The Morale Effects of Pay Inequality

THE MORALE EFFECTS OF PAY INEQUALITY

EMILY BREZA†, SUPREET KAUR‡, AND YOGITA SHAMDASANI∓

Abstract. The idea that worker utility is affected by co-worker wages has potentiallybroad implications for the labor market—for example, through wage compression, wagerigidity, firm boundaries, and the distribution of earnings. We use a field experiment withIndian manufacturing workers over a one-month period to test whether relative pay com-parisons affect effort and labor supply. In this setting, workers are paid a flat daily wage.They perform individual production tasks, but are organized into distinct teams—definedby the type of product they produce. We randomize teams to receive either compressedwages (where all teammates earn the same random wage) or heterogeneous wages (whereeach team member is paid a different wage according to his baseline productivity rank).For a given absolute wage level, workers reduce output by 0.31 standard deviations if theyare on a team where they are paid less than their peers. In contrast, workers do notincrease output when they are paid more than their peers. The perceived justification forpay differences plays an important role in mediating negative morale effects. Specifically,lower relative pay does not cause effort reductions when co-worker output is highly observ-able, or when one’s higher-paid co-workers are substantially more productive than oneself(in terms of baseline productivity). These findings suggest that relative pay concerns areanother channel through which observability and verifiability of output could affect paypolicies and equilibrium wage structure.

JEL Classification Codes: J3, O15, D03, E24Keywords: Reference dependence, peer comparisons, wage compression

Date: October 2015.We thank James Andreoni, Dan Benjamin, Pascaline Dupas, Edward Glaeser, Uri Gneezy, Seema Jayachan-dran, Lawrence Katz, David Laibson, Ulrike Malmendier, Bentley MacLeod, Sendhil Mullainathan, MarkRosenzweig, Bernard Salanie, and Eric Verhoogen for their helpful comments. Arnesh Chowdhury, MoharDey, Piyush Tank, and Deepak Saraswat at JPAL South Asia provided excellent research assistance. Wegratefully acknowledge financial support from the National Science Foundation, the IZA Growth and La-bor Markets in Low Income Countries (GLM-LIC) program, and the Private Enterprise Development forLow-Income Countries (PEDL) initiative.†Columbia Business School. Email: [email protected].‡Department of Economics and SIPA, Columbia University. Email: [email protected].∓Department of Economics, Columbia University. Email: [email protected].

1

mailto:[email protected]



THE MORALE EFFECTS OF PAY INEQUALITY 2

1. Introduction

In traditional agency models, workers care about only their own wage levels when makingeffort and labor supply decisions. However, a long tradition in economic thought—as wellas in psychology, sociology, and human resource management—has advanced the notionthat individuals also care about their pay relative to that of their co-workers.1 This impliesthat relative pay may be a compensating differential—affecting utility and therefore thewillingness to accept work at a given absolute pay level. In addition, these utility effectscould translate into changes in worker effort, and therefore output. A growing literatureemphasizes the potential importance of such “morale” effects in the presence of incompletecontracting (Bewley 1999, Fehr et al. 2009).2

If relative pay concerns affect effort and labor supply, this could influence many features ofthe labor market. For example, it could help explain why wage compression—when wagesvary less than the marginal product of labor—appears prevalent in both poor and richcountries (Dreze and Mukherjee 1989, Fang and Moscarini 2005, Charness and Kuhn 2007).Such considerations have also been proposed as a micro-foundation for wage rigidity, withconsequences for unemployment and volatility (Akerlof and Yellen 1990). In addition, theycould affect how heterogeneous workers are sorted into firms, possibly leading some firmsto specialize in higher or lower ability workers (e.g., Frank 1984). They could also influencewhether labor is contracted through the external market or organized within firm boundaries(e.g., Nickerson and Zenger 2008).3 Finally, the effects of relative pay concerns on theseaspects of the labor market could themselves be dependent on features of production–forexample, if the observability or verifiability of output affects whether pay differences areperceived as acceptable (Bracha et al. 2015).

These implications rest on the presumption that workers do indeed care about relativepay. In this paper, we empirically test the validity of this view using a field experiment withmanufacturing workers. We conceptualize relative pay concerns as reference dependence inco-worker wages, with effort effects coming from reciprocity under incomplete contracting.Under most prominent formulations of reference dependence, being paid less than one’sreference point will decrease utility and effort. Earning more than one’s reference point mayweakly increase effort—for example, under loss-aversion (Kahneman and Tversky 1979). Wedevelop a design to enable comparisons of workers who earn the same absolute wage, sothat potential reference dependence effects arise only from changes in co-worker wages.We formulate primary tests of our predictions by constructing cases where the relationship

1See, for example, Marshall (1890), Veblen and Almy (1899), Hicks (1932),Duesenberry (1949), Easterlin(1974), Hamermesh (1975).2Many employment arrangements are characterized by some degree of incomplete contracting; very fewoccupations are solely governed by explicit performance incentives such as piece rates (MacLeod and Parent1999). Bewley (1999) documents that firm managers consider relative pay concerns to be important forworker motivation.3See Fehr et al. (2009)for an overview of how fairness considerations may affect the labor market.


between own wage and the reference point is clear: when a worker earns strictly more or lessthan all of his peers. We use additional tests to better understand the nature of referencedependence. For example, we examine what happens when a worker earns the median wage.In addition, we explore whether perceived justifications for pay differences affect whetherfairness violations are triggered.

In the experiment, 378 workers in Odisha, India are employed full-time for one month inseasonal manufacturing jobs—a prominent source of employment in the area. They work insmall factories, where they are organized into distinct teams, with three workers per team.All team members produce the same exact product (e.g. rope), while every team withina factory produces a different product (e.g. rope vs. brooms). One’s teammates thereforeconstitute a natural and salient reference group for pay comparisons.4 Note that there isno joint production in teams—production is an individual activity, enabling us to measureeach worker’s individual output.5 All workers are paid a flat daily wage, in accordance withthe typical pay structure in the area.

To induce exogenous variation in co-worker pay, each team is randomized into one of twopay structures. In the Heterogeneous pay condition, each team member is paid a differentwage—wHigh, wMed, or wLow—in accordance with his respective productivity rank withinthe team (as determined by baseline productivity levels). These relative pay differences arefairly modest: the difference between each wage level is less than 5%. In the Compressedpay condition, all teammembers are paid the exact same wage; we randomly assign the levelof this wage to be wHigh, wMed, or wLow. Each team is assigned to either the Heterogeneousor one of the three Compressed wage treatments.6

To test whether perceived justifications mediate morale effects, we incorporate two ad-ditional sources of variation into our design. First, while wage differentials are fixed at 5%,underlying baseline productivity is continuous. This induces variation in “actual fairness”:the extent to which pay differences among co-workers overstate productivity differences.Second, we generate variation within and across treatments in “perceived fairness”: whethera team engages in a production task for which it is easy to observe co-worker output.7

4This is consistent with Card et al. (2012), for example, who found that relative pay comparisons werestronger within university departments than across departments.5In order to accomplish this, we pay to hire extra staff to measure each worker’s output at the end of eachday.6At the beginning of the baseline (i.e. “training”) period, workers are told that they will receive a wageincrease on a pre-specified date, and that the size of this increase may depend on their baseline productivity.Once they are randomized into their wage treatment on this date, no additional future wage changes arepossible. This shuts down dynamic incentive effects; see below for a discussion of this.7We quantified the observability of each of the ten possible production tasks ex ante using a pilot. Adifferent sample of workers–all of whom were on teams with Compressed wages–were asked after three weeksof working together to rank their output relative to that of their teammates. We use the mean accuracy ofthese responses for a given task as the observability value for that task. In the experiment, we stratify wagetreatments by production task, ensuring variation in task observability within each wage treatment cell.


We find that for a given absolute pay level, output declines by 0.31 standard deviations(approximately 29%) on average when a worker is paid less than both his co-workers.8 Thisis accompanied by a 9 percentage point decrease in attendance—indicating that employeesgive up approximately 9% of their earnings to avoid a workplace where they are paid lessthan their peers. We estimate that the attendance decrease accounts for 50% of the overalldecline in output. These negative effects persist over the duration of the employment period,with some evidence that they become stronger in later weeks. In contrast, we find littleevidence that performance improves if a worker is paid more than both his peers: averageeffects on output are statistically indisguishable from zero, and attendance actually declines.In addition, we similarly cannot reject that there is no impact on the output or attendanceof Heterogeneous pay workers who receive the median wage on their team.

