Inequality as an Incentive

Inequality as an Incentive

Jeffrey V. ButlerUniversity of California, Merced

Eric CardellaTexas Tech University ∗

This version: March 13, 2019

Abstract

We study the incentive effects of wage and income inequality in a laboratory experiment.We randomly assign wage levels on a real-effort task. Across treatments, we exogenously varyinformation about wage and (experimental) income inequality as well as work content. Inall treatments, we provide participants with subsequent opportunities to behave pro- or anti-socially. We specifically design all tasks to be ecologically valid. We formulate a novel hypothesis,based on previous research into Just World Beliefs, about how salient inequality interacts withcontextual factors to affect subsequent behavior. We find that behavior and post-experimentsurvey data are largely consistent with our preferred hypothesis.

JEL Classification:

Keywords: Inequality; Incentives; Just World Beliefs

∗We wish to thank

Preliminary and Incomplete. Please do not cite without permission

1 Introduction

Scholars have become increasingly interested in the economic consequences of inequality. While

early investigations concentrated on how inequality directly enters individuals’ preference func-

tions (Akerlof and Yellen 1990; Bolton and Ockenfels 2000; Charness and Rabin 2002; Fehr and

Schmidt 1999; Konow 2003; Koszegi 2014), a more recent literature considers indirect channels

through which inequality may affect behavior and outcomes. In this vein, a handful of recent

papers investigate how salient wage and income inequality may undermine employee satisfaction

and morale, reducing productivity or increasing turnover (Breza, Kaur, and Shamdasani 2018;

Card, Mas, Moretti, and Saez 2012; Dube, Giuliano, and Leonard 2019; Godechot and Senik 2015).

Understanding both the direct and indirect channels through which inequality affects behavior is

especially important for economists because many economic models prescribe wage and earnings

inequality, with optimal compensation schemes often involving a substantial random component.1

In this study we implement a carefully designed laboratory experiment involving over 420 par-

ticipants. Our focus is on understanding the indirect channels linking inequality to productivity.

We formulate and provide evidence for novel hypotheses in which context interacts with inequality

to affect the indirect incentives provided by both wage and income inequality. Our hypotheses

proceed from contextual variation in Just World Beliefs (Lerner 2013, 1965), hereafter JWBs, a

motivated belief thought by economists to be particularly economically important Benabou and

Tirole (2006); Benabou and Tirole (2016); Butler (2014). JWBs are thought to be universal and

entail believing that people “generally get what they deserve” (Lerner 2013, pg. 11), so that, e.g.,

“effort pays and crime does not” (Benabou and Tirole 2006, pg. 710).

We have two competing conjectures about the mechanism(s) through which context interacts

with inequality to ultimately affect behavior. Our first, primary, conjecture follows in the vein of

many previous studies, from Akerlof and Yellen (1990) to Breza et al. (2018), that have interpreted

the indirect incentives provided by inequality through the lens of Equity Theory (Adams 1963).

At the heart of Equity Theory is a proportional fairness standard: the ratio of productivity to

compensation. When these equity ratios are substantially different across workers, individuals

1In the context of unobservable or non-contractable action to, e.g., overcome moral hazard, theoretically optimalcontracts often tie earnings to outcomes which may induce substantial randomness ex-post in compensation. As aresult, even identical workers performing identical tasks may end up with substantially different earnings. Whenemployees are not identical, the optimal compensation scheme may entail ex ante wage inequality as well, assigningdifferent wages to workers with different appetites for risk or effort costs. In this case, an additional source ofearnings inequality could be employees’ endogenous responses to their (different) optimal wages. Other analysesjustify wage inequality on efficiency grounds as a way for high-ability workers to credibly signal their ability (Spence1973). The use of tournaments or other types of performance-based bonuses are also well justified by traditionaleconomic theory (e.g., Gibbons 1987; Lazear 2000; Stiglitz 1975), and obviously capable of generating a great deal ofsalient inequality. For recent overviews of the literature on incentives in personnel economics and in economics moregenerally, see Lazear and Oyer (2007); Prendergast (1999) and Lazear (2018).

2

experience inequity, which is aversive. Behavioral consequences result from individuals trying to

ameliorate inequity by affecting their own equity ratio, which can be accomplished by increasing

or decreasing own effort or by, perhaps nefariously, increasing own compensation. Extending this

literature, we note that since the components of own and others’ equity ratios are often imperfectly

observed and subjectively defined, experienced inequity and its behavioral consequences may depend

on myriad, often subconscious, cognitive processes that are themselves shaped by contextual factors.

In particular, previous research suggests that one such contextual factor, work content, may

induce cognitive processes reconciling wage inequality with JWBs when work is ability-intensive

but not when it is effort-intensive (Butler 2014). Ostensibly, ability-intensive environments are close

enough to meritocratic settings in which fairness requires wage inequality that, to avoid cognitive

dissonance (Akerlof and Dickens 1982; Festinger 1957), individuals color their beliefs about the

basis of wage inequality, believing wages to be allocated according to merit rather than by chance.2

Consequently, our primary conjecture is that work content may interact with salient inequality

to provide an implicit justification for inequality, thereby ameliorating the negative responses to

inequality documented elsewhere (cf, Bracha, Gneezy, and Loewenstein 2015; Breza et al. 2018)

Our second, competing, conjecture builds on the cognitive process just outlined. We conjec-

ture that an end result of the process of subconsciously reconciling inequality and equity, salient

inequality may actually become compelling “evidence” in support of a just world, strengthening

JWBs. Strengthening JWBs should, in turn, reinforce the set of (indirect) incentives they provide.

In particular, Benabou and Tirole (2006) describe JWBs as enhancing self-reliance and providing

intrinsic effort motivation, which may help to overcome moral hazard but, at the same time, reduce

pro-sociality. Heuristically, if people get what they deserve and deserve what they get, then it

makes little difference whether effort is observable – it only matters that effort generally deserves

to be rewarded. The same logic reduces the urgency for other-regard, charity and altruism, as

others (automatically) get what they deserve. Therefore, our competing conjecture is that wage

inequality in an ability-intensive context will reduce the negative consequences of moral hazard but

also reduce other-regarding behavior.

To provide evidence on our conjectures, we conduct a laboratory experiment in which we ex-

ogenously vary work content as well as wage levels and information about wage and (experimental)

income inequality before providing participants subsequent opportunities to react positively, by

2An illustrative example might be the finals of an Olympic track meet, the Olympics being a quintessentiallymeritocratic and ability-intensive environment. Individuals, even the competitors themselves, tend to accept unequalrewards – a gold medal is more valuable than a bronze medal – as equitable, and would perceive equal rewards asinequitable. However, with as little as a few hundredths or even thousandths of a second often separating the first-and third-place finisher, it is typically difficult to objectively describe finishing order as depending on anything otherthan chance.

3

donating to charity, and negatively, by lying and shirking (undetectably) for financial gain. We

specifically design all experimental tasks to be ecologically valid, i.e., familiar and appropriate for

our student subject pool. In addition to observing behavior, through a post-experiment survey we

collect a measure of JWBs as well as a variety of demographics. Several design features ensure

participant anonymity to reduce the possibility of peer effects or experimenter demand effects.

As a preview of our results, we find that disclosing wage inequality in an ability-intensive

context increases effort provision and reduces charitable donations. We find no evidence for any

behavioral response to wage inequality when work is effort-intensive and little evidence for an

effect of disclosing income inequality in either context. In our post-experiment survey, we find

that wage inequality disclosure in an ability-intensive context is associated with stronger JWBs,

while JWBs are weaker when wage inequality occurs in an effort-intensive context. Finally, we find

explicit lying for financial gain to be surprisingly rare despite many design features guaranteeing

anonymity. However, in line with past results conditional on lying, liars lie little (Dufwenberg and

Dufwenberg 2018; Fischbacher and Follmi-Heusi 2013; Gneezy, Kajackaite, and Sobel 2018; Mazar,

Amir, and Ariely 2008).

