the daily life. Exemplarily, yearly global tax evasion ranges at an abstruse $3.1 trillion or 5.1% of world GDP (The New York Times, 2011), over $1 trillion is estimated in bribes paid yearly around the globe (The World Bank, 2013), and some 210 million people use illicit drugs each year (United Nations Office on Drugs and Crime, 2011). Here, one might credibly argue that from a purely rational selfmaximizing perspective we should observe way less illicit behavior than we actually do. From a game theoretic perspective, in some illicit deals that involve trust-related actions such as bribery do not represent a subgame perfect Nash equilibrium, which is true for both one-shot and finitely repeated contexts. Thus, in many situations the occurrence of illicit behavior is already surprising (for a discussion, see Dimant & Schulte (forthcoming)). A solution to this conflict is the recognition of, among others, peer effects. As has been thoroughly argued throughout the paper, the incorporation of behavioral contagion allows us so explain why observed behavior goes seemingly beyond clear-cut self-maximization, but is rather embedded in and the result of a social context. Consequently, both aspects raise the following question: how do these findings translate into a real world setting and what to do about it? While being careful at drawing concrete inferences from a laboratory setting and relating them directly to the outside world, our results conclusively indicate the spillover of unethical behavior to be much more likely than 137 the spillover of ethical behavior. That is, getting people to start donating solely based on peer effects (e.g. through observing others giving to charity) has a long way to go compared to having them do something unethical. What we see, however, is the individual’s responsiveness to social identification, in particular on the side of females. A potential solution to have people donate more to a good cause is to provide them information beyond simple statistics on what other people do, i.e. amount of money that has been collected so far (like Wikipedia). Instead, some research already indicates that the inception of social norms especially for settings that most closely match the individual’s immediate situational circumstances have the strongest effect on compliance (Goldstein, et al., 2008). Our results suggest to go one step further and provide information that allows us to draw inferences on the social proximity to the peers, thus increasing the saliency of social identification. Exemplarily, a message of the form “People in your neighborhood / with similar demographic characteristics have donated an average of $...” would lead to pick-up rates of this behavior that are higher than when resorting to a simple statistic. In all likelihood, a similar approach could be applied to make people refrain from behaving unethically. Research on slippery slope indicates that once behavior is spoiled, even honest people converge quickly to a steady state with a plethora of unethical behaviors (cf. Gino & Bazerman (2009), Welsh et al. (2015)). Even more worrisome, recidivism rates for convicts are normally very high, leading to what is called a recidivism nightmare. In the US recidivism rates are up to 80% of re-arrests within the first 3 to 5 years after their release from prison (National Institute of Justice, 2014). This is particularly detrimental from a welfare perspective, as the US spends approximately $75 billion on incarceration and $260 billion on prevention, detection, and prosecution of crime on a yearly basis (Khadjavi, 2015). In terms of effective countermeasures to unethical behavior, our experimental results indicate that exposing individuals to unfaithful of socially close people is likely to trigger repulsion and less unethical behavior, thus contributing to a positive transformation.48 48 For an approach along similar lines, see Pennsylvania State University's Justice Center for Research on desistance from crime (Pennsylvania State University's Justice Center for Research, 2014). 144 A3: Amount Revised (%) as a function of social identification and observed unethicality. Model 1 and 5 (Model 2 and 6) tests for the effect of the treatments (of observed unethicality) while controlling for initial behavior and the observed adaptation gap. Model 3 (Model 7) tests for both treatment effects and observed unethicality simultaneously. Model 4 (Model 8) adds controls for gender and some interaction terms as well as the number of interest. In order to rule out any endogenous concerns and stress the effectiveness of exogenous variation of social proximity we also add dummies for the dating website questions. None of the dummies turn out significant in neither model specification, thus emphasizing the robustness of our results. 145 A4: Examination of the adaptation gap (%) using Logit estimations. Participants who did not revise their initial decision are treated as adaptation gap = 100%. 146 A5: Examination of the drivers of behavioral contagion using a multinomial logistic regression. Behavior of participants who revised their initial decision but into the opposite direction of what they have observed from their peer is treated as anti-contagion (repulsion). 