You are on page 1of 15

Journal of Economic Behavior & Organization 117 (2015) 233–247

Contents lists available at ScienceDirect

Journal of Economic Behavior & Organization
journal homepage: www.elsevier.com/locate/jebo

Merit pay and wage compression with productivity
differences and uncertainty夽
Till Gross b,∗ , Christopher Guo c , Gary Charness a
a
b
c

University of California, Santa Barbara, United States
Carleton University, Canada
Rand Corporation, United States

a r t i c l e

i n f o

Article history:
Received 11 November 2014
Received in revised form 19 May 2015
Accepted 12 June 2015
Available online 27 June 2015
JEL classification:
C91
J31
M52

a b s t r a c t
This paper experimentally investigates wage setting and effort choices in a multi-worker
setting when there is heterogeneity in worker productivity and managers’ perception of
this productivity is imperfect. Worker ability is assigned via an aptitude test and, in an
innovative design, manager uncertainty concerning this ability is related to the manager’s
own test performance. We propose a merit-pay hypothesis, that higher-ability workers
will reduce their effort if they are not paid more than coworkers with lower ability, but not
vice versa. Based on a simple model, we also predict that the higher the uncertainty about
employee ability levels, the more managers compress wages between perceived high- and
low-ability workers. We find strong experimental support for both hypotheses.
© 2015 Elsevier B.V. All rights reserved.

Keywords:
Wage inequality
Wage compression
Gift-exchange game
Reciprocity
Fairness
Merit pay

1. Introduction
Wage inequality within countries and within firms is increasing and is at historically high levels. For example, the average
ratio between the highest and lowest salary in firms is much higher today than fifty years ago. Structural changes such as
skill-biased technological change, see for instance Bekman et al. (1998), or capital-skill complementarity as proposed by
Krusell et al. (2000), could lead to higher wage inequality through performance pay (Lemieux et al., 2009). We are interested
in the question of how wage inequality within a firm affects employees’ motivation and their work effort and how this can
in turn affect wage inequality.
When firms attempt to optimally set wages, they must take into account the incentive effects of these wages on employee
work effort. (Campbell and Kamlani, 1997, p. 780) survey firms and find that: “Overall, firms appear to believe that wages

夽 We would like to thank Peter Kuhn and Rod Garratt for their thoughts. We appreciate comments from anonymous referees, seminar participants at
the University of Ottawa, and participants at the 2014 meetings of the Canadian Economic Association. All remaining errors are of course our own. We
acknowledge financial support from the Department of Economics at the University of California Santa Barbara.
∗ Corresponding author at: Department of Economics, Carleton University, C870 Loeb Building, 1125 Colonel By Dr, Ottawa, ON, Canada K1S 5B6.
Tel.: +1 613 520 2600x3773.
E-mail addresses: till.gross@carleton.ca (T. Gross), cguo@rand.org (C. Guo), charness@econ.ucsb.edu (G. Charness).
http://dx.doi.org/10.1016/j.jebo.2015.06.009
0167-2681/© 2015 Elsevier B.V. All rights reserved.

To investigate this hypothesis. reports USA Today (Kay. but we find that this changes in our richer environment.1 Previous work has found little or no effect of relative pay on worker performance. Our paper makes three main contributions. (2001) provide some evidence that a person’s input contributions. Gross et al. Wourio (2011) at Microsoft Business argues that “many businesses approach the concept of job promotions with something less than a studied eye” and that “the consequences of an ill-advised promotion can be nothing short of cataclysmic. Second. and so we would not observe very many violations of merit pay. Our hypothesis also holds when we define merit pay as receiving at least as much as less-productive coworkers (instead of strictly more). neither have manager uncertainty of worker ability. The idea is related to the dynamic learning model in Harris and Holmstrom (1982) on the true level of productivity of a worker. we asked participants to fill out a questionnaire. random wage assignments in Bartling and von Siemens (2011) do not seem to affect effort in a team production environment. 1966. The Financial Post (Deveau. wage assignments should still be attributable to the competence of the managers. 2 This is in line with what Cappelen et al.7 1 In fact. Gantner et al. but the productivity levels are common information. while less productive workers will not react much to relative wages. 3 For instance.4 Third. We wish to assess whether the relative wage of a worker matters for choosing effort when productivities are different. resulting in higher-ability workers receiving a lower wage in 26% of the cases. when deciding on bonuses and wage increases. we needed to address three important issues. Charness and Kuhn (2005) find that. and one of the greatest sources of employee frustration are “idiot” bosses. with more-able managers receiving a more precise signal in this regard. But what constitutes a fair wage? Are fair wages the same across coworkers. as otherwise workers are not likely to punish the manager for a wage they perceive as unfair (Ku and Salmon. When there is uncertainty. if workers “earn” their productivity by their own doing.9% of the cases and Charness and Kuhn (2007) observe shares of 6. a higher wage dispersion is cost-effective for inducing more effort from higher-productivity workers (Adams. as argued by Akerlof and Yellen (1990). / Journal of Economic Behavior & Organization 117 (2015) 233–247 have a strong effect on effort by affecting workers’ attitudes toward their employer. or do they depend on each worker’s effort.5% and 8. However. We show this in a simple theoretical model. 7 Specifically. so we instead assign ability in the experiment according to the results of an aptitude test at the beginning of the experiment. more productive workers who believe in merit pay will reduce their effort when not receiving a higher wage than their less productive coworkers. Lazear and Rosen. even with no uncertainty about imposed productivity. then they are more likely to expect merit pay.2 At the end of the experiment.234 T. there are very few studies that do not simply assign productivity to workers randomly.6 A related issue is that of wage compression. firms have an incentive to compress wages so as to guard themselves against the possibility of being wrong in their assessment of worker ability. First. whereas impersonal market features such as prices do not. While previous work has found evidence of wage compression when differences in worker productivity are clear. However. and so our setting may have more external validity.5 If workers care about equality. Frank (1984) reports that firms compress wages relative to productivity. we feel that there is considerable uncertainty about such differences in the field environment. 5 Every organization has its share of incompetent managers. 2013) reports that 69% of employers complain that bad hires lowered company productivity and affected morale. or for promotion. . we develop a novel mechanism for generating uncertainty about a worker’s productivity. even when controlling for ability (which was randomly assigned). the manager is very unlikely to intentionally pay a higher wage to the lower-ability worker. our hypothesis thus predicts that more capable workers would expect higher compensation or else reduce effort. or on each worker’s output? Cappelen et al. (2007) find that a person’s effort and productivity strongly influence perceptions of a fair wage. we are the first to investigate the effect of relative pay on worker performance when productivity is assigned on the basis of test performance and the firm only receives a signal concerning the productivity of each worker. if workers differ in their productivity.1% in two treatments. For this subsample. We therefore create a mechanism that induces an endogenous degree of uncertainty into the managers’ perception of worker ability. In our mechanism. This reflects the fact that there is heterogeneity in the field regarding managerial competence with respect to differentiating the productivity of different workers – whether at hiring. Rivas (2009) and Bolton and Werner (2012) do assign productivity to workers according to a test. they worry that high-ability workers will not provide effort when they receive a low wage – even when workers do not respond to relative wages.” 6 In a similar vein. 2012). To the best of our knowledge. an arbitrary assignment of productivity may not give high-productivity workers a sense of entitlement. which included the question: “Do you think the higher-ability worker deserves to be paid more than the lower-ability worker?” Fifty-two out of 66 assigned workers (79%) answered this question positively. We hypothesize that many people believe that merit should be rewarded and therefore more-able workers should receive higher remuneration and we therefore postulate our merit-pay hypothesis: Ceteris paribus.” There are reasons to believe that not only absolute wages matter. 2013). including productivity. which are reflected in the contributions listed above. the precision with which the manager can distinguish between higher and lower ability workers depends on her performance in the aptitude test.3 Second. then higher wage dispersion could result in lower worker effort (Milgrom and Roberts. Charness and Kuhn (2007) find no statistically or economically significant effect of coworker wages on effort. 1981). are correlated with one’s demands in a bargaining game. managers did compress wages significantly more when wages are public than when they are not. Bewley (1999) shows the importance of fairness in determining effort in response to wages. 4 Bolton and Werner (2012) report this in 4. Similarly. 1992). but also that the relative remuneration in comparison to one’s coworkers matters. Third. (2010) call meritocracism. we use this mechanism to investigate the effect of uncertainty about productivity on whether managers choose to compress wages. The mechanism reflects the fact that managers differ in their ability to assess worker skill and performance.