Our study makes several contributions. First of all, we contribute to the literature seeking

to understand the indirect incentives associated with salient wage and income inequality (Abeler,

Altmann, Kube, and Wibral 2010; Akerlof and Yellen 1990; Angelova, Guth, and Kocher 2012;

Bartling and Von Siemens 2011; Bolton and Werner 2016; Bracha et al. 2015,?; Breza et al. 2018;

Butler 2014; Card et al. 2012; Charness, Cobo-Reyes, Lacomba, Lagos, and Perez 2016; Charness

and Kuhn 2007; Clark, Masclet, and Villeval 2010; Cohn, Fehr, Herrmann, and Schneider 2014;

Dube et al. 2019; Gachter and Thoni 2010; Gill, Prowse, and Vlassopoulos 2013; Godechot and

Senik 2015; Greiner, Ockenfels, and Werner 2011; Gross, Guo, and Charness 2015; Hennig-Schmidt,

Sadrieh, and Rockenbach 2010; Nosenzo 2013). In this growing literature, we are among the first to

consider how the behavioral consequences of inequality may depend on the content of employment,

i.e., whether it is ability- or effort-intensive.3 This is an important distinction, as personal and

societal economic success increasingly depends on skill- or ability-intensive employment sectors. If

the consequences of inequality differ qualitatively and predictably across this dimension, optimal

policies on earnings secrecy may vary across this dimension as well.

We also contribute to understanding the relationship between pro-social and anti-social behav-

ior. Some previous research has argued for a positive relationship, characterized as moral licensing

or moral cleansing (Branas-Garza, Bucheli, Espinosa, and Garcıa-Munoz 2013; Gneezy, Imas, and

Madarasz 2014). Other research, characterizing morality as a muscle that gets stronger with use,

3The only other paper we are aware of is Butler (2014).

4

suggests a negative relationship (Baumeister and Juola Exline 1999). Intuitively, if pro-sociality

is a trait then one might expect a negative relationship as well, with more pro-social types be-

ing generally less prone to anti-social behaviors. We provide evidence that the relationship may

be highly context-dependent and document a specific context inducing a positive relationship at

the aggregate level: shirking and charitable donations are both reduced by the revelation of wage

inequality in an ability-intensive context.4

A third contribution of our paper is methodological. We chose our primary tasks to be ecolog-

ically valid, i.e., appropriate for and familiar to our student population. Throughout the course

of a typical day, students may need to proofread their own or others’ work, to take an (ability-

dependent) quiz or to bubble Scantrons. They are also likely to be confronted with an opportunity

to donate to a well-known charity. The purpose of this design feature is to provide properly

incentivized evidence on pro- and anti-social behavior complementary to much of the related ex-

perimental literature, which typically uses decontextualized and unfamiliar tasks such as (abstract)

dictator games or rolling dice for money. By comparing behavior in our experiment with behavior

from conceptually related experiments, we generate evidence on the external validity of previous

findings. Our data suggest some patterns carry over (limited extent of lying), while others do not

(prevalence of lying).

Finally, we contribute to the literature on motivated beliefs. Researchers have long understood

that in order to avoid cognitive dissonance (Festinger 1957) individuals may subconsciously color

their beliefs and this process may alter economic incentives (Akerlof and Dickens 1982; Benabou

and Tirole 2016). However, the literature on the determinants of motivated beliefs is scant. We

formulate and test a conjecture about how the economic environment can interact with inequality

to affect motivated beliefs (cf, Di Tella, Galiant, and Schargrodsky 2007).

The remainder of the paper proceeds as follows. First, we present our experimental design in

detail, before formally stating several hypotheses including the two conjectures mentioned above.

In Section 4 we provide empirical results. In the penultimate section we revisit closely related

literature, putting our findings in context. In the final section we summarize and conclude.

2 Experimental Design and Procedures

We conducted a laboratory experiment involving a real-effort work task followed by a subsequent

cheating task as well as a charitable giving opportunity. All experimental sessions were conducted

at the Rawls College of Business at Texas Tech University in the Spring and Fall of 2018 and were

programmed in oTree (Chen, Schonger, and Wickens 2016). All participants were recruited through

4At the individual level, while the relationship is positive it is non-significant.

5

a college maintained subject pool. In total, we conducted 31 experimental session in which 423

individuals participated (44% were female; average age was about 21). Our experiment consisted

of eight treatments implemented using a between-subjects design. Sessions lasted approximately 1

hour and average compensation was $15.01. Participant instructions for all parts of the experiment

are provided in an Appendix.

There were three primary phases in the experiment, which are described in more detail below. In

Phase 1, participants completed one of two possible real-effort tasks and accrued earnings according

to one of two possible piece-rate compensation schemes. Marginal monetary incentives were iden-

tical in the two pay schemes. The differing pay schemes combined with differences in productivity

in the real-effort task created the potential for substantial (experimental) income inequality.

After completing the Phase 1 real-effort task, participants were randomly assigned to a feedback

condition where they received information about the differing pay schemes, the income distribution

in their session, both or neither. It is through this exogenous variation in feedback that we are able

to identify how wage and income inequality affect subsequent behavior.

In particular, in Phase 2 participants completed another real-effort task, this time without wage

inequality. In this second task, earnings were based on self-reported productivity. Since it was not

logistically possible to verify either the quantity or quality of individuals’ production (more on

this below), the Phase 2 task presented participants with two forms of cheating opportunities: by

inflating their self-reported production quantity (lying), participants could directly increase their

earnings; by producing lower quality items (shirking), participants could either produce a given

quantity with less effort or produce a higher quantity in a given amount of time, increasing either

the income-to-effort ratio or their effective hourly wage.

In Phase 3, participants were presented with an opportunity to privately and anonymously

donate a portion of their total experimental earnings to charity. Phase 3 was followed by a time-

preferences elicitation task, which we do not analyze, as well as a post-experiment survey where we

measure demographics and attitudes, including JWBs.

2.1 Phase 1: Initial Real-Effort Task

To induce wage and income inequality, participants first completed an incentivized real-effort task.

As part of the experimental design, we consider two distinct tasks, which differ in the extent to

which performance credibly depends on skill or ability versus effort alone. The first task, which we

refer to as the ability-intensive (AI) task, consists of 48 Raven’s Advanced Progressive Matrices,

each of which requires selecting a picture that best completes a given pattern. Raven’s matrices are

designed to measure the test taker’s reasoning ability, considered an important component of general

6

intelligence, a fact which we convey to our participants in the description of the task. We therefore

consider it a plausible assumption that participants perceive the task as being ability-intensive.

Alternatively, participants could be assigned an effort-intensive (EI) task. The EI task is a sim-

plified proofreading task in which participants count the number of typos appearing in a sequence

of 48 sentences. We deliberately constructed the sentences to contain only typos that were easily

recognizable, e.g., duplicated words, missing words, or numbers that replaced letters, requiring lit-

tle (reading) ability but a reasonable amount of effort. Consequently, our maintained assumption

is that performance on the EI task depends primarily on effort and that, moreover, participants

perceived this to be the case.

Each participant performed only one of these two tasks. They were constructed to be as parallel

as possible, other than their content (e.g., both involved 48 multiple-choice questions, each with

eight possible responses). However, because the two tasks appear quite different on a computer

screen – one involves pictures, the other text – to minimize the possibility of participants realizing

there were different tasks, all participants in a given session were assigned the same task. That is

to say, task manipulation was accomplished across sessions.

As a way to induce explicit wage inequality we implemented two different piece-rate pay schemes:

Low Pay: $0.20 for each correct response; $0.00 for each incorrect response.

High Pay: $0.30 for each correct response; $0.10 for each incorrect response.

Importantly, the marginal (monetary) incentive for correctly answering a question is constant

at $0.20 across both pay schemes, thus mitigating the possibility of effort changes in response

to monetary incentive effects (see Butler 2014). Both tasks were timed, with a time limit of 10

minutes. Prior to beginning the task, each participant was informed about the task they would

complete, their own pay scheme, the total number of questions possible as well as the 10 minute time

limit. Importantly, however, they were not informed about the two possible pay schemes before

completing the task, eliminating any scope for relative pay concerns to affect Phase 1 behavior.

2.2 Phase 1.5: Relative Pay and Income Disclosure

After completing the 10-minute Phase 1 task, participants were informed of their own performance

and their own income from the task.5 In addition to these basic pieces of information, as part of

5We inserted filler questions between Phase 1 and Phase 1.5 which we do not analyze, but simply control for.Participants received one of three possible sets of questions, with the sets being randomly assigned at the individuallevel. Specifically, they were asked to state their political party or gender and then describe briefly an event whichmade them feel connected to their gender or party. The third possible set of questions asked them about the primarymedium through which they receive television programming (e.g., cable or the internet), and to briefly describe theirreason for selecting that medium.