Perceived justifications play an important role in mitigating the negative treatment effectof low relative pay. When teammates’ baseline productivity levels are farther apart—so thatdifferences in productivity swamp differences in wages—we find no evidence for negativeeffects of being paid less than one’s peers. The same pattern holds for production taskswhere co-worker output is more observable—so that workers can easily see that their higherpaid peers are more productive than themselves.

These findings raise the question of whether the source of fairness violations comes fromcomparisons of a worker’s own ratio of pay/productivity relative to that of referent others(Adams 1963). Under this view of the reference point, no difference in the level of pay isrequired to trigger a violation – individuals may resent earning the same wage for differentlevels of productivity. However, we find no evidence that within Compressed teams, workersbehave any differently when the relative baseline productivity gap between themselves andtheir peers changes in magnitude. These findings suggest that in our particular setting,workers appear to compare differences in pay in levels. When lower relative pay levelstrigger a potential fairness violation, this is mitigated if lower pay is clearly justified byrelative productivity.9

Finally, we augment our findings on morale effects with supplementary endline activities.On the last day of employment, workers play cooperative team games, which have no benefitto the firm. The games require coordination to solve puzzles in pairs, and performance isincentivized with a piece rate prize at the pair-level. Workers may be randomly paired withsomeone from their own team or from another team. Compared to workers from Compressedteams, workers from Heterogeneous teams perform worse if they are paired with someone8This estimate is based on comparing low-rank workers in Heterogeneous (who are paid wLow) with low-rankworkers in Compressed teams where everyone is paid wLow. In general, all relative pay effects are identifiedoff pairwise comparisons of workers in Heterogeneous who receive a given wage and those on Compressedteams who have the same absolute wage level and same rank (i.e. average productivity).9Consistent with this, the (lack of) treatment effects of Heterogeneous for workers earning wHigh andwMedium, respectively, are not affected by productivity differentials or task observability. If there is nofairness violation in pay levels for these groups, then perceived justifications therefore do not play anyadditional mediating role.


from their own team (relative to being paired with a stranger from another team). Thissuggests a decrease in the cohesion and cooperativeness of Heterogeneous team members.

While our findings indicate that pay differences can cause substantial effort reductions inour setting, one cannot draw conclusions about optimal pay structure. One potential benefitto firms of differential pay is dynamic incentives: workers know that if they work hard now,it could lead to higher pay in the future. Our study design shuts down this channel becauseafter the baseline period, there is no further chance of wage changes. This is important forour specific purpose: cleanly isolating the morale effects of relative pay differences. It is alsoa realistic feature of seasonal and other contract jobs—a very common form of employmentamong the workers in our study. More generally, in deciding on optimal pay structure, afirm would weigh any potential costs of differential pay (e.g. morale reductions) against thepotential benefits (e.g. dynamic incentive or selection effects). Our findings indicate thatworkers’ relative pay concerns could affect this calculus.

This study builds on the literature on relative pay comparisons in the workplace. Tworecent experimental studies with workers have examined relative pay concerns. First, Cardet al. (2012) document that University of California employees report higher job dissatis-faction on surveys when they find out that they are paid less than their co-workers. Second,Cohn et al. (2012) show that relative random pay cuts matter more than absolute pay cutsfor effort; these effects persist strongly over a six-hour period. Our results are consistentwith those in both these studies.10 In addition, our work relates to the broader literatureon the effect of fairness preferences on effort provision under incomplete contracting, suchas gift-exchange (Akerlof 1982).11

Our study advances the literature by finding substantial impacts of relative pay compar-isons on effort and labor supply. In our experiment, wage differences reflect interpersonal10A small number of laboratory studies have explored relative pay comparisons using gift exchange games,with mixed results (Charness and Kuhn 2007, Gatcher and Thoni 2010, Bartling and von Siemens 2011).Notably, Bracha et al. (2015) find support for relative pay concerns, and for the importance of perceivedjustifications. Related laboratory experiments have examined the effects of rank (Brown et al. 2008, Clarket al. 2010, Kuziemko et al. 2011). In addition, several studies examine relative pay concerns using ob-servational data. Dube, Guliano, and Leonard (2015) document an increase in quitting behavior when anindividual’s pay increase is lower than that of her co-workers. Mas (2015) offers evidence that mandatorypay disclosure for municipal employees led to pay cuts and subsequent quits for high earners; he interpretsthis as public aversion to high compensation. A number of other studies are consistent with a relationshipbetween relative pay and worker satisfaction or behavior (Levine 1993, Pfeffer and Langton 1993, Clark andOswald 1996, Hamermesh 2001, Kwon and Milgrom 2008, Mas 2008, Rege and Solli 2013, Dube et al. 2015).Related work has explored links between relative income and other outcomes, such as happiness (Frey andStutzer 2002, Luttmer 2005), health (e.g. Marmot 2004), and reward-related brain activity (e.g. Fliessbacket al. 2007).11These studies test for reference points that are determined by a worker’s own absolute past wage (orexpected wage). A large number of studies find gift exchange in the lab (e.g., Fehr et al. 1993). Fieldevidence, however, is more limited; existing field experiments generally find limited support for positivereciprocity (i.e., effort increases from wage increases), while the evidence on negative reciprocity (i.e. effortreductions from wage cuts) is more mixed over a one-day period (Gneezy and List 2006, Kube et al. 2013,Esteves-Sorenson & Macera 2015, DellaVigna et al. 2015). For reviews of this literature, see Fehr et al.(2009), List (2009), and Charness and Kuhn (2011).


differences in worker productivity; this reflects why wage differences may arise in the labormarket and is important given laboratory findings that justifications can undo fairness vio-lations (Falk et al. 2008, Bracha et al. 2015). We augment these findings with the first pieceof field evidence that perceived justifications play an important role in mediating moraleeffects. This has bearing on understanding, for example, why wage compression may arisein some settings or occupations and not in others. Finally, workers make decisions for a jobfrom which they derive full-time earnings over the one-month study period. This helps ver-ify that impacts from reference dependence do not disappear once the novelty of treatmentswears off (Gneezy and List 2006, Levitt and List 2007).

While our results indicate that relative pay concerns can affect output in large magni-tudes, they also suggest that negative morale effects can be mitigated when the justificationfor differential pay is extremely transparent. These findings suggest that firms may haveseveral potential tools at their disposal to manage morale in the presence of pay disper-sion. For example, technologies that make it easier to quantify worker productivity couldhave aggregate output benefits—not just through a reduction in moral hazard, but alsothrough improved morale. Firms could also potentially alter the organizational structure ofthe workplace itself—through job titles, physical co-location of similar workers, or the con-struction of “teams” (as we did in the experiment)—to affect who a worker views as beingin her reference group. The extent to which firms can and do make use of such strategieshas the potential to affect wage compression, wage rigidity, and firm boundaries. Whilespeculative, such possibilities suggest a variety of ways through which relative pay concernscould affect pay structure, organizational arrangements, unemployment, and other labormarket outcomes.

The remainder of the paper proceeds as follows. Section 3 describes our empirical settingand experimental design, Section 4.2 presents our results, Section 5 discusses threats tovalidity, and Section 6 concludes.

2. Framework

2.1. The Worker’s Maximization Problem. We build on the model of DellaVigna et al.(2015), which allows workers to have social preferences toward the firm. We extend theirmodel by allowing altruism to the firm to be a function of peer wages in additional to aworker’s own wage.

We assume that a worker i receives a take-it-or-leave-it wage offer wi from the firm andmakes two decisions: a) whether to work that day si ∈ {0, 1} and b) if si = 1, how mucheffort ei ≥ 0 to expend working. Effort is not contractible by the firm. If the worker choosesnot to work (i.e., si = 0), then he receives a stochastic outside option Ri = R + εi. Theworker’s payoff from working is

V (wi,w−i, ei) = wi − c (ei) +A (wi,w−i) ei.


This payoff is a function of the worker’s own wage, wi, the wages of the worker’s peers,w−i, and the effort choice ei. We assume that the worker pays a convex effort cost c (ei)also experiences altruism toward the firm, A (wi,w−i), that scales linearly with effort.

