Heterogeneity in Conformity and Cooperation : Two Experiments and Statistical Software

Volltext

(1)

Heterogeneity in Conformity and

Cooperation

Two Experiments and Statistical

Software

Dissertation

zur Erlangung des akademischen Grades eines

Doktors der Wirtschaftswissenschaften (Dr. rer. pol.)

vorgelegt von Fabian Dvorak

an der

Sektion Politik – Recht – Wirtschaft Fachbereich Wirtschaftswissenschaften

Konstanz, 2018

(2)

Tag der m¨undlichen Pr¨ufung: 28. November 2018 1. Referent: Prof. Dr. Urs Fischbacher

(3)
(4)
(5)

Danksagung

Ich danke Urs Fischbacher f¨ur die wertvolle Unterst¨utzung, die ich ¨uber den gesamten Zeitraum der Entstehung dieser Arbeit erhalten habe. Ohne ihn w¨urde es diese Dissertation nicht geben. Ich bin dankbar f¨ur seine Bereitschaft, diese Dissertation als Betreuer zu begleiten, und f¨ur die vielen Dinge, die ich in den letzten Jahren von ihm gelernt habe.

Ausdr¨ucklich m¨ochte ich mich auch bei Susumu Shikano f¨ur die Betreuung dieser Arbeit be-danken. Von seinen Anmerkungen und denen seiner Forschungsgruppe hat vor allem das dritte Forschungspapier sehr profitiert.

Außerdem danke ich Armin Falk f¨ur die Begutachtung dieser Arbeit. Vor allem aber bin ich dankbar f¨ur den interessanten und lehrreichen Forschungsaufenthalt in Bonn, den er mir erm¨oglicht hat.

Bei meinen Koautoren, Sebastian Fehrler und Katrin Schmelz, und bei Irenaeus Wolff m¨ochte ich mich f¨ur die angenehme und lehrreiche Zusammenarbeit bedanken. Alle drei haben auf ihre Weise maßgeblich zur Entstehung dieser Arbeit beigetragen.

Bedanken m¨ochte ich mich auch bei allen anderen Mitgliedern des Lehrstuhls f¨ur angewandte Wirtschaftsforschung der Universit¨at Konstanz f¨ur hilfreiche Gespr¨ache und Anregungen sowie bei Yongping Bao, Felix Beck, Steffen Guth und Malin Pimper und f¨ur ihre Mitarbeit.

Ein besonderer Dank gilt der Graduiertenschule Entscheidungswissenschaften, dem Thurgauer Wirtschaftsinstitut und dem Zukunftskolleg der Universit¨at Konstanz f¨ur die finanzielle Un-terst¨utzung dieser Dissertation. Außerdem m¨ochte ich mich bei Barbara Monstein, Madeleine Hafner und Jutta Obenland f¨ur die administrative Unterst¨utzung bedanken.

In der Entstehungszeit dieser Arbeit sind mir viele Menschen begegnet, bei denen ich mich f¨ur die sch¨one Zeit, die ich mit ihnen verbringen durfte, bedanken m¨ochte. In besonderem Maße gilt dies f¨ur: Dominik Bauer, Christian Neumeier, Carl Maier, Nathalie Popovic, Jan Hausfeld, Thomas Hattenbach, Konstantin von Hesler, David Dohmen, Franziska Deutschmann, Anja Folli, Timo Dimitriadis, Baiba Renerte, Marco Menner, Michael D¨orsam und Michal Marenˇc´ak. Zuletzt m¨ochte ich mich bei den Menschen bedanken, die mich immer unterst¨utzt haben und mir besonders am Herzen liegen. Ich danke meinen Eltern, meinem Bruder und dir, Tasja.

(6)
(7)

Contents

Summary 1

Zusammenfassung 5

1 Incentives for Conformity and Disconformity 11

1.1 Introduction . . . 13

1.2 Related Literature . . . 15

1.3 Experimental Design . . . 17

1.3.1 Domains and treatments . . . 18

1.3.2 Hypotheses . . . 19 1.3.3 Implementation . . . 21 1.4 Experimental Results . . . 22 1.4.1 Evaluation behavior . . . 22 1.4.2 Treatment effects . . . 24 1.4.3 Heterogeneity . . . 28 1.4.4 Questionnaire validation . . . 29 1.5 Discussion . . . 29

A Appendix of Chapter One . . . 32

A.1 Analysis of heterogeneity . . . 32

A.2 Study materials . . . 33

2 Negotiating Cooperation: Communication in Noisy, Indefinitely Repeated Interactions 39 2.1 Introduction . . . 41

2.2 The Repeated Prisoner’s Dilemma with Noise . . . 44

2.2.1 Predictors of cooperation . . . 45

2.2.2 Communication . . . 47

2.3 Experimental Design . . . 48

2.3.1 Experimental parameters . . . 50

2.3.2 Research questions and methods . . . 51

2.4 Experimental Results . . . 53

2.4.1 Cooperation . . . 54

2.4.2 Strategy choice and communication content . . . 56

2.5 Discussion and Conclusion . . . 64

B Appendix of Chapter Two . . . 66

B.1 Belief-free equilibria . . . 66

B.2 Renegotiation-proof & truthful communication equilibria . . . 73

B.3 Strategy inference . . . 76

B.4 Communication content categories . . . 77

(8)

3 stratEst: An R Package for Strategy Estimation

in Experimental Economics 85

3.1 Introduction . . . 87

3.2 An Introductory Example . . . 88

3.3 Deterministic Finite-State Automata . . . 91

3.3.1 General definition . . . 92 3.3.2 Matrix representation . . . 92 3.4 Strategy Estimation . . . 94 3.4.1 Model definition . . . 95 3.4.2 Algorithm . . . 96 3.4.3 Parameter estimation . . . 97 3.5 Model Selection . . . 98

3.5.1 Restrictions on strategy parameters . . . 99

3.5.2 Selection of the number of strategy parameters . . . 99

3.5.3 Selection of the number of strategies . . . 100

3.5.4 Information criteria . . . 100

3.6 Latent Class Regression . . . 101

3.6.1 The multinomial latent class model . . . 101

3.6.2 Parameter estimation . . . 102

3.7 Standard Errors . . . 102

3.7.1 Analytic standard errors . . . 102

3.7.2 Analytic standard errors with parameter restrictions . . . 104

3.7.3 Bootstrapped standard errors . . . 104

3.8 Using stratEst . . . 105

3.8.1 Input objects . . . 105

3.8.2 Output objects . . . 107

3.8.3 Example 1: Simulated data . . . 108

3.8.4 Example 2: DalBo & Frechette, 2011 . . . 111

3.9 Limitations & Future Directions . . . 114

C Appendix of Chapter Three . . . 115

Abgrenzung 121

(9)
(10)
(11)

Summary

This dissertation consists of three independent research papers. The three papers make distinct contributions to different strains of the economic literature. The unifying element of all three papers is the experimental method. The first two research papers report the results of two laboratory experiments that have been conducted at the University of Konstanz in the years 2016 and 2017. The third research paper introduces a statistical software package for the analysis of experimental data.

A topic that appears in all three papers is the analysis of heterogeneity of behavior. Hetero-geneity is an empirical regularity in most economic environments. In contrast, most of the contemporary theoretical work in economics assumes homogeneous agents. One reason for this limitation is that the diversity in economic decision-making is not well understood. Laboratory experiments can be used to study differences in decision-making under controlled conditions. The controlled laboratory environment promises insights into the nature of heterogeneity that are difficult to obtain based on observational data.

The first research paper of this dissertation documents individual differences in the propensity to conform to the behavior of a majority. In an environment that offers the opportunity to receive a monetary reward, we find that some experimental participants strategically deviate from the majority behavior while most others conform. Interestingly, the monetary incentives of the environment are such that disconformity is profitable. In a different environment, where disconformity is not profitable, we do not find heterogeneity in behavior, as all participants conform.