7

the experimental design we manipulated the feedback participants received about the existence of

alternative pay schemes and the distribution of incomes in their sessions. Within each session, each

participant was equally likely to be informed of the existence of the two different pay schemes or

to receive no such information. The former condition can be thought of as involving visible wage

inequality, which we denote by VW, while the latter condition features invisible wage inequality,

denoted IW.

The other domain of feedback we manipulated was relative income. Specifically, in the visible

income inequality (VI) condition participants were provided with a binned relative frequency chart

of experiment incomes in their session. In the invisible income inequality (II) condition partici-

pants received no such feedback. Through this manipulation, we reveal comprehensive information

about the income inequality within a session, allowing participants to locate themselves within

the session’s income distribution. Because the VI condition involves a large chart appearing on

participants’ screens, to minimize the possibility of participants in the II condition being made

inadvertently aware of the VI condition, this manipulation was implemented across sessions. All

participants in a particular session were assigned to either the VI condition or the II condition.

Feedback about wage inequality and income inequality in the Phase 1 task were provided to

participants ex post, i.e., after they completed the task. As a result, effort provision in the Phase

1 task should not be affected by our wage and income inequality manipulations. This enables

us to cleanly identify how knowledge of wage and income inequality impact subsequent behavior,

including anti-social behavior (lying and shirking) and pro-social behavior (charitable donations).

Overall, our study features a 2 (wage inequality feedback) x 2 (income inequality feedback) x 2

(task type) full factorial design with eight distinct treatments. For ease of exposition, we refer to

each treatment using the format [VW/IW]-[VI/II]-[AI/EI]. As an example, VW-VI-AI refers to the

treatment with visible wage inequality, visible income inequality and an ability-intensive Phase 1

task. We summarize our treatments in Table 1.

2.3 Phase 2: Cheating Task

After completing Phase 1 task and receiving any treatment-specific feedback on wage or income

inequality, participants proceeded to the Phase 2 cheating task. From Phase 2 on, all experimental

features were identical for all participants in all treatments.

In the spirit of ecological validity, the Phase 2 cheating task was framed as a second real-effort

task where participants could earn additional compensation. Specifically, participants would pre-

pare multiple versions of a Scantron answer key, corresponding to different versions of a statistics

exam. Participants were provided with a mock exam closely resembling an actual exam admin-

8

Table 1: Summary of Treatments

TreatmentName

TaskWage InequalityFeedback

Income InequalityFeedback

N

IW-II-EI Proofreading No No 54IW-II-AI Raven’s Matrices No No 56VW-II-EI Proofreading Yes No 53VW-II-AI Raven’s Matrices Yes No 58IW-VI-EI Proofreading No Yes 52IW-VI-AI Raven’s Matrices No Yes 49VW-VI-EI Proofreading Yes Yes 53VW-VI-AI Raven’s Matrices Yes Yes 48

Notes: [1] Raven’s matrices refer to Raven’s Advanced Progessive matrices, commonly thought to be aculture- and language-free measure of general intelligence.measure. [2] For the proofreading task, we ran-domly inserted simple forms of typogaphical errors into sentences and participants had to count the numberof typos appearing in each sentence. [3] Task format was constructed to be as similar as possible; eachitem of each task involved selecting the correct answer from among eight choices, and was approximatelythe same size and format on participants’ computer screens.

istered in a statistics class taught by one of the authors. Participants were also provided with

a master answer key for the original (mock) exam, a sheet listing the fifteen desired version ma-

nipulations (labeled generically as Version A – Version O), and fifteen Scantrons with which to

create the answer key versions.6 The Phase 2 task was again timed: participants had 15 minutes

to complete as many versions from the list as they could. They were instructed that at the end of

the 15 minutes, they would (self-report) how many Scantrons they completed and would be paid,

based on this self-report, $1 for every completed version.

After the 15 minutes had elapsed, participants placed all their Scantrons in a box at the back

of the room which was not monitored by the (lone) experimenter. After submitting their Scantrons

and returning to their carrels, each participant filled out an anonymized payment slip. It was

on this slip that production was self-reported and upon which payment from Phase 2 was based.

Because participants were explicitly informed we would pay them according to their self-reports,

there were opportunities to engage in cheating. By inflating the number of completed Scantrons,

participants could cheat in a very deliberate manner: lying for financial gain.7 A more subtle form

6To facilitate statistical identification, all participants received the same exam copy, master answer key and versionvariations. The versions were created by taking random permutations of question orderings on the master key. Toenhance ecological validity, through opacity in our instructions we created an environment where participants likelyperceived the task as regular economically valuable work (Falk and Ichino 2006). At the same time, to avoid deceptionparticipants were not told anything about the intended use of the answer keys they would be preparing. They weresimply informed that they were to make answer keys for different versions of the exam based on the exam master keyprovided.

7We went to great lengths to convey to participants, both explicitly and implicitly, that their self-reports wereanonymous. With only one experimenter in the room, it would have been logistically impossible at the time toexamine each self-report for accuracy, which should have been apparent to participants. Moreover, we paid usingcash in envelopes marked with code numbers which could not be linked to names and we did not have participants sign

9

of cheating was also possible, which we term “shirking.” By completing versions less accurately,

which presumably requires less time and effort, participants could earn a given amount of money

with less effort or complete more Scantrons in the allotted time than if they filled out versions

accurately.

Importantly, it was obvious that for logistical reasons both forms of cheating – lying or shirking

– were imperceptible by the experimenter during the session. Great pains were taken to ensure,

and to convey implicitly to participants, that cheating could only be detected after sessions were

completed and could never be attributed to a particular name, but rather only to the participant’s

carrel number, and that we could never match names to carrel numbers.

2.4 Phase 3: Charitable Giving Opportunity

After completing Phase 2, participants received an envelope of cash in the amount of their cu-

mulative earnings from the Phase 1 and Phase 2 tasks. They were then given the opportunity to

anonymously donate any amount they wished to two specific charities. The two charitable giving

options were: South Plains Food Bank, a local charity; or the American Red Cross, a national

charity. We conveyed to participants that this was a credible donation and that they would receive

by email at the conclusion of the study a link where they could view a receipt for the total amount

of money donated to each charity (across all participants). Participants were provided with a brief

description of each charity and then instructed to fill out an anonymous donation slip indicating how

much of their earnings (if any) they wanted to donate to each charity. Participants were instructed

to leave any cash they wanted to donate along with their completed slip, which only indicated only

how the experimenters should allocate the money in the envelope between the two charities, in the

envelope and to leave the envelope at their carrel when the left the experiment.8 This procedure

ensures that donations were anonymous to the experimenters as well as to participants’ peers.

After completing phase 3, participants filled out an anonymous survey. Through the survey we

gathered general demographic data and psychological measures, as well as self-reported beliefs and

attitudes on a variety of relevant topics including, importantly, JWBs. We summarize the phases

of the experiment in Figure 1.

Well after each session was completed, we matched outcome measures from all phases of the

experiment. Matching was accomplished by carrel number, which was automatically recorded

receipts. Overall, this cheating opportunity was designed to be as close as possible to the more standard dice-in-cupsprocedure which has been the focus of a voluminous literature (Abeler, Nosenzo, and Raymond 2016; Dufwenbergand Dufwenberg 2018; Fischbacher and Follmi-Heusi 2013; Gneezy et al. 2018; Mazar et al. 2008) while still beingfamiliar and appropriate for our subject pool and, importantly, allowing for precise identification of cheating.

8As an added step to ensure anonymity, even participants who chose to donate nothing were asked to place acompleted donation slip in their envelope and leave the envelope at their carrel.

10

Figure 1: Timeline

Phase 1

AI/EI task

10 min limit

Phase 1.5

Info revelation

Phase 2

Scantron task

15 min limit

Phase 3

Donations

Phase 4

Survey

for the computerized portions of the experiment (Phase 1 and the post-experiment survey). For

the other phases of the experiment, participants noted their carrel numbers on materials they

submitted. It is not possible for the experimenters to match participants’ names to carrel numbers

or even payment amounts as, e.g., no receipts were collected.