We conceptualize relative pay concerns as referencedependence in utility, where co-worker pay enters as an argument into the worker’s reference wage, wR (wi,w−i). Weincorporate this into worker utility in the following way

A (wi,w−i) = α1wi<wR(w−i) + β1wi>wR(w−i) + f (wi)

f (wi) captures the non-peer-dependent contributors toward firm altruism such as gift ex-change.12 If the worker is paid less than the reference wage, 1wi<wR(w−i) = 1, then there isan effect of α on the worker’s utility per unit of effort provided. Conversely, β captures theeffect on a worker’s utility per unit of effort from being paid more than his reference wage.

Prior work conceptualizing relative pay comparisons predicts that α < 0, that is, indi-viduals dislike being paid less than their peers (i.e., Adams 1963, Akerlof and Yellen 1990).The prediction on the sign and relative magnitude of β, however, varies. Preferences forstatus or advantageous inequality generate β > 0. Further, under loss aversion |α| < |β|.Inequality aversion, on the other hand, would lead to β < 0.

2.2. Labor Supply and Effort Decisions. For ease of exposition, we assume that c (e) =c e

2

2 , εi ∼ U (a, b), and α+ f (w) > 0, where w is the minimum wage paid by the firm. Thislast assumption on α ensures an interior choice of ei.

Lemma 2.1. Optimal effort and labor supply decisions are functions of both wi and w−i.

e∗ (wi ,wR (w−i)) =

[α1

wi<wR(w−i)+β1wi>wR(w−i)+f(wi)

]c and

s∗i (wi, wR (w−i)) = 1

wi +

(α1

wi<wR(w−i)+β1wi>wR(w−i)+f(wi)

)2

2c −R ≥ εi

.These optimal effort and labor supply policy functions give rise to our main predictions.

Proposition 2.2. If wR (w−i) = w̃ > wi then e∗ (wi , w̃)−e∗ (wi , wi) = αc and P (s∗

i (wi, w̃) = 1)−P (s∗

i (wi, wi) = 1) = α2+2αf(wi)2c(b−a) . Similarly, if wR (w−i) = ˜̃w > wi then e∗ (wi , ˜̃w

)−

e∗ (wi , wi) = βc and P

(s∗i

(wi, ˜̃w

)= 1

)− P (s∗

i (wi, wi) = 1) = β2+2βf(wi)2c(b−a) .

The proposition implies that when α < 0, workers respond to small positive deviations inthe reference wage relative to own wage by decreasing both effort and labor supply. Whenβ > 0, workers respond to small negative deviations in the reference wage relative to ownwage by increasing both labor supply and effort.

12Alternately, f (wi) could capture other drivers of effort including monitoring.


Our primary goal is to identify the signs of α and β using our experiment. Note thatif wR (w−i) were fully observable, then one could fully identify the signs of α and β byobserving the responses to small changes in peer wages that trigger 1wi<wR(w−i) = 1 and1wi>wR(w−i) = 1. While we do not take a strong ex ante stand on the functional form ofwR (w−i), we discuss below how we identify instances where, under most reasonable cases,1wi<wR(w−i) = 1 and 1wi>wR(w−i) = 1.

3. Experimental Design and Data

3.1. Experimental Design. We construct a design to test the above predictions withmanufacturing workers employed in small factories (see details below). In this setting,there is incomplete contracting on effort: in accordance with the typical pay structure inthe area, all workers are paid a flat daily wage for each day they come to work. This providesthem with some latitude to select both attendance (with implications for earnings) as wellas effort (with implications for output).

In order to test for reference-dependence in co-worker pay, we must first define, for eachworker, a clear reference group of peers. To accomplish this, within each factory, workersare organized into “teams” of three workers each. All team members produce the sameexact product (e.g. rope), while every team within a factory produces a different product(e.g. rope vs. brooms). Production is an individual activity—teammates sit together butdo not do any work jointly. Because each worker’s two teammates are the only other peopleat the factory making the same product, they are likely the most salient reference groupfor wage comparisons.

To construct tests for our core predictions, we design wage treatments that allow us tofix workers’ absolute pay levels, while creating variation in co-worker pay. Using baselineproductivity data, we rank each worker as the lowest, medium, or highest productivityworker within his respective team. Each team is then randomized into one of four wagestructures, as shown in Table 1:

• Heterogeneous: Each team member is paid according to his productivity rank withinthe team, where the rank is based on workers’ baseline productivity level. Thewages for the lowest, middle, and highest productivity workers are wL, wM , andwH , respectively.• Compressed_L: All team members are paid the same daily wage of wL.• Compressed_M : All team members are paid the same daily wage of wM .• Compressed_H : All team members are paid the same daily wage of wH .

These wage differences are fairly modest: the difference between each of the three wagelevels is approximately 5%. For each of the three ranks, this design enables us to comparegroups of workers who have the same average productivity levels and are paid the sameabsolute wages, but differ in the distribution of their co-workers’ wages.


To test for the role of justifications, we cross-cut the wage treatment with two additionalsources of variation. First, we vary actual fairness—the extent to which pay differentialsoverstate productivity differentials—by randomizing workers into teams. Because output iscontinuous while productivity rankings are discrete, this generates variation in how mucha worker’s productivity level differs from that of his teammates. This, in turn, enables usto examine how effects vary with changes in the ratio of {wage difference}/{productivitydifference} within and across wage treatments.

Second, we vary perceived fairness—the extent to which workers can observe co-workerproductivity. The ten production tasks in the factories differ in how easy it is to observethe output of one’s teammates. To ex-ante quantify the observability of each task, we usedpilot trials to measure whether workers could accurately rank their output relative to that oftheir teammates after three weeks of working together. In these trials, all teammates werepaid the same wage, so that wage was not a signal of productivity rank. We stratify wagetreatments by production task, enabling us to test for the effects of observability within andacross wage treatments. The randomization design is summarized in Figure 1.

3.2. Predictions. To test our first core prediction—a strict decrease in morale when wi <wR—we compare outcomes for Low rank workers in Heterogeneous with those in Com-pressed_L. Low rank workers in Heterogeneous are paid strictly less than all their team-mates. Under virtually any reference point that depends on co-worker pay levels, theywill feel more aggrieved than their counterparts in Compressed_L—who receive the sameabsolute pay of wL, but whose teammates earn the same as they do.

To test our second core prediction—asymmetric effects from deviations from the referencepoint—we compare High rank workers in Heterogeneous with those in Compressed_H. Wepredict a weak increase in effort and attendance for High rank workers in Heterogeneouswith those in Compressed_H.

Note that there is no clear ex ante prediction on the behavior of Medium rank workersin Heterogeneous relative to those in Compressed_M. Examining effects for this group willhelp provide insight into the nature of the reference point in our setting.

Finally, if the justification for pay differences matters for fairness violations, then theeffects will be mediated by the perceived and actual fairness of pay differences. The mag-nitude of treatment effects will be smaller when differences between a co-worker and hishigher paid peers is large, and when these differences are observable.13

3.3. Time Line, Recruitment and Survey Instruments. We detail the implementa-tion of our experiment in Figure 2. Workers are males between the ages of 18 and 55. Werecruit workers from villages surrounding rural factory sites in Odisha, India. We neverrecruit from the same village more than once. Each round requires hiring 30 workers; if

13Note that these predictions are consistent with those of the model in Fang and Moscarini (2005).


more than the required number of workers apply for the job within a village, we randomlyselect among applicants.

After recruitment, workers are randomly assigned into teams of 3 workers each at thestart of the round, and each team is assigned to a unique production task. These teamassignments are stable for the duration of the employment period. The round begins withtraining period for each of the tasks. During the first three days of training, factory stafffocus on making sure that the workers fully understand how to complete their tasks andhow to ensure a baseline level of quality demanded in the market. Typically after day four,output has reached a level of quality that can be sold in the market, and this is the time atwhich we begin recording individual output per worker.14

Although output is salable by day 4, we prolong the “training” period to 14 days toobtain accurate measures of baseline productivity. On day 10, workers are given individual,private feedback by the factory manager on their rank relative to the other members of theteam.15

On day 14, we randomly assign teams to treatments and inform each worker in privateof his new wage. We again deliver this message in private and remind each worker that thisis the wage that will be paid for the remainder of the contract period, that there will be nofuture opportunities for wage changes, and that there will be no future job opportunitiesafter the end of the contract period. The factories then run as usual under the treatmentwages until day 34. On day 35, the workers are surveyed and participate in laboratoryactivities, described below.