The second research paper of this dissertation documents heterogeneity in the propensity to co-operate with an anonymous partner in a repeated interaction. We find substantial heterogeneity in the strategies of experimental participants. While some participants seek to establish long-term cooperative relationships, others refrain from cooperation, and some even try to exploit attempts to establish cooperation. We also analyze the evolution of heterogeneity in strate-gies over time. In some experimental conditions, strategy choice converges to a state where nearly all participants follow few cooperative strategies. In other experimental conditions, the heterogeneity in strategy choice remains constant over time.

(12)

Chapter 1: Conformity

Conformity, the act of adopting the behavior to the majority, can be the reason why people shy away from innovative practices, are susceptible to group think, and communicate in echo chambers or filter bubbles. Behavior which is the opposite of conformity is far less studied and evidence is mostly anecdotal. While disconformity can help to trigger innovation or break undesirable routines, it reduces coordination and predictability of behavior which can be costly for society.

So far, little is known about how certain economic environments create incentives for confor-mity or disconforconfor-mity. In the first research paper of this dissertation we provide experimental evidence that shows that environments with punishment and reward create incentives for con-formity and disconcon-formity respectively. Our experimental design represent a methodological innovation since it allows us to simultaneously detect conformity and disconformity on the aggregate but also on the individual level. We use the design to study environments where one member of a group of individuals is selected by a third party and either receives a reward or a punishment.

The results show that reward decreases the level of conformity in participants’ preferences and judgments and induces disconformity in a minority of participants. The results suggest that the benefit of acting differently in environments with reward is not to get noticed - as often suggested. Rather the benefit is to get rewarded for unusual behavior by people with unusual preferences. In contrast, we find that environments with punishment increase conformity to a dramatic level. The results suggest that some economic environments fundamentally affect the level of conformity and disconformity in social groups.

Chapter 2: Cooperation

Many social and economic relationships are characterized by repeated interactions in which the behavior of partners is observable only with noise. Two examples are teamwork arrangements in which workers repeatedly produce goods for each other, and cartels with repeated price-setting by its members. How much effort a worker exerts in the production of the good cannot be observed directly but only inferred from the good itself. Likewise, whether or not other cartel members stick to a collusive agreement cannot be observed directly but only inferred from noisy signals like own sales or the market price.

(13)

3

We find that with repeated communication, where subjects can chat with their partner before every round of the supergame, cooperation rates are high and stable over rounds, also when monitoring is imperfect. With pre-play communication, where subjects can chat with their partner before the first round of the repeated game, cooperation rates start high in the first round of the repeated game but decline rapidly if monitoring is imperfect. In contrast, co-operation rates do not decline under pre-play communication if monitoring is perfect and the past behavior of the partner can be observed without noise. The results highlight the impor-tance of repeated communication for the stability of cooperative relationships under imperfect monitoring.

Chapter 3: Heterogeneity

Laboratory experiments offer the opportunity to study heterogeneity in behavior under con-trolled laboratory conditions. A prerequisite for the analysis of heterogeneity in economic ex-periments is statistical software that can accommodate the peculiarities of experimental games. In economic games, behavior is characterized in terms of strategies. Strategies are complete action plans that prescribe a behavioral response for every potential situation in a game. The structure of strategies can become very complex especially if a game is repeated, which is of-ten the case in economic experiments. In order to characterize the behavior of experimental participants as a mixture of individual strategies, the structural assumptions strategies impose on behavior have to be accounted for in the estimation process.

(14)
(15)

Zusammenfassung

Diese Dissertation besteht aus drei voneinander unabh¨angigen Forschungspapieren. Jedes der drei Forschungspapiere leistet einen Beitrag zu einem anderen Strang der ¨okonomischen Fachlit-eratur. Das verbindende Element der drei Arbeiten ist die experimentelle Methode. Die ersten beiden Arbeiten berichten die Resultate zweier Laborexperimente, die mit Studierenden an der Universit¨at Konstanz durchgef¨uhrt worden sind. Das dritte Forschungspapier beschreibt ein statistisches Softwarepaket, welches auf die Analyse von experimentellen Daten zugeschnitten ist.

Ein Thema, welches in allen drei Forschungsarbeiten auftaucht, ist die Analyse von Hetero-genit¨at im menschlichen Entscheidungsverhalten. Unterschiede im Entscheidungsverhalten zwischen Personen sind eine empirische Konstante, welche in den allermeisten ¨okonomischen Umgebungen beobachtet werden kann. Im Unterschied zu den empirischen Befunden nehmen die meisten theoretischen Modelle in der ¨Okonomie jedoch ein homogenes Entscheidungsver-halten der Akteure an. Ein Grund hierf¨ur liegt darin, dass die Ursachen unterschiedlichen Verhaltens noch nicht hinreichend erforscht sind. Laborexperimente bieten die M¨oglichkeit, Heterogenit¨at im Entscheidungsverhalten der Teilnehmer unter kontrollierten Bedingungen zu beobachten. Die Bedingungen im Labor erm¨oglichen Erkenntnisse ¨uber in die Natur und die Determinanten von heterogenem Verhalten, welche auf der Grundlage von Beobachtungsstu-dien nur ¨außerst schwierig zu gewinnen sind. Die Analyse von Heterogenit¨at in experimentellen Spielen ben¨otigt spezialisierte Software, welche das Entscheidungsverhalten der Teilnehmer als eine Mischverteilung unterschiedlicher Strategien darstellt.

Das erste Forschungspapier dieser Dissertation dokumentiert Heterogenit¨at in der menschlichen Neigung, das eigene Verhalten der Mehrheit anzugleichen. Die experimentellen Ergebnisse zeigen, dass in einer Umgebung, in der eine monet¨are Belohnung in Aussicht gestellt wird, einige wenige Teilnehmer gezielt vom Mehrheitsverhalten abweichen. Die allermeisten Teilnehmer gleichen ihr Verhalten hingegen der Mehrheit an. Interessanterweise zeigen die Ergebnisse deutlich, dass das abweichende Verhalten profitabler ist. In einer anderen Umgebung wiederum, in der sich abweichendes Verhalten nicht auszahlt, l¨asst sich keine Heterogenit¨at im Verhalten nachweisen.

(16)

Strategien benutzen. In anderen Versuchsbedingungen wiederum, bleibt die Heterogenit¨at in der Strategienwahl ¨uber die Zeit erhalten.

Das dritte Forschungspapier dieser Dissertation beschreibt eine statistische Software zur Anal-yse von Heterogenit¨at in experimentellen Spielen. Es existieren bislang keine Softwarel¨osungen f¨ur die Analyse von Heterogenit¨at im Verhalten, die auf die besonderen Anforderungen ex-perimenteller Spiele zugeschnitten sind. Die Analyse von Heterogenit¨at in experimentellen Spielen wird von theoretischen ¨Uberlegungen angeleitet, welche das Verhalten der Teilnehmer in Form von Strategien vorhersagen. Diese Strategien zeichnen sich durch strukturelle Annah-men aus, welche Abh¨angigkeiten im Verhalten der Teilnehmer in unterschiedlichen Situationen beschreiben. Um das Verhalten der Teilnehmer als eine Mischverteilung von individuellen Strategien darzustellen, ist es notwendig, die strukturellen Annahmen w¨ahrend des Sch¨ atzver-fahrens zu ber¨ucksichtigen. Das dritte Papier stellt eine Software vor, anhand derer Hetero-genit¨at im Verhalten in vielen experimentellen Spielen mit diskreten Entscheidungen analysiert werden kann.

Kapitel 2: Konformit¨

at

Konformit¨at kann als das Angleichen des eignen Verhaltens an die Mehrheit verstanden werden und kann der Grund sein, warum Menschen von innovativen Ideen ablassen oder in sogenannten Echokammern oder Filterblasen leben. Verhalten, welches als das Gegenteil von Konformit¨at bezeichnet werden k¨onnte, ist weit weniger untersucht. W¨ahrend Diskonformit¨at Neuerungen bef¨ordern oder ung¨unstige gesellschaftliche Zust¨ande aufl¨osen kann, reduziert diskonformes Verhalten gleichzeitig die Koordination innerhalb sozialer Gruppen und die Berechenbarkeit von Verhalten. Beides kann unerw¨unschte Konsequenzen f¨ur eine Gesellschaft haben.