Loewenstein et al. In their “fair wage-effort hypothesis. find results in line with incentive considerations. the issue of equality versus inequality is of central importance. see Charness and Kuhn (2007). 1984). wage inequality is thought to negatively affect worker effort. Gift-exchange games have been argued to be well-suited to detect equity-motivated effort effects because there are no other influences on worker behavior besides fairness considerations and because it is costless to punish firms for unfair wages. In Bartling and von Siemens (2011). In Section 3 we explain the experimental setup.” Akerlof and Yellen (1990) note that within a labor-wage setting where effort is not perfectly contractible. Their focus is less on effort levels and more on coordination issues between .and white-collar workers in Austrian firms. social information. in the form of bilateral gift-exchange games (GEG) between a manager and a single agent (Fehr et al.T. Nosenzo (2010) compares public versus secret wage treatments and finds effort choices decrease when wages are public (in a threeperson GEG with one manager and two equal-ability workers). In the next section. (2012) observe an asymmetric response to wage inequality: employees who are paid less than the average tend to be less satisfied with their job and their pay when aware of their coworkers’ salaries. and this is particularly true when it is possible to punish non-contributors (Carpenter. attention has focused on horizontal wage comparisons within the workforce and the effect of social information on possibly unequal wages amongst workers. Campbell and Kamlani. Winter-Ebmer and Zweimüller. 2001). On the other hand. 2007). with multiple workers. both in field studies (Leonard. 1993. and when payment is not linked to an individual’s marginal product. 1992. 2002). leading to morale problems and diminished teamwork (Milgrom and Roberts. inequality reflects the unequal human capital or productive capabilities of different workers (Adams. Bandiera et al. 2000. 2000). 2004. we discuss the related literature. 2004. and find that the reciprocity relationship is not affected by the number of workers per firm. Eriksson. Incentive arguments in tournament models claim a greater wage inequality stimulates effort as workers aspire to superior positions (Lazear and Rosen. operating under a number of unobservable real world forces such as social pressure. According to equity theory. and the authors do not find any effect of wage inequality on effort choices. In reality though.. while those who are paid more than the average do not react noticeably. Brown. Experimental evidence supports the idea that information about the behavior and treatment of others systematically influences an individual’s behavior in various settings. 2. 2006. between firms. Winter-Ebmer and Zweimüller (1999). (2010) are interested in the question when to pay unequal wages to workers of the same ability. people regularly contribute non-zero amounts. information-based wage comparison effect. especially if a worker’s effort induces a positive externality on the effectiveness of other workers’ efforts. Gill and Stone (2010) consider notions of fairness and desert in their model of providing effort in tournaments. Under the premise that wages are commensurate with their contributions to the firm. then fairness considerations should not adversely affect effort from workers (Goerg et al. learning. even with strong incentives to free-ride and contribute nothing. however. Within organizational management. wages are assigned exogenously with no managers in the experiment. 1999. a theoretical model by Winter (2004) suggests that it might even be preferable to pay workers of equal ability unequally. Goerg et al.. 1981). Carpenter and Seki. 2010. People make social comparisons with their coworkers as their reference point for a fair wage (Falk and Knell. 1966. Related literature A common theme in the work on social preferences is that people not only pursue their own self-interest but also care about the well-being of others. These field studies of wage dispersion and firm performance were conducted in vastly different contexts. 2005) and in laboratory experiments. 1996. Dufwenberg and Kirchsteiger (2004) provide a theory of reciprocity in sequential games. this allows for free-riding behavior. researchers have turned to laboratory experiments in an attempt to isolate a pure. 1989. Furthermore. 1995). Frank. 1997. Holmstrom (1982) points out that equal wages do not account for heterogeneity in agent’s ability and performance. Martins. Bardsley and Sausgruber. 1992). responders turn down offers with positive probability. On the one hand. 2000) or also intention-driven (Rabin. 2010. Card et al. 2008. depending on the shape of the production function. then present the data. unequal wages could positively affect worker effort. 1993. / Journal of Economic Behavior & Organization 117 (2015) 233–247 235 Our experimental results lend strong support to the merit hypothesis and there is a significant resulting decrease in effort when high-ability workers are paid less than low-ability workers in an environment where productivity is not assigned randomly by experimenters. Section 6 deals with wage compression and the final section concludes.. Maximiano et al. In public goods experiments. (2007) extend the bilateral gift exchange game to a situation where a firm has multiple workers.. In consequence. Much of the seminal experimental research into efficiency wages is based on vertical inequality. Agell and Bennmarker. Gross et al. labor settings are often complex social systems. internal wage equity is critical for worker effort and firm productivity (Pfeffer and Langton. We also find strong experimental evidence that wage compression due to uncertainty about true ability is indeed taking place. and varying ability levels. 1998). Hibbs and Locking. and explore in Section 5 the merit-pay hypothesis. More recently. intrinsic motivation. Main et al. Charness and Rabin. Cowherd and Levine. Clark and Oswald. 1999. beyond possible concerns about equality. whether this sense of fairness is strictly payoff-driven (Fehr and Schmidt. doing comparisons within and across groups of lower-level employees and upper management. Empirical research exploring the relationship between wage inequality and performance has produced mixed results. Clark and Senik. workers’ effort is affected by the perceived fairness of their salary. each earning the same wage. knowledge spillovers. Mas (2006) shows how pay raises for policemen below a fair wage reference point reduce job performance. 1989. etc. and if social comparisons are important for teamwork and morale.. 1993. using panel data of blue. In the ultimatum game. Bolton and Ockenfels. 1999. 2005). For example. especially if the offer is low (Knez and Camerer. 1993. Konow. and workers withhold effort if their actual wage is less than their fair wage.