On the first day of employment, workers are shown calendars (which they see at workeveryday) that outline the dates of each of these events. They are also told on the first daythat their post-training wage may depend on their baseline productivity.

An implication of our design is that within a factory, different teams have differing paystructures and average pay levels. This is not odd since every team within a factory pro-duces a unique product, in conjunction with a distinct contractor. Also, note that factorymanagers maintain pay secrecy—each individual is privately told only his own wage; to theextent that we observe effects of relative pay differences, it is through self-disclosure amongteam members.

We collect information about the workers at several points in time. First, when wecompile a list of interested workers after the village meeting, we record information abouthousehold size and landholdings. Second, once workers have reported to the worksites for thefirst day of training, we collect a very short baseline survey to capture worker demographiccharacteristics including age, literacy, employment history, and basic information about

14Throughout the contract period, we hire extra workers to maintain accurate records of individual-leveloutput on a twice-daily basis.15This helps underscore to workers that we are paying attention to productivity. It also helps ensure oursubsequent wage treatment effects are not confounded by information revelation by the factory.


household assets. Third, throughout the period of employment, we collect daily measuresof worker attendance, production, as well as a subjective measure on the quality of workeroutput. If workers are absent we record the reason for the absence when they return.Fourth, on the final day of employment, we record information on worker activity andearnings on days for which they were absent, as well as on the labor market activities of theother members in the household. We also use a survey instrument to map out their socialnetworks with other workers at the worksite.

In Panel A of Table 2, we briefly describe the workers employed at our factories.16 Theyare all males between the ages of 18-55 who engage primarily in casual labor. 54% of theseworkers own land, with average landholdings of 0.7 acres. In addition, 70% sharecrop land,with an average land size of 1.2 acres of land. While many workers do own land or sharecropland, nevertheless, the land holdings are too small to generate year-round income. All ofthe workers primarily supply their labor to the daily labor market.

In Panel B of Table 2, we briefly describe the workers’ collective labor market experiences.71% of workers have worked on piece rates in the past. While 72% of workers report everreceiving wages that differ from the prevailing agricultural wage in their village (largely dueto piece rate work such as stone cutting or non-agricultural work such as construction),only 17% of workers report ever receiving a wage different from that of other laborers inthe village for the same task.

3.4. Regression Specifications. To test our key predictions, we compare outcomes be-tween individuals in the Heterogeneous and Compressed teams, holding fixed a worker’sproduction ranking rank and wage. Recall, from Table 1 that the most direct comparisonsare between the Low rank Heterogeneous worker with the Low rank Compressed_L worker,the Medium rank Heterogeneous worker with the Medium rank Compressed_M worker, andthe High rank Heterogeneous worker with the High rank Compressed_H worker. We referto this set of six worker treatment types as the “relevant group”. We begin by presenting asimplified regression specification in Equation 3.1 where we only include observations fromso-called “relevant” individuals. We then present an augmented regression in Equation 3.2where we identify the main treatment effects off of the relevant group, but use the otherworker treatment types in the regression to estimate the fixed effects and other controls.

The differences-in-differences regression equation restricting the sample to only the rele-vant worker treatment types can be written:

yi,j,t = βHetτG,j1(t≥0) + βM,HetτG,j1r(i)=M1(t≥0) + βH,HetτG,j1(r(i)=H)1(t≥0)(3.1)

+δHetτG,j + δM,HetτG,j1r(i)=M + δH,HetτG,j1(r(i)=H)

+ηk(j),t + ηw(j),t + ηr(i),t + εi,j,t

16Table 2 is based on a small subset of responses from a representative sample of workers. The majority ofthe baseline and endline surveys are still being entered.


Each observation captures information for worker i on team j on day t of the experimentalround. We rescale t such that t = 0 corresponds to the first day of the wage treatmentin each of the worksite-rounds. We focus on two outcome variables yi,j,t, attendance andproduction. attendancei,j,t is a binary variable capturing whether worker i is present onday t. The output variable, productioni,j,t measures standardized production in units of onestandard deviation of task-level production. To harmonize production across all ten tasksin the worksites, we use the pre-treatment data for each task to demean and standardizethe raw production data.

While treatments are randomized at time t = 0, we estimate differences-in-differencesregressions to economize on power. All key terms in the regression are interacted with thepost-treatment indicator 1(t≥0).

The four variables τl,j indicate the treatment status of team j, where l ∈ {G,CL,CM,CH}.G denotes the Heterogeneous wage treatment, and CL, CM , and CH denote CompressedLow, Medium, and High, respectively. It is also useful to define the vector of indicators(k (j) , w (j) , r (i)). Here, k (j) indexes the task k produced by team j; w (j) indexes theworksite-round w in which team j works; r (i) ∈ (H,M,L) denotes the pre-period rankingof worker i.

Our main prediction is that βHet < 0, that is Low rank workers in Heterogeneous teamswill decrease production and attendance relative to the Low rank workers in Compressed_Lteams. We also predict βH,Het > 0 and βHet + βH,Het ≥ 0, that the High-ranked work-ers in the Heterogeneous teams will produce weakly more than their counterparts in theCompressed_H teams. Recall that we do not have a strong prediction on βM,Het.

We also include fixed effects for task × experience(ηk(j),t

), worksite-round × time(

ηw(j),t), and worker ranking × pre/post

(ηr(i),t

)in our main regression specification.

Rather than discard the “irrelevant” observations, we instead specify the augmentedregression:

yi,j,t = βHetτG,j1(t≥0) + βM,HetτG,j1r(i)=M1(t≥0) + βH,HetτG,j1(r(i)=H)1(t≥0)(3.2)

+δHetτG,j + δM,HetτG,j1r(i)=M + δH,HetτG,j1(r(i)=H)

+θpostRel relevanti,j1(t≥0) + θpostRel,M (τG,j + τCM,j) 1(r(i)=M)1(t≥0)

+θpostRel,M (τG,j + τCH,j) 1(r(i)=H)1(t≥0) + θpreRelrelevanti,j

+θpreRel,M (τG,j + τCM,j) 1(r(i)=M) + θpreRel,M (τG,j + τCH,j) 1(r(i)=H)

+ηk(j),t + ηw(j),t + ηr(i),t + εi,j,t

The key regressors and fixed effects are identical to those in Equation 3.1, however, we alsoadd a set of indicators (i.e., the θ terms) for the relevant comparison groups both pre- andpost- wage change. We define relevanti,j to include all members of the relevant comparison


set:relevanti,j = 1

(τG + τCL ∗ 1r(i)=L + τCM ∗ 1r(i)=M + τCH ∗ 1r(i)=H = 1

)Note that the (τG,j + τCM,j) 1(r(i)=M) terms denote the Medium rank workers in Hetero-geneous or Compressed_M teams, and the (τG,j + τCH,j) 1(r(i)=H) terms denote the Highrank workers in the Heterogeneous or Compressed_H teams. We present the results ofestimating Equation 3.2 in Section 4.2.

.

4. Results

In what follows, we present results from fourteen worksite rounds, employing a total of378 workers. Because the experiment is still in the field, data collection is ongoing.

4.1. Knowledge of Co-worker Wages. Since managers maintained pay secrecy, the wagetreatments will only have power if workers discussed wages with each other. Using the end-line survey, we asked workers about their knowledge of their co-workers’ wages. AmongCompressed teams, 100% of the workers reported that each of their teammates had thesame wage as themselves. Among Heterogeneous teams, 92% of workers stated that theirteammates were paid differently than themselves, and could accurately report their team-mates wages 77% of the time. These findings indicate that overall, there was a substantialamount of discussion of wages within teams. In addition, the difference in information shar-ing between the Compressed and Heterogeneous teams, while only suggestive, is consistentwith awkwardness of discussing pay with teammates when there are wage differences.