Bislang blieb weitgehend unerforscht, ob unterschiedliche ¨okonomische Umgebungen Anreize f¨ur konformes oder diskonformes Verhalten erzeugen. Das erste Forschungspapier dieser Disser-tation berichtet die Ergebnisse eines Laborexperiments, welche nahelegen, dass Umgebungen mit Bestrafung Anreize f¨ur konformes Verhalten erzeugen, wohingegen Umgebungen mit Be-lohnung Anreize f¨ur diskonformes Verhalten erzeugen. Unser experimentelles Design stellt insofern eine methodische Innovation dar, als dass sich Konformit¨at und Diskonformit¨at gle-ichzeitig und sowohl im durchschnittlichen als auch im individuellen Verhalten messen lassen. Wir nutzen dieses experimentelle Design, um Umgebungen zu untersuchen, in denen ein Grup-penmitglied einer Gruppe ausgew¨ahlt wird, um eine monet¨are Belohnung oder, in einer anderen Versuchsbedingung, eine monet¨are Bestrafung zu erhalten.

(17)

CONTENTS 7

sondern die M¨oglichkeit, von Gleichgesinnten belohnt zu werden. Unsere Resultate legen nahe, dass manche ¨okonomischen Umgebungen die H¨aufigkeit von Konformit¨at und Diskonformit¨at stark beeinflussen – mit allen Chancen und Problemen, die diese Verhaltensweisen f¨ur die Gesellschaft mit sich bringen.

Kapitel 3: Kooperation

Viele ¨okonomische Beziehungen bestehen aus einer Kette von Interaktionen unbestimmter L¨ange. Oft kann das Verhalten des Gegen¨ubers dabei nur indirekt beobachtet werden kann. Zwei Beispiele sind wirtschaftliche Beziehungen in denen wiederholt G¨uter untereinander ausge-tauscht werden oder Kartelle, welche wiederholt Preisabsprachen treffen. Wieviel Anstrengung eine Partei unternimmt, um das Ergebnis f¨ur die andere zufriedenstellend zu gestalten, bleibt in beiden F¨allen oft im Unklaren. Jedoch existieren in beiden Beispielen Hinweise, welche mit Un-sicherheit behaftete Information ¨uber die Anstrengungen des Gegen¨ubers darstellen. Beispiel-sweise korreliert die Qualit¨at der ausgetauschten G¨uter mit der Anstrengung, die zu ihrer Erzeugung aufgewendet wurde, auch wenn es sich hierbei nicht um einen perfekten Zusammen-hang handelt. Ebenso enthalten die von den Unternehmen erwirtschafteten Ums¨atze Hinweise dar¨uber, ob sich das jeweils andere Unternehmen an die getroffenen Absprachen h¨alt.

Wie Kooperation in solchen Umgebungen mit imperfekter Information aufrechterhalten wer-den kann, ist das zentrale Thema der Forschung zu unbestimmt wiederholten Spielen der let-zten dreißig Jahre. Das zweite Forschungspapier dieser Dissertation berichtet die Ergebnisse eines Laborexperiments, welches die Effekte verschiedener Varianten von Kommunikation in solchen Umgebungen untersucht. Das Experiment nutzt das unbestimmt wiederholte Gefan-genendilemma und variiert, ob die Teilnehmer sich vor jeder Runde, nur vor der ersten Runde oder gar nicht auszutauschen k¨onnen. Jede der drei Kommunikationsstrukturen wird jeweils in einer Umgebung mit perfekter Information und zwei Umgebungen mit imperfekter Information implementiert. In den imperfekten Umgebungen erhalten die Teilnehmer jeweils nur ein mit Fehlern behaftetes Signal ¨uber das Verhalten des Partners in der vergangenen Runde.

Die Ergebnisse zeigen, dass wiederholte Kommunikation hohe Kooperationsraten erzeugt, die ¨

uber die Zeit stabil auf hohem Niveau bleiben. Dies gilt auch f¨ur die Umgebungen mit imper-fekter Information. Wenn die Teilnehmer hingegen lediglich vor der ersten Runde mit ihren Partnern kommunizieren k¨onnen, kommt es ebenfalls zu anf¨anglich hohen Kooperationsraten. Allerdings sinkt in Umgebungen mit imperfekter Information die Bereitschaft der Teilnehmer zu kooperieren deutlich ¨uber die Zeit hinweg. Im Gegensatz hierzu bleibt die Kooperationsrate in Umgebungen mit perfekter Information weitgehend stabil. Die Ergebnisse unterstreichen die Wichtigkeit von wiederholten Kommunikationsgelegenheiten f¨ur stabile kooperative Beziehun-gen in UmgebunBeziehun-gen mit imperfekter Information.

Kapitel 4: Heterogenit¨

at

(18)

Auswertung von Heterogenit¨at in experimentellen Daten ist statistische Software, welche an die Besonderheiten des vorliegenden Experiments angepasst werden kann. In experimentellen Spielen, werden spieltheoretische Vorhersagen meist auf der Grundlage von Strategien for-muliert. Strategien sind vollst¨andige Handlungsanweisungen f¨ur jede erdenkliche Situation in einem Spiel. Aufgrund dieser Tatsache ist es nicht verwunderlich, dass Strategien ¨außerst komplexe Formen annehmen k¨onnen. Dies ist vor allem gegeben, wenn das zugrunde liegende Spiel wiederholt gespielt wird, da sich in diesem Fall die Anzahl der m¨oglichen Spielsitua-tionen um ein Vielfaches erh¨oht. Welche Strategien von welchen Teilnehmern unter welchen Versuchsbedingungen verfolgt werden, stellt die Kernfrage bei der Analyse von Heterogenit¨at in experimentellen Spielen dar.

(19)
(20)
(21)

Chapter

1

Incentives for Conformity

and Disconformity

Fabian Dvorak

Graduate School of Decision Sciences, University of Konstanz

Urs Fischbacher

Department of Economics, University of Konstanz

and

Katrin Schmelz

Department of Economics, University of Konstanz

Abstract

We experimentally study the role of environments with punishment and reward for conformity and disconformity in preferences and judgments, resulting in a 3 (reward vs. punishment vs. control) × 2 (preferences vs. judgments) design. To implement punishment and reward, the choices of a group of participants are shown to a non-group member, who assigns either a cost or a benefit to one participant depending on the treatment. We find that punishment leads to more conformity and reward leads to less conformity in participants’ decisions. Disconformity is rare and only exists in a minority of participants in the reward treatment. Our results suggest that economic environments with punishment create incentives for conformity and economic environments with reward create incentives for disconformity in social groups.

(22)
(23)

1.1. INTRODUCTION 13

1.1

Introduction

What do you recommend wearing to a job interview? I would argue that wearing jeans and a t-shirt is your dominant strategy: If you are a good student, then a department that will not give you a job because of your ”sloppy” appearance does not deserve to have you. If you are mediocre, then there are many other candidates like you and dressing casually is the only way for you to get noticed. (Rubinstein, 2013, pp. 195-196)1

The quote from Ariel Rubinstein (2013) presumes that being unique or special can be an advantage in situations where one person is selected from a group to receive a benefit. Given that such situations are frequent in every day life, one would expect to observe people who intentionally separate themselves from the majority based on their actions or opinions on a daily-basis. However, most of the existing research on social decision-making focuses on the contrary behavior.

Ever since the seminal studies of Asch (1952, 1956), conformity is generally considered to be a strong behavioral motive. Examples of conformity abound and can also be found in the economic sphere. When being informed that most others will do so, people are more likely to pay taxes (Bobek et al., 2007; Coleman, 2007), save energy (Allcott, 2011; Allcott and Rogers, 2014; Nolan et al., 2008; Schultz et al., 2007) or donate to a charity (Alpizar et al., 2008; Smith et al., 2015). At the same time, conformity can be the reason why people shy away from innovative practices, are susceptible to group think, and communicate in echo chambers or filter bubbles. Behavior which is the opposite of conformity is far less studied and evidence is mostly anecdotal. The terms disconformity, anticonformity and counterconformity have been used to describe deliberate behavior which is in opposition to the behavior of the majority. While disconformity can help to trigger innovation or break undesirable routines, it reduces coordination and predictability of behavior and can therefore be both beneficial or costly for society.

As the examples illustrate, both conformity and disconformity can have desirable but also undesirable consequences. So far, little is known about how different economic environments give rise to incentives for conformity or disconformity. As the initial quote suggests, some environments can create incentives to deviate from the majority. What are the defining features of such environments? Is to get noticed the underlying mechanism which incentivizes separation from the majority as suggested in the quote? And, do other environments exist which create incentives to act in line with the majority?

In this paper we provide experimental evidence that shows that environments with punishment and reward create incentives for conformity and disconformity respectively. The results of our experiment suggest that the benefit of acting differently is not to get noticed per se as suggested by the quote. Rather the benefit is to get rewarded for unusual behavior by people with unusual preferences. Simply put, if the probability that the committee of professors favors wearing jeans and a t-shirt themselves is low but greater than the frequency of job candidates who show up

1

(24)

in jeans and t-shirt, you should wear jeans and a t-shirt.

The environments we study experimentally have the following characteristics: One member of a group of individuals is selected. The consequence of the selection is defined by the environment and either a benefit (reward) or a cost (punishment) for the selected individual. Group members have incomplete information about the preferences or judgments of the people who select. The job interview discussed in the quote can serve as an example for an environment with reward. The faculty appoints one or several professors who sequentially interview several graduate students and select one of them for the vacant position. Hence, the consequence of the selection is fixed and implies a benefit for the selected candidate. Candidates have incomplete information about the preferences or judgments of the professors but may possess information about the behavior of other candidates in the interview. To give an example for an environment with punishment, consider a team-meeting where a necessary but rather tedious duty is assigned to one of the team members. Intuitively, it becomes clear, that Rubinstein’s advice does not apply in this environment. In contrast, it is advisable to behave in line with the majority in this environment.

Following Cialdini and Goldstein (2004), we define conformity as a shift in the decisions of an individual towards the majority decision of a group. We use the term disconformity to characterize the opposite behavior and define it as a shift in the decisions of an individual away from the majority decision of a group. In the following, we refer to these changes in behavior with the terms conformity and disconformity. Instead, when we speak of preferences for con-formity or disconcon-formity, we refer to the underlying intrinsic motivation for these behaviors. We consider conformity and disconformity as one dimensional concept with independence as intermediate state in which no shift in behavior occurs when exposed to social information (see Willis, 1965, for a different conceptualization).

We derive hypotheses for the effects of punishment and reward on conformity and disconformity based on the assumption that people prefer to allocate benefits to individuals with similar preferences and costs to individuals with different preferences. Evidence in support of this assumption can be found in the literature on ingroup favoritism (see Hewstone et al., 2002, for a review). As reflected in the initial quote, ingroup favoritism is not the only way to derive our hypotheses. As a minority often possesses the properties of a focal point (Sugden, 1995) in the sense of Schelling (1960), the allocation of benefits and costs based on salience generates the same predictions. The intuition behind our hypotheses is that members with similar preferences or judgments share the consequences if benefits and costs are allocated based on preferences or judgments. As a result, it can be optimal to strategically express the preferences or judgments of the majority in environments with punishment and to express the preferences or judgments of the minority in environments with reward.

(25)

1.2. RELATED LITERATURE 15

and selects one group member based on her decision. In the reward treatment, the selected group member receives a bonus in the form of a fixed amount of experimental currency. In the punishment treatment, the selected group member loses the same amount. In the control treatment decisions are not evaluated. We find that punishment increases conformity and reward reduces conformity in choices compared to the control treatment. Conformity is higher in judgments as compared to preferences. Disconformity is rare and only exists in a minority of participants in the reward treatment. Finally, reward reduces the frequency of unanimity among group members to a level which would be expected in the absence of conformity. The remaining paper is structured as follows. Section 1.2 highlights other theoretical and experimental research on conformity and disconformity. Section 1.3 contains the details of the experimental design, the experimental implementation, and the hypotheses. The experimental results are presented in Section 1.4. Section 1.5 concludes with a discussion of the results.

1.2

Related Literature

People conform if (perceived) pressure makes them act differently from what they would do in the absence of this pressure. Cialdini and Goldstein (2004) integrate these notions into a condensed definition: ”Conformity refers to the act of changing one’s behavior to match the response of others.” (p. 606).

The underlying motives for conformity can be manifold. One driver of conformity is social learning. People acquire information about the utility of choices by observing others’ actions. Since others usually select what they consider to be the best option, the majority choice has some intuitive appeal. The models of Banerjee (1992) and Bikhchandani et al. (1992) suggest that belief learning can also lead to suboptimal choices as the dependency in the publicly observed choices inhibits information aggregation. Anderson and Holt (1997) find that social learning can indeed explain why groups make homogeneous choices in the laboratory but find less homogeneity than predicted by theory.

Several studies suggest that people also learn about their own preferences by observing oth-ers’ actions. Goeree and Yariv (2015) find that people follow othoth-ers’ actions even when it is clear that others are completely uniformed, i.e. they do not possess any private or public information about the usefulness of the alternatives (see also Bernheim and Exley, 2015, for more experimental evidence of preference learning). Models of frequency dependent learning can explain this behavior. Frequency dependent learning postulates that individuals tend to adopt behavior which is common in the population (Boyd and Richerson, 1982). Populations with homogeneous individuals can emerge if the probability that an individual adopts a cer-tain behavior is greater than its actual frequency in the population for common behaviors and smaller for uncommon behaviors. In an experiment, Efferson et al. (2008) find that some but not all of their participants exhibit frequency dependent learning.

(26)

of the actions varies across the type space, individuals can signal their type through actions. When status is sufficiently important relative to the direct utility from actions, most individ-uals will neglect their preferences and conform. The model can explain why norms emerge endogenously in a population as deviations in behavior will seriously harm social status. The mechanism which induces conformity in the model points towards the (often rational) expec-tation that most others will approve the behavior of the majority. Experimental research from differential psychology suggests that fear of negative evaluation increases conformity (Wright et al., 2009) as those who have greater fear of negative evaluation shy away from standing out in a group. Jones and Linardi (2014) show theoretically and experimentally that some people are averse to standing out both negatively and positively.

Evidence for disconformity is mostly anecdotal and few theoretical accounts of the phenomenon exist. An exception is Hornsey and Jetten (2004). The authors explain disconformity based on an essential human need to differentiate oneself from others in order to maintain a favor-able self-concept. Despite the lack of theoretical concepts, experimental findings related to disconformity exist. Among the oldest experimental findings is Argyle (1957). It shows that individual judgments become more extreme if publicly challenged, which has been attributed to reactance (Brehm, 1966). Other studies that document disconformity focus on cultural dif-ferences. Kim and Markus (1999) let participants choose item from sets which include both common and unique elements. While East Asians target the common items more often, Euro-pean Americans more frequently choose the unique items. The authors explain the difference by a positive (freedom, independence) connotation of disconformity in Western cultures and negative (deviance) in eastern cultures. Yamagishi, Hashimoto, and Schug (2008) show that the findings of Kim and Markus (1999) are sensitive to the context. When others are affected by the individual choice due to an insufficient availability of items, Americans choose majority items as rarely as Japanese. Moreover, Japanese went as often for the minority items as Amer-icans when others were not affected by the choice. These findings suggest that the observed cultural differences do not go back to different preferences but to different default strategies across cultures.

(27)

1.3. EXPERIMENTAL DESIGN 17

the opinion of the reviewer is known beforehand, individuals strategically align their opinion with the opinion of the reviewer. Ingratiation is rewarded by the reviewers but less when it is common knowledge that opinions can be adapted strategically.

1.3

Experimental Design

The core elements of our experimental design are sets which consist of three options, e.g. A, B, and C. Accordingly, three binary choices are possible for each set (A vs. B, B vs. C, and A vs. C). Individuals face these three binary choices in a specific order. First, they make two decisions (e.g. A vs. B,B vs. C) without information about others’ decisions. We will refer to these decisions as unconditional decisions since participants cannot condition their behavior on information about others’ decisions in the same choice. We use the unconditional decisions in two ways. We use the own unconditional decisions of an individual to predict decisions and the unconditional decisions of the other individuals for social influence. In the remaining conditional decision (in the example A vs. C), individuals obtain information about the unconditional decisions of two other participants facing the same choice. An example of a decision screen for an unconditional and for a conditional decision can be found in Appendix A.2.

Table 1.1: Timing of unconditional and conditional decisions

t1 t2 t3

unconditional unconditional unconditional 2

individual 1 Avs. B B vs.C Avs.C

individual 2 Avs.C B vs.C Avs. B

individual 3 Avs. B Avs.C B vs.C

Table 1.1 shows the timing of the unconditional and conditional decisions for the item set (A, B, C) in a group of three individuals. As time advances from left to right, individual 1 starts with the unconditional decision A vs. B followed by the unconditional decision B vs. C. By the time individual 1 faces her conditional decision (A vs. C), individuals 2 and 3 did already face the same choice as an unconditional decision in t1 and t2 respectively. Table 1.1 illustrates that this property holds for each of the three conditional decisions in the group of three.

(28)

C  B, we predict the conditional decision based on self-reports of the strength of preferences in the unconditional decisions. For example, if individual 2 strongly prefers A over C but only weakly prefers B over C, we would predict that individual 2 prefers A over B in the conditional decision.

Assume we predict that individual 2 prefers A over B in the conditional decision. Individual 2 could potentially face three different scenarios is the conditional decision. We introduce the following notation for the three scenarios:

A|BB (1)

A|AB (2)

A|AA (3)

In the first scenario, individual 2 is informed that individuals 1 and 3 preferred B in their unconditional decisions. The notation we use to refer to this scenario is A|BB since we predict the individual prefers A conditional on observing two B choices. If individual 2 violates tran-sitivity and chooses B instead of A, we consider this as intrantran-sitivity in line with conformity and use the notation A → B|BB. In the third scenario, individuals 1 and 3 both preferred A in the unconditional decisions. If individual 2 makes an intransitive choice in this scenario, we consider this as intransitivity in line with disconformity and use the notation A → B|AA. We use the second scenario to obtain a benchmark for intransitivity. In scenario A|AB, one individuals preferred A over B and the other B over A. Since intransitivity cannot be explained based on conformity or disconformity when information about others’ decisions is mixed, we speak of intransitivity without any motive and use the notation A → B|AB.

1.3.1

Domains and treatments

(29)

1.3. EXPERIMENTAL DESIGN 19

We study choices across two domains: preferences and judgments. Participants first indicate their decisions for a block of 10 item sets from one domain and subsequently for 10 item sets from the other domain. The order of the domains is balanced across all sessions of one treatment. In each block, participants first face the 20 unconditional decisions. Thereafter participants face the 10 conditional decisions. To study preferences, we use sets of paintings. Each set consists of three paintings with similar motives from the same or different painters. We incentivize participants’ decisions by handing over one painting in the form of an art postcard at the end of the experiment. To study judgments, we use knowledge questions which have an objectively correct answer. Each set consists of one question and three different potential answers to the same question. For each possible piece-wise comparison of two answers, one answer is correct. To give an example, one set consists of answers to the question: ”In which language does the letter a occur more often? (in percent)”. Potential answers were ”German”, ”French”, and ”English”. If the possible answers are ”German” and ”French”, ”French” is the correct answer. If the answers are ”English” and ”French”, ”English” is the correct answer. Lists of the sets of paintings and questions we used in the experiment can be found in Appendix A.2. The item sets are pre-selected based on pilot sessions. We use item sets where each option is sufficiently often selected in each pairwise comparison of the items. In this sense, the decisions we use in the experiment are difficult to predict.

1.3.2

Hypotheses

In the following, we outline our main hypotheses for the effects of reward and punishment on conformity and disconformity. We begin by sketching a simple behavioral model to illustrate our hypotheses. Consider a choice between option A and option B, where A and B could be two postcards or two answers to the same question. Let ∆i = ui(A) − ui(B) denote the utility

difference between A and B for individual i. After observing the two unconditional decisions of the other group members, individual i updates her prior belief about the distribution of the utility differences. Let σAi denote the posterior belief of individual i that a randomly selected individual from the population will prefer A over B. In addition to the utility from choices, individuals receive utility ci if their choice matches the majority choice in their group.

The parameter ci reflects the intrinsic preference of individual i to conform to the majority.

Individual i is a conformists if ci > 0, a disconformists if ci < 0 and independent if ci = 0. To

derive predictions for the conditional decisions, we assume that the distributions of ∆i and ci

are independent and have full support on the real line.

(30)

choice is selected with probability s due to its higher salience. With probability 1 − s, the process unfolds as discussed above. The assumptions regarding the evaluation behavior do not follow endogenously from the model but can be tested in our experiment. The selected group member receives utility r. The selection induces reward if r > 0, punishment if r < 0 and has no consequence if r = 0.

Our main hypotheses concern the conditional decisions and are:

H1: The frequency of conditional choices in line with the majority is higher under punishment and lower under reward as compared to the control treatment.

H2: Reward decreases the level of conformity and increases the level of disconformity. Punish-ment increases the level of conformity and decreases the level of disconformity.

In the following, we show that the hypotheses follow from the model if at least one of the following statements is true:

S1: Individuals’ posterior beliefs are sufficiently close to one half. S2: The effect of salience is strong.

For the conditional decision, three different situations are possible. (I) Either both other group members have chosen A in their unconditional decisions or (II) one group member has chosen A and the other group member has chosen B or (III) both group members have chosen B. We first analyze the effect of reward in situations (I), (II), and (III) and then the effect of punishment.

With reward (r > 0) individual i will prefer A in situation I, II and III respectively, if:

ui(A) + ci+ r/3 − ui(B) + s + (1 − s) 1 − σAi  r > 0 (AA)

ui(A) + ci+ (1 − s)σiAr − ui(B) + ci+ (1 − s) 1 − σiA r > 0 (AB)

ui(A) + s + (1 − s)σiA r − (ui(B) + ci+ r/3) > 0 (BB)

The frequency of A choices under reward is determined by the frequency of situations in which the conditions outlined above hold. The equations simplify to:

∆i+ ci+ (1 − s)σAi − 2/3 r > 0 (R I)

∆i+ (1 − s)(2σiA− 1)r > 0 (R II)

∆i− ci+ (s + (1 − s)σAi − 1/3)r > 0 (R III)

Consider Equation (R I) first. If either σiA< 2/3 ∀ i ∈ {1, · · · , N } or s > 1/3, the left hand side of the inequality decreases in r. Introducing reward then has two effects. First, as A choices become less frequent, the frequency of choices in line with the majority decreases in situation (AA). Second, reward will prevent intransitive decisions B → A|AA in line with conformity for some conformists with ∆i < 0 and induces intransitive decisions A → B|AA in line with

disconformity for some conformists and disconformists with ∆i > 0.

(31)

1.3. EXPERIMENTAL DESIGN 21

the frequency of choices in line with the majority decreases in situation (BB). Also, reward will prevent intransitive decisions A → B|BB in line with conformity for some conformists with ∆i > 0 and induce intransitive decisions B → A|BB in line with disconformity for some

conformists and disconformists with ∆i < 0.

Finally, Equation (R II) reveals that the effect of r > 0 in situation (AB) depends on the posterior belief of individual i. The effect of reward in situation (AB) is small if σiA≈ 0.5. To adapt Equations (AA), (AB), and (BB) to punishment, every occurrence of σAi in (AA), (AB), and (BB) has to be exchanged by (1 − σA

i ), since evaluators who prefer A select B for

punishment and vice versa. After simplification, we obtain the following results for punishment: Individual i will prefer A in situations (I), (II) and (III) respectively, if:

∆i+ c + (1 − s) 1 − σAi  − 2/3 r > 0 (P I)

∆i+ (1 − s)(2 1 − σAi  − 1)r > 0 (P II)

∆i− c + (s + (1 − s) 1 − σAi  − 1/3)r > 0 (P III)

Consider Equation (P I) first. If either 1/3 < σA

i ∀ i ∈ {1, · · · , N } or s > 1/3, the left hand side

of the inequality decreases in r. Introducing punishment has two effects. First, as A choices become more frequent, the frequency of choices in line with the majority increases in situation (AA). Second, punishment prevents intransitive decisions A → B|AA in line with disconformity for some disconformists with ∆i> 0 and induces intransitive decisions B → A|AA in line with

conformity for some conformists and disconformists with ∆i < 0.

A similar result can be obtained for Equation (P III). If either σAi < 2/3 ∀ i ∈ {1, · · · , N } or s > 1/3, the left hand side of the inequality increases in r. As B choices become more frequent under punishment, the frequency of choices in line with the majority increases in situation (BB). Also, punishment prevents intransitive decisions B → A|BB in line with disconformity for some disconformists with ∆i < 0 and induces intransitive decisions A → B|BB in line with

conformity for some conformists and disconformists with ∆i > 0.

Equation (P II) reveals that the effect of r < 0 in situation (AB) depends on the posterior belief of individual i. The effect of punishment in situation (AB) is small if σAi ≈ 0.5.

1.3.3

Implementation

We conducted 6 experimental sessions between February and May 2017 with students of the University of Konstanz, Germany. Participants were recruited using ORSEE (Greiner, 2015) and the sessions were conducted with z-Tree (Fischbacher, 2007). The sample consists of 168 students with a mean age of 22.5 years. 52 % of the participants were female.

(32)

three different options for each domain. The 10 sets of paintings were selected from 15 sets with four options each. For the questions we selected 10 sets from a larger list of 18 sets. Lists of the sets of paintings and questions can be found in Appendix A.2. In total, each participant faced 60 choices and evaluated members of a different group 10 times using the strategy method. To elicit the strength of preferences and judgments, we use the following method. In the preference domain, participants indicate how strongly they prefer one option over the other based on a slider which appears on the screen after every choice. In the judgment domain, participants indicate how certain they are that the answer they have chosen is correct using the same slider. An example of how the slider was displayed to participants can be found in Appendix A.2.

In the punishment or reward sessions, one of the 30 choices in each domain was randomly selected for each group at the end of the experiment. Every selected individual received 10 Euros or lost 10 Euros depending on the treatment. To incentivize the decisions in the preference domain, one of the 30 choices was randomly selected for each group at the end of the session and postcards of the chosen paintings were handed over to participants. We also displayed the correct answers of the knowledge questions at the very end of the experiment. Participants received a flat payment for participation which varied based on the treatment (Control: 16 Euro, Reward: 20 Euro, Punishment: 30 Euro). The flat payment was adjusted such that expected earnings were similar across the three treatments.

1.4

Experimental Results

In the following, we will present most of the main results graphically. The robustness of the results is assessed based on logistic regression models. Results which require predictions based on transitivity, use the self-reported strength of preferences and judgments whenever it is not possible to make predictions based on the unconditional decisions alone. The results remain qualitatively the same if we use only the subset of conditional decisions which can be predicted based on choices.

1.4.1

Evaluation behavior

(33)

1.4. EXPERIMENTAL RESULTS 23

individuals with the same preferences or judgments.

The right panel shows the relative frequency of selecting the unique decision, i.e. the painting (or answer) that was only chosen by one member of the evaluated group. As a benchmark, the relative frequency of selecting the unique decisions is shown for the case where evaluators strictly select according to their own preference. The dotted grey line additionally depicts the benchmark of one third for a random selection of one of the group members. The differences between the observed values and the benchmark for the own preference are small. Overall, the results of Figure 1.1 suggest that evaluation behavior is mainly driven by the own preference or judgment of the evaluator and is not influenced by salience. The results depicted in Figure 1.1

Figure 1.1: Evaluation behavior across treatments

0.0 0.2 0.4 0.6 0.8 1.0 selection (relativ e frequency) reward punishment benchmark

questions pictures questions pictures

own preference unique choice

0.0 0.2 0.4 0.6 0.8 1.0

Notes: The figure shows the relative frequency of selecting the own preference (left panel) or unique choice (right panel) across treatments. Decisions without unanimity considered (e.g. A, A, B). Decisions were elicited using the strategy method. In the left panel, benchmark indicates the average expected frequency based on random selection of either an A or a B decision. In the right panel the benchmark indicates the average frequency based on the selection of the own preference. The dotted line indicates the benchmark of one third of randomly choosing one of the group members. Whiskers indicate 95% CIs, based on 10000 block bootstrap samples (participant ID).

are confirmed by the logit regression models depicted in Table 1.2. The table shows the average marginal effects of the two dummy variables across the reward and punishment treatments for observations where group choices were mixed (e.g. A, A, B). The dependent variable is an indicator variable which is one if the choice A was selected or not. The variable own preference is a dummy variable that indicates whether choice A was preferred by the evaluator in her unconditional decisions between A and B. The variable unique choice is a dummy variable that indicates whether A is chosen by as single group member only. The coefficients of the dummy variables indicate that the own preference is a significant predictor of evaluation behavior. Whether or not the choice is unique does not play a role for evaluation behavior. The first result of the experiment is:

(34)

Table 1.2: Logit models of evaluation behavior

reward punishment

questions paintings both questions paintings both own preference -∗∗∗0.62∗∗∗ ∗∗∗0.54∗∗∗ -∗∗∗0.58∗∗∗ ∗∗∗-0.48∗∗∗ ∗∗∗-0.56∗∗∗ ∗∗∗-0.52∗∗∗ unique choice ∗∗∗-0.14∗∗∗ 0.03 -0.05 -0.09 -0.06 -0.07 obs. 648 648 1296 648 648 1296

N 54 54 54 54 54 54

Notes: Table shows average marginal effects. Dependent variable is a dummy that indicates whether option A is selected. Own preference is a dummy variable that is one if A was the own choice. Unique choice is a dummy variable that is one if A was the unique choice. Decisions were elicited using the strategy method. ∗∗∗ Significant at the 1 percent level. ∗∗Significant at the 5 percent level. ∗Significant at the 10 percent level. Standard errors are clustered on participant ID.

1.4.2

Treatment effects

In the control treatment, the relative frequency to decide in line with the majority in a con-ditional decision is 80 percent. The relative frequencies in the reward treatment and in the punishment treatment are 65 percent and 87 percent respectively. Table 1.3 shows average marginal effects of covariates on an indicator variable which is 1 if the individual chooses the majority decision in the conditional decision and zero otherwise. The coefficients of the

treat-Table 1.3: Logit models for conditional choices in line with the majority

questions paintings both reward ∗∗∗-0.11∗∗∗- ∗∗∗-0.17∗∗∗- ∗∗∗-0.14∗∗∗ -punishment ∗∗∗0.07∗∗∗ ∗0.09∗ ∗∗∗0.08∗∗∗ strength -0.02 0.06 -0.00 -sequence -0.01 0.00 -0.00 -male -0.01 -0.02 -0.01 -age 0.00 0.01 0.00 birth position -0.03 0.03 -0.00 -obs. 960 930 1890 N 168 168 168

Notes: Strength is the predicted strength value for the conditional decision. Sequence is a dummy which indicates if decisions were first made in the preference domain. ∗∗∗ Significant at the 1 percent level. ∗∗ Significant at the 5 percent level. ∗Significant at the 10 percent level. Standard errors are clustered on participant ID.

ment dummies indicate that the differences in the frequency of choosing the majority between the reward and the control treatment and between the punishment and the control treatment are statistically significant. The second result is:

(35)

1.4. EXPERIMENTAL RESULTS 25

A → B|BB is higher compared to A → B|AB across all treatments. This suggests that on average, individuals have an intrinsic preference for conformity. The comparative statics of punishment and are as hypothesized. Punishment increases the frequency of intransitivity A → B|BB while reward decreases this frequency.

Figure 1.2: Frequency of intransitivity across treatments

0.0 0.2 0.4 0.6 0.8 1.0 intr ansitivity (relativ e frequency) A→BIBB A→BIAB A→BIAA

reward control punishment reward control punishment

questions pictures 0.0 0.2 0.4 0.6 0.8 1.0

Notes: The figure shows the frequency of intransitivity in the conditional choices across treatments. Predictions for the conditional choices are based on participants’ self-reported strength of preferences and judgments in the unconditional decisions if no prediction is possible based on the unconditional decisions alone. Whiskers indicate 95% CIs, based on 10000 block bootstrap samples (participantID).

Figure 1.2 shows that the frequency of intransitive decisions A → B|AA is very low. The comparative statics of the model also hold for this scenario as reward induces more intransitivity and punishment induces less intransitivity - even though disconformity is already close to zero in the control treatment. However, the frequency of intransitivity A → B|AA is not bigger than the benchmark A → B|AB in any of the treatments. This implies that diconformity is not observed in average behavior.

The right panel of Figure 1.2 shows a very similar pattern for the conditional decisions in the preference domain. The main difference is that the general propensity towards conformity is not as strong as in the judgment domain. In the reward treatment, the levels of intransitivity are not different across the three scenarios. The results of the logit regression models provided in Table 1.4 confirm the graphically illustrated results. Table 1.4 shows average marginal effects of covariates on a indicator variable that is one if the conditional choice was intransitive. The coefficients of the dummies for the scenario A|BB and A|AA show that intransitivity A → B|BB is more frequent than A → B|AB, especially in the control an the reward treatment. In contrast intransitivity in line with disconformity is less frequent as compared to the benchmark scenario. Table 1.4 also reveals that the predicted strength value of the conditional decision is negatively related to the frequency of intransitivity which is in line with our model.

(36)

Table 1.4: Logit models for conformity and disconformity

reward control punishment

questions pictures questions pictures questions pictures

A|BB ∗∗∗0.21∗∗∗ ∗0.10∗ ∗∗∗0.41∗∗∗ ∗∗∗0.23∗∗∗ ∗∗∗0.46∗∗∗ ∗∗∗0.38∗∗∗ A|AA ∗∗-0.15∗∗- 0.04 ∗∗∗-0.25∗∗∗- ∗∗∗-0.11∗∗∗- ∗∗∗-0.27∗∗∗- ∗∗∗-0.20∗∗∗ -strength ∗∗∗-0.37∗∗∗- ∗∗∗-0.33∗∗∗- ∗∗∗-0.20∗∗∗- ∗∗∗-0.48∗∗∗- ∗∗∗-0.26∗∗∗- ∗∗∗-0.20∗∗∗ -sequence 0.01 0.02 0.01 0.04 -0.01- -0.03 -male -0.03- -0.04- 0.01 -0.02- 0.03 0.02 age 0.00 0.00 0.00 0.00 ∗∗-0.01∗∗- 0.00 birth pos. 0.02 0.00 -0.01- 0.00 -0.02- 0.02 obs. 540 540 600 600 540 540 N 54 54 60 60 54 54

Notes: The table shows average marginal effects of logit regressions with the frequency of intransitivity as dependent variable. A|BB and A|AA are dummies for these scenarios. Strength is the predicted strength value of the conditional decision based on the self-reported strength of preference or judgments in the unconditional decisions. Sequence is a dummy which indicates if decisions were first made in the preference domain. ∗∗∗ Significant at the 1 percent level.∗∗ Significant at the 5 percent level. ∗Significant at the 10 percent level. Standard errors are clustered on participant ID.

Table 1.5: Logit models for treatment effects on conformity and disconformity

questions paintings both

A → B|BB A → B|AA A → B|BB A → B|AA A → B|BB A → B|AA

reward ∗∗∗-0.20∗∗∗- ∗0.07∗- -0.10- ∗∗∗0.20∗∗∗ ∗∗∗-0.16∗∗∗- ∗∗∗0.13∗∗∗ punishment ∗∗0.11∗∗ -0.02- ∗∗0.17∗∗ -0.05- ∗∗∗0.14∗∗∗ -0.04 -strength ∗∗∗-0.28∗∗∗- ∗∗-0.09∗∗- ∗∗∗-0.31∗∗∗- ∗∗∗-0.24∗∗∗- ∗∗∗-0.32∗∗∗- ∗∗∗-0.14∗∗∗ -sequence -0.03- -0.03- 0.03 -0.01- 0.01 -0.01 -male -0.04- -0.03- -0.00- -0.00- -0.02- -0.03 -age -0.00- -0.00- 0.01 -0.01- 0.00 -0.01 -birth pos. -0.03- ∗0.03∗ ∗0.07∗ 0.00 0.02 0.02 obs. 390 570 400 530 790 1100 N 168 168 168 168 168 168

Notes: The table shows the average marginal effects of treatment dummies for reward and punishment on a indicator variable for intransitivity. Strength is the predicted strength value of the conditional decision. Sequence is a dummy which indicates if decisions were first made in the preference domain. ∗∗∗ Significant at the 1 percent level. ∗∗

(37)

1.4. EXPERIMENTAL RESULTS 27

dummy for punishment reveals the opposite effect. The decrease of intransitivity in line with disconformity is not statistically significant. The third result is:

R3: Punishment increases conformity and reward decreases conformity. Disconformity is not observed in average behavior.

Next, we turn to the frequency of unanimity in the randomly formed groups which means that all three group members choose the same option. Given that the treatments do not change behavior in the unconditional decisions, reward should decrease unanimity and punishment should increase unanimity in groups. Figure 1.3 depicts the relative frequency of unanimity in groups across treatments. To include a benchmark for the frequency of unanimity, Figure 1.3 also shows a benchmark the level of unanimity which would follow by imposing transitivity on the conditional decisions. The benchmark can be seen as an estimate for unanimity that controls for the effects of treatments and preferences for conformity. The left panel shows that the observed frequency of unanimity in the judgment domain is smaller in the reward treatment and higher in the punishment treatment compared to the control treatment. The differences to the benchmark are substantial for the control and punishment treatment but not for the reward treatment. The right panel shows a similar pattern for the preference domain. The benchmark of the reward treatment indicates that the frequency of unanimity observed with reward is similar to the frequency expected based on imposing transitivity. The results

Figure 1.3: Frequency of unanimity across treatments

unanimity (relativ e frequency) 0.0 0.2 0.4 0.6 0.8 1.0 reward control punishment benchmark questions paintings

Notes: The figure shows the average frequency of unanimity in group choices across treatments. Benchmark indicates the average frequency of unanimity based on predictions based on transitivity. Whiskers indicate 95% CIs, based on 10000 block bootstrap samples (participant ID).

of regression models provided in Table 1.6 confirm the the result depicted in Figure 1.3. The table shows average marginal effects of treatment dummies for reward and punishment. The binary dependent variable is an indicator variable which is one if a group displays unanimity. The coefficient of the treatment dummy for reward indicates that the frequency of unanimity is significantly lower in the reward treatment as compared to the control treatment. The fourth result is:

(38)

Table 1.6: Logit models for unanimity

questions paintings both reward ∗∗∗-0.14∗∗∗- ∗∗∗-0.13∗∗∗- ∗∗∗-0.13∗∗∗

-punishment -0.01 0.03 0.01 obs. 1680 1680 3360 groups 56 56 56

N 168 168 168

Notes: ∗∗∗Significant at the 1 percent level. ∗∗Significant at the 5 percent level. ∗Significant at the 10 percent level. Standard errors are clustered on group ID.

1.4.3

Heterogeneity

To explore heterogeneity in the conditional choices, we characterize participants’ decisions by three probabilities. The individual probability to make an intransitive decision in line with conformity A → B|BB (pA→B|BB), the probability to make an intransitive decision in line

with disconformity A → B|AA (pA→B|AA), and the benchmark (pA→B|AB). We estimate

several mixture models with different numbers of parameters and report the simplest model which is complex enough to capture the meaningful heterogeneity in our data.

The results indicate no heterogeneity in the control and punishment treatment where a single conformist type with three distinct probabilities pA→B|BB > pA→B|AB > pA→B|AA survives the selection procedure. In contrast, the behavior in the reward treatment is best summarized by two behavioral types, one conformist and the other disconformist, each characterized by three probabilities. Table 1.7 depicts maximum-likelihood estimates of the population shares and situation specific probabilities for intransitivity for the two behavioral types. The

over-Table 1.7: Heterogeneity in the reward treatment

behavioral type conformist disconformist

-population share 0.86 0.14

-pA→B|BB 0.52 0.06

-pA→B|AB 0.29 0.27

-pA→B|AA 0.14 0.66

-≈ N 46 8

Notes: The table shows the best mixture model for the data of the reward treatment from both domains according to ICL (Celeux & Goveart, 2000). pA→B|BB indicates the type specific probability for intransitivity in line with

conformity, pA→B|AA the type specific probability for intransitivity in line with disconformity, and pA→B|AB the

benchmark without a conformity or disconformity motive. The negative log likelihood of the model is 622 and ICL, the penalized BIC criterion, is 639. The result of the selection procedure is similar if executed from the data of each domain separately.

(39)

1.5. DISCUSSION 29

1.4.4

Questionnaire validation

We use post-experimental questionnaire data to check whether the heterogeneity in behavior can be explained based on individual differences. The questionnaire includes a psychometric conformity measure (Mehrabian & Stefl, 1995), a 10 item measure of the Big-Five personality dimension (Gosling, Rentfrow, Swann, 2003), and socio-demographic questions. To validate the result of the mixture model analysis, we classify the participants of the reward treatment according to their posterior probability of belonging to one of the two behavioral types. By comparing the test scores of the psychometric conformity measure of participants classified as conformists to the scores of participants classified as disconformists, we find that the average score of the conformist type is higher (−0.62 vs. −1.48, higher values indicate conformity, t-test, df = 53, p-value = 0.05).

Table 1.8: Correlation of individual behavior to questionnaire data

pA→B|BB pA→B|AA neuroticism -0.11- -0.17 -extraversion -0.01- -0.04 -openness -0.07- -0.08 -agreeableness -0.19- -0.18 -conscientiousness -0.01- -0.01 -conformity score -0.15- -0.20 -birth position -0.05- -0.11

-Notes: -Notes: The table shows Pearson correlation coefficients for the relationship between behavior observed in the experiment and questionnaire data. pA→B|BB indicates the type specific probability for intransitivity in line with

conformity, pA→B|AA in line with disconformity. Higher conformity scores suggest a greater individual disposition

towards conformity. Birth position is the self-reported birth position of the participants where 1 indicates being the first-borne child, 2 the second-born child, and so on.

As exploratory part of the analyses, Table 1.8 shows Pearson correlation coefficients of the relationship between the individual probabilities for intransitivity in both domains across all treatments and the questionnaire data. The relationship of individual behavior and the psy-chometric conformity measure is as expected but small. Individuals who score higher in the conformity measure also have a higher probability of intransitivity in line with conformity and a smaller probability of intransitivity in line with disconformity. A similar result holds for the Big-Five personality trait agreeableness. Birth position does not correlate with behavior in our experiment.

1.5

Discussion

(40)

to more conformity and reward leads to less conformity in both preferences and judgments. The experiment also reveals that the necessary behavioral assumption regarding the way how people allocate punishment and reward is valid. In the experiment, people reward those who display similar preferences or judgments and punish those who display different preferences or judgments. Also in line with our predictions, we observe fewer choices in line with the majority preference or opinion of the own group under reward and more choices in line with the majority preference or opinion of the own group under punishment. The effect of reward on individual choices can also be observed on the group-level, as reward reduces the frequency of unanimity in groups.

Our results suggest that the level of conformity is closely tied to the domain in which decisions are made. We observe a higher degree of conformity in judgments compared to preferences. One important difference between the domains we study is that there exists an objectively correct answer for the type of judgments we use in our experiment. The higher level of conformity observed in judgments could thus be explained based on belief learning. Adopting one’s own decision to the decisions of others can be rational if individuals posses private information which is on average correct - with the associated problems outlined by Banerjee (1992) and Bikhchandani et al. (1992). Another explanation of the higher level of conformity in judgments could be the fear of being the only one who answered incorrectly. Being perceived as the only one who gives the wrong answer might outweigh the benefit of being the only one who gives the correct answer. We are not able to disentangle these explanations based on our experiment but think of them as an interesting direction for future research.

The results of our experiment are generally in line with the findings of Griskevicius et al. (2006) despite some important methodological differences between the two studies. In our experiment, we measure conformity and disconformity without using deception and induce the anticipation of punishment and reward by monetary incentives. The results of our reward treatment complement the findings of Robin et al. (2014). The crucial difference in their experiment is that the preferences and judgments of the person distribution monetary reward are known to all group members. In their case, monetary reward induces conformity due to ingratiation.

(41)

1.5. DISCUSSION 31

about the selection behavior during the experiment, our experimental design is not able to address the effect of disconformity on the future behavior of other participants.

Some aspects of the evaluation behavior do also not turn out to be as expected. We assumed that the evaluation process is influenced by salience but find that salience only plays a minor role for evaluation behavior. A potential explanation might be the small group size in our experiment. Three participants might simply not be enough to make the single preference or judgment particularly salient.

An important limitation of our experiment with respect to external validity is that participants do not receive feedback about the evaluation behavior. Hence, participants in the punishment and reward treatments are not able to update their beliefs about the evaluation behavior during the experiment and the treatment differences are driven by participants’ prior beliefs about the evaluation decisions. We think of our study as a reasonable starting point to study the effects of punishment and reward. Our design can be extended to study the effects of punishment and reward when belief learning is possible. We expect that the treatment differences we observe will further increase if participants receive feedback about the evaluation behavior during the experiment.

Finally, we would like to highlight a property which makes environments with reward interesting for mechanism design. Implementing an optimal degree of diversity in groups of individuals has been identified as an important challenge of organizations (March, 1991; Kets and Sandroni, 2016). In some cases, the relevant degree of diversity is defined by the distribution of preferences or opinions in the population. Conformity can bias the diversity of groups downwards even if groups consist of independent samples from the population. However, incentivizing diversity directly is not possible if the distribution of preference or opinions is unknown or changes frequently. The incentives in environments with reward possess the interesting feature that the choices and opinions eventually expressed in the group mirror the most likely sample of preferences and opinions given their distribution in the population. To give an example consider the case of voting, assume that candidates are independently sampled from the population and receive a benefit from receiving votes in a single-vote system. If the benefit is high enough, the publicly expressed statements of the candidates should mirror the preferences of the electorate even if the distribution of preferences of the candidates does not. A negative consequence of the core mechanism which guarantees this property has been recognized by political scientists as the problem of vote-splitting2. Our results suggest that reward can restore the diversity of publicly expressed opinions in groups negatively affected by conformity or group-think.

(42)

A

Appendix of Chapter One

A.1

Analysis of heterogeneity

The exact procedure to obtain the mixture models reported in Table 1.7 is the following: For each treatment, we estimate 10 mixture models with K ∈ {1, · · · , 10} distinct behavioral types where each type k ∈ K is characterized by up to three distinct probabilities corresponding to scenarios A|BB, A|AB and A|AA. To select the number of distinct probabilities, we assume that the behavior of types is characterized by s ∈ Sk internal states which determine the

probability of an intransitive response and a mapping of the scenarios A|BB, A|AB, A|AA into internal states s ∈ Sk. Hence, total number of distinct probabilities across types is S =

P

k∈KSk and reflects the complexity of the behavioral responses of the types.

(43)

A. APPENDIX OF CHAPTER ONE 33

A.2

Study materials

Figure A.1: Unconditional and conditional decision screens

Notes: The screen depicted in on the left shows an unconditional decision without information about others’ behavior. After participants select one option and confirm their selection, the slider below the paintings appears. The screen depicted on the right illustrates a conditional decisions where the decisions of the two other group members are depicted as the paintings on the left and right in the top line. The painting in the middle represents the participant’s conditional choice.

Figure A.2: Evaluation screen

(44)
(45)
(46)
(47)
(48)

Abbildung

Updating...

Referenzen

Updating...

Verwandte Themen :