Top Banner
Games and Economic Behavior 74 (2012) 285–298 Contents lists available at ScienceDirect Games and Economic Behavior www.elsevier.com/locate/geb Whom do you distrust and how much does it cost? An experiment on the measurement of trust Bill McEvily a , Joseph R. Radzevick b , Roberto A. Weber c,a Rotman School of Management, University of Toronto, Canada b Department of Management, Gettysburg College, United States c Department of Economics, University of Zurich, Switzerland article info abstract Article history: Received 3 June 2010 Available online 30 June 2011 JEL classification: C72 C90 D03 Keywords: Trust Trustworthiness Altruism Experiment We advance the measurement of trust in economics in two ways. First, we highlight the importance of clearly identifying the target of trust, particularly for obtaining concordance between attitudinal and behavioral measures of trust. Second, we introduce a novel behavioral measure of (dis)trust, based on individuals’ willingness to pay to avoid being vulnerable to the target of trust. We conduct an experiment in which we vary the target of trust among passersby at several locations around a city, measuring both behavioral distrust and trust attitudes towards these varying targets. We find that subjects discriminate based on perceived characteristics of different targets in determining whether to trust, in a manner consistent with trust elicited using attitudinal measures and with actual trustworthiness. Risk aversion and altruism do not correlate highly with our measure of distrust. © 2011 Elsevier Inc. All rights reserved. 1. Introduction Trust is often regarded as an important concept for understanding economic, financial, organizational, and social activ- ity (Arrow, 1974; Putnam, 1993; Knack and Keefer, 1997; McEvily et al., 2003; Guiso et al., 2004). Trust is viewed as a mechanism whereby potentially beneficial exchanges can occur while overcoming the presence of moral hazard. Despite concurrence on the potential importance of trust, there is less agreement across the social sciences on how it should be de- fined or, especially, on how it should be measured (Glaeser et al., 2000; McEvily and Tortoriello, 2011). The measurement of trust in economics relies primarily on behavioral measures, consistent with an emphasis on using choice as an appropriate way to infer preference. In contrast, other social sciences (i.e., psychology, sociology) primarily rely on attitudinal measures of trust that elicit unincentivized beliefs or expectations about the motives and intentions of another party. Although there are several advantages to the reliance on behavioral measures in economics, there are also some po- tentially significant shortcomings. Behavior in the “investment game” (Berg et al., 1995), the most widely used behavioral measures in economics, often fails to correlate with other (attitudinal) measures of trust (Glaeser et al., 2000). This alone We gratefully acknowledge funding from a Carnegie Mellon Berkman Faculty Development Grant. Weber also gratefully acknowledges support from the National Science Foundation (SES-0433152) and from the research priority program at the University of Zurich “Foundations of Human Social Behavior.” We also thank participants at several seminars and conferences, as well as an anonymous editor and two anonymous reviewers, for helpful comments and suggestions. We are also very thankful to John Hamman for valuable research assistance, the Pittsburgh Experimental Economics Laboratory (PEEL) for access to resources, and David Laibson for granting us access to his data. Finally, one of the authors would like to thank his wife, Stephanie, for valuable support and help during the completion of this project. * Corresponding author. E-mail address: [email protected] (R.A. Weber). 0899-8256/$ – see front matter © 2011 Elsevier Inc. All rights reserved. doi:10.1016/j.geb.2011.06.011
14

Whom do you distrust and how much does it cost? An experiment on the measurement of trust

Apr 24, 2023

Download

Documents

Alireza Nouri
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: Whom do you distrust and how much does it cost? An experiment on the measurement of trust

Games and Economic Behavior 74 (2012) 285–298

Contents lists available at ScienceDirect

Games and Economic Behavior

www.elsevier.com/locate/geb

Whom do you distrust and how much does it cost? An experiment onthe measurement of trust ✩

Bill McEvily a, Joseph R. Radzevick b, Roberto A. Weber c,∗a Rotman School of Management, University of Toronto, Canadab Department of Management, Gettysburg College, United Statesc Department of Economics, University of Zurich, Switzerland

a r t i c l e i n f o a b s t r a c t

Article history:Received 3 June 2010Available online 30 June 2011

JEL classification:C72C90D03

Keywords:TrustTrustworthinessAltruismExperiment

We advance the measurement of trust in economics in two ways. First, we highlight theimportance of clearly identifying the target of trust, particularly for obtaining concordancebetween attitudinal and behavioral measures of trust. Second, we introduce a novelbehavioral measure of (dis)trust, based on individuals’ willingness to pay to avoid beingvulnerable to the target of trust. We conduct an experiment in which we vary thetarget of trust among passersby at several locations around a city, measuring bothbehavioral distrust and trust attitudes towards these varying targets. We find that subjectsdiscriminate based on perceived characteristics of different targets in determining whetherto trust, in a manner consistent with trust elicited using attitudinal measures and withactual trustworthiness. Risk aversion and altruism do not correlate highly with our measureof distrust.

© 2011 Elsevier Inc. All rights reserved.

1. Introduction

Trust is often regarded as an important concept for understanding economic, financial, organizational, and social activ-ity (Arrow, 1974; Putnam, 1993; Knack and Keefer, 1997; McEvily et al., 2003; Guiso et al., 2004). Trust is viewed as amechanism whereby potentially beneficial exchanges can occur while overcoming the presence of moral hazard. Despiteconcurrence on the potential importance of trust, there is less agreement across the social sciences on how it should be de-fined or, especially, on how it should be measured (Glaeser et al., 2000; McEvily and Tortoriello, 2011). The measurement oftrust in economics relies primarily on behavioral measures, consistent with an emphasis on using choice as an appropriateway to infer preference. In contrast, other social sciences (i.e., psychology, sociology) primarily rely on attitudinal measuresof trust that elicit unincentivized beliefs or expectations about the motives and intentions of another party.

Although there are several advantages to the reliance on behavioral measures in economics, there are also some po-tentially significant shortcomings. Behavior in the “investment game” (Berg et al., 1995), the most widely used behavioralmeasures in economics, often fails to correlate with other (attitudinal) measures of trust (Glaeser et al., 2000). This alone

✩ We gratefully acknowledge funding from a Carnegie Mellon Berkman Faculty Development Grant. Weber also gratefully acknowledges support from theNational Science Foundation (SES-0433152) and from the research priority program at the University of Zurich “Foundations of Human Social Behavior.”We also thank participants at several seminars and conferences, as well as an anonymous editor and two anonymous reviewers, for helpful commentsand suggestions. We are also very thankful to John Hamman for valuable research assistance, the Pittsburgh Experimental Economics Laboratory (PEEL) foraccess to resources, and David Laibson for granting us access to his data. Finally, one of the authors would like to thank his wife, Stephanie, for valuablesupport and help during the completion of this project.

* Corresponding author.E-mail address: [email protected] (R.A. Weber).

0899-8256/$ – see front matter © 2011 Elsevier Inc. All rights reserved.doi:10.1016/j.geb.2011.06.011

Page 2: Whom do you distrust and how much does it cost? An experiment on the measurement of trust

286 B. McEvily et al. / Games and Economic Behavior 74 (2012) 285–298

may not be problematic, but when one also considers that investment game behavior frequently correlates with altruism(Cox, 2004; Ashraf et al., 2006; Capra et al., 2007) and risk seeking (Karlan, 2005; Schechter, 2007) it raises the question ofhow accurately behavioral methods such as the investment game capture trust.1 Therefore, we begin from the need for ad-ditional research that tests the relationship between behavioral and attitudinal trust measures and that explores alternativebehavioral methods for studying trust.

In this paper, we report an experiment that advances behavioral measurement of trust in economics in the above twoways. First, and most importantly, we emphasize the need to clearly identify the target, or object, of trust when comparingthe reliability of different trust measures. To clearly specify the target of trust, laboratory subjects in our experiment areeach matched with several field participants at different locations in the same city. This allows us to measure how trustchanges as the target changes, and the extent to which target-specific attitudinal and behavioral measures are related. Wefind that subjects’ trust varies depending on the target and that such variation is consistent with the actual trustworthinessof the target. More importantly for our purposes, we also demonstrate that, when one clearly specifies and holds constantthe target of trust, behavioral and attitudinal measures are correlated.

Second, to address the need for alternative behavioral methods for measuring trust, to serve as complements to theinvestment game, we introduce a simple novel behavioral measure. In the game, we measure subjects’ distrust towarda particular target individual by eliciting the costs that the subject is willing to incur to limit her vulnerability to thetarget. We find that this behavioral measure, which correlates with actual trustworthiness of the target and with attitudinaltrust measures, is not strongly confounded with other closely related concepts such as altruism and risk seeking in ourexperiment.