4.2. Effects of Pay Disparity. Table 3 presents the estimation of Equation 3.2 on the fullsample of workers in each round. Columns (1)-(4) measure the effects on standardized pro-duction, while Columns (5)-(8) measure effects on attendance. Workers decrease productionand attendance when they are paid less than their team-mates, holding their absolute wagelevels fixed (βHet < 0). The output of Low wage workers declines by 0.372 standard devia-tions in response to the Heterogeneous treatment (significant at the 1% level). This effectis robust to the inclusion of team fixed effects (Col. 2) and individual fixed effects (Col.3, our preferred specification). The 0.311 standard deviation decrease in Col. 3 is equiv-alent to about a 29% reduction in output relative to the Compressed (control) treatmentmean. In addition, there is some evidence that on average, relatively lower-paid workersare 9 percentage points less likely to come to work (on a base of 92.8% attendance in theCompressed groups). Combining our administrative earnings data with endline survey dataon workers’ outside employment activities and wages when they were absent, we estimatethat these workers give up about 9% of their earnings to avoid a workplace where they arepaid less than their peers. This attendance and earnings effect is strengthened when we


limit analysis to non-paydays (Col. 8), since on paydays most employees come to work inorder to collect their payments.

In contrast, we find little evidence that performance improves when workers are paid morethan their peers. In Cols. (1)-(4), we cannot reject that there is no impact on high rankworkers in Heterogeneous teams relative to their counterparts on Compressed_H teams.The point estimates (given by βHet + βH,Het) are actually negative, though statisticallyindistinguishable from zero, as reported in the F-test p-values at the bottom of the table.In fact, the effects on attendance for these workers are negative, and significant at the10% level (Cols. 5-8). In addition, there is no evidence that medium rank workers on theHeterogeneous teams have different average output or attendance than their counterpartson the Compressed_M teams (given by βHet+βM,Het). Our results indeed suggest that theeffort and attendance responses from being paid less than one’s co-workers are much largerthan any positive effects from being paid more than one’s co-workers.

One natural question is whether the large negative impacts on the Low rank workerspersist or instead wear off over time. In Appendix Table 4, we separately estimate Equation3.2 over the first and second halves of the post-treatment period.17 We find large, negative,and statistically significant effects of the Heterogenous treatment on the Low rank workersin both samples. There is some evidence that effects actually strengthen over time.

Attendance and Effort Decomposition. We have documented that there are large, negativeeffects of Heterogeneous wages on production when a worker is paid less than his peers.This deleterious effect on production can occur through both the extensive (attendance)and intensive margins (effort conditional on attendance). The attendance effect poses prob-lems for identifying the intensive margin effect on effort. If dis-advantageous peer wagecomparisons affect some types of workers more than others, then running the regressionin Equation 3.2 conditioning on attendance may introduce a potentially severe selectionproblem. Thus, we do not run the conditional regression, but instead rely on two differentstrategies that provide suggestive support that both channels are at play.

First, we use a back-of-the-envelope calculation to decompose these effects. The meanoutput conditional on attendance for Low rank workers in the Compressed_L team is 1.69.The effect of Heterogeneous wages for the Low rank workers on attendance is -0.0925 per-centage points. If the full treatment effect on production were coming through attendance,then we would predict an output decrease of −0.0925× 1.69 = −0.156 standard deviations.This corresponds to 50.2% of the total effect on production.

Second, we estimate effects limiting analysis to only paydays. On paydays, workers almostalways come to work in order to collect their payments (overall attendance is 96%); reasonsfor missing work on those days are usually idiosyncratic such as illness (based on workerself-reports). If treatment status does not differentially affect attendance on paydays, then

17The full pre-treatment period is included in both regressions.


output effects on those days are less likely to be affected by treatment effects on attendance.We find that output effects hold robustly on paydays as well. While both these approachescertainly have limitations–due to the assumptions they impose–together, they suggest thatin our setting, morale effects operate through both the labor supply and effort channels.

4.3. The Role of Perceived Justifications. Table 5 examines whether morale effectsof pay differences are mediated by “actual fairness”–the difference in peers’ productivitylevels. For each low and medium rank worker in each team, we compute the differencein mean baseline productivity between that worker and his next higher-ranked peer. Wethen add interaction terms of measures of baseline productivity differences into the mainspecification. In Cols. 1-2, the interaction term is the continuous linear productivity dif-ference; in Cols. 3, it is a binary measure of whether the baseline productivity difference isabove the mean. All 3 columns paint a similar picture. Compared to their counterparts onthe Compressed_L teams, low rank workers in Heterogeneous strongly lower output whenthey are closer in productivity to their higher-paid medium rank peers. However, when thedifference in baseline productivity is large, we cannot reject that there is no effort reductionfor these workers. Col. 4 shows that these results are robust to including interactions oftreatment status with the worker’s own baseline productivity–so that the effects of interestare identified off of changes in co-workers’ productivity (rather than one’s own productiv-ity). Col. 5 shows that the results are robust to excluding the bottom decile of low-rankworkers (in terms of baseline productivity). This indicates that the results are not drivenby the fact that the least productive low-rank workers hit floor effects, and that is whatdrives the interaction effects of interest. Cols. 6-10 show that the effects on attendancefollow a similar pattern.

We next check for mediating effects of “perceived fairness”–whether co-worker outputis observable. Some of our tasks, by nature, are much more observable than others. Wequantify observability using data from 3-week pilot rounds with a different sample of workers(conducted before the start of this experiment). On the last day of these pilot rounds,we asked workers to rank their co-workers by productivity. In Figure 3, we present thecorrelations between the actual and survey rankings of the workers for each productiontask. These correlations constitute our measure of task observability.18 The correlationsrange from negative (bottom-coded at zero) to 0.87.

Table 6 presents a pattern of results similar to that in Table 5. Col. 1 adds interactionswith the continuous measure of observability, and Col. 2 adds interactions with a dummy forwhether the production task has above-average observability. The results in both columnsindicate that negative morale effects for low-rank Heterogeneous workers are concentratedin production tasks where it is more difficult to observe co-worker productivity. When this

18At the end of the pilots, we added two additional tasks. We followed a similar procedure with a separatesample of workers to quantify observability for these two additional tasks, and these were run concurrentlywith the experiment.


is highly observable, we cannot reject that there are no negative morale effects of being paidless than one’s peers. Taken together, the results suggest that the reactions of workers topay inequality do depend on the underlying circumstances.19

4.4. Team Cohesion Games. Finally, we augment our findings on morale effects withsupplementary endline activities. On the last day of employment, workers play cooperativeteam games, which have no benefit to the firm. The games require coordination to solvepuzzles in pairs, and performance is incentivized with a piece rate prize at the pair-level.Workers may be randomly paired with someone from their own team or from another team.Compared to workers from Compressed teams, workers assigned to Heterogeneous teamsperform worse if they are paired with someone from their own team (relative to beingpaired with a stranger from another team) (Table 7). This suggests a decrease in thecohesion and cooperativeness of Heterogeneous team members.

5. threats to Validity and Discussion

Internal validity concerns. Could an explanation other than relative pay comparisons ex-plain our findings? One potential confound is career concerns. Suppose that—even thoughwe stress to workers that this is a one-time seasonal temp job—workers supply effort partlyin hopes of increasing the probability of future employment. When a worker in Heteroge-neous observes he is paid less than his co-workers, he may believe the firm is less likely to hirehim in the future and therefore decrease effort. However, our design generates additionalpredictions that are not consistent with career concerns. First, we find that workers thatare close in productivity to their higher paid colleagues—and therefore more valuable to thefirm—are more likely to decrease effort. In contrast, in a career concerns model, workersthat are relatively further behind their colleagues should be more likely to believe theirchances of future employment are low, generating the opposite prediction. Second, giventhat we find large extensive margin effects on attendance, it is difficult to explain under acareer concerns model why workers are willing to give up full-time earnings (due to poorattendance) and sit at home unemployed. Similar arguments apply to the potential concernthat lower-paid Low rank workers in Heterogeneous decrease effort because the wage is asignal that helps them learn about their own type, affecting their future expectations.

Another potential issue is possible gift exchange effects from the fact that all workersreceive a wage increase after training. Such effects should be common across all treatments,since all workers receive a pay raise. If peer effects from more higher-paid (and thereforemore productive) peers increase own output, this would make it harder to detect our maineffects.

19Appendix Table 8 verifies that the productivity difference results (Table 5) and observability results(Table 6) are identified off of different sources of variation. There is substantial non-overlap among thesetwo measures in the sample.


Our design relies on the presumption that each worker’s reference group is comprisedof his two teammates. If workers instead compared themselves to those in other teams,it could create contamination across treatment groups. However, this should decrease thepotency of our treatments and make it harder to find our hypothesized effects. Given ourexperimental design, we believe that it is reasonable to expect that for someone makingrope, the other 2 people making rope (who sit with and work next to him daily) are a moresalient comparison group than those making incense sticks or brooms (who have their ownunique seating area and production task). This is consistent with the findings of Card et al.(2012), where workers cared about pay relative to others in their particular departments,and less about other departments in the same workplace.