3 Hypotheses

Having described the experiment in detail, we are now in a position to state several formal hy-

potheses concerning the mechanism(s) through which inequality might affect behavior. Our first

hypothesis constitutes a simple specification check as well as a confirmation of standard incentive

theory. In the absence of indirect incentives from salient inequality, performance should theoret-

ically depend primarily on marginal monetary incentives. Because marginal monetary incentives

are identical in our two pay schemes – $0.20 for each additional correct answer – we expect no

significant variation in Phase 1 task performance across pay schemes. Since, furthermore, neither

wage nor income inequality was disclosed until after the Phase 1 task was completed, neither type

of inequality should affect Phase 1 task performance.

Hypothesis 1: There will be no significant differences in the average number of correct re-

sponses on the Phase 1 task across the following experimental factors: pay scheme; wage inequality

visibility; income inequality visibility.

Our next three hypotheses consider the effects of disclosing wage inequality, in particular. We

leave the consideration of income inequality disclosure to a later set of hypotheses.

Hypothesis 2 assumes that the primary mechanism through which salient inequality interacts

with context to affect behavior is by providing an implicit justification for observed inequality,

i.e., our primary conjecture. We also assume that negative reactions to inequity are manifested

as subsequent anti-social behavior (lying and shirking), that positive responses are manifested as

subsequent pro-social behavior (charitable donations). Given these assumptions, because previous

research suggests that wage inequality tends to be rationalized as fair in an ability-intensive envi-

ronment, but not in an effort-intensive environment, we expect the disclosure of wage inequality to

increase subsequent lying and shirking to a lesser extent when the Phase 1 task is ability-intensive

11

than when it is effort-intensive. Similarly, revealing wage inequality in an ability-intensive context

should decrease charitable donations to a lesser extent than in an effort-intensive environment.

Hypothesis 2: The effect of revealing wage inequality on subsequent behavior will differ quanti-

tatively by Phase 1 task type. Wage inequality disclosure related to an ability-intensive Phase 1 task

will increase anti-social behavior and decrease pro-social behavior to a lesser extent than disclosing

wage inequality related to an effort-intensive Phase 1 task.

Hypotheses 3 and 4 relate to our competing conjecture about the mechanism relating wage in-

equality to behavior. In this alternative mechanism, to avoid cognitive dissonance arising from the

juxtaposition between JWBs and objectively unjustifiable inequality, individuals subconsciously

reconcile inequality and inequity when feasible. Prior research suggests that such reconciliation

is more feasible in ability-intensive contexts, perhaps because they are reminiscent of meritocratic

environments in which inequality is equitable, than in effort-intensive contexts. Through the rec-

onciliation process, wage inequality in an ability-intensive environment may become “evidence”

supporting JWBs, thereby strengthening them.

Hypothesis 3: Disclosing wage inequality will strengthen JWBs when the Phase 1 task is ability

intensive, but not when the Phase 1 task is effort-intensive.

If JWBs are strengthened sufficiently, there should be detectable consequences for behavior.

Our next hypothesis builds on previous research suggesting that JWBs provide a particular set of

indirect incentives, incentives which intrinsically motivate effort provision and discourage other-

regarding behavior. Intuitively, although previous research is silent on this aspect, JWBs might

also discourage lying: if people get what they deserve and lying deserves to be punished, then

JWBs provide a disincentive. We hypothesize that JWBs will be strengthened sufficiently by the

revelation of wage inequality in an ability-intensive to affect behavior in a way consistent with belief

in a just world.

Hypothesis 4: Following an ability-intensive Phase 1 task, revealing wage inequality will reduce

shirking, lying and charitable donations.

Note that Hypothesis 4 predicts a qualitatively different interaction between wage inequality and

work content than Hypothesis 2. In particular, Hypothesis 2 predicts less of a decrease in charitable

donations when revealed wage inequality is associated with an ability-intensive task than when it

is associated with an effort-intensive task; Hypothesis 4 seemingly predicts the opposite.

12

Our final pair of hypotheses concern the effect of revealing income inequality. We expect income

inequality to be more easily rationalized as being due to performance differences, and therefore

equitable, particularly absent knowledge of wage inequality. If, as in our primary conjecture, the

main effect of salient inequality is to engender negative responses to perceived inequity, then one

might expect the effect of revealing income inequality to be qualitatively similar to, just more muted

than, the effect of revealing wage inequality.

Hypothesis 5: The revelation of income inequality will increase lying and shirking and decrease

charitable donations to a lesser extent than the revelation of wage inequality, irrespective of Phase

1 task.

If, on the other hand, the primary effect of revealing inequality is to induce cognitive processes

reconciling inequality with fairness, strengthening JWBs which, in turn, affect behavior (our com-

peting conjecture), then revealing income inequality may have no effect at all on behavior by itself

as it poses no direct threat to JWBs. Income inequality can easily be reconciled with a just world

through beliefs about performance differences. On the other hand, revealing income inequality in

addition to wage inequality may produce countervailing effects to the latter because the combina-

tion of wage and income inequality poses less of a threat to JWBs than wage inequality alone, and

weaker threats are less likely to trigger the subconscious reconciliation processes responsible for

behavioral responses to inequality.

Hypothesis 6: Revealing income inequality in isolation will have no effect on behavior; re-

vealing income inequality in addition to wage inequality will generally weaken the effects of wage

inequality on behavior, i.e., produce less of reduction in shirking and lying and less of a reduction

in charitable donations than wage inequality alone.

4 Results

4.1 Descriptive statistics

We begin by describing a simple balance check. We collected a limited set of demographics on

the post-experiment survey, including age, gender and self-reported measures of family income and

GPA, both categorical. For each of these variables, separately, we conduct a Chi-square with the

null hypothesis of independence across treatments. For only one of these tests, the test associated

with GPA (p = 0.084), can the null hypothesis be rejected at a 10 percent significance level. We

take this as evidence that randomization into treatments was generally successful, but in our formal

econometric estimates below we control for demographics.

13

To provide an overview of the data, in Table 2 we report raw means of our primary outcome

variables by treatment and pay scheme. For Phase 1, we report (experimental) income and the

number of correct responses on the assigned task (score). For Phase 2, we report the quantity

and quality of Scantrons produced. Our quantity measure is the number of Scantrons submitted,

irrespective of quality, which hypothetically takes values from 0 to 15. Our quality measure is the

proportion of correctly bubbled items (out of 20) averaged over all of the Scantrons the participant

submitted. This can be interpreted as the inverse of shirking – more (less) effort should increase

(decrease) quality – and takes values from 0 to 1. Our measure of lying is an indicator variable

taking the value of one whenever the participant’s self-reported number of Scantrons completed

does not match the actual number he or she submitted.9 For charitable donations, for simplicity

we report only the total amount donated to both possible charities.

From the raw means we can glean a few patterns. First of all, we were successful in generating

substantial income inequality on the Phase 1 task across pay schemes. The income difference across

pay schemes is highly significant within each of the eight treatments (p < 0.001 always; two-tailed

t-test). Pooling all treatments, participants assigned the Low-Pay scheme earned $4.90 from this

task, while High-Pay scheme participants earned $8.97, or 83% more (p < 0.001; two-tailed t-test).

This substantial income heterogeneity should aid our ability to detect any behavioral reactions to

the revelation of income inequality.

Secondly, considering Phase 1 task performance (score) we see that although average perfor-

mance is generally higher in the effort-intensive task than in the ability-intensive task,10 within each

task there is little variation in performance by treatment. This appearance is confirmed by sepa-

rate Chi-square tests, one in which we pool all observations from the EI sessions (p = 0.224) and

the other where we pool observations from the AI sessions (p = 0.568). Testing for a relationship

between pay scheme and Phase 1 task performance yields similar conclusions.11 Consequently, our

data pass an important specification check which formed the basis of our first hypothesis. On the

whole, our data support the notion that marginal monetary incentives are decisive determinants of

Phase 1 task performance absent information about inequality and that participants did not antic-

ipate the experimental manipulations to follow. This set of null findings permits the attribution of

9As with the more prevalent dice-in-cups procedure (Fischbacher and Follmi-Heusi 2013; Mazar et al. 2008), wefind that participants who lie do not lie as much as they can. Lies here typically take the form of inflating one’sreport by one rather than reporting the maximum possible of 15, which few achieved or reported. Therefore, we loselittle by using an indicator variable rather than the difference between self-reports and actual production.

10This would be consistent with, e.g., individuals exerting full effort in both tasks but the ability component holdingback performance in the ability-intensive task.

11Pooling all AI-session, a Chi-square test for a relationship between pay scheme and score is non-significant(p = 0.209). The same procedure for EI-session observations yields a p-value of 0.495. Moreover, testing by payscheme within each of the eight treatments separately yields only one significant difference. The approximatelyfour-item difference in the VW-VI-EI treatment by pay scheme yields p = 0.014 using a two-sided t-test

14

differences in subsequent outcome variables to the revelation of inequality.

Result 1: Hypothesis 1 is supported. For each of the two possible Phase 1 tasks, performance

does not vary by treatment or by pay scheme.

Moving on to the Phase 2 (Scantron) task, Table 2 documents little variation in the (actual)

quantity of Scantrons produced across treatments or pay schemes. Average quantity ranges nar-

rowly from about seven to about eight. Pooling across all treatments, those assigned the High-Pay

scheme on the Phase 1 task produced 7.70 Scantrons on average in Phase 2, while Low-Pay scheme

participants produced marginally fewer 7.36 (p = 0.103; two-sided t-test); pooling over pay schemes,

participants in the AI condition also submitted marginally fewer scantrons on average (7.36 in AI

vs. 7.71 in EI (p = 0.084; two-sided t-test) while, overall, a Chi-square test does not reject the null

hypothesis of no relationship between Scantron quantity and treatment (p = 0.407).

The remaining columns repeat the same theme – little evidence of substantial variation across

treatments or pay schemes. Pooling across pay schemes, a Chi-square tests suggests the quality

of Scantrons submitted, i.e., the inverse of “shirking,” varies little across experimental factors

(p = 0.773).12 Similar Chi-square tests for lying and for charitable donations also fail to reject their

null hypotheses (lying : p = 0.704; donations : p = 0.698).

4.2 Econometric estimates

While descriptive statistics are useful in providing an overview of the behavioral patterns in our

data, the relatively large number of treatments involved complicates our ability to say anything

more detailed using just the raw averages. In particular, while a handful of individually significant

relationships are present in the data, singling them out runs the risk of identifying false positives.

More formal econometric estimates will lessen this concern and, at the same time, allow us to

control for potentially important confounding factors such as random variation in demographics or

other observables across treatments.

In Table 3 we report our primary set of econometric estimates, using simple OLS for ease of

interpretation. We relegate alternative specifications and other robustness checks to an appendix.

The estimates in each column include a full set of experimental factors and their interactions as

well as demographics, however we report only factors of primary interest for ease of exposition.13

12The primary exception is the eight point drop in quality associated with the ability-intensive task in the absenceof any inequality feedback. Pooling across pay schemes, quality is 0.98 on average in treatment IP-II-EI, and 0.90 inIP-II-AI, a difference which is highly significant (p = 0.003; two-sided t-test).

13We do not report coefficients related to triple and quadruple interactions among our experimental factors, asthey are numerous and almost never significant. we also control the filler questions inserted between Phase 1 andPhase 1.5 and described mentioned in footnote 5, but do not report the coefficients.

15

In the table HP is an indicator for the High-Pay scheme; VW denotes visible wage inequality; VI

denotes visible income inequality; while AI denotes the ability-intensive Phase 1 task. We control

for each participant’s own Phase 1 task income, which they all learn in Phase 1.5, to account for

the possibility of a negative reaction to falling short of a subjective reference level of earnings.

The demographics we control for are: age, gender and categorical measures of grade point average

and family income. For all estimates we cluster standard errors by session to allow for arbitrary

within-session correlation of behavior.14

In the first column, labeled quantity, we report the effect of our experimental manipulations on

the number of Scantrons submitted, unadjusted for quality. As in the simple means, there is little

evidence that our factors had any effect on this most basic level of productivity. The coefficient on

HP is marginally significant and negative, while the coefficient on HPxAI is marginally significant

and positive, indicating that being highly paid for an effort-intensive task reduces subsequent

production even controlling for the effect of increased (Phase 1) income, but that this negative

effect is erased by an ability-intensive Phase 1 task. The coefficient on Phase 1 income is also

marginally significant and positive, controlling for pay scheme, suggesting there is some common

component, perhaps intrinsic motivation, explaining productivity on both the Phase 1 and Phase 2

tasks. The magnitude of the pay scheme effect in an effort-intensive environment is non-negligible:

the coefficient on HP suggests a reduction in Scantron quantity of about 12% of the unconditional

average quantity, which is 7.53. However, the coefficient on the interaction HPxAI is positive,

(marginally) significant, and of sufficient magnitude to overturn the implication of negative pay

scheme spillovers in an ability-intensive environment. Because there is evidence of a mild direct

effect of our experimental factors on Scantron quantity which varies across treatments, in the

remaining columns we control for it.

In the next column (quality) we examine how our experimental factors affected the quality of

submitted Scantrons. Because it ostensibly requires less effort to bubble Scantrons less accurately,

we interpret this measure as being negatively related to shirking – the more subtle form of cheat-

ing available to our participants. We find little evidence for an effect of pay scheme or inequality

revelation when the (Phase 1) task is effort-intensive: the coefficients associated with HP, VW and

VI are all small in magnitude and non-significant. The situation changes drastically in an ability-

intensive environment, however. The negative and highly significant coefficient on AI indicates that

14Also not reported is an OLS estimate of Phase 1 performance, which we conducted to confirm that our datasupport Hypothesis 1. Specifically, using Phase 1 task score as the dependent variable, none of the coefficients onexperimental factors are either large in magnitude or significant with the (anticipated) exception of the coefficienton AI, which has a coefficient of −8.38 and is highly significant (p < 0.001). Among the demographic controls, ageis negative and marginally significant (p = 0.068) while the dummy for having a high GPA is positive and significant(p = 0.024). The latter would be consistent with, e.g., there being some common trait that explains productivityirrespective of task and academic performance, which would not be entirely surprising.

16

an ability-intensive Phase 1 task induces significantly more shirking in the absence of inequality

revelation. The small and non-significant coefficient on HPxAI indicates this shirking effect has

little to do with pay scheme. At the same time, the positive and significant coefficients on the

interactions VWxAI and VIxAI suggest that revealing the presence of wage or income inequality,

respectively, significantly reduces shirking. The magnitudes of the effects implied by these coeffi-

cients, particularly the former which represents about 12% of the unconditional mean of quality,

are non-trivial. The relative magnitudes of the coefficients suggest that revealing wage inequality

has a (salutary) effect on shirking that is nearly twice as large as the effect of revealing income

inequality.

In the column labeled lying we examine our more blatant measure of cheating – lying about the

number of Scantrons submitted. Recall, participants were paid based on their self-reports and we to

great pains to convey both explicitly and implicitly that these self-reports were anonymous, so that

this cheating measure shares much in common with the more prevalent dice-in-cups paradigm (Fis-

chbacher and Follmi-Heusi 2013; Mazar et al. 2008). We find that none of our experimental factors

affect lying for financial gain. All of the associated coefficients are both non-significant and small

in magnitude.

We turn next to charitable donations, our measure of pro-sociality and the primary behavioral

distinction between our competing hypotheses. The negative, large in magnitude, and highly signif-

icant coefficient on the interaction term VWxAI implies that revealing wage inequality associated

with an ability-intensive Phase 1 task substantially lowered contributions. The magnitude of the

coefficient, −2.065, represents three-quarters of the unconditional average of charitable donations

($2.73).15

Considering more general patterns in donations, most of our other experimental factors have

little effect. Surprisingly being assigned to the High-Pay scheme does not translate into higher

donations, even though High-Pay participants earned roughly twice as much on average from the

initial task, and roughly 25% more overall.16 There is also surprisingly little evidence for a direct

effect of being assigned the ability-intensive Phase 1 task, even though earnings were substantially

lower there irrespective of pay scheme.

Considering all of these patterns together, behavior appears to be more consistent with our

primary conjecture based on inequality being implicitly justified, than our competing conjecture,

15The coefficient on the triple interaction HPxVWxAI is non-significant (p = 0.281), suggesting that the reductionin donations was not concentrated on those who were disadvantaged by wage inequality. The only other significantcoefficients were associated with: age (0.025; p = 0.002); being male (−1.12; p = −.010); and having a low GPA(−1.00; p = 0.053). Thus, older participants were slightly more generous, while males and low-performing studentswere substantially less generous.

16High-Pay scheme participants made $16.77 on average from the Phase 1 and Phase 2 tasks combined, comparedwith $12.42 for their Low-Pay scheme counterparts (p < 0.001; two-tailed t-test).

17

based on strengthened JWBs. The effect of revealing wage inequality differs substantially across the

content of work, a prediction shared by both conjectures. Revealing wage inequality does increase

anti-social behavior by less following an ability-intensive task than following an effort-intensive task,

another common feature of both our primary hypothesis and our competing conjectures. However,

the contextual variation in charitable donations is consistent with our competing conjecture but

not our primary conjecture.

Result 2: Behavior is more consistent with our competing conjecture (Hypothesis 4) than with

our primary conjecture (Hypothesis 2), suggesting contextual variation interacts with inequality to

affect behavior by strengthening JWBs and the associated set of indirect incentives.

As additional evidence in support of our competing conjecture, we can examine contextual vari-

ation in JWBs directly. In the final column of Table 3 we use as the dependent variable answers

from a post-experiment survey question asking participants how much they agree with the state-

ment “people generally get what they deserve.” Responses were collected on a scale ranging from 1

(totally disagree) to 7 (totally agree). Although the data are merely suggestive – we cannot know

whether JWBs affected behavior or, rather, whether JWBs ex post rationalized behavior – they

clearly support our competing hypothesis.17 The highly significant, large-in-magnitude and posi-

tive coefficient on VWxAI reveals that JWBs are stronger when wage inequality is revealed than

when wage inequality is concealed in an ability-intensive context. On the other hand, in an effort-

intensive environment, where wage inequality is more difficult to square with a just world, JWBs

are lower when wage inequality is revealed than when it is concealed, as evidenced by the negative

and significant coefficient on VW. The coefficient on the triple interaction HPxVWxAI (omitted for

readability) is actually negative and marginally significant (coeff = −0.89; p = 0.095) suggesting

that, if anything, the salutary effect of revealing wage inequality on JWBs is stronger for those

assigned the Low-Pay scheme than those assigned the High-Pay scheme.18

Result 3: Hypothesis 4, the (motivated) beliefs component of our competing conjecture, is

17That is to say, one may worry that it is the act of shirking or lying, rather than our experimental factors directly,that causes people to adjust their JWBs in order to ex-post rationalize behavior. Since we do not observe JWBsin between Phase 1 and Phase 2, but only after the experiment was concluded, we cannot definitively rule out thispossibility. One way we can partially address this concern is by simply inserting controls for shirking and lying intoour estimates of JWBs in the last column of Table 3. We omit these further estimates for space considerations,but simply note that inserting our measure for shirking or lying separately, or inserting both simultaneously doesnot change our findings. For instance, the coefficient on VWxAI always remains above 1.54 and is always highlysignificant.

18Other patterns of note include that being in an ability-intensive context with no knowledge of wage inequality isgenerally associated with lower JWBs, particularly among the disadvantaged. The magnitude of the coefficient on AIrepresents about one-quarter of the unconditional average of responses (4.41; sd = 1.39), or about 70% of a standarddeviation; while the magnitude and sign of the positive coefficient on HPxAI indicates that this pattern is almostcompletely offset by the High-Pay scheme in the same situation.

18

supported. Disclosing wage inequality in an ability-intensive context strengthens JWBs.

Finally, we consider the effect of revealing income inequality. We find that income inequality

revelation by itself has little effect on JWBs directly, in either the ability-intensive or the effort-

intensive environment, consistent with the notion that this type of inequality is easy to justify

as being due to performance heterogeneity. This evidence seems contrary to Hypothesis 5, but

consistent with Hypothesis 6. Moreover, rather than compounding the salutary effect of wage

inequality in an ability-intensive environment, revealing income inequality actually counteracts it,

just as predicted in Hypothesis 6. The triple interaction VWxVIxAI is negative, large in magnitude

and significant (coefficient = −1.66; p = 0.030).19 Overall, the data seem more consistent with our

competing conjecture, embodied here in Hypothesis 6, than with our primary conjecture.

Result 4: The effects on behavior of income inequality revelation is more consistent with

Hypothesis 6 than with Hypothesis 5, lending further support to our competing conjecture.

Summing up, our econometric estimates provide compelling evidence that revealing wage in-

equality in an ability-intensive context can affect subsequent pro-social and anti-social behavior.

Both behavior and (just world) beliefs are consistent with our competing conjecture about the

mechanism through which these affects occur. Specifically, our data support the notion that when

work is ability-intensive, revealing wage inequality affects behavior by inducing cognitive processes

that ultimately strengthen JWBs and the associated set of indirect incentives.

5 Relation to Existing Literature

Our study is closely related to the growing body of experimental literature examining how rela-

tive wage comparisons (i.e., salient exposure to wage inequality) impact contemporaneous effort

provision and productivity. While some studies document effort responses to horizontal wage com-

parisons in a chosen-effort paradigm (e.g., Abeler et al. 2010; Angelova et al. 2012; Charness et al.

2016; Clark et al. 2010; Gachter and Thoni 2010; Gross et al. 2015; Nosenzo 2013) as well in a

real-effort paradigm (e.g., Bracha et al. 2015; Cohn et al. 2014), several other studies find little

effort response in either of these settings (Bartling and Von Siemens 2011; Bolton and Werner 2016;

Butler 2014; Charness and Kuhn 2007; Greiner et al. 2011; Hennig-Schmidt et al. 2010). Overall,

the results are somewhat mixed. In contrast, the focus of our study is on spillovers to subsequent

tasks. That is to say, we examine how exposure to inequality impacts effort provision, pro-social

19Among demographics, only the indicator for high (family) income is (even marginally) significant (coefficient= 0.41; p = 0.058), indicating that students from wealthier families are more likely to believe the world is just andfair.

19

and anti-social behavior in subsequent tasks even when those subsequent tasks do not feature wage

inequality.

Our study also relates to a smaller literature focusing on how current inequality affects subse-

quent attitudes and behavior, with a particular focus on those disadvantaged by inequality. This

literature consistently documents negative effects of inequality, typically in terms of job satisfaction

and turnover. Notably, Card et al. (2012) conducts a field experiment with public (university) em-

ployees in which a random subset were informed about the existence of website where they could

learn their peers’ salaries and finds that acquiring information about peers’ salaries lowers job and

pay satisfaction and may increase turnover for below-median earners. Godechot and Senik (2015)

document a similar negative association between wage rank within a firm and employees’ reported

satisfaction. Dube et al. (2019) exploit variation in raises at a large US retailer creating essentially

random and transparent wage inequality, estimating a large causal effect of inequality on turnover.

They also find that comparisons to peers’ wages, rather than to a market wage, explain most of

the effect. In one of the few laboratory experiments in this vein, Gill et al. (2013) randomly vary

pay levels and knowledge of pay levels for performance on an effort-intensive task, finding that

knowing about pay differences significantly and uniformly increases anti-social behavior (lying for

financial gain) on a subsequent task. Differently from this literature, we examine both subsequent

pro-social and anti-social behavior, focus on a novel contextual factor, work content, and investigate

the mechanisms through which context may interact with inequality to affect behavior.

Finally, our paper is closely related to two recent studies investigating how justifications for pay

inequality can ameliorate subsequent negative reactions. In a laboratory experiment, Bracha et al.

(2015) show labor supply reacts negatively to knowledge of being paid less than others but that

even very thin explicit justifications can eliminate this negative response. In a field experiment

involving manufacturing workers in India, Breza et al. (2018) also document a negative labor

supply response to knowledge of disadvantageous pay inequality and then show that an implicit

justification, observing higher productivity workers being paid more, also drives out this negative

reaction. Differently from these two studies, we study how a particular contextual factor, the

content of work, can interact with both wage inequality and income inequality to induce particular

cognitive processes capable of ameliorating the negative consequences of inequality.

6 Concluding Remarks

Summing up, our data paint an interesting picture, one in which the content of employment is a

crucial factor in understanding the incentive effects of inequality. When work is ability-intensive,

by making pay inequality salient well-documented subconscious psychological processes may spring

20

into action to reconcile unequal treatment with justice and fairness, even when this entails self-

discriminatory beliefs about one’s performance or ability Benabou and Tirole (2006); Butler (2014);

Lerner (1965). We document a three-fold pattern associated with revealing pay inequality in an

ability-intensive environment: a reduction in shirking; a reduction in charitable donations; and a

strengthening of JWBs. This pattern makes sense: in a world in which people generally get what

they deserve, caring about others’ welfare is less urgent but deserving one’s own outcomes become

more imperative.