2. Advancing behavioral measurement of trust in economics

In this section, we discuss the contribution of our experiment in the context of the preceding literature on the mea-surement of trust. We note the importance of clearly identifying, and holding constant, the target of trust when testing therelationship between different trust measures. We also discuss the need for alternate behavioral trust measures and presentthe novel game we use in our experiment.

2.1. Identifying the target of trust

Sociologists (Blau, 1964; Luhmann, 1979; Lewis and Weigert, 1985) and social psychologists (Johnson-George and Swap,1982; Rempel and Holmes, 1986; Lewicki and Bunker, 1996) have long noted the importance of the target, or object, of trustand the ability of individuals to discriminate the trustworthiness of different targets. For instance, individuals may exhibithigh trust towards their family members but not towards the government; an employee may fully trust her direct supervisorbut not the company CEO. From this perspective, trust is less a stable attribute, or disposition, of the person placing trustand more a property of each specific trustor–trustee interaction. Given this, sociologists and social psychologists approachthe measurement of trust as a feature of a relationship with a specific individual, in a specific context. For instance, ina review of the most widely adopted attitudinal measures in social psychology, Wrightsman (1991; 375) highlights the“Specific Interpersonal Trust Scale” (Johnson-George and Swap, 1982) as an exemplar since it was “designed to measuretrust of another person under particular circumstances.”

If we accept that individuals can exhibit different levels of trust for different targets, considering targets is critical to themeasurement of trust. Researchers must first decide whether generalized trust (i.e., a trustor’s relatively stable disposition)or target-specific trust (i.e., directed towards a particular individual or entity) is relevant to their research questions. In thelatter case, researchers should further identify the specific target(s) of interest. Many economic studies, however, ignore thetarget of trust and instead treat an individual’s willingness to trust as a stable trait that can be adequately measured using asingle behavioral measure with a single target.2 Thus, when an experimental game reveals a laboratory subject i to exhibita level of trust xi , we must be cautious to note that what we have really measured is xi( j), where j represents the target(usually another anonymous subject in the experiment). Of course, this observed behavior is a useful measure of some kindof trust, but the researcher should be careful not to attempt to generalize this as a measure of trust that necessarily extendsto other targets.

1 However, Eckel and Wilson (2004) find only a weak relationship with risk attitudes and Ben-Ner and Halldorsson (2010) find a significant relationshipbetween investment game behavior and altruism but not risk-seeking.

2 The few studies in economics that consider features of the target of trust are consistent with our contention of its importance, though none highlightsthe general importance of target-specificity for the measurement of trust to the extent we do here. Specifically, studies explore how trust varies basedon target characteristics such as attractiveness (Eckel and Wilson, 2006; Scharlemann et al., 2001), “social distance” from the trustor (Buchan et al., 2002;Ben-Ner et al., 2009), ethnic origin (Fershtman and Gneezy, 2001), and on whether the target is an individual or group (Kugler et al., 2007; McEvily et al.,2006). None of these studies explores how behavioral and attitudinal measures are related when accounting for the target of trust, as we do here. Twostudies make a point closer to ours. Holm and Danielson (2005) distinguish between “thick” trust (towards those with whom one has close social ties) and“thin” trust (towards people in general). Yet, by attempting to correlate “thin” attitudinal measures (General Social Survey questions) with behavioral trustwithin potentially “thick” populations (students in the same economics classes) they do not hold constant the specific target of trust. Finally, Knack andKeefer (1997) note a distinction between “generalized” trust (towards people in general) and “specific” trust (towards “people one has repeated interactionswith” (p. 1258)), but this discussion of target-specificity is tangential in their paper.

Page 3: Whom do you distrust and how much does it cost? An experiment on the measurement of trust

B. McEvily et al. / Games and Economic Behavior 74 (2012) 285–298 287

While identifying the target of trust is important for trust measurement in general, we believe it is particularly importantfor comparing and integrating research on trust across the social sciences, especially in cases where different approachesare used to measure trust. To make valid comparisons, different measures must address the same target. For example, if wecompare individual i’s behavior in an experimental game measuring trust towards target j, si( j), to an attitudinal measureof trust towards target k, ai(k), it should not be surprising if the two measures do not correlate. To test the reliability of thetwo measures, we should hold the target constant, comparing si( j) to ai( j).3

The importance of identifying the target of trust helps to explain the lack of correspondence across different measuresof trust, as notably observed by Glaeser et al. (2000). The primary behavioral measures used by Glaeser et al. consisted ofan “envelope drop,” in which the targets were strangers at different locations around Boston, and the investment game, inwhich subjects played either against an anonymous other student – a “stranger” – or someone they knew. Glaeser et al.compared the degree of trust exhibited by subjects in these contexts to 12 different attitudinal measures, ranging from abroad General Social Survey (GSS) item to more specific questions. Of the 12 attitudinal measures, only two significantlypredicted trusting behavior in either game.

However, Glaeser et al. primarily tested the reliability of different measures of trust across targets. For instance, theycompared the very general GSS question, which measures trust in “most people,” to behavior in an investment game inwhich some subjects were paired with someone they knew.4 If a subject thought that people in general are trustworthy butknew that the particular matched person was not, then it is unsurprising to find no relationship between the measures.

An accurate comparison across trust measures requires holding the target constant. In fact, some aspects of Glaeser etal.’s data, which we re-analyzed, support our claim that such measures are related if the target of trust is held constant.Specifically, the two attitudinal measures that correlate significantly with the behavioral measures identify “strangers” asthe targets of trust (see also Fehr et al., 2003). Subjects who express trust in strangers in these two attitudinal questionsstate significantly higher reservation prices for the envelope drop (indicating greater trust). Similarly, subjects who expressgreater trust of strangers in the attitudinal measures send significantly more money in the investment game, but only whenplaying against an anonymous opponent. Thus, when the target of trust is held constant to “strangers” – anonymous individualsone encounters in passing – Glaeser et al.’s results reveal strong positive relationships between behavioral and attitudinalmeasures.

Our experiment directly addresses the importance of the target of trust by explicitly varying the counterpart in a be-havioral game. We compare trust elicited via a novel behavioral measure (described below) with attitudinal trust measures(each collected at different times, as in Glaeser et al.’s study). We find that when accounting for variation in the target,there exists a positive relationship between attitudinal and behavioral measures.

2.2. Developing alternative behavioral trust measures

In addition to emphasizing the importance of the target of trust, we also highlight the need to develop alternative,complementary, behavioral measures of trust. While the investment game has proved a highly versatile and valuable methodfor economists interested in measuring trust, it is not without problems. For instance, behavior in the investment game isoften confounded with other closely related concepts, such as altruism or fairness (Cox, 2004; Ashraf et al., 2006; Capra etal., 2007) or risk-seeking (Karlan, 2005; Schechter, 2007).

Our experiment develops a different approach to measuring trust through behaviors. While behavioral measures such asthe investment game approach the measurement of trust by attempting to elicit the act of trusting (i.e., by observing howmuch money subjects send to another player), we note that the phenomenon of trust is often evident and easily measuredby observing the extent to which individuals distrust others. When trust is absent or incomplete, individuals frequentlyincur costs to mitigate their vulnerability. We view the willingness to incur such costs as an inverse indicator of trust.For example, if one individual fully trusts another, she is willing to enter into an exchange with no safeguards to mitigateher vulnerability to the other. However, when trust is absent, individuals often pay for mechanisms – such as contracts,monitoring, or enforcement – to mitigate such vulnerability. Thus, we propose to measure trust/distrust through individuals’willingness to incur costs to mitigate vulnerability to the target.5

To formalize our procedure and present the game we use in our experiment, suppose that two individuals, i and j, canpotentially enter into an exchange that yields a surplus W , which they have agreed to divide evenly. Suppose also that jpossesses agency over the final distribution of W , meaning that he can appropriate any portion (x) of W , leaving i withthe remainder (W − x).

3 In fact, the social psychology, sociology, and organizational literatures further suggest that the context in which trust occurs is important. Individual imay trust j with a monetary investment (as one might with a financial advisor), but the same individual i may exhibit almost no trust towards j in othercontexts, such as caring for i’s children.

4 Karlan (2005) similarly compared behavior in non-anonymous investment game pairings with GSS trust questions.5 Our measure is related to work that explores subjects’ willingness to adopt contracts in situations involving trust. Ben-Ner and Putterman (2009) report

an investment game experimental treatment in which both parties may agree to bear an exogenous fixed cost to create a contract that penalizes playersfor deviating from strategies to which both players have previously agreed. They find that few pairs of subjects adopt this kind of contract. Other relatedresearch explores the extent to which binding contracts have detrimental effects on voluntary trust (e.g., Malhotra and Murnighan, 2002). Our approach isnew in that we explicitly use subjects’ willingness to pay for a kind of contract as a measure of distrust.