Finally, our experiment is not well-suited to precisely disentangle the psychological mech-anism that drives effort reductions. For example, unfairness and envy are different emotionsthat could trigger a decrease in morale, and could micro-found reference dependence inutility. We do not take a stance on the underlying psychology—what matters for our inter-pretation is that the mechanism is something that operates through reference-dependencein co-worker pay.

Humiliation from being identified as a low productivity type in front of one’s peers,for example, is one competing explanation. In this world, the effects on output might stilloperate through a loss of worker morale, but not through reference dependence in co-workerpay. However, this class of mechanism is also unlikely to explain our full results. Under astory of humiliation, workers that are close in productivity to their higher-paid colleaguesshould experience less shame and motivation to decrease output. We find the oppositeresult. In addition, note that we maintain a policy of pay secrecy; if workers disclose thatthey are lower paid, then they do so voluntarily.

External validity concerns. Two important external validity concerns stem from whetherthe wage treatments appear unusual to the workers. First, since we have selected tasksin which output is measurable, firms could consider paying piece rates or some other formof explicit incentives. However, whether this makes sense will depend on the cost of themonitoring technology. In the experiment, we bear the considerable expense to hire extrastaff to measure each worker’s output daily. In addition, in the local context in which ourexperiment takes place, it is common for workers to receive flat wages even when output ismeasurable. For example, many retail goods are produced under both piece rates and underflat wages by firms in the study region. Similarly, some employers pay piece rates whileothers in the same village pay fixed daily wages to harvest a given crop. It is also the casethat under explicit incentives, quantity may improve but at the expense of quality—suchmultitasking problems are well documented.

Second, workers may have found it odd that some teams were paid based on baselineproductivity while others were paid equal wages. We developed our design to mitigate this


concern to the full extent possible. This is one of the driving reasons for having each teamproduce a unique task, which in turn is associated with its own unique contractor. Therewas thus no opportunity to compare one’s own wages with those of other teams producingthe same output in the same worksite.

A related issue is whether it is reasonable for the firm to pay differential wages based ontraining output (rather than ex-post output). This is also common in many settings. Forexample, firms usually set the pay of short-term consultants based on expected productivity.Even for salaried workers, pay is usually based on ex-ante expectations, with stickinessthroughout a worker’s tenure at the firm (Fehr et al. (2009))—this is not adjusted with newinformation on ex-post performance, but rather re-negotiated at infrequent intervals. Moregenerally, explicit incentives like piece rates based on ex-post output are not that commonin poor or rich countries (e.g. Dreze and Mukherjee 1989, MacLeod and Parent 1999).

One potential benefit to firms of differential pay is dynamic incentives: workers knowthat if they work hard now, it could lead to higher pay in the future. Our study designshuts down this channel since after the training period, there is no further chance of wagechanges. However, the objective of our study is not to isolate the optimal pay policy forfirms. Rather, our objective is to test whether relative pay comparisons affect effort—atopic on which there is limited field evidence, and which is currently ignored in mainstreamagency models of pay structure. The optimal pay policy for a firm would depend on weighingthe potential costs of differential pay (e.g. morale reductions) against the potential benefits(e.g. dynamic incentives). In addition, evidence on when differential pay is most likely todamage morale—for example when output is harder to precisely quantify or less observableby co-workers—can enhance our understanding of why we observe differential pay in someoccupations and not in others.

6. Conclusion

We find that when workers are paid less than their peers, they reduce output and arewilling to give up substantial earnings through decreases in attendance. The perceivedjustification for these pay differences plays an important role in mediating these effects.Our findings provide support for reference dependence in co-worker pay, and indicate thattransparency about the firm’s rationale for pay is important for fairness perceptions andoutput.

The results suggest that optimal pay for a given worker will potentially be a function ofco-worker pay. This could help us understand why wage compression—when wages varyless than the marginal product of labor—is so prevalent. For example, in many occupa-tions—from tollbooth attendants to supermarket cashiers—all workers in a firm are paidthe same fixed hourly wage even though managers are aware of their productivity differ-ences. In casual daily labor markets—for example among agricultural day laborers in Indiaor California—an employer usually pays all workers the same prevailing daily wage, despite


knowing which are more productive than others (e.g. Dreze and Mukherjee 1989). Whilesuch behavior is hard to reconcile under neoclassical agency theory, if relative pay is im-portant, then it may be profit maximizing for firms to compress wages. Our results mayalso have bearing on explaining the conditions under which differential pay will arise—forexample, when it is easy to observe and quantify co-workers’ relative productivity. Thiscould help explain why workers accept earnings dispersion under piece rates or within sportsteams (where performance statistics reflect productivity), but not among clerical workersat the University of California (Card et al. 2012).

Wage compression could have potentially important effects on labor market outcomes.For example, Akerlof and Yellen (1990) tie this to wage rigidity: if firms cannot cut pay afterindividual adverse shocks and therefore fire workers instead, this will increase unemploymentand business cycle volatility. Wage compression may also have distributional consequences.If the wage for all laborers is the same, better quality workers will be hired first and worsequality ones may be more likely to face involuntary unemployment. This implies that smallproductivity differences may lead to large earnings differences, exacerbating inequality andamplifying the adverse effects of shocks like illness. Thus, the rationing mechanism mayhurt the most vulnerable, generating a rationale for targeting in unemployment programs.

Relative pay concerns may also have relevance for the organization of production and firmboundaries. For example, they could influence whether workers of heterogeneous ability areorganized within a firm or contract their labor through the external market. Consistentwith this, Nickerson and Zenger (2008) argue that pay differences across firms serve as ahindrance to firm mergers. Similarly, firms may “specialize” in hiring workers of a givenproductivity level to avoid pay discrepancies. Relative pay concerns also have bearing onhuman resource policies—for example, they could help explain why about one-third of USfirms require employees to sign nondisclosure contracts that forbid them from discussingtheir pay with their co-workers (Card et al. 2012).

In addition, our findings suggest that firms may have several potential tools at theirdisposal to manage morale in the presence of pay dispersion. For example, technologiesthat make it easier to quantify worker productivity could have aggregate output benefitsnot just through increased monitoring, but also through improved morale. Firms could alsopotentially alter the organizational structure of the workplace itself—through job titles,physical co-location of similar workers, or the construction of “teams”—to affect who aworker views as being in her reference group. Indeed, our experimental design leveragesthe insight that the organization of production can be manipulated to affect the referencegroup for relative pay comparisons.

While speculative, the above possibilities suggest a variety of ways through which rela-tive pay concerns could affect pay structure, organizational arrangements, and other labormarket outcomes. These possibilities are a promising direction for further research.


References

Adams, J. S. (1963): “Towards an understanding of inequity.” The Journal of Abnormaland Social Psychology, 67, 422.

Akerlof, G. (1982): “Labor contracts as partial gift exchange,” The Quarterly Journalof Economics, 97, 543–569.

Akerlof, G. A. and J. L. Yellen (1990): “The fair wage-effort hypothesis and unem-ployment,” The Quarterly Journal of Economics, 255–283.

Bartling, B. and F. A. von Siemens (2011): “Wage inequality and team production:An experimental analysis,” Journal of Economic Psychology, 32, 1–16.

Bewley, T. F. (1999): Why wages don’t fall during a recession, Harvard University Press.Bracha, A., U. Gneezy, and G. Loewenstein (2015): “Relative Pay and Labor Sup-ply,” Journal of Labor Economics, 33, 297–315.

Brown, G. D., J. Gardner, A. J. Oswald, and J. Qian (2008): “Does Wage RankAffect Employees? Well-being?” Industrial Relations: A Journal of Economy and Society,47, 355–389.

Card, D., A. Mas, E. Moretti, and E. Saez (2012): “Inequality at work: The effectof peer salaries on job satisfaction,” American Economic Review, 102, 2981–3003.

Charness, G. and P. Kuhn (2007): “Does pay inequality affect worker effort? Experi-mental evidence,” Journal of Labor Economics, 25, 693–723.

——— (2011): “Lab labor: What can labor economists learn from the lab?” Handbook ofLabor Economics, 4, 229–330.

Clark, A., D. Masclet, and M.-C. Villeval (2010): “Effort and Comparison Income:Experimental and Survey Evidence,” Industrial and Labor Relations Review, 63.

Clark, A. E. and A. J. Oswald (1996): “Satisfaction and comparison income,” Journalof Public Economics, 61, 359–381.

Cohn, A., E. Fehr, B. Herrmann, and F. Schneider (2012): “Social Comparison inthe Workplace: Evidence from a Field Experiment,” IZA Discussion Paper No. 5550.

DellaVigna, S., G. Rao, and U. Malmendier (2015): “Gift Exchange at Work,”Working Paper.

Dreze, J. and A. Mukherjee (1989): “Rural Labor Markets in India,” The Balance Be-tween Industry and Agriculture in Economic Development, Vol, 3, Manpower and Trans-fers.

Duesenberry, J. S. (1949): Income, saving, and the theory of consumer behavior, vol. 87,Harvard Univ Pr.

Easterlin, R. A. (1974): “Does economic growth improve the human lot? Some EmpiricalEvidence,” Nations and Households in Economic Growth: Essays in Honor of MosesAbramowitz. P. A. David and M. W.Reder.


Falk, A., E. Fehr, and U. Fischbacher (2008): “Testing theories of fairness?Intentionsmatter,” Games and Economic Behavior, 62, 287–303.

Fang, H. and G. Moscarini (2005): “Morale hazard,” Journal of Monetary Economics,52, 749–777.

Fehr, E., L. Goette, and C. Zehnder (2009): “A behavioral account of the labormarket: The role of fairness concerns,” Annual Review of Economics.

Fehr, E., G. Kirchsteiger, and A. Riedl (1993): “Does fairness prevent market clear-ing? An experimental investigation,” The Quarterly Journal of Economics, 108, 437–459.

Frank, R. (1984): “Are Workers Paid Their Marginal Products?” American EconomicReview, 74, 549–571.

Frey, B. S. and A. Stutzer (2002): “What can economists learn from happiness re-search?” Journal of Economic literature, 40, 402–435.

Gatcher, S. and C. Thoni (2010): “Social comparison and performance: Experimentalevidence on the fair wage–effort hypothesis,” Journal of Economic Behavior and Organi-zation.

Gneezy, U. and J. A. List (2006): “Putting behavioral economics to work: Testing forgift exchange in labor markets using field experiments,” Econometrica, 74, 1365–1384.

Hamermesh, D. (2001): “The Changing Distribution of Job Satisfaction,” Journal ofHuman Resources, 36, 1–30.

Hamermesh, D. S. (1975): “Interdependence in the labour market,” Economica, 42, 420–429.

Hicks, J. (1932): The Theory of Wages.Kahneman, D. and A. Tversky (1979): “Prospect theory: An analysis of decision underrisk,” Econometrica: Journal of the Econometric Society, 263–291.

Kube, S., M. A. Maréchal, and C. Puppe (2013): “Do wage cuts damage work morale?Evidence from a natural field experiment,” Journal of the European Economic Associa-tion, 11, 853–870.

Kuziemko, I., R. Buell, T. Reich, and M. Norton (2011): “Last-place Aversion:Evidence and Redistributive Implications", NBER Working Paper No. 17234,” NBERWorking Paper.

Levine, D. I. (1993): “What do wages buy?” Administrative Science Quarterly, 462–483.List, J. A. (2009): “Social preferences: Some thoughts from the field,” Annu. Rev. Econ.,1, 563–579.

Luttmer, E. F. (2005): “Neighbors as negatives: Relative earnings and well-being,” TheQuarterly Journal of Economics, 120, 963–1002.

Marmot, M. (2004): “Status syndrome,” Significance, 1, 150–154.Marshall, A. (1890): Principles of Economics.Nickerson, J. A. and T. R. Zenger (2008): “Envy, comparison costs, and the economictheory of the firm,” Strategic Management Journal, 29, 1429–1449.


Pfeffer, J. and N. Langton (1993): “The effect of wage dispersion on satisfaction,productivity, and working collaboratively: Evidence from college and university faculty,”Administrative Science Quarterly, 382–407.

Rege, M. and I. Solli (2013): “Lagging Behind the Joneses: The Impact of RelativeEarnings on Job Separation,” .

Veblen, T. and C. Almy (1899): The theory of the Leisure Class: An Economic Studyin the Evolution of Institutions, Macmillan & Company, Limited.


Figures

Figure 1. Randomization Design

Randomize workers into teams of 3

Randomize teams into

tasks

Randomize into wage treatments (stratify by task)

Workers of heterogeneous

ability Variation in

relative productivity within teams

(Actual fairness)

Variation in observability of

co-worker output (Perceived fairness)

Figure 2. Time Line

1 Day

4 9 36 15

Job begins

Output is sellable

Feedback on rank

“Training” period (baseline output) Recruitment Treatment period

Teams randomized

into wage treatments

Endline survey &

lab activities


Figure 3. Task Observability: Actual vs. Survey Correlations

0

0.2

0.4

0.6

0.8

1

A B C D E F G H I J

Correla'

on Actua

l vs. Survey

Task ID

Produc'vity Correla'ons by Task

Figure plots the correlation between actual productivity rankings and perceived rankings by the workers(reported in endline surveys) for eight of the production tasks. Note that this data come from four pilotrounds where the research team did not inform workers of their production rankings. In our analysis, wesplit the tasks at the median level of observability (0.5 correlation).


Tables

Table 1. Treatments and Relevant Comparisons

Worker Type Low productivity

Medium productivity

High productivity

Heterogeneous

wLow

wMedium

wHigh

Compressed_L

wLow

wLow

wLow

Compressed_M

wMedium

wMedium

wMedium

Compressed_H

wHigh

wHigh

wHigh

Table 2. Summary Statistics

Panel A: Demographic Characteristics MeanOwn any land 0.54Sharecrop any land 0.70Land Owned (Acres) 0.68Land Leased Out (Acres) 0.04Own Land Cultivated (Acres) 0.67Land Sharecropped In (Acres) 1.17Female HH members 2.08Male HH members 2.37Female HH members engaged in labor force 0.81Male HH members engaged in labor force 1.79N 145

Panel B: Labor Market Experience MeanReceived wage different from prevailing wage 0.72Received wage different from other laborers in village 0.17Ever worked on piece rates 0.71N 313

TH

EM

OR

ALE

EF

FE

CT

SO

FPA

YIN

EQ

UA

LITY

26

Table 3. Effects of Pay Disparity

Full sample

Full sample

Full sample

Non-paydays

Full sample

Full sample

Full sample

Non-paydays

(1) (2) (3) (4) (5) (6) (7) (8)

Post x Heterogeneous -0.372*** -0.281** -0.311** -0.316** -0.0880 -0.0766 -0.0925* -0.120**(0.119) (0.126) (0.125) (0.127) (0.054) (0.053) (0.052) (0.052)

Post x Heterogeneous x Med wage 0.518** 0.465* 0.500** 0.584** 0.0453 0.0312 0.0441 0.0682(0.228) (0.241) (0.250) (0.263) (0.073) (0.073) (0.075) (0.074)

Post x Heterogeneous x High wage 0.203 0.181 0.236 0.257 -0.0154 -0.0236 -0.0104 0.0199(0.215) (0.215) (0.211) (0.220) (0.072) (0.072) (0.073) (0.075)

Team fixed effects? No Yes No No No Yes No NoIndividual fixed effects? No No Yes Yes No No Yes YesF-test pvalue: (Post x Het) + (Post x Het x Med) = 0 0.465 0.386 0.392 0.252 0.365 0.347 0.356 0.277F-test pvalue: (Post x Het) + (Post x Het x High) = 0 0.405 0.611 0.693 0.760 0.0491 0.0576 0.0499 0.0581

Post-treatment Control Mean -0.0266 -0.0266 -0.0266 -0.0460 0.928 0.928 0.928 0.917R-squared 0.252 0.220 0.187 0.194 0.206 0.200 0.198 0.209N 7755 7755 7755 6169 7755 7755 7755 6169

Dependent variable: Dependent variable:Output (standard dev.) Attendance

Notes: Difference in differences regressions. Post is an indicator that equals 1 if the day is after workers have been randomized into wage treatments, and 0 during the baseline training period. Regressions include task fixed effects, day*round fixed effecst, and controls for neighboring teams. All coefficients are identified off comparisons of workers who earn the same absolute wage and have the same productivity rank within their team (see regression specification in text). Standard errors clustered by team.


Table 4. Standardized Production and Attendance Over Time

First half of post period

Second half of post period

First half of post period

Second half of post period

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

Post x Heterogeneous -0.221* -0.467*** -0.0422 -0.193**(0.133) (0.169) (0.0562) (0.0789)

Post x Heterogeneous x Med wage 0.327 0.828*** -0.0321 0.195*(0.256) (0.304) (0.0723) (0.115)

Post x Heterogeneous x High wage 0.177 0.323 -0.0599 0.0833(0.229) (0.249) (0.0833) (0.0969)

Individual fixed effects? Yes Yes Yes YesR-squared 0.169 0.208 0.190 0.243N 5,697 5,109 5,697 5,109

Dependent variable: Dependent variable:

Output (standard dev.) Attendance

Notes: Difference in differences regressions. Regressions include task and day*round fixed effects. Standard errors clustered by team.

TH

EM

OR

ALE

EF

FE

CT

SO

FPA

YIN

EQ

UA

LITY

28

Table 5. Perceived Justifications: Mediating Effects of Relative Productivity Differences

Productivity difference measure Production difference

Production difference

Above mean

difference

Above mean

difference

Above mean

differenceProduction difference

Production difference

Above mean

difference

Above mean

difference

Above mean

difference(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

Post x Heterogeneous -0.483*** -0.294** -0.539*** -0.460** -0.581*** -0.113* -0.107* -0.213*** -0.203*** -0.207***(0.157) (0.149) (0.197) (0.204) (0.205) (0.0598) (0.0577) (0.0651) (0.0643) (0.0664)

Post x Heterogeneous x Prod difference 0.480* 0.128 0.604** 0.566** 0.676** 0.107 0.077 0.289*** 0.311*** 0.298***(0.262) (0.223) (0.277) (0.284) (0.281) (0.0742) (0.0697) (0.0858) (0.0934) (0.0868)

Post x Heterogeneous x Med wage 0.827*** 0.651** 0.956*** 0.807*** 0.987*** 0.111 0.0980 0.203*** 0.197*** 0.197**(0.273) (0.280) (0.300) (0.279) (0.304) (0.0779) (0.0774) (0.0755) (0.0749) (0.0766)

Post x Heterogeneous x Med wage x Prod difference -0.982** -0.634 -1.210*** -1.088*** -1.283*** -0.210* -0.156 -0.347*** -0.368*** -0.357***(0.414) (0.409) (0.397) (0.387) (0.390) (0.125) (0.122) (0.123) (0.130) (0.124)

Post x Heterogeneous x High wage 0.313 0.216 0.460* 0.393 0.509* 0.00854 0.00310 0.110 0.0956 0.104(0.223) (0.218) (0.255) (0.263) (0.259) (0.0717) (0.0715) (0.0727) (0.0765) (0.0733)

Individual fixed effects? No Yes Yes Yes Yes No Yes Yes Yes YesControls for own baseline prodn x treatment x rank x post? No No No Yes No No No No Yes NoDropping bottom 10% of low-rank workers? No No No No Yes No No No No YesR-squared 0.254 0.189 0.190 0.192 0.198 0.210 0.200 0.203 0.204 0.203N 7,755 7,755 7,755 7,755 7,512 7,755 7,755 7,755 7,755 7,512

Dependent variable: Output (standard dev.) Dependent variable: Attendance

Notes: Regressions include task fixed effects, day*round fixed effects, and neighboring team controls. Standard errors clustered by team.


Table 6. Perceived Justifications: Mediating Effects of Output Observability

Observability correlation

Above mean observability

Observability correlation

Above mean observability

(1) (2) (3) (4)Post x Heterogeneous -0.840*** -0.503*** -0.120 -0.116

(0.234) (0.178) (0.081) (0.070)

Post x Heterogeneous x Observability measure 1.205*** 0.565* 0.0910 0.106(0.414) (0.289) (0.117) (0.080)

Post x Heterogeneous x Med wage 0.964** 0.741** 0.131 0.0680(0.407) (0.339) (0.107) (0.088)

Post x Heterogeneous x Med wage x Observability measure -1.127* -0.679 -0.212 -0.107(0.626) (0.436) (0.195) (0.133)

Post x Heterogeneous x High wage 0.552* 0.202 -0.00875 -0.0309(0.318) (0.285) (0.095) (0.099)

Post x Heterogeneous x High wage x Observability measure -0.851 0.0230 -0.0596 0.00785(0.551) (0.411) (0.139) (0.110)

F-test pvalue: (Post x Het) + (Post x Het x Med) = 0 0.729 0.424 0.877 0.348F-test pvalue: (Post x Het) + (Post x Het x High) = 0 0.334 0.227 0.0341 0.0361Post-treatment Control Mean -0.0266 -0.0266 0.928 0.928R-squared 0.193 0.191 0.200 0.201Number of observations (worker-days) 7755 7755 7755 7755Notes: Regressions include task fixed effects, day*round fixed effects, individual fixed effects, and neighboring team controls. Standard errors clustered by team.

AttendanceDependent variable:Dependent variable:

Output (standard dev.)


Table 7. Effects on Team Cohesion

(1) (2)

Both workers from same team 0.440 0.613**(0.287) (0.290)

Both workers from same team x Heterogeneous -0.929** -0.888*(0.464) (0.465)

At least one Heterogeneous worker in pair 0.411 0.383(0.345) (0.334)

At least one low rank worker in pair -0.820***(0.269)

At least one medium rank worker in pair -0.599**(0.281)

Observations: Number of pair-games 1,870 1,870Dependent variable mean 4.329 4.329R-squared 0.199 0.207

Heterogeneous team -11.26** -10.41**(4.40) (4.83)

Compressed_High team 2.52(6.34)

Observations: Number of teams 73 73Dependent variable mean 54.13 54.13R-squared 0.066 0.059

(Dependent variable: Tower height)

(Dependent variable: Number correct)

Notes: This table shows results from cooperative team building games at endline. In the Panel A games, workers were organized into pairs and incentivized to solve puzzles together. The dependent variable is the number of items correct within each pair-game. Both workers from same team is a dummy that equals 1 if both members of the pair were on the same production team during the experiment. Panel B shows results from a game where each production team was given materials to build as high a tower as possible. Standard errors are clustered by team.

Panel A: Cooperative Games in Pairs

Panel B: Tower Building in Teams


Table 8. Sample Overlap Between Perceived Justification Measures

Correlationofobs3anddiffnext_mean

Overlap between perceived justification measures: Percentage of observations in each cell

Below mean production difference

Above mean production difference

Below mean observability 30.56 21.03

Above mean observability 24.60 23.81

Notes: Tabulations of overlap between the two sources of variation for perceived justifications: baseline productivity differences between a worker and his higher ranked peer, and the observability of co-worker output. The binary splits for each measure shown here are those used in the tables in the analysis. The table shows the percentage of observations in each cell.


Appendix A. Supplemental Figures

.2.4

.6.8

11.

2K

erne

l Den

sity

0 .5 1 1.5 2Raw Production

(a) Correlation = 0

0.0

01.0

02.0

03K

erne

l Den

sity

100 200 300 400 500 600Raw Production

(b) Correlation = 0.09

0.0

05.0

1.0

15.0

2.0

25K

erne

l Den

sity

20 40 60 80 100Raw Production

(c) Correlation = 0.42

0.0

002

.000

4.0

006

Ker

nel D

ensi

ty1000 1500 2000 2500 3000 3500

Raw Production

(d) Correlation = 0.45

.02

.04

.06

.08

.1.1

2K

erne

l Den

sity

0 5 10 15Raw Production

(e) Correlation = 0.47

.05

.1.1

5.2

.25

.3K

erne

l Den

sity


(f) Correlation = 0.63

0.0

02.0

04.0

06.0

08K

erne

l Den

sity


(g) Correlation = 0.74

0.0

005

.001

.001

5.0

02.0

025

Ker

nel D

ensi

ty


(h) Correlation = 0.81

0.0

5.1

.15

.2K

erne

l Den

sity

0 2 4 6 8 10Raw Production

(i) Correlation = 0.81

0.0

5.1

.15

.2K

erne

l Den

sity


(j) Correlation = 0.87

Figure 4. Density of Raw Production by Task, Ranked by Output Observability

Documents

The Morale Effects of Pay Inequality