147 Following the regression results, our results also speak to the idea that behavioral contagion seems to facilitate the magnitude of (un)ethical behavior rather than changing individual behavior to the better or worse, respectively. That is, those who behaved (un)ethically in the first place become even more (un)ethical after being exposed to their peers. This is true for both revision of one’s initial amount and the reduction of the adaptation gap to one’s peer. In more detail and in line with our previous results, this relation is more pronounced the more salient social identity to one’s peers is and in particular in the unethical domain. This is additional support for our hypothesis H1 that unethical behavior is more contagious than ethical behavior. We find similar results for those who decided to keep the fair equal split in the beginning, while the behavioral change of those who donated initially seem not to be susceptible to changes in social identity. The figure below illustrates results. A6: The upper (lower) graph depicts the amount revised (%) (adaptation gap (%)) conditional on initial behavior broken down by social proximity. This figure depicts the percentage of initial behavior revised as a function of the participant’s initial behavior and broken down into different social identity categories. Essentially, the figure illustrates the magnitude and direction of behavioral contagion after observing the passive peer conditional on one’s initial decision prior to having observed a peer. 148 B: Social Identity Statements 1. I am a reliable person. 2. I am interested in politics and/or economics. 3. Money is important to me. 4. I am an honest and sincere person. 5. I am a cinephile. 6. I am interested in sports. 7. I am a religious person and faith is important to me. 8. I am fond of animals. 9. I am interested in art and/or cultures. 10. I am an active and adventurous person. 11. I am interested in cars and/or technology. 12. I am fond of children and family-oriented. 13. I am interested in foreign languages and/or countries. 14. I am a warmhearted and helpful person. 15. I am a tolerant person. 16. I like to gossip. 17. I am a faithful person. 18. I play an instrument. 19. I like to go out and dance. 20. I am a goal-oriented person. 21. I spend a lot of time in front of the TV. 22. I am a sociable person and like to be among people. 23. I like to play videogames. 24. I am a humorous and entertaining person. 25. I am a strong-willed person. Average amount of chosen statements (across all treatments): 15.8 (63%) 149 C: Experimental Instructions General Information on the Experiment  First of all, we would like to thank you very much for participating in this experiment. Please read the instructions carefully. The experiment will last for about 45-60 minutes.  During the entire experiment, no communication is allowed. If there is something you do not understand or if you have any questions, now or at some point during the experiment, please raise your hand and remain seated. One of our colleagues will come to you and answer your question.  During the experiment, you have the possibility to earn money. The amount you will receive at the end of the session depends on how many “Taler” you earn during the experiment.  At the end of the experiment, the amount of “Taler“ that you have earned will be converted into real money at an exchange rate of 20 Taler = 1 Euro.  All decisions you make during this experiment will remain anonymous. None of the participants gets to know the identity of other participants in the experiment and decisions cannot be linked to a specific participant. Moreover, you will be paid anonymously at the end of the experiment. 150 Order of Events:  The experiment consists of a list of statements that you will receive at the beginning and further decisions. Explanations and information related to these decisions will be given as the experiment progresses. You will make these decisions once.  Both you as well as a charitable organization of your choice (i.e. an officially registered charity organization) will be provisionally assigned a monetary amount of 300 Taler each.  During the experiment you will have to decided on whether you want to…  … take a part or all of the money from the charitable organization.  … leave the division of the sum of money as it is.  … give a part or all of your money to the charitable organization.  In case you decide to take money from the charitable organization, the respective amount of money will be transferred to your individual cash account and exactly the same amount will be deducted from the cash account of the charitable organization.  Should you decide to give money to the charitable organization of your choice, the respective amount of money will be deducted from your individual cash account and given to the charity. The experimenter will double all ECUs remaining in the charity’s account at the end of the experiment.  Your decision remains anonymous and neither the other participants of the experiment nor the experimenters have the possibility to assign your choices to your identity.  At the end of the experiment, one participant will be chosen at random and his or her choice will be implemented and count towards the charity (i.e. that choice will be rel- 151 evant for the payment). In particular, we will double the respective amount and donate it to the charity after the experiment ends. The receipt of this donation will be published on the homepage of the BaER-Lab (www.baer-lab.org) in a timely manner. All other participants will receive 150 Taler (including the show-up fee) at the end of the experiment.  The total payoff of the participants:  In case you are the randomly chosen participant 300 Taler +/- the amount of money that has been given to/taken from the cash account of the charitable organization  In case you are not the randomly chosen participant 150 Taler  The total payoff of the charitable organization:  (Amount of money in the cash account of the charitable organization of the randomly chosen participant) × 2  At the end of the experiment, the relevant information on the payment will be made visible to each participant on his or her screen.  After the actual experiment concludes, we will ask you to fill out a questionnaire. Please fill out the questionnaire carefully and truthfully. 152 D: Screenshots of Decision Screens 1. List of statements: generates the proximity measure in all treatments 153 2. First decision: behaving (un)ethically (Exemplarily for the taking away decision)