References

Abeler, J., S. Altmann, S. Kube, and M. Wibral (2010): “Gift exchange and workers’

fairness concerns: when equality is unfair,” Journal of the European Economic Association, 8,

1299–1324.

Abeler, J., D. Nosenzo, and C. Raymond (2016): “Preferences for truth-telling,” .

Adams, J. S. (1963): “Towards an understanding of inequity.” The Journal of Abnormal and

Social Psychology, 67, 422.

Akerlof, G. A. and W. T. Dickens (1982): “The Economic Consequences of Cognitive Disso-

nance,” The American Economic Review, 72, 307–319.

Akerlof, G. A. and J. L. Yellen (1990): “The fair wage-effort hypothesis and unemployment,”

The Quarterly Journal of Economics, 105, 255–283.

Angelova, V., W. Guth, and M. G. Kocher (2012): “Co-employment of permanently and

temporarily employed agents,” Labour Economics, 19, 48–58.

Bartling, B. and F. A. Von Siemens (2011): “Wage inequality and team production: An

experimental analysis,” Journal of Economic Psychology, 32, 1–16.

Baumeister, R. F. and J. Juola Exline (1999): “Virtue, personality, and social relations:

Self-control as the moral muscle,” Journal of personality, 67, 1165–1194.

Benabou, R. and J. Tirole (2006): “Belief in a just world and redistributive politics,” The

Quarterly journal of economics, 121, 699–746.

Benabou, R. and J. Tirole (2016): “Mindful economics: The production, consumption, and

value of beliefs,” Journal of Economic Perspectives, 30, 141–64.

21

Bolton, G. and P. Werner (2016): “The influence of potential on wages and effort,” Experi-

mental Economics, 19, 535–561.

Bolton, G. E. and A. Ockenfels (2000): “ERC: A theory of equity, reciprocity, and competi-

tion,” American economic review, 90, 166–193.

Bracha, A., U. Gneezy, and G. Loewenstein (2015): “Relative pay and labor supply,”

Journal of labor economics, 33, 297–315.

Branas-Garza, P., M. Bucheli, M. P. Espinosa, and T. Garcıa-Munoz (2013): “Moral

cleansing and moral licenses: experimental evidence,” Economics & Philosophy, 29, 199–212.

Breza, E., S. Kaur, and Y. Shamdasani (2018): “The morale effects of pay inequality,” The

Quarterly Journal of Economics, 133, 611–663.

Butler, J. V. (2014): “Inequality and relative ability beliefs,” The Economic Journal, 126, 907–

948.

Card, D., A. Mas, E. Moretti, and E. Saez (2012): “Inequality at work: The effect of peer

salaries on job satisfaction,” American Economic Review, 102, 2981–3003.

Charness, G., R. Cobo-Reyes, J. A. Lacomba, F. Lagos, and J. M. Perez (2016): “Social

comparisons in wage delegation: Experimental evidence,” Experimental Economics, 19, 433–459.

Charness, G. and P. Kuhn (2007): “Does pay inequality affect worker effort? Experimental

evidence,” Journal of labor economics, 25, 693–723.

Charness, G. and M. Rabin (2002): “Understanding social preferences with simple tests,” The

Quarterly Journal of Economics, 117, 817–869.

Chen, D. L., M. Schonger, and C. Wickens (2016): “oTree–An open-source platform for

laboratory, online, and field experiments,” Journal of Behavioral and Experimental Finance, 9,

88–97.

Clark, A. E., D. Masclet, and M. C. Villeval (2010): “Effort and comparison income:

Experimental and survey evidence,” ILR Review, 63, 407–426.

Cohn, A., E. Fehr, B. Herrmann, and F. Schneider (2014): “Social comparison and effort

provision: Evidence from a field experiment,” Journal of the European Economic Association,

12, 877–898.

22

Di Tella, R., S. Galiant, and E. Schargrodsky (2007): “The formation of beliefs: evidence

from the allocation of land titles to squatters,” The Quarterly Journal of Economics, 122, 209–

241.

Dube, A., L. Giuliano, and J. Leonard (2019): “Fairness and Frictions: The Impact of

Unequal Raises on Quit Behavior,” American Economic Review, 109, 620–663.

Dufwenberg, M. and M. A. Dufwenberg (2018): “Lies in disguise–A theoretical analysis of

cheating,” Journal of Economic Theory, 175, 248–264.

Falk, A. and A. Ichino (2006): “Clean evidence on peer effects,” Journal of labor economics,

24, 39–57.

Fehr, E. and K. M. Schmidt (1999): “A theory of fairness, competition, and cooperation,” The

quarterly journal of economics, 114, 817–868.

Festinger, L. (1957): A theory of cognitive dissonance, vol. 2, Stanford university press.

Fischbacher, U. and F. Follmi-Heusi (2013): “Lies in disguise–an experimental study on

cheating,” Journal of the European Economic Association, 11, 525–547.

Gachter, S. and C. Thoni (2010): “Social comparison and performance: Experimental evidence

on the fair wage–effort hypothesis,” Journal of Economic Behavior & Organization, 76, 531–543.

Gibbons, R. (1987): “Piece-rate incentive schemes,” Journal of Labor Economics, 5, 413–429.

Gill, D., V. Prowse, and M. Vlassopoulos (2013): “Cheating in the workplace: An ex-

perimental study of the impact of bonuses and productivity,” Journal of Economic Behavior &

Organization, 96, 120–134.

Gneezy, U., A. Imas, and K. Madarasz (2014): “Conscience accounting: Emotion dynamics

and social behavior,” Management Science, 60, 2645–2658.

Gneezy, U., A. Kajackaite, and J. Sobel (2018): “Lying Aversion and the Size of the Lie,”

American Economic Review, 108, 419–53.

Godechot, O. and C. Senik (2015): “Wage comparisons in and out of the firm. Evidence from a

matched employer–employee French database,” Journal of Economic Behavior & Organization,

117, 395–410.

Greiner, B., A. Ockenfels, and P. Werner (2011): “Wage transparency and performance:

A real-effort experiment,” Economics Letters, 111, 236–238.

23

Gross, T., C. Guo, and G. Charness (2015): “Merit pay and wage compression with produc-

tivity differences and uncertainty,” Journal of Economic Behavior & Organization, 117, 233–247.

Hennig-Schmidt, H., A. Sadrieh, and B. Rockenbach (2010): “In search of workers’ real

effort reciprocity–a field and a laboratory experiment,” Journal of the European Economic As-

sociation, 8, 817–837.

Konow, J. (2003): “Which is the fairest one of all? A positive analysis of justice theories,” Journal

of economic literature, 41, 1188–1239.

Koszegi, B. (2014): “Behavioral contract theory,” Journal of Economic Literature, 52, 1075–1118.

Lazear, E. P. (2000): “The power of incentives,” American Economic Review, 90, 410–414.

——— (2018): “Compensation and Incentives in the Workplace,” Journal of Economic Perspec-

tives, 32, 195–214.

Lazear, E. P. and P. Oyer (2007): “Personnel economics,” Tech. rep., National Bureau of

economic research.

Lerner, M. (2013): The Belief in a Just World: A Fundamental Delusion, Springer Science &

Business Media.

Lerner, M. J. (1965): “Evaluation of performance as a function of performer’s reward and at-

tractiveness.” Journal of Personality and Social Psychology, 1, 355.

Mazar, N., O. Amir, and D. Ariely (2008): “The dishonesty of honest people: A theory of

self-concept maintenance,” Journal of marketing research, 45, 633–644.

Nosenzo, D. (2013): “Pay secrecy and effort provision,” Economic inquiry, 51, 1779–1794.

Prendergast, C. (1999): “The provision of incentives in firms,” Journal of economic literature,

37, 7–63.

Spence, M. (1973): “Job Market Signaling,” The Quarterly Journal of Economics, 87, 355–374.

Stiglitz, J. E. (1975): “Incentives, risk, and information: notes towards a theory of hierarchy,”

The Bell Journal of Economics, 552–579.

24

Tables and Figures

Table 2: Descriptive Statistics

Treatment Pay scheme NPhase 1 Phase 2 Phase 3

income score quantity quality lying donations

IP-II-EILow Pay 27 5.76 28.81 7.67 0.99 0.04 2.33High Pay 27 10.06 28.07 7.67 0.98 0.04 2.52

IP-II-AILow Pay 32 4.18 20.91 6.91 0.88 0.16 3.13High Pay 24 7.29 19.38 7.75 0.92 0.13 3.42

VP-II-EILow Pay 26 6.00 30.00 7.88 0.96 0.08 3.04High Pay 27 10.60 29.89 8.26 0.96 0.11 3.04

VP-II-AILow Pay 24 4.03 20.17 7.75 0.96 0.25 1.63High Pay 34 7.70 19.03 7.74 0.96 0.09 3.44

IP-VI-EILow Pay 22 5.74 28.68 7.59 0.97 0.09 3.73High Pay 30 10.30 29.07 7.93 0.95 0.03 1.87

IP-VI-AILow Pay 28 4.01 20.04 7.64 0.94 0.14 3.18High Pay 21 7.52 20.00 6.76 0.95 0.10 2.33

VP-VI-EILow Pay 30 5.53 27.67 6.73 0.96 0.10 3.23High Pay 23 10.66 31.30 8.09 0.94 0.04 2.35

VP-VI-AILow Pay 20 3.85 19.25 6.85 0.92 0.10 1.90High Pay 28 7.63 19.79 7.25 0.96 0.11 2.14

Notes: [1] Treatments are denoted using a triple (V/I)P-(V/I)I-(E/A)I, where the first component refersto whether P(ay) inequality was visible or not, the second component refers to the visibility of I(ncome)inequality and the last component refers to the task type – A(bility) or E(ffort) intensive. [2] Columns 4-8present averages of the variables listed in the column headings. “Phase 1 income” refers to participants’earnings from the phase 1 task only; “Phase 1 score” refers to the number of correct questions (out of amaximum of 48) on the phase 1 task; “Phase 2 quantity” refers to the number of scantrons submitted;“Phase 2 quality” refers to the average accuracy of submitted scantrons; “Phase 2 lying” is the proportionof participants who misreported the number of Scantrons they submitted; “Phase 3 donations” refers tothe total amount donated across both charity options (American Red Cross and the South Plains FoodBank).

25

Table 3: OLS Estimates

Phase 2 Phase 3JWB

quantity quality lying donations

HP -0.907* -0.045 0.072 -0.504 -0.600*(0.5197) (0.0293) (0.0718) (1.021) (0.309)

VW 0.170 -0.033 0.042 0.799 -1.024**(0.6641) (0.0292) (0.0491) (0.790) (0.391)

VI -0.089 -0.009 0.062 1.480 -0.579(0.4384) (0.0154) (0.0677) (1.314) (0.616)

AI -0.504 -0.091*** 0.076 1.042 -1.114***(0.5747) (0.0276) (0.0590) (0.725) (0.315)

HPxAI 1.265* 0.050 -0.022 0.565 0.974***(0.7122) (0.0416) (0.0779) (1.350) (0.331)

VWxAI 0.760 0.113** 0.065 -2.065** 1.540***(0.7933) (0.0436) (0.1178) (0.851) (0.407)

VIxAI 0.811 0.066* -0.075 -1.297 0.930(0.6836) (0.0378) (0.0978) (1.666) (0.623)

Phase 1 income 0.163* 0.009 -0.020 0.120 -0.0543(0.0905) (0.0067) (0.0122) (0.153) (0.0501)

Phase 2 quantity 0.008 -0.005 -0.101 0.0250(0.0074) (0.0075) (0.0766) (0.0277)

Constant 5.715*** 0.936*** 0.291** -2.109 3.891***(1.0967) (0.0931) (0.1275) (2.325) (0.894)

Observations 419 419 419 419 416R-squared 0.121 0.073 0.064 0.083 0.104

Notes: [1] Each column reports an OLS regression with the dependent variable labeled inthe column heading. “Phase 2 quantity” refers to the number of scantrons submitted; “Phase2 quality” refers to the average accuracy of submitted scantrons; “Phase 2 lying” is the pro-portion of participants who misreported the number of Scantrons they submitted; “Phase3 donations” refers to the total amount donated across both charity options (American RedCross and the South Plains Food Bank); “JWB” is the participant’s response to the Just WorldBeliefs question. [2] Experimental controls are as follows: HP is an indicator variable for thehigh pay scheme, VW is an indicator for visible pay inequality, VI is an indicator variable vis-ible (experimental) income inequality, while AI indicates the ability-intensive task. Indicatorvariables take the value of 1 or 0. [3] Included in each estimate, but not reported for readabil-ity, are additional interactions terms: HPxVWxAI, HPxVIxAI, VWxVIxAI, HPxVWxVIxAI.None of these terms is consistently significant, except as described in the text. [2] Specifica-tions in all columns include additional controls that are not reported for readability. Thesecontrols are: gender, age, gpa and family income as well as dummies for the type of fillerquestions inserted between stages 1 and 1.5, which were randomly assigned and asked aboutgender, political affiliation or a purchasing decision. [4] Robust standard errors clustered bysession appear in parentheses, *** p < 0.01, ** p < 0.05, * p < 0.10.

26

Table 4: OLS Estimates EI Only

Phase 2 Phase 3JWB

quantity quality lying donations

HP -1.169 -0.0562 0.0995 -0.397 -0.0857(0.701) (0.0367) (0.0835) (1.010) (0.437)

VW 0.164 -0.0317 0.0525 0.853 -0.973**(0.613) (0.0332) (0.0458) (0.830) (0.399)

VI -0.0888 -0.0160 0.0684 1.489 -0.584(0.440) (0.0142) (0.0705) (1.263) (0.631)

HPxVW 0.307 0.00808 0.0249 -0.384 1.024***(0.675) (0.0186) (0.0885) (0.778) (0.297)

HPxVI 0.496 -0.00516 -0.0525 -1.663 0.511(0.755) (0.0189) (0.0615) (1.640) (0.629)

VWxVI -0.963 0.0130 -0.0659 -1.669 1.178(0.909) (0.0464) (0.116) (1.624) (0.689)

HPxVWxVI 0.533 -0.0103 -0.0146 1.597 -1.317(1.020) (0.0447) (0.133) (1.889) (0.869)

Phase 1 income 0.233 0.0104 -0.0247 0.0964 -0.174**(0.156) (0.00938) (0.0146) (0.178) (0.0807)

Phase 2 quantity -0.00759 -0.000271 -0.122 -0.0322(0.00785) (0.00624) (0.0742) (0.0296)

Constant 4.230** 0.984*** 0.335** -1.596 5.516***(1.442) (0.0987) (0.141) (2.929) (0.850)

Observations 212 212 212 212 209R-squared 0.118 0.056 0.063 0.108 0.161

27

Table 5: OLS Estimates AI only

Phase 2 Phase 3JWB

quantity quality lying donations

HP 0.707 -0.00932 0.0668 0.280 -0.144(0.582) (0.0604) (0.106) (1.612) (0.259)

VW 0.996** 0.0748** 0.118 -1.420** 0.412***(0.417) (0.0328) (0.117) (0.520) (0.122)

VI 0.745 0.0528 -0.01000 0.00347 0.286(0.513) (0.0371) (0.0762) (1.219) (0.205)

HPxVW -1.166*** -0.0423 -0.160 1.339 0.207(0.340) (0.0420) (0.113) (1.316) (0.469)

HPxVI -1.736** -0.00686 -0.0441 -1.609 0.148(0.634) (0.0437) (0.113) (1.853) (0.292)

VWxVI -1.466** -0.0836 -0.131 0.0409 -0.221(0.613) (0.0564) (0.149) (1.447) (0.389)

HPxVWxVI 2.236** 0.0491 0.191 0.140 -0.843(0.948) (0.0873) (0.172) (1.799) (0.710)

Phase 1 income 0.0742 0.0101 -0.0214 0.0740 0.0784(0.0809) (0.00839) (0.0219) (0.271) (0.0517)

Phase 2 quantity 0.0264** -0.0127 -0.0879 0.120**(0.0105) (0.0136) (0.155) (0.0425)

Constant 7.168*** 0.714*** 0.328 -0.868 0.368(1.156) (0.161) (0.262) (3.345) (1.349)

Observations 207 207 207 207 207R-squared 0.162 0.190 0.074 0.087 0.117

28