their results regarding merit pay are perfectly in line with standard economic reasoning (without any behavioral fairness considerations) that agents reduce their effort when their remuneration depends not only on their own effort. worker remuneration includes a performance component and most of their subjects do not exhibit any significant deviations from materialistically optimal behavior. Abeler et al. In our study. upon observing their efforts. making it difficult to measure the impact of fairness (violations of merit pay do not occur at all in their data). we do not allow for information on coworker effort because introducing public effort comparisons on top of public wage comparisons could confound the interpretation of the results. . and results are not consistent across trials or for low. with the second worker having the benefit of additional information on the first worker’s effort choice. does not control for individual fixed effects and overall “generosity” of the contract. There are thus two identical cycles in every experiment. The novelty of our experimental design lies in the mechanism that injects uncertainty about the true ability of a worker. but through their own fault. even though the ratios and differences between wage levels are much lower. we had two periods for each information level. normal. but is to prevent workers from responding negatively to relative wages in these studies. the aptitude test questions. which remains constant throughout the experiment. 9 While the design in Abeler et al. In each period. 8 The authors show a preliminary result that agents favored by fairness provide higher effort. as in other studies. we find evidence of non-linear effects of own wages on effort. we exogenously vary the manager information level in each period. Experimental design As in many other experimental studies. productivity is assigned via an aptitude test. 10 The exact instructions. (2010) is simpler than ours. we also refer to them as levels 1 to 5 respectively. The composition of firms changes every period. Each worker. In Rivas (2009). The game consists of ten periods. We privately tell workers their type. (2001) investigate fairness within a game with one principal and two agents and find that public wage contracts lead to wage compression. Wage compression has been studied. Gross et al. In contrast to our study. finding workers’ effort choices do not depend on coworkers’ wages. In contrast to our study. (2010) invert the traditional GEG order. for instance by Güth et al. assign lower wages to higher-ability workers than to the lower-ability co-workers. the manager pays them a wage. and do not examine our merit-pay hypothesis. which allows for incomplete contracts and reputation-building. He assigns wages to the perceived high. / Journal of Economic Behavior & Organization 117 (2015) 233–247 workers in a repeated game. though. They also know the actual high-ability worker generates twice as much money for the manager as the actual-low ability worker for the same effort. However. and no information. this wage compression is not due to uncertainty about worker ability as in our study. their merit criterion is based on effort. It endogenously creates instances where managers mistakenly. However the manager is uncertain who the actual high-ability worker is. After this. higher ability workers can never receive a lower wage in her experimental setting.9 3. but concede that further data collection is required: the statistical analysis for this result suffers from potential endogeneity. both consisting of each of the five different information levels. In this paper. first having the workers provide effort and then.8 Charness and Kuhn (2007) study a pure gift exchange setup where a firm offers various wages to two workers whose productivities differ. all participants take a twenty-question aptitude test which consists of quantitative and verbal test questions.236 T. not on ability. which we would publicly announce at the beginning of the period. good.and high-productivity workers.10 For the workers. Gächter and Sefton (2012) include sequential effort decisions. They find that forcing firms to pay workers equal wages reduces worker effort and argue that this effect is due to fairness/merit considerations. the labor market is represented by a gift-exchange game between managers and workers. we would like to emphasize the commonalities and differences with the literature. (2001) also investigate the impact of relative wages on effort provision. All information concerning the rules of this game is common to all participants. To sum up how this paper fits into the literature.and low-ability workers. Out of the ten time periods in each game. poor. automatically knows the ability of his coworker. Managers are not told their aptitude test results. We use a standard two-worker gift-exchange game. the top 50% are classified as high ability and the lower 50% are classified as low ability. but it is not clear if her results are driven by non-linear responses to own wages – in some wage regimes the highest wage is 30 times as high as the lowest. a three-person firm is formed by randomly and anonymously matching a manager with one actual high-ability worker and one actual low-ability worker. In preparation. Gächter and Thöni (2010) use a three person GEG to explore whether wage comparison effects are due to intentional wage discrimination or due to mere payoff differences. They find very few instances of higher wages for lower-ability workers. She finds that the more unequal wage assignments result in lower average effort. Bolton and Werner (2012) also assign productivity via a test and find that unequal wage assignments do not lead to a reduction in effort. None of the aforementioned horizontal wage comparison papers take into account possible differences in worker ability. in which heterogeneous ability is assigned non-randomly through a test. considers only one-sided t-tests. because he knows his own ability level. we designate a third of the participants to be managers and two-thirds to be workers by random drawings. and a detailed description of the experimental procedure can be found in Appendix. Güth et al. The higher the manager scored on the aptitude test. It allows us to consider the merit-pay hypothesis and wage compression due to uncertainty. the better the chances his perception is accurate. We label them as perfect. Some papers explicitly do so: Güth et al. In addition. and the perceived high-ability worker is the one with the higher signaling score (details are in the next section and Appendix). (2001) and Charness and Kuhn (2005). but also their coworker’s (similar to public good contributions). similar to Charness and Kuhn (2005). though.