Page 4: Whom do you distrust and how much does it cost? An experiment on the measurement of trust

288 B. McEvily et al. / Games and Economic Behavior 74 (2012) 285–298

In the above situation, the extent to which i trusts j matters for whether the exchange occurs and under what conditionsit occurs. If i fully trusts j to divide the surplus evenly (x = 1

2 W ), then she should agree to the exchange with no additionalsafeguards.6 However, suppose instead that i exhibits some distrust towards j, and expects him to appropriate a largershare of the surplus (x > 1

2 W ). Then i should be willing to pay some positive amount for an instrument to mitigate hervulnerability to j – or, put differently, for some mechanism that constrains j’s behavior.7

In particular, suppose i can pay some amount, p, to ensure an equal division of W . Doing so eliminates j’s agency overW , and guarantees the final payments to be πi = 1

2 W − p and π j = 12 W . The maximum price that i is willing to pay to

avoid vulnerability to j, p∗ , measures i’s distrust towards j. If i fully trusts j then p∗ = 0, while if i fully distrusts j thenp∗ = 1

2 W . Our game elicits, in an incentive-compatible manner, this maximum price, which we interpret as a measure ofdistrust.

We believe this alternative approach is valuable for a couple of reasons. First, it corresponds to how one frequentlyobserves trust/distrust outside the laboratory, with economic agents incurring costs to mitigate their vulnerability to others.Such costs are often identifiable and measurable – managers’ costly monitoring of employees, alliance partners’ willingnessto pay lawyers to write detailed contracts, and suspicious spouses’ willingness to pay for surveillance are all quantifiablemeasures of distrust by one party towards specific targets. Second, our method for measuring (dis)trust potentially avoidsimportant problems with the investment game. For instance, in the investment game, the act of trusting involves “giving”to someone else. If individuals derive utility from the mere act of giving to others (Andreoni 1989, 1990), then behavior asmeasured by the investment game may confound two motives (see Cox, 2004). Conversely, the behavior in our measure –indicating a willingness to pay to mitigate vulnerability to another – is fundamentally distinct from the act of giving andmay therefore exhibit less correlation with altruism. Moreover, in the investment game, distrustful behavior (investing zero)penalizes trustworthy and untrustworthy targets alike. However, in our game the act of distrust only penalizes the otherplayer if he intended to appropriate more than the fair share. Therefore, we propose that our measure possesses featuresthat make it a potentially valuable complement to the investment game in trust research.

2.3. Our experiment

Our experiment creates the exchange above, with respect to varying targets, j, and then elicits from player i the maxi-mum price she is willing to pay to mitigate her vulnerability to the other player, or p∗ . Our hypothesis is that this procedurewill accurately measure distrust, evidenced both by a strong (negative) correlation with attitudinal measures of trust whenholding the target ( j) fixed and by weak correlations with other behaviors such as altruism and risk aversion.

We used the above game with an endowment (W ) of $10. Player i was always a laboratory subject. To explore theimportance of the target of trust in obtaining correlation between attitudinal and behavioral trust measures, we variedthe role of j across several non-laboratory participants and elicited reservation prices from each laboratory subject withrespect to each target (p∗

i j). We obtained targets with varying characteristics by pairing laboratory subjects with randomly-selected field participants passing by several locations around the city of Pittsburgh. Each laboratory subject in the role ofPlayer 1 (i) played the game five times independently for payment, each time with a new random field participant in therole of Player 2 ( j) from one of five specified locations. Other research demonstrates that subjects playing the investmentgame discriminate in their trusting behavior between people at different locations in a city, in a manner consistent withthe targets’ actual trustworthiness (Falk and Zehnder, 2007).8 Therefore, we expect subjects to exhibit different degrees ofbehavioral distrust towards the varying targets, and we explore the extent to which attitudinal measures of trust explainthis variation in behavior.

We selected five locations that varied considerably in population characteristics (socioeconomic status, ethnicity, educa-tion). To measure these characteristics and ensure that we obtained variability in the targets, field participants in the roleof Player 2 completed a questionnaire eliciting several demographic variables. We also measured the perceptions of thesecharacteristics by laboratory subjects in the role of Player 1, to assess the accuracy of their expectations about the target andto explore the determinants of trust. Finally, to compare our behavioral distrust measure to attitudinal measures, as well as

6 They could in principle agree to any other division, against which the trustworthiness of j can be evaluated. For example, the two agents may agreethat j will divide the surplus according to the proportion of costly inputs borne by each player (Yaari and Bar-Hillel, 1984; Konow, 2000; Tungodden etal., 2007). For our purposes, it is sufficient that there is some such agreed upon division that corresponds to the amount i would expect from a fullytrustworthy individual j, and the 50–50 split is often focal in these contexts (Andreoni and Bernheim, 2009).

7 Falk and Kosfeld (2006) demonstrated that a principal exerting limited control over an agent’s actions can negatively influence the degree of generosityexhibited by the agent. This differs from our game, where the only kind of control that can be exerted fully determines the second player’s choices.Moreover, while we focus on the first mover’s willingness to pay to constrain the second mover, Falk and Kosfeld focus on the response by the secondplayer.

8 We developed our methodology of pairing laboratory subjects with people from different parts of a city independently of Falk and Zehnder, and learnedof their work after conducting our experiment at roughly the same time. While this aspect of the methodology is similar, the two papers are motivated byentirely different research questions and yield unrelated primary conclusions. Our research focuses on the relationship between behavioral and attitudinalmeasures when accounting for the target of trust, and we use this “citywide” design to create the variability in targets necessary to test this relationship.In addition, we focus more on specific characteristics of the individual target – rather than of the general population at the target location – for instance,by measuring laboratory subjects’ beliefs regarding the specific characteristics of the target individual and using those expected characteristics to predictdistrust towards the target.

Page 5: Whom do you distrust and how much does it cost? An experiment on the measurement of trust

B. McEvily et al. / Games and Economic Behavior 74 (2012) 285–298 289

to behaviors unrelated to trust, laboratory subjects completed other tasks measuring risk seeking, target-specific altruism,and trust attitudes.9

3. Experimental design

Our experiment paired laboratory subjects – students and staff recruited from the University of Pittsburgh’s ExperimentalEconomics Laboratory subject pool – with field participants. Each laboratory subject participated in two separate sessions,about three to four weeks apart, and received most of their payment at the end of the second session. To recruit fieldparticipants we went to five different locations in the city of Pittsburgh and asked people passing by, one at a time, to stopand participate in a brief experiment for pay.

3.1. Laboratory subjects – Session 1

Upon arriving, laboratory subjects received a private participant number to be used in both sessions and instructions(see Appendix A), which were handed to them and also read aloud.

3.1.1. The behavioral distrust measureThe instructions stated that laboratory subjects (Player 1) would play a game in which they would be paired with

someone (Player 2) from a location in the city of Pittsburgh that would be specified later. The laboratory subjects also sawthe instructions that were to be shown to field participants as part of the game.

The field participant would be told that he or she was matched with a laboratory subject in an experiment at theUniversity of Pittsburgh, and that the pair had been allocated a combined $10, of which $5 was designated for each person.The field participants would then specify how much of the total wealth, including the laboratory subjects’ $5, he or shewould like to keep ($10 � x � $5). The field participant would receive either this amount or $5, determined by whetherthe paired laboratory subject had paid to bind the field participant’s behavior (a decision unknown to the field participantwhen selecting an amount to keep).

Laboratory subjects could specify how much they were willing to pay out of their own $5 share to not play the gameand instead guarantee the remainder of their $5. That is, they could pay an amount between $0 and $5 to eliminate thefield participant’s discretion over the division of the endowment. Laboratory subjects recorded their choices on a “DecisionSheet” (see Appendix A) containing 51 rows, each corresponding to an amount $0.00 � p � $5.00, in ten cent increments.For each row, laboratory subjects specified whether they would prefer to pay the corresponding amount or “play the game”by selecting one of two boxes. One of the rows would be selected at random and the subject’s choice on that row woulddetermine whether the field participant would have the opportunity to keep more than $5. More precisely, if a row corre-sponding to an amount p was selected, then if Player 1 selected to pay that amount the payoffs would be π1 = $5 − p andπ2 = $5, while if for that amount Player 1 selected to play the game then the payoffs would be determined by Player 2’schoice of x (π1 = $10 − x and π2 = x). The row on the Decision Sheet at which a laboratory subject switches from one col-umn to the other indicates the most she is willing to pay to avoid vulnerability to that target.10 Regardless of the outcome,laboratory subjects received their earnings from the game at the second session.

To facilitate comprehension, laboratory subjects first completed a practice Decision Sheet, for which they were pairedwith a hypothetical field participant. An experimenter publicly drew a number from 1 to 51 to select a row (price) andexplained publicly what would happen for each of the two possible choices for that row based on the behavior of thehypothetical passerby. Next, laboratory subjects completed a quiz to ensure understanding of the game and procedures.

Laboratory subjects then played five rounds of the game, corresponding to each of the five locations in the city ofPittsburgh. The top of each Decision Sheet included a written description of the location (the name of the neighborhoodand an intersection for two major roads) and the back included both a map of Pittsburgh with an arrow pointing to thelocation and a picture of the location. We privately counterbalanced the order of locations between subjects. Laboratorysubjects were not told in advance how many times they would play the game.

3.1.2. LocationsThe five locations consisted of two universities (the student centers at the University of Pittsburgh and Carnegie Mel-

lon), one upper-income neighborhood (Shadyside), one middle-income neighborhood (the Southside), and one lower-incomeneighborhood (the Hill District). The three specific non-university locations were commercial areas, but all three were withina block of residential areas. Appendix B, Table B1 provides summaries of 2000 U.S. Census demographic data for each non-university location.

9 By allowing several weeks to pass between the collection of the behavioral and attitudinal measures, we employ a similar method to that used byGlaeser et al. (2000).10 This procedure is structurally identical to the Becker et al. (1964) lottery-elicitation procedure for incentive-compatible valuation. A subject (i) deciding

whether to trust a target ( j) should identify the maximum price he or she is willing to pay to not play the game, p∗i j , by selecting to pay at any value

p � p∗i j and selecting to play the game for any value p > p∗

i j .

Page 6: Whom do you distrust and how much does it cost? An experiment on the measurement of trust

290 B. McEvily et al. / Games and Economic Behavior 74 (2012) 285–298

3.1.3. Risk seekingAfter playing the game five times, laboratory subjects completed a behavioral task measuring risk attitudes. We used the

same task as Schechter (2007), in which subjects bet on the outcome of a die roll, because she found a relationship betweenrisk seeking and trust measured by the investment game. Appendix A includes the instructions for this task.

3.1.4. AltruismTo measure altruism or interpersonal concern toward each target, laboratory subjects played a $10 dictator game (see

Hoffman et al., 1994) with a new field participant (someone who had not played the distrust game) from each of the fivelocations. Subjects were informed that one of these locations would be selected at random at the end of the session. Forthe selected location, the experimenters would recruit a new person passing by and would give this person the allocatedportion of the $10. Any portion of the $10 not allocated to this second field participant would be paid to the laboratorysubject at the second session.

3.1.5. Prediction of targets’ demographic characteristicsFor the final task in the first session, laboratory subjects received five copies, one per location, of a one-page demographic

questionnaire that was to be completed by the same field participant at each location with whom they would be paired forthe distrust game. Laboratory subjects completed each questionnaire, attempting to match the responses of the paired fieldparticipant at the specified location. At the end of the session, one location and one question would be selected at random.If the laboratory subject matched the response on this item to that of the paired field participant, the laboratory subjectwould receive an additional $5 in the second session.

3.1.6. End of first sessionOnce laboratory subjects completed all the above tasks, an experimenter conducted several random draws to determine

outcomes for the various measures. Following public announcement of the random draws, laboratory subjects were paid a$6 participation fee and their earnings from the risk seeking task.

3.2. Laboratory subjects – Session 2

After approximately three weeks, we brought laboratory subjects back to complete additional questionnaires and re-ceive their earnings from the five distrust games, the dictator game, and any bonus for correctly predicted demographiccharacteristics.

3.2.1. Attitudinal trust measuresThe first part of the questionnaires asked laboratory subjects to indicate the extent to which they agreed or disagreed

with several statements reflecting trust attitudes towards people at each of the five locations from which we selected fieldparticipants. Each sheet corresponded to a particular location, which was clearly indicated at the top of the sheet. For eachlocation, we asked three sets of questions, which varied in how specific they were regarding the precise target and thecontext of the distrust game.

The first set of questions dealt with trust towards the specific target in the context of the game from Session 1. Laboratorysubjects rated agreement with five statements about the field participant with whom they were paired at the specifiedlocation (e.g., “I believe that this person would take advantage of my vulnerability.”). The statements are representative ofthe kind frequently used by social psychologists, sociologists, and organizational researchers to measure trust (Cummingsand Bromiley, 1996; Mayer and Davis, 1999).

The next set of questions dealt with trust towards related targets in a context similar to the game. Laboratory subjects wereasked to consider two scenarios in which they had to rely on people at the specified locations (specifically, because theyhad left their wallet containing $60 in a store or they had mailed a check for the purchase of sold out event tickets). Theythen rated agreement with statements concerning their perceived vulnerability to individuals at the specified locations.

Finally, we asked one question regarding the general trustworthiness of individuals at the particular location. The questionread “I believe that people at [location] are generally trustworthy.”

We expect the behavioral measure of distrust towards people at a particular location to be negatively related to the threeattitudinal trust measures for that location. Also, since we vary the degree to which the attitudinal measures are target andcontext specific, we predict that the correlation will be strongest for the attitudinal measure that deals with trust in thespecific target/context and weakest for the attitudinal measure that deals with general trustworthiness.

3.2.2. Familiarity with locationWe elicited laboratory subjects’ familiarity with each location. Subjects stated how familiar they were with the location,

how often they visited the location, and the number of people from the location with whom they regularly interacted.

3.2.3. Demographics and general backgroundIn the final part of the session, subjects provided their own demographic information (identical to the items asked of

field participants). We also asked several questions related to subjects’ general background and experiences. This includedone measure of general trust in strangers.

Page 7: Whom do you distrust and how much does it cost? An experiment on the measurement of trust

B. McEvily et al. / Games and Economic Behavior 74 (2012) 285–298 291

3.3. Field participants

Before the second laboratory session, the experimenters went to each of the five locations. Upon encountering a passerby,the experimenter identified himself as a university researcher and asked if the individual would participate in a briefexperiment, lasting about two minutes, in return for at least $5 and possibly more.

3.3.1. Behavioral untrustworthiness measureWe handed field participants a brief instruction sheet which the experimenter also read aloud.11 Field participants were

told that they were paired with a laboratory subject at the University of Pittsburgh, that the pair had been allocated $10,and that the field participant would receive his or her $5 and could also claim up to $5 of the laboratory subject’s allocation.The experimenter then handed field participants a sealed envelope and stated that the field participant might receive allor a portion of the laboratory subject’s $5 they claimed, depending on the contents of the envelope. Field participants thenindicated the amount they wished to keep ($5 � x � $10) on a decision sheet. They then found out whether they wouldreceive the portion of the laboratory subject’s $5 they claimed by opening the envelope and reading whether or not theirpaired laboratory subject had opted to play the game.

3.3.2. Demographic questionnaireWhile the experimenter prepared the payment, field participants completed a one-page questionnaire with five multiple-

choice demographic questions (age, gender, ethnicity, education, and socio-economic status). These were the same questionsthat laboratory participants had tried to predict in their first session.

3.3.3. Dictator allocationsDepending on the outcome of the random draws in prior laboratory sessions, we also distributed dictator game earnings

at the relevant locations to a different set of field participants than those participating in the distrust game. This occurredonly after collecting distrust game and demographic data at that location.

4. Results

We obtained data from 60 laboratory subjects and 300 field participants (60 at each location).12 Laboratory subjectsearned $43.51, on average, across both sessions. Table B2, in Appendix B, provides self-reported demographic information onlaboratory and field participants at each location, as well as laboratory subjects’ predictions of field participants’ responsesto the same demographic questions. Appendix C provides a list of the main variables used in the analyses along with adefinition and summary statistics for each variable.

Our primary dependent variable, distrust, consists of the price at which a laboratory subject switched from preferring topay to avoid the game to playing the game. In the majority of cases, laboratory subjects switched from the column “pay,don’t play” to the column “don’t pay, play” only once. In these cases the amount corresponding to the row at which alaboratory subject switched to the “don’t pay, play” column was that subject’s threshold price for that location (p∗

i j).In a few cases, laboratory subjects indicated inconsistent responses by changing between columns more than once on

a Decision Sheet. Out of 300 completed Decision Sheets (by 60 laboratory subjects), this occurred a total of 10 times (by4 different laboratory subjects). Responses of this sort do not provide us with a clear threshold price. Nevertheless, wecan still use information on the Decision Sheet to infer a roughly equivalent statistic measuring how many times, and forwhat values, the laboratory subject selected boxes in each column.13 This statistic tells us something about how much thatlaboratory subject is generally willing to pay to avoid vulnerability to the field participant and is comparable to the thresholdprice used for the rest of the analysis.

Fig. 1 shows the distributions of maximum buy-out prices, in the form of cumulative densities, separately for each lo-cation. They range from $0 (complete trust, five instances) to $5 (complete distrust, two instances), with a mean of $2.61,median of $2.60, and standard deviation of $0.97. While there is a central tendency in the data – the lines all increasesubstantially at $2.50 – the majority of responses (51 percent) lie outside of the interval [$2, $3], revealing substantial

11 The instructions given to field participants were very simple and concise – appearing on a single page (see Appendix A). We therefore felt it unnecessaryto add a comprehension quiz. After reading the instructions, participants were prompted to ask any clarifying questions. In conducting the experiment, wefound that field participants had little difficulty in understanding their instructions.12 We also distributed 60 dictator recipient envelopes, but collected no data from these field recipients.13 More precisely, we represent subject i’s choice to either pay (1) or not pay (0) in each of the 51 rows of the Decision Sheet by xik ∈ {0,1}, where

k = $0.00,$0.10, . . . ,$5.00. Then, we compute the following statistic as an imputed measure of the cutoff, based on the number of choices in each column:

p̂∗i j = $0.10 + 2

(∑k

xikk

)/∑

k

xik.

For a subject choosing “consistently” (only switching between columns once), this formula returns the first value in the “don’t pay, play” column. We alsoconducted our analysis excluding the individuals who provided at least one inconsistent set of responses. The results are generally unchanged by omittingthese four individuals. We discuss specifically how they change in instances where this is the case.

Page 8: Whom do you distrust and how much does it cost? An experiment on the measurement of trust

292 B. McEvily et al. / Games and Economic Behavior 74 (2012) 285–298

Fig. 1. Distributions of buy-out prices (distrust) by location.

dispersion. The lines also reveal some differences in distrust across locations: buy-out prices for some locations, e.g., Shady-side, tend to be considerably lower than for others, e.g., the Hill District.14 We return to a more thorough exploration ofdifferences in distrust towards different locations in Section 4.2.

4.1. Individual-level differences among laboratory subjects

To illustrate the value of measuring trust with respect to a specific target we first briefly explore the data at the individualtrustor level (i.e., average maximum buy-out prices for laboratory subjects) and then disaggregate the data by target. Foreach laboratory subject we compute the average maximum price across all five locations (p̂∗

i = (∑5

j=1 p∗i j)/5). We can

interpret this as an individual’s propensity to distrust passersby at public locations throughout the city of Pittsburgh. Thispropensity ranges from highly trusting ($0.51) to highly distrusting ($4.68), with a mean of $2.61 and standard deviation of$0.84.

Table 1 reports regressions that explore the relationship between individual-level distrust and three individual-level be-haviors or attitudes. First, we construct a broad measure of general individual-level trust attitudes, Overall trust attitude,using all of a subject’s responses to the attitudinal trust questions.15 Average altruism is the mean amount that a labora-tory subject shared in the five dictator games. Risk attitude is the amount bet by a laboratory subject on the die roll. Forthese regressions, we standardize all three measures by subtracting the mean and dividing by the standard deviation. Weinclude as controls four demographic characteristics of laboratory subjects that, in separate analyses, we found to predict apropensity to distrust.

A few results emerge from Table 1.16 First, we find little relationship between risk attitude and distrust, either in model 3or in model 4, which also simultaneously accounts for the influence of altruism and overall trust attitude on distrust. Second,altruism and overall trust attitude have very similar relationships with distrust. The coefficients indicate less behavioraldistrust when a subject exhibits greater attitudinal trust or greater altruism, by a roughly $0.20 lower maximum buy-outprice for each standard deviation of either variable (models 1 and 2), but the effects are both statistically modest (bothcoefficients are significant at p = 0.06). When we estimate the joint effects of all three measures in model 4, neither Overalltrust attitudes nor Average altruism achieve statistical significance.17

14 Applying a non-parametric Kolmogorov–Smirnov test to the distributions in Fig. 1, the difference in distributions between Hill District and Shadyside isstatistically significant (p = 0.005). The distributions for all locations, except for the University of Pittsburgh, are at least marginally significantly different(p < 0.1) from the distribution for the Hill District. The distribution for University of Pittsburgh is marginally significantly different from that for Shadyside.15 Recall that for each of the five locations, subjects provided 10 attitudinal responses concerning their beliefs and expectations regarding trustworthiness

of people at that location. We first average an individual’s responses, across the five locations, separately for each of these questions. We then add togetherall of the 10 averaged responses and also the subject’s agreement with the statement, “You can’t count on strangers anymore,” (we reversed the sign of thethree items for which agreement with a statement reflects distrust).16 Excluding the four individuals who provided inconsistent responses increases the significance and magnitude of the coefficient for Age in all models,

slightly increases the magnitude and the significance of the coefficient for Overall trust attitude (p = 0.04) in model 1, and slightly decreases the magnitudeand significance of the coefficient for Altruism (p = 0.12) in model 2.17 The joint effects may be limited by the high correlation between Overall trust attitude and Average altruism (0.44, p < 0.001), suggesting a relationship

between a general attitudinal propensity to trust and behavioral altruism. A test of the restriction that the coefficients for Overall trust attitude and Averagealtruism both equal zero in model 4 is marginally rejected (p < 0.1).

Page 9: Whom do you distrust and how much does it cost? An experiment on the measurement of trust

B. McEvily et al. / Games and Economic Behavior 74 (2012) 285–298 293

Table 1Factors influencing individual-level distrust (OLS).

Dependent variable: average maximum buy-out price (distrust)

(1) (2) (3) (4)

Attitudes / behaviorsOverall trust attitude (standardized) −0.19* −0.13

(0.10) (0.11)

Average altruism (standardized) −0.20* −0.14(0.10) (0.12)

Risk attitude (standardized) 0.05 0.01(0.10) (0.10)

Constant 3.01*** 2.97*** 3.07*** 2.95***

(0.17) (0.17) (0.17) (0.18)

Own demographicsAge (above 24) −0.51 −0.32 −0.53 −0.39

(0.31) (0.32) (0.33) (0.34)

Female −0.62*** −0.58*** −0.70*** −0.55**

(0.21) (0.21) (0.21) (0.21)

Race (non-white and non-black) −0.95*** −0.78*** −0.92*** −0.84***

(0.26) (0.28) (0.27) (0.28)

Education (college degree) 0.79*** 0.60** 0.75** 0.69**

(0.29) (0.29) (0.31) (0.31)

R2 0.36 0.36 0.32 0.38Obs. 60 60 60 60

Note: Standard errors in parentheses.* p < 0.1.

** p < 0.05.*** p < 0.01.

Table 2Distrust, altruism, attitudes, and untrustworthiness by location.

Targetlocation

Player 1(laboratory subject)

Player 2(field participant)

Avg. buy-outprice(distrust)

Avg. amountshared(altruism)

Average attitudinal trust (standardized) Avg. excessamount kept(untrustworthiness)

Specifictarget/context

Relatedtarget/context

Generaltrust-worthiness

(a) U of Pitt. $2.75 $1.64 0.13 0.50 0.48 $1.89(b) CMU $2.61 $1.23 0.31 0.58 0.53 $1.90(c) Shadyside $2.35 $1.08 0.28 0.27 0.23 $1.28(d) Southside $2.54 $1.48 −0.16 −0.35 −0.34 $1.62(e) Hill District $2.80 $2.29 −0.55 −1.00 −0.90 $3.38

Wilcoxon Matched-pairs sign-rank Rank-sump < 0.10 a–e a–c a-b a–c, b–cp < 0.05 a–d, b–e b–d a–b b–cp < 0.01 c–d a–c, a–e, c–d a–d a–c, b–cp < 0.001 a–c, b–c, c–e, d–e a–b, b–e, c–e, d–e a–e, b–d, b–e, c–d,

c–e, d–ea–d, a–e, b–d, b–e,c–d, c–e, d–e

a–d, a–e, b–d, b–e,c–d, c–e, d–e

a–e, b–e, c–e, d–e

4.2. Determinants of target-specific trust

We now consider how behavioral distrust, trust attitudes, and altruism vary across different targets, which is the primaryfocus of our experiment. Table 2 presents summary statistics, by location, of behaviors and attitudes from both laboratorysubjects and field participants. The variables in the table are the average buy-out price (our behavioral measure of distrustby the laboratory subject towards the field participant), amount shared in the dictator game (altruism by the laboratorysubject towards a different field participant), the three attitudinal measures of trust by the laboratory subject towardspeople at that location, and average amount kept in excess of $5 (untrustworthiness) by the field participant. The bottom ofthe table reports non-parametric (Wilcoxon) tests of pair-wise differences between locations.

Average buy-out prices (distrust) generally exhibit the same pattern as average amount kept (untrustworthiness). HillDistrict field participants were generally both the most distrusted and the most untrustworthy, while those in Shadysidewere generally the least distrusted and the least untrustworthy. However, average buy-out prices vary less by location thanexcess amounts kept.

On average, if laboratory subjects are risk neutral and have rational expectations about the behavior of field participants,then to maximize their earnings they should be willing to pay up to the average excess amount kept. Against this standard,

Page 10: Whom do you distrust and how much does it cost? An experiment on the measurement of trust

294 B. McEvily et al. / Games and Economic Behavior 74 (2012) 285–298

Table 3Relationship between behavioral distrust and attitudes toward target (OLS).

Dependent variable: maximum buy-out price (distrust)

(1) (2) (3) (4) (1a) (2a) (3a) (4a)

Attitudes toward targetTrust – Specific Target/Context(standardized)

−0.23*** −0.21***

(0.07) (0.07)

Trust – Related Target/Context(standardized)

−0.12** −0.12**

(0.05) (0.06)

Trust – General Trustworthiness(standardized)

−0.10** −0.07*

(0.05) (0.04)

Altruism toward Target(standardized)

0.04 0.02(0.09) (0.08)

Target-specific controls (predicted target demographics)Predicted Target Age −0.33*** −0.34*** −0.31*** −0.29***

(above 24) (0.08) (0.08) (0.08) (0.07)

Predicted Target SES 0.27*** 0.30*** 0.35*** 0.41***

(Low & Low-Mid) (0.07) (0.09) (0.07) (0.08)

Predicted Same Age −0.17* −0.13 −0.14 −0.13(0.09) (0.09) (0.09) (0.09)

Predicted Same SES 0.17** 0.19** 0.19** 0.20***

(0.07) (0.07) (0.07) (0.07)

Familiarity with Target 0.02 0.03 0.03 0.02(0.04) (0.04) (0.04) (0.04)

Decision order 0.03 0.03 0.03 0.02 0.02 0.02 0.02 0.02(0.02) (0.02) (0.02) (0.02) (0.02) (0.02) (0.02) (0.02)

Constant 2.53*** 2.53*** 2.53*** 2.54*** 2.65*** 2.63*** 2.60*** 2.56***

(0.07) (0.07) (0.07) (0.07) (0.11) (0.10) (0.12) (0.10)

R2 0.771 0.761 0.759 0.752 0.792 0.784 0.783 0.781(partial r2 for attitude variable) (0.081) (0.037) (0.032) (0.002) (0.052) (0.018) (0.012) (0.000)

Obs. 299 (60) 299 (60) 300 (60) 300 (60) 298 (60) 298 (60) 299 (60) 299 (60)

Note: All regressions include subject fixed effects; robust standard errors clustered by subject (in parentheses).* p < 0.1.

** p < 0.05.*** p < 0.01.

laboratory subjects generally distrusted too much. Mean buy-out prices are usually higher than excess amounts kept. Interms of expected value, laboratory subjects paid $0.60 per location ($3.00 in total across locations) too much to distrust($2.61 average buy-out price across locations vs. $2.01 average excess amount kept). For the first four locations in Table 2,they paid on average $0.89 too much, while for the Hill District they paid $0.58 too little.18

Comparing across locations (as opposed to across individual laboratory subjects as in Table 1), the average amount sharedin the dictator game (altruism) is generally positively related to distrust (the correlation between the first two columns inTable 2 is 0.83). That is, the populations generally distrusted (i.e., for which there are high buy-out prices) are also the oneswith which the most is shared. This contrasts with the investment game, where high “trust” generally corresponds to highaltruism (e.g., Cox, 2004; Ashraf et al., 2006).

The attitudinal measures exhibit similar patterns to buy-out prices. For example, the least trusted group for all attitudinalmeasures is the Hill District, which is also the group most distrusted and most untrustworthy in the context of the game.Shadyside is generally trusted attitudinally, mirroring low distrust and low untrustworthiness in the game. However, thereare also differences. Southside people are generally not trusted in the attitudinal measures but are distrusted relatively littlein the game. They also prove to be relatively trustworthy in the game.19

To more closely examine the relationship between these behaviors and attitudes, we estimate regressions in which alaboratory subject’s maximum buy-out price for a location (in contrast to the average buy-out price across locations used inTable 1) is the dependent variable. These regressions provide target-specific analyses of the determinants of trust. Since the

18 Comparing the distributions of buy-out prices and excess amounts kept separately for each location (using a rank-sum test), we reject equality ofdistributions for every location. If we use the average excess amount kept at a location as a standard against which to compare whether particularlaboratory subjects distrusted too much or too little, individual subjects generally distrusted too much (222 of 300 buy-out prices are higher than thecorresponding average excess amount kept for that location). The same pattern holds separately for all locations (between 49 and 55 out of 60) except forthe Hill District, where subjects distrusted too little (only 16 of 60 buy-out prices are higher than $3.38).19 There is also variation between the different attitudinal measures. For instance, people at the University of Pittsburgh are not considered particularly

trustworthy in the specific context of the game (the average rating of 0.13 is significantly lower than the ratings for both CMU and Shadyside). However, inrelated contexts they are considered much more trustworthy.

Page 11: Whom do you distrust and how much does it cost? An experiment on the measurement of trust

B. McEvily et al. / Games and Economic Behavior 74 (2012) 285–298 295

models include repeated observations for each laboratory subject, we include subject fixed effects,20 thereby also controllingfor the individual characteristics in Table 1. We report these regressions in Table 3.21

Each regression in Table 3 explores the relationship between some attitude held by a laboratory subject toward thefield participant and the laboratory subject’s distrust toward the field participant. The first four models report the directrelationship, controlling only for possible effects in decision order, while the latter four models include additional target-specific controls, which we describe below.22 In addition to reporting the R2, for each model we also present the squaredpartial coefficient of correlation for the attitudinal variable, which provides a measure of the proportion of variance in thedependent variable explained by the attitudinal measure, after eliminating the effect of all the other variables in the model(including the subject-specific fixed effect).23

The controls are primarily based on laboratory subjects’ predictions of field participants’ demographic characteristics,which were generally accurate (see Table B2 in Appendix B).24 In separate analyses using similar regressions, which wedo not report here, we explored the relationship between all of the predicted target demographic characteristics (includingmeasures of the laboratory subjects’ predicted similarity with the field participant) and distrust. Across several regressions,only the variables included as controls in Table 3 were ever statistically significant at p < 0.1 or lower. We also includefamiliarity with the location as a control variable.25

Recall that the first attitudinal trust measure (specific target/context) dealt with trust towards the specific target in thecontext of the game, the second measure (related target/context) dealt with trust in similar hypothetical contexts towardsrelated targets, and the final measure (general trustworthiness) dealt with overall perceptions of the trustworthiness of peopleat that location. All three measures are significantly (negatively) related to behavioral distrust in the game, indicating greaterdistrust when laboratory subjects have low (attitudinal) trust toward the target. As we predicted, the relationship is strongestin both magnitude and significance for specific, less so for related, and least so for general. All three measures continue to benegatively related to distrust (with comparable magnitudes) when we include control variables.26

Model 4(a) explores the relationship between target-specific altruism and distrust. The additional explanatory variable isthe (standardized) amount that a laboratory subject shared in the dictator game with a different person at that location.There is a very small, statistically insignificant, positive relationship between distrust towards a field participant and theamount shared with a similar person from the same location. Therefore, while in Table 1 we found a relationship betweensubjects’ overall propensities to share (altruism) and distrust when pooling across locations, we find virtually no relationshipbetween altruism and distrust towards specific targets.27 This finding is a particularly vivid illustration of how analysis atthe level of the target of trust can materially alter the findings.

To summarize the main results from Table 3, we find that attitudinal measures of trust are strong predictors of distrusttoward a specific target. We also find that the specificity of the attitudinal measure toward the specific target affects thestrength of this relationship.28 However, we find little relationship between target-specific altruism and distrust as measuredusing our game.

Table 4 explores the relationship between the distrust for targets at specific locations and the average excess amountkept at those locations (i.e., average untrustworthiness at that particular location, from Table 2). In the first regression, thereis a significant relationship between distrust and actual mean target untrustworthiness, but the coefficient is less than 1. Thisconfirms our earlier finding, from a visual inspection of Table 2, that laboratory subjects’ distrust discriminated based on

20 The fixed effects account for a significant proportion of the variance (between 70 and 74 percent) in all models.21 One participant failed to answer the predicted demographics for one location (University of Pittsburgh). Therefore, we have 299 predicted demographics

questionnaires completed by 60 laboratory participants. Another participant did not complete one of the attitudinal questionnaires. This means that wehave complete data (behavioral and attitudinal measures) for 298 targets, by 60 laboratory participants.22 Across the five decisions, average buy-out prices range from $2.56 (Decision 3) to $2.69 (Decision 5). The differences are never statistically significant

in pairwise tests. Aside from decision-order effects, we also tested for session-specific effects on distrust by running a regression using distrust (buy-outprice) as the dependent variable and session-specific binary variables as explanatory variables (clustering standard errors by subject). We fail to reject therestriction that all of the binary variables are jointly equal to zero (F (6,59) = 1.68).23 More precisely, this measure is the squared correlation between the residuals of separate regressions of behavioral distrust and the relevant attitudinal

measure on the remaining variables in the model.24 Predicted Target Age is coded as 0 for the “18–24” category and 1 otherwise. Predicted Target Socioeconomic Status (SES) is included as a binary variable

indicating a response of “Lower”/“Lower-Middle.” Predicted Same Age and Predicted Same SES each indicate that the laboratory subject predicted the target tohave the same demographic characteristics as him or herself.25 For each location, we summed the three questionnaire responses measuring familiarity and then standardized the resulting variable to construct a

measure of Familiarity with Target. The results do not change if we instead use the disaggregated familiarity variables as controls.26 Excluding the four individuals who provided inconsistent responses slightly weakens the statistical significance of the coefficient for General Trustwor-

thiness in model 3a.27 One possible explanation for the null relationship between distrust and altruism in Table 3 is that the negative relationship exists only for some targets,

specifically those with (predicted) low SES. To explore this possibility, we added an interaction term between Altruism and Predicted Target SES (Low &Low-Mid) to model 4a in Table 3. Neither the coefficient for Altruism nor for the interaction is significant in this model, suggesting that this is not anadequate explanation.28 We also estimated models in which we simultaneously include all four of the attitudinal measures in Table 3 as explanatory variables, both with and

without controls. In both models, the only attitudinal measure that has a statistically significant coefficient is Specific Target/Context, which remains highlystatistically significant (p < 0.01 in both models). The fact that the other two coefficients lose statistical significance is perhaps not surprising, as the threeattitudinal trust measures are correlated (r > 0.36 in all pair-wise comparisons). A test of the restriction that the coefficients for all three attitudinal trustmeasures equal zero is strongly rejected (F (3,228) = 4.70, p < 0.005).

Page 12: Whom do you distrust and how much does it cost? An experiment on the measurement of trust

296 B. McEvily et al. / Games and Economic Behavior 74 (2012) 285–298

Table 4Relationship between distrust toward targets and target untrustworthiness (OLS).

Dependent variable: maximum buy-out price (distrust)

(1) (2) (3) (4) (5)

Mean target untrustworthinessAvg. Excess Amount Kept at location 0.18*** 0.13 0.11 0.11 0.13

(0.05) (0.09) (0.08) (0.08) (0.09)

Trust attitudes toward targetSpecific Target/Context (standardized) −0.20***

(0.07)

Related Target/Context (standardized) −0.11*

(0.05)

General Trustworthiness (standardized) −0.07*

(0.04)

Target demographics: predictedPredicted Target Age (above 24) −0.25*** −0.30*** −0.30*** −0.28***

(0.09) (0.09) (0.09) (0.09)

Predicted Target SES (Low & Low-Mid) 0.23 0.12 0.15 0.17(0.15) (0.13) (0.16) (0.13)

Predicted Same Age −0.13 −0.17* −0.14 −0.14(0.09) (0.09) (0.09) (0.09)

Predicted Same SES 0.19*** 0.17** 0.19** 0.19**

(0.07) (0.07) (0.07) (0.07)

Familiarity with Target 0.02 0.02 0.04 0.04(0.04) (0.04) (0.04) (0.04)

Decision order 0.03 0.02 0.02 0.02 0.02(0.02) (0.02) (0.02) (0.02) (0.02)

Constant 2.17*** 2.33*** 2.46*** 2.42*** 2.38***

(0.12) (0.22) (0.22) (0.20) (0.23)

R2 0.769 0.784 0.794 0.787 0.786Obs. 300 (60) 299 (60) 298 (60) 298 (60) 299 (60)

Note: All regressions include subject fixed effects; robust standard errors clustered by subject (in parentheses).* p < 0.1.

** p < 0.05.*** p < 0.01.

the actual untrustworthiness of target field participants, but that buy-out prices did not fully respond to variation in (mean)actual untrustworthiness.

The relationship between distrust and untrustworthiness becomes smaller and statistically insignificant, after we in-troduce the predicted demographic variables that we found to be significant predictors of behavioral distrust in Table 3(model 2). Thus, while the actual untrustworthiness of field participants forms a basis for the behavioral distrust by labora-tory subjects, most of this relationship can be accounted for by the laboratory subjects’ perceptions of targets’ characteristics.

The remaining regressions in Table 4 add the three target-specific attitudinal variables. All three attitudinal variablespredict behavioral distrust, beyond what is explained by the predicted demographic variables and actual average untrust-worthiness. Thus, there is a significant component of our behavioral distrust measure that is related to attitudinal trustmeasures, even after controlling for important perceived and actual target characteristics.

4.3. Determinants of untrustworthiness

Even though our focus is on the trust/distrust exhibited by the laboratory subjects towards the target field participants,we also analyzed the relationship between the field participants’ demographic characteristics and the amount of money theykept in excess of $5. We find two factors that reliably predict untrustworthiness across locations: older people and femalesare generally less untrustworthy (more trustworthy). Note that one of these relationships is consistent with laboratorysubjects’ expectations, where we found that laboratory subjects generally exhibited less distrust towards older individuals(see Table 3).

5. Conclusion

Although trust has long been a concept of interest in the social sciences, quantifying and measuring such an elusivephenomenon has proven challenging. Different methods and approaches are used across disciplines, and previous researchoften identifies problems with existing measures. Moreover, comparisons of the behavioral approaches that dominate eco-nomics with attitudinal approaches favored in other social sciences often find little relationship. We contribute to this effortby showing a relationship between attitudinal and behavioral measures, and by proposing a new behavioral paradigm for

Page 13: Whom do you distrust and how much does it cost? An experiment on the measurement of trust

B. McEvily et al. / Games and Economic Behavior 74 (2012) 285–298 297

measuring trust that is based on measuring distrust, or individuals’ willingness to incur costs to mitigate vulnerability toothers.

We believe this study advances the measurement of trust in economic research in two ways. First, by explicitly account-ing for the target we draw attention to an important, but often ignored, factor in the measurement of trust. When we holdconstant the target of trust, we observe a strong relationship between our behavioral and attitudinal measures. In contrast,when we aggregate across targets (i.e., by calculating each laboratory subjects’ average trust across all five locations), weobserve a weaker relationship between behavioral and attitudinal measures of trust. These findings underscore the impor-tance of identifying the target of trust. Our results are consistent with the view, widely held in other social sciences, thattrust is often best conceived as a property of specific trustor–trustee interactions, as opposed to a stable, relatively invarianttrait of the person placing trust.

Second, we introduce a novel behavioral method for measuring trust, which focuses on measuring distrust, rather thantrust, by observing the costs individuals are willing to bear to eliminate their vulnerability to a counterpart. Our experimentreveals this measure to have no relationship with risk attitudes, and very little relationship with altruism (i.e., a modestrelationship as a subject-specific trait and none across specific targets). Moreover, our new measure of trust demonstratespredictive validity – laboratory subjects’ distrust towards field participants accurately discriminates untrustworthiness, basedon perceived characteristics of targets. The emphasis on paying to avoid being vulnerable to another has the further advan-tage of external applicability; the principle underlying our laboratory procedure could also be used to measure distrust inother, non-laboratory settings – for example by measuring how much individuals spend to safeguard themselves againstothers’ moral hazard.

It is important to be cautious, however, in comparing our method to the widely-used investment game. At this point weare unable to assess the relative superiority of our measure of distrust and the investment game, as such an assessmentwould require a direct comparison between the two methods, using clear standards regarding what they are attemptingto measure. This lies beyond the scope of our paper, but would represent valuable future research. Rather, we view ourcontribution in this respect as introducing an alternative behavioral measure of (dis)trust. We would further suggest thatour new behavioral measure be considered as a complement, rather than a substitute, for the investment game. Accordingly,we recommend that future research identify the kinds of situations for which one or the other measure is better, allowingresearchers to use both kinds of games, depending on the context and research question.

Of course, our primary findings also await further work before more definitive conclusions can be drawn regardingthe relationship between attitudinal and behavioral trust measures. For example, as prior research demonstrates, there areclearly cases in which attitudinal and behavioral measures of trust are unrelated. Here, we motivate and present a case inwhich the two types of trust measures are related. Thus, understanding when different approaches to measuring trust arerelated, and how this interacts with target- and context-specificity, is an important research agenda for developing trust asa useful predictive tool in economics. By providing a means of linking research on trust in economics with related researchin other social science disciplines, we create an opportunity for cumulating and integrating findings to a much greaterextent than has been possible previously. Given the wide-ranging importance of trust to economic, social, organizational,and political activity, and the inherent difficulty of precisely measuring such an elusive concept, we believe that our effortsto advance the measurement of trust are of fundamental importance.

Supplementary material

The online version of this article contains additional supplementary material – Appendices A–C.Please visit doi:10.1016/j.geb.2011.06.011.

References

Andreoni, James, 1989. Giving with impure altruism: Applications to charity and Ricardian equivalence. J. Polit. Economy 97 (6), 1447–1458.Andreoni, James, 1990. Impure altruism and donations to public goods: A theory of warm-glow giving. Econ. J. 100, 464–477.Andreoni, James, Bernheim, B. Douglas, 2009. Social image and the 50–50 norm: A theoretical and experimental analysis of audience effects. Economet-

rica 77 (5), 1607–1636.Arrow, Kenneth J., 1974. The Limits of Organization. Norton, New York.Ashraf, Nava, Bohnet, Iris, Piankov, Nikita, 2006. Decomposing trust and trustworthiness. Exper. Econ. 9, 193–208.Becker, Gordon M., DeGroot, Morris H., Marschak, Jacob, 1964. Measuring utility by a single-response sequential method. Behav. Sci. 9 (3), 226–232.Ben-Ner, Avner, Halldorsson, Freyr, 2010. Trusting and trustworthiness: What are they, how to measure them, and what affects them. J. Econ. Psych. 31 (1),

64–79.Ben-Ner, Avner, Putterman, Louis, 2009. Trust, communication and contracts: An experiment. J. Econ. Behav. Organ. 70, 106–121.Ben-Ner, Avner, McCall, Brian P., Stephane, Massoud, Wang, Hua, 2009. Identity and in-group/out-group differentiation in work and giving behaviors:

Experimental evidence. J. Econ. Behav. Organ. 72 (1), 153–170.Berg, Joyce, Dickhaut, John W., McCabe, Kevin A., 1995. Trust, reciprocity, and social history. Games Econ. Behav. 10 (1), 122–142.Blau, Peter M., 1964. Exchange and Power in Social Life. Wiley, New York.Buchan, Nancy R., Croson, Rachel T., Dawes, Robyn M., 2002. Swift neighbors and persistent strangers: A cross-cultural investigation of trust and reciprocity

in social exchange. Amer. J. Sociology 108 (1), 168–206.Capra, C. Monica, Lanier, Kelli, Meer, Shireen, 2007. Attitudinal and behavioral measures of trust: A new comparison. Working paper.Cox, James C., 2004. How to identify trust and reciprocity. Games Econ. Behav. 46, 260–281.Cummings, L.L., Bromiley, Philip, 1996. The Organizational Trust Inventory (OTI): Development and validation. In: Kramer, Roderick M., Tyler, Tom R. (Eds.),

Trust in Organizations: Frontiers of Theory and Research. Sage, Thousand Oaks, CA, pp. 261–287.

Page 14: Whom do you distrust and how much does it cost? An experiment on the measurement of trust

298 B. McEvily et al. / Games and Economic Behavior 74 (2012) 285–298

Eckel, Catherine C., Wilson, Rick K., 2004. Is trust a risky decision? J. Econ. Behav. Organ. 55, 447–465.Eckel, Catherine C., Wilson, Rick K., 2006. Judging a book by its cover: Beauty and expectations in a trust game. Polit. Res. Quart. 59 (2), 189–202.Falk, Armin, Kosfeld, Michael, 2006. The hidden costs of control. Amer. Econ. Rev. 96 (5), 1611–1630.Falk, Armin, Zehnder, Christian, 2007. Discrimination and in-group favoritism in a citywide trust experiment. IEW-working paper No. 318, University of

Zurich.Fehr, Ernst, Fischbacher, Urs, von Rosenbladt, Bernhard, Schupp, Jurgen, Wagner, Gert, 2003. A nation-wide laboratory: Examining trust and trustworthiness

by integrating behavioral experiments into representative surveys. IZA discussion paper No. 715.Fershtman, Chaim, Gneezy, Uri, 2001. Discrimination in a segmented society: An experimental approach. Quart. J. Econ. 116 (1), 351–377.Glaeser, Edward L., Laibson, David I., Scheinkman, Jose A., Soutter, Christine L., 2000. Measuring trust. Quart. J. Econ. 115 (3), 811–846.Guiso, Luigi, Sapienza, Paola, Zingales, Luigi, 2004. Does local financial development matter? Quart. J. Econ. 119 (3), 929–969.Hoffman, Elizabeth, McCabe, Kevin, Shachat, Jason, Smith, Vernon, 1994. Preferences, property rights and anonymity in bargaining games. Games Econ.

Behav. 7, 346–380.Holm, Hakan J., Danielson, Anders, 2005. Tropic trust versus Nordic trust: Experimental evidence from Tanzania and Sweden. Econ. J. 115, 505–532.Johnson-George, C., Swap, W., 1982. Measurement of specific interpersonal trust: Construction and validation of a scale to assess trust in a specific other.

J. Person. Social Psych. 43 (6), 1306–1317.Karlan, Dean, 2005. Using experimental economics to measure social capital and predict financial decisions. Amer. Econ. Rev. 95 (5), 1688–1699.Knack, Stephen, Keefer, Philip, 1997. Does social capital have an economic payoff? A cross-country investigation. Quart. J. Econ. 112 (4), 1251–1288.Konow, James, 2000. Fair shares: Accountability and cognitive dissonance in allocation decisions. Amer. Econ. Rev. 90 (4), 1072–1091.Kugler, Tamar, Bornstein, Gary, Kocher, Martin G., Sutter, Matthias, 2007. Trust between individuals and groups: Groups are less trusting than individuals

but just as trustworthy. J. Econ. Psych. 28 (6), 646–657.Lewicki, Roy J., Bunker, Barbara B., 1996. Developing and maintaining trust in work relationships. In: Kramer, Roderick M., Tyler, Tom R. (Eds.), Trust in

Organizations: Frontiers of Theory and Research. Sage, Thousand Oaks, CA, pp. 114–139.Lewis, J. David, Weigert, Andrew, 1985. Trust as a social reality. Social Forces 63, 967–985.Luhmann, Niklas, 1979. Trust and Power. John Wiley, New York.Malhotra, Deepak, Murnighan, J. Keith, 2002. The effects of contracts on interpersonal trust. Administr. Sci. Quart. 47 (3), 534–559.Mayer, Roger C., Davis, James H., 1999. The effect of the performance appraisal system on trust for management: A field quasi-experiment. J. Appl. Psych. 84

(1), 123–136.McEvily, Bill, Tortoriello, Marco, 2011. Measuring trust in an organizational context: Review and recommendation. J. Trust Res. 1, 23–63.McEvily, Bill, Perronne, Vincenzo, Zaheer, Akbar, 2003. Trust as an organizing principle. Organ. Sci. 14, 91–103.McEvily, Bill, Weber, Roberto, Bicchieri, Christina, Ho, Violet, 2006. Can groups be trusted? An experimental study of collective trust. In: Bachmann, Reinhard,

Zaheer, Aks (Eds.), The Handbook of Trust. Edward Elgar Publishing, Cheltenham, UK.Putnam, Robert D., Leonardi, Robert, Nanetti, Rffaella Y., 1993. Making Democracy Work. Princeton University Press, Princeton.Rempel, John K., Holmes, John G., 1986. How do i trust thee? Psych. Today 20 (2), 28–34.Scharlemann, Jorn P.W., Eckel, Catherine C., Kacelnik, Alex, Wilson, Rick K., 2001. The value of a smile: Game theory with a human face. J. Econ. Psych. 22

(5), 617–640.Schechter, L., 2007. Traditional trust measurement and the risk confound: An experiment in rural Paraguay. J. Econ. Behav. Organ. 62, 272–292.Tungodden, Bertil, Cappelen, Alexander, Hole, Astri D., Sørensen, Erik Ø., 2007. The pluralism of fairness ideals? Amer. Econ. Rev. 97 (3), 818–827.Wrightsman, L.S., 1991. Interpersonal trust and attitudes toward human nature. In: Robinson, John P., Shaver, Phillip R., Wrightsman, Lawrence S. (Eds.),

Measures of Personality and Psychological Attitudes. Academic Press, San Diego, CA, pp. 373–412.Yaari, M.E., Bar-Hillel, M., 1984. On dividing justly. Soc. Choice Welfare 1 (1), 1–24.