such as worker reputation.00 $4. Workers have to internalize their role as a higher/lower ability worker.00 $1. The keys to this mechanism are that worker productivity is determined by performance on a quiz and that the manager only receives a signal about the worker’s productivity. as well as the different information treatments. / Journal of Economic Behavior & Organization 117 (2015) 233–247 237 Table 1 Worker effort cost function. but increased the amount from $5 to $7 since we set the minimum wage to $1.12 All of this is common information. In studying the effect of endogenous wages on effort. 3.30 $0. Each period. Effort level 1 2 3 4 5 6 Cost $0. the manager knows exactly who the high-ability worker is. Due to the different combinations of workers and managers. Charness and Kuhn (2007) convincingly argue that random and anonymous assignment each period. we had three main goals in creating an experimental mechanism: 1. A worker who scored in the top (bottom) half of all workers was assigned higher (lower) 11 We chose an uneven number of dollars in accordance with Charness and Kuhn (2007).10 Table 2 Firm revenue produced by worker type. The minimum wage is one lab dollar. The information about worker types should be related to the manager’s competence. details are in the following section. even though this makes the experiment somewhat less natural. 2. 3. It has been observed that wage responses are very sensitive to the source and intention of the wage assignment (Charness.00 $6. In the good (poor) information level.1. the variation of wages is not exogenously or randomly imposed by the experimenters but is instead determined endogenously by managers who anticipate effort decisions by workers. With no information.60 $1. It is in order to avoid confounding the merit-pay hypothesis with possible other systematic interactions between type and effort choices. Effort level 1 2 3 4 5 6 Higher ability worker output Lower ability worker output $3. This difference should not affect wage comparisons across workers. Finally. the manager has seven lab dollars total to spend on wages of both workers (in integer dollar amounts).00 $9. as well as controlling for worker fixed effects. addresses these potential endogeneity concerns. Managers should not have perfect information about worker types. 12 The exchange rate from lab-dollars into actual earnings is the same for all workers. Firm profit is simply firm revenue plus any potential left-overs from the wage assignment. We chose a lower exchange rate for managers as compared to workers to increase the incentives for managers to assign more than the minimum wage to both workers. We had each participant take an aptitude test at the beginning of the experiment. To achieve these goals. at the cost shown in Table 1. with wages being endogenously determined. it is not clear that workers would respond reciprocally. The novelty of the design is that the strength of the signal is determined by the manager’s own score on the same quiz. By the nature of the GEG design. Ku and Salmon. in order to gauge the competence of both workers and managers. the firm generates revenue (shown in Table 2) depending on the actual ability level of the workers (not the levels perceived by the manager) as well as their efforts provided. Design rationale and mechanism details This experimental design relies on the well-established three-person gift-exchange game as the foundation onto which we will build more realistic characteristics of manager uncertainty. they choose how much effort to provide on a scale from one to six.00 $7. 13 We have purposefully set the effort decision orthogonal to the ability test.10 $0.00 $1. Workers are paid 1 dollar for every 2 lab-dollars.50 $12.13 Moreover. the manager is given a random guess.50 $18.00 $9. . which is what matters for our purposes. we increase (decrease) the accuracy of managers’ perception of worker ability.00 With a perfect information level. but differs from the managers’ exchange rate. though. Gross et al.T. 2012). Earnings accumulate over the course of the session. and managers are paid 1 dollar for every 5 lab-dollars of cumulative firm profits.11 Once workers in each firm learn their own wages as well as their coworker’s wages.00 $15. the perception of which person is the high-ability worker may change.00 $3. Managers are not informed of the worker effort choices at the end of each period nor of their own payoffs. Charness and Levine. there is a possibility that the wages assigned might be influenced by unobserved characteristics. 2007. that also affect effort. The signal is always informative unless the manager somehow believes that the number of her correct answers is lower than what would be expected by random guessing. so they cannot infer anything about how well they did on the aptitude test. 2004.50 $2. and the manager keeps money not assigned to workers.50 $6. we needed to develop a new mechanism for assigning relative productivity to each worker. as mentioned in Section 1. This feature is important because unless the wage offers come from a real person with a personal financial stake.

Importantly. The accuracy of the signal depends on the manager’s performance on the aptitude test. First. assigning a probability p < 1/2 to one answer being right is equivalent to assigning a probability of 1 − p > 1/2 for the other answer to be right. Therefore. we recapitulate why we believe it is necessarily complicated to answer the question at hand. Fourth. Second. independently of how one reads the signal. It is possible that some workers react to violations of the merit-pay norm only if they perceive them as the intention of the manager. from his perspective the signal is (almost) completely random.15 We matched the answers given by the paired workers and manager. Intuitively. We conducted six sessions at the University of California. assigning a probability p > 1/2 to one answer being right increases the expected signal score compared to a random guess. The mechanism thus addresses the third point. the true type is revealed). we introduce uncertainty about worker productivity in order to generate a sizable number of observations in which the wage is higher for the low-productivity worker. 19 However.(poor-) information treatment. It is thus not possible to improve the signal quality by trying to minimize the score by reversing the signal. 16 In the perfect-information treatment. Similarly. the manager does not know how well he scored on the test. Thus. with this signal based on which worker matched more of the manager’s answers. Consider first a manager who knows that he has answered every question correctly. we do not find any evidence for this. which is again better than a random guess. but since managers never find out about the test score. another is that it allows us to study wage compression due to uncertainty about worker productivity (which relies on a comparison of the dispersion of wages across information levels). productivity is assigned non-randomly so that workers may feel they deserve a higher wage for higher productivity. see Section 5. in expectation. we report descriptive statistics of our data.17 Is it possible to game the mechanism? If a manager knew he had performed poorly on the test. In our experiment. Therefore.238 T. From the manager’s perspective. 17 Presuming that one’s random guesses are worse than expected must mean one is feeling unlucky. Since the experimental mechanism is somewhat more complicated than most comparable studies. Managers know they are paired in each round with one higher-ability and one lower-ability worker. since the higher-ability worker has answered more questions correctly than the lower-ability worker. a worker who matches more of the manager’s answers should always be evaluated as having higher productivity. Of course. But if manager intentions mattered for our merit-pay hypothesis. 18 Even if some managers do reverse the signal. as just mentioned. with precision increasing with performance. Santa Barbara. he is not aware of it at any time during the experiment. It should not affect wage-compression either. the probability that the signal is correct is actually slightly higher than 1/2 (since in the case of a tie. 4. Gross et al. the probability of a correct signal is increasing in the number of questions to which the manager knows the correct answer. Participants were 14 While the assumption of perfect knowledge of type and productivity among workers might not be the most realistic. we expect that not paying high-ability workers more than their co-workers should be more attributable to the competence of managers than to intentions. we converted all of the manager’s answers into correct ones. the more a worker believes that the manager can identify the actual high-ability worker. However.18 Overall we are confident that subjects understood the mechanism.4. having some kind of information is therefore always associated. this should not affect how a worker reacts to the wages for herself and her co-workers. Now consider an ignorant manager who randomly guesses the answer to each question.19 This is one reason we included the exogenous variation in information levels.14 However. In the good. we assigned a correct/incorrect answer to each question with a probability of one half each. as the accuracy of the signal decreases in the lower-information levels. the uncertainty is partly endogenous for workers to blame the managers for “unfair” wage assignments. with three sessions of 18 students and three sessions of 15 students. We discuss the possibility of purposefully answering questions incorrectly below. instead of telling managers who is who. In the no-information treatment. We offer some examples and intuition below. . one obviously needs to know the correct answer in order to purposefully give the incorrect answer. This addresses the second point. the worst ex ante expected signaling is still slightly better than random. for more details see Appendix B. We also had a manager comprehension quiz at the beginning of the experiment. with a better performance than if he guesses randomly. the uncertainty is partly exogenous to separate our hypothesis from manager intentions and to investigate wage compression. we send them an imprecise (except in the perfect-information treatment) signal about worker ability. This implies that even though a manager might perform very poorly on the test (even worse than the expectation with random guesses). it allows us to focus on the merit-pay hypothesis. from an ex ante perspective (that is without knowing the test results) one cannot do worse than guessing. for a total of 99 participants. a manager never learns her test score. which all managers were able to answer correctly. His signal will always be correct. we converted three of the manager’s incorrect (correct) answers into the opposite.16 This signal does relate to the manager’s competence. 15 Since our test has binary answers. even in this (worst) case the probability that the signal is correct is greater than 50%. Third. The manager received a signal concerning which worker had high productivity. the more likely that the worker should punish a violation of merit pay. based on their behavior outlined in the next section and our conversations with participants after the experiment. / Journal of Economic Behavior & Organization 117 (2015) 233–247 productivity. Data In this section. We think this serves as a reasonable proxy for workers’ perception of real ability (especially as compared to randomly assigning ability) and addresses the first point. then he should invert the signal ex-post.

The average payout was $19 per person. Gross et al. wH < wL ) happens much more often for wage assignments in terms of the actual productivity rather than the perceived productivity: 26% of the time for all information environments. Wage perc. Whatever the reason. We had to eliminate six observations due to a small initial glitch in the computer program that led to a false signal in the perfect information environment for one manager in sessions 1. 2.and low-ability workers (Table 5 shows the wages received by the actual high. managers are more than eight times more likely to assign a higher wage to the high-ability worker than vice versa. The third row indicates how often managers assign a higher wage to higher-ability workers. Table 4 shows in the first row the number of instances in which the manager assigned to the perceived low-ability worker a wage (wL ) that is higher than the wage assigned to the perceived high-ability worker (wH ) for each information environment.e. high ability $1 $2 $3 $4 $5 $6 Total 2 4 6 11 4 30 57 2 11 18 29 43 0 103 0 3 26 51 0 0 80 0 2 14 0 0 0 16 0 1 0 0 0 0 1 1 0 0 0 0 0 1 5 21 64 91 47 30 258 Table 4 Frequency of merit and non-merit wages for perceived ability by information. it is always optimal for the manager to follow the guidance of the signal. We begin by examining the frequency of the combinations of wage offers to the two workers in each firm. since the designation of the perceived high-ability worker is arbitrary in this case. good-. and 31% excluding perfect information 20 The observations in the perfect information environment only add up to 60 (66 for the other information levels). and 4. No one could participate in more than one session. Wage act. and normal-information environments. wH < wL wH = wL wH > wL Perfect Good Normal Poor Random 0 2 58 5 3 58 6 11 49 12 23 31 8 29 29 Table 5 Frequency of wage-offer combinations for actual high. the two instances in the perfectinformation treatment where the lower-ability worker was assigned a higher wage both happened in the first period. thus. it excludes the no-information environment. In the perfect-. All sessions lasted less than 90 min.and low-ability workers). only the significance level of a few robustness checks changes from 1% to 5%. with a maximum of $30 and a minimum of $12. Without uncertainty and given a concave worker-effort function and a linear revenue function. managers assign a higher wage to the perceived lower-ability worker (we exclude the no-information treatment). In the second row. 27% excluding no information. Possible reasons for non-optimal wage assignments include having an arbitrary “hunch” concerning which worker actually has higher ability or due to some form of equity or fairness considerations..20 While non-optimal wage assignments (based on perception) in Table 3 are rare. / Journal of Economic Behavior & Organization 117 (2015) 233–247 239 Table 3 Frequency of wage-offer combinations for perceived high and low ability workers. Table 3 shows manager wage assignments to the perceived high. even with uncertainty (as pointed out above). we also observe that the frequency with which managers pay equal wages increases with uncertainty.9%). which gives us an indication of wage compression. managers offered the perceived high ability worker a higher wage. In fact. low ability Wage act. while in 23 cases (8. low ability $1 $2 $3 $4 $5 $6 Total Wage perc. For example. it is optimal for the manager to pay the high-ability worker a strictly higher wage. in 196 of the 258 wage-offer combinations (76%). these cases are both infrequent and clearly related to the amount of uncertainty in the information environment. where such mistakes would be most common. . Table 5 shows that this (i. high ability $1 $2 $3 $4 $5 $6 Total $1 $2 $3 $4 $5 $6 Total 2 3 2 8 3 27 45 3 11 16 19 39 0 88 4 5 26 44 0 0 79 3 12 21 0 0 0 36 1 5 0 0 0 0 6 4 0 0 0 0 0 4 17 36 65 71 42 27 258 recruited through an email to the general university population.T.and low-ability workers. Including these leaves results essentially unchanged. It could also be a mistake or a misunderstanding.

83 0.17 2. in the regression analysis we do not find evidence to suggest that higher. Second.21 0. five out of 66 workers provided the minimum effort in every period while one worker provided average effort that was more than 90% of the maximum possible).53 0. 3. we test our merit-pay hypothesis. Baseline result We regress effort on wages (linearly in regression 1.77 0. which we do not report either. and 5 is the no-information environment. We used as explanatory variables worker ability.08 2.g. Gross et al. Merit-pay hypothesis In this section.06 in the second. 42%. and 4) and the dummy variable Ifless*High*Merit (which is one if the person received a wage lower than or equal to the coworker’s. we add the dummy variable Ifless*Low (which is one if the person received a wage lower than or equal to the coworker’s 21 A simple regression of wages on information indeed confirms that there is a statistically significant decrease in wages with information.21 As Table 6 shows. Indeed.81 with less information.3 shows that this hypothesis holds for the full population of high-ability workers.14 2. Err.29. our hypothesis focuses on workers who answered the merit-pay question positively (‘Do you think the higher-ability worker deserves to be paid more than the lower-ability worker?”).19 2.15 2.and lower-ability workers respond differently to the same wages. high-ability workers are paid more on average per period than are low-ability workers.70 0. Section 5.02 in the first cycle and 3. High-ability workers who believe in merit pay will ceteris paribus provide a lower level of effort when paid no more than low-ability workers. at 2.78 vs.” exactly the same proportion as for high-ability workers.88 0. and answered the merit-pay question positively).75 0.13 2. 3. and with a dummy for every wage level in regressions 2. . 4 is the poor. Higher ability worker mean effort Std.16 3.64 0. one might argue that the response to this question could be endogenous to the participants’ experience in the experiment. This does seem to track wages. logit.17 0. 4.88 0. has high ability. Given the heterogeneity observed for effort across workers (e.21 3. 5. A t-test shows no significant difference in wages across cycles. 3 is the normal.43. at 10% with a two-sided and 5% with a one-sided test. We see a decline in average wages as information decreases. 25 of the 33 low-ability workers answered “Yes.10 1.35 0. to 2. controlling for individual fixed effects. Average effort is also higher for high-ability than for low-ability workers. / Journal of Economic Behavior & Organization 117 (2015) 233–247 Table 6 Average wage and effort by worker type and information. we ran several regressions to determine whether the answer to this question is correlated to the outcomes of the experiment and found that none of these come close to a statistically significant correlation. This is due to our uncertainty mechanism. respectively).59 0. All our regressions also include a constant.35 0.29 0. 23 This implies that the characteristics “High” and “Merit” are redundant. Err. 2. which creates cases of an inept manager mistakenly assigning the high ability worker a lower wage.15 0. formalized below: Hypothesis 1.1.00 0. Average wages do not decrease from the first to the second cycle (we had two cycles with the five different information environments).16 1.17 3.92 0. Low-ability workers. similarly. 22 We chose to ask this question at the end of the experiment rather than at the beginning in order to avoid potentially affecting behavior during the experiment.09 2. 3. Information level 1 2 3 4 5 Overall Higher ability worker mean wage Std. We believe that the answer to the merit-question is unlikely to depend on the outcomes of the experiment for two reasons: First.11 with normal information. as an inspection of the table below indicates.44 0. 2 is the good. and probit regression models.09 Information level 1 is the perfect.83 with no information. average effort. average wage compensation.19 2. Err. Lower ability worker mean wage Std. and the proportion of periods the worker was paid no more (or strictly less) than the coworker in a variety of different combinations.44 compared to 2.14 4. averages are 3.13 2. we generally include worker-fixed effects in our regression analysis below.77 0. we also show that it holds for people who believe in merit pay when a high-ability worker is paid strictly less than the paired low-ability worker. will ceteris paribus not change their effort level.78 0. (the corresponding numbers for wH ≤ wL are 47%.19 2. since they are included in the individual fixed effects (which we do not report). Since one would not expect workers who do not believe in merit pay to have their effort levels affected by relative wages (and indeed they do not).48 0.64 to 2. Err.26 in the perfect information environment.20 2. since it is of no particular interest.06 2.02 0.64. over 3. when paid no more than high-ability workers..16 3.44 0. However. and 56%. 2. Average effort on the other hand decreased slightly from 2. from 3. Lower ability worker mean effort Std. which is exactly what we would like to observe in order to statistically identify our merit-pay hypothesis. average coworker wage compensation.12 2.23 In regressions 3 and 4.17 3.21 with good information.14 2. 5.240 T. whereas average effort is significantly lower in the second cycle.05 0. using OLS.22 Nevertheless.15 2.

02 0.28** 0. We have also used robust errors clustered on individuals in regressions 24 One additional dollar leads to 0.17 2.25 −0. while the average effort was 2. managers pay higher wages to actual than perceived high-ability workers. In the normal-information environment the worker with the most answers in common with the manager is signaled as high-ability. our definition of Ifless is now a strictly negative relative wage and also applies to Ifless*Low.25 2.46*** 0..13 1. The asterisks (1.2.14 1. we add the dummy variable Ifless*High*NMerit (which is one if the person received a wage lower than or equal to that of the coworker. We have also used a linear wage specification. Gross et al.59–0.67 The dependent variable is worker effort.26* 0. leaving out the dummy variable Ifless*High*Nmerit does not change the significance of the results. which is when the high-ability worker receives strictly less than the co-worker (i. At the same time.70*** 0.64*** 0.68*** 0. in regression 4. Including time-period dummies in regression (iii) or information and cycle dummies in regression (iv) leaves results virtually unchanged as compared to the baseline. 5. as shown by wage compression across information levels in Section 6. the coefficient for Ifless*Low is not significantly different from zero (Table 7).54 = 1. 2.66 0.15 0. here we use wH < wL instead of wH ≤ wL ). but remain roughly the same as in Table 7. and 1% level.33*** 0. Indeed. which are not reported. Regression 1 Linear wage Wage = 2 Regression 3 −0.21 2.04 0. respectively. we should observe the highest wage differences in the normal-information environment.14 1. All regressions include dummy variables for each individual and a constant.66*** 0.34*** 0. .79*** 0.79 points. according to regression (1) a violation of merit pay is equivalent to reducing the wage by 1.67 Wage = 3 Wage = 4 Wage = 5 Wage = 6 Ifless*High*Merit Regression 2 Ifless*Low Ifless*High*Nmerit R2 Regression 4 0. and answered the merit-pay question negatively). is high ability.20 2. We believe this explanation does not apply in our experiment for the following reasons: First. The coefficient for Ifless*High*Merit is highly significant in all four regressions and large: violating merit pay reduces effort by 0. Since costs are lower than the effort levels. As has been found in many other experimental studies.28 0.34*** 0. which reinforced results in every specification. the merit-pay hypothesis still holds for all high-ability workers grouped together. The coefficient for Ifless*Low is tiny and insignificant. / Journal of Economic Behavior & Organization 117 (2015) 233–247 241 Table 7 Merit pay and effort.02*** 0. If social-identity considerations were driving the results.18 0. 26 All results in the table are from regressions with wage dummies. Alternatively. Moreover.25 We focus on high-ability workers who answered the merit-pay question positively. in terms of cost rather than the effort level.16 0. then there is no reason why one should react to not receiving a higher wage than the lower-ability co-worker. but do not report them here. there is a strong and significant effect of workers’ wages on the effort they supply.79 units of effort due to a violation of merit pay is equivalent to reducing the pay by 0. Robustness checks We have conducted several robustness checks which are summarized in Table 8.79/0.54 more units of effort.78*** 0.23 2.02*** 0. The coefficient on Ifless*High*Merit is somewhat smaller than before.14 1. the coefficient for Ifless*High*Nmerit is not significantly different from zero. There are 648 observations. Robust standard errors below the estimates. which signals the higher-ability worker as the one who has the most answers in common with the manager.26* 0. and has low ability). There could evolve an ingroup/outgroup thinking stemming from our mechanism.28 −0. In regression (ii) we employ a different definition of effort.54*** 0. since if one does not believe that higher-ability workers deserve a higher wage. 5%.20 0. then decreasing towards more and less information.11 0.46 dollars. whereas in the perfectand no-information environment the manager does not have any indication of which worker shared more answers with him or her on the test. Second.e.27 −0. as is apparent from regression (4). the coefficient for Ifless*High*Merit is of course correspondingly smaller but remains highly significant. the workers have no information whatsoever about the basis on which managers assign wages and whether they are in the ingroup or outgroup. Formally.04 and the minimum is 1.14 1. or 3) refer to results significant at the 10%. 25 One might think that social-identity considerations are driving these results instead of (or jointly with) our merit-pay hypothesis.26 In regression (i) we use a different definition of a violation of merit-pay. and perhaps more importantly. but still significant at the 1%-level.45*** 0. so −0.32*** 0.25 2.T.67 0.14 1.37 0. given that the mean wage was 3.19 0.59*** 0.03*** 0. Moreover.46.53 and the minimum possible is 1.17 0.24 which is quite substantial. as we show in Section 5.3.

60*** 0. / Journal of Economic Behavior & Organization 117 (2015) 233–247 Table 8 Robustness checks. Robust standard errors below the estimates. . clustered around 66 individuals.4.242 T. 5. or information treatments. since not only the means. The asterisks (1.15 −0. or 3) refer to results significant at the 10%. respectively. Ifless*High only −0. and 1% level. 5.19 0. or 3) refer to results significant at the 10%.02 0. controlling for individual 27 The number of clusters should not be too low. The asterisks (1. In regressions (III) and (IV) we also include the dummy Ifless*Low. Moreover. we have delegated results with clustered errors to the robustness checks.02 0. There are 648 observations. even if high-ability workers who believe in merit pay reduce their effort when merit-pay is violated.15 −0.24 (ii) Effort definition (iii) Period dummies (iv) Info and cycle dummies (v) Clustered errors (vi) Clustered errors and period dummies (vii) Clustered errors and info and cycle dummies The dependent variable is worker effort. controlling for individual fixed effects.59*** 0. There are 648 observations. The results are shown in Table 9. 5%.55*** 0.11 0. we still find a strong and significant (albeit smaller) effect.14 −0. does it matter for the overall provision of effort? We note that 25 out of 33 high-ability workers answered the merit-pay question positively.09 0. Manager intentions and responses to merit pay violations Finally.19 −0.09 0.06 −0. where applicable.60** 0.07 −0. Regression Ifless*High*Merit Ifless*Low (i) Strictly less −0.20 (II) Wage dummies. we wish to address the question of whether we can link responses to violations of merit pay to the intentions of the manager.3. Regression Ifless*High Ifless*Low (I) Linear wages. Table 9 Effects of merit-pay violations for all workers. 2.17 −0. and 1% level. We regress effort on wages – linearly in regressions (I) and (III) and with dummies in regressions (II) and (IV) – and the dummy variable Ifless*High. All regressions include dummy variables for each individual and a constant. Since we believe that we cannot cluster errors properly.60** 0.24 −0. In Table 10.18 0. All regressions include wage dummies and dummy variables for each individual and a constant.24 −0. While it is in principle best to cluster at the most aggregate level (although the question would remain which of the three is best). If managers’ intentions matter.18 −0.17 −0. 10. when we estimate the effect of a violation of merit pay on the entire population of high-ability workers. or 5 observations seems likely to make matters worse than not clustering at all.01 0. They refer – only somewhat in jest – to 42 as the minimum number of clusters.60*** 0. we report the results of regressions of effort on wages – linearly in regression (a) and with dummies in regression (b) – and the dummy variables Ifless*High*Merit*Info for each (non-perfect) level of information. both Ifless*High &Ifless*Low The dependent variable is worker effort. Gross et al. Size of effects One might wonder. Ifless*High only (III) Linear wages.08 0.68*** 0.24 0. as Angrist and Pischke (2008) point out. respectively. Robust standard errors below the estimates.13 −0. but also the error terms might be correlated for each worker.25 −0.40*** 0. 2.45*** 0. both Ifless*High &Ifless*Low (IV) Wage dummies. which is why we do not cluster errors around sessions. whereas the coefficient for Ifless*Low remains insignificant. clustering on 6. (v)–(vii).14 −0.09 0.27 We have also combined robust clustered errors and time-period dummies in regression (vi) or information and cycle dummies in regression (vii). 5%.24 −0.08 0.22*** 0.40** 0. the coefficient for Ifless*High*Merit retains about the same magnitude and remains always significant at least at the 5% level. time periods. then the reduction in effort should be strongest when the manager has the most information. We test whether violations of merit pay are “punished” differently depending on the information the manager had.

Ifless*High*Merit*Info2 Ifless*High*Merit*Info3 Ifless*High*Merit*Info4 Ifless*High*Merit*Info5 Regression (a) linear wages Regression (b) wage dummies −0. the corresponding coefficient is naturally not significantly different from zero.70*** 0.20 −0. We fail to reject the hypothesis that all the coefficients are equal to each other (the corresponding joint F-tests in regressions (a) and (b) have p-values of 0.24 The dependent variable is worker effort. Model Our formal hypothesis regarding wage compression is: Hypothesis 2. Gross et al.71*** 0.80*** 0. The asterisks (1.e. 6. which we do not observe.19 −0. independent of fairness considerations. There are 648 observations. . With a linear production function. responses to violations of merit pay should be stronger. the more information the manager has (since the ability of the manager is then more likely revealed). only in combination with uncertainty about true worker ability do they affect wage compression. We exclude the possibility of wage compression due to fairness concerns of workers or due to merit-pay considerations for two reasons.1.84*** 0. First. we would presume that both low. This also indicates that workers do not resent managers with presumably low ability per se. both in the first period of a session.23 −0. in fact. so managers compress wages compared to the full-information benchmark.26 −0.80*** 0. We assume that these beliefs are of the following nature: the effort function e(w) is increasing and strictly concave in wages w. Second. in our empirical analysis we can isolate wage compression from fairness motives by managers or workers. We then test whether the degree of uncertainty does indeed result in wage compression in our experimental data.and high-productivity workers react negatively to a low-ability manager. as in Charness and Kuhn (2005). the higher the wage compression.989 respectively). wage compression here does not rely on the behavioral component expressed in the merit-pay hypothesis. given some beliefs about how workers respond to wage payments with effort. If this were the case. Given the very limited data. merit-pay considerations by themselves are not enough to result in wage compression. respectively. 4 is the poor. and 5 is the no-information environment.19 −0.79*** 0.28 The coefficients are large and negative and have roughly the same magnitude for each of the information levels. Moreover.T.75*** 0. the lower will be the difference between the wages assigned to the two workers in a firm. as there were only two observations of violations of merit pay. the expected payoff is: E() = p{H e(wH ) + L e(wL )} + (1 − p){H e(wL ) + L e(wH )} − wL − wH (1) 28 We exclude the perfect-information treatment (Info1) from the regression.995 and 0. / Journal of Economic Behavior & Organization 117 (2015) 233–247 243 Table 10 Manager intentions and punishment of merit-pay violations. This suggests that the degree to which high-ability workers reduce effort as a consequence of the violation of merit pay does not vary across information conditions. only strictly concave effort functions give a reason to pay both workers. The concavity assumption seems to be in line with results reported in previous studies on workers providing effort and the fact that managers assign wages to both the more and less productive workers. All regressions include dummy variables for each individual and a constant. in the form stated before. 3 is the normal. We can derive this hypothesis by using a simple model in which managers aim to maximize expected profits. Interestingly.20 −0. and the derivative tends to infinity as w tends to zero. Wage compression In this section. 2. 6.68*** 0. Assuming risk-neutrality for managers (risk-aversion would reinforce the effect). fixed effects.27 −0. 5%. The higher the degree of uncertainty about worker ability. the hypothesis being true does not affect wage compression at all. and 1% level. The assumption of lim e(w) = ∞ simply assures an w→0 interior solution. Robust standard errors below the estimates. Uncertainty about worker types provides an incentive for wage compression. Information level 2 is the good. At the end of this section we briefly discuss how the merit-pay hypothesis would affect results. or 3) refer to results significant at the 10%. we first present a simple model of wage compression due to uncertainty about worker ability. The expected productivity of the perceived high (low) ability worker is thus lower (higher) when uncertainty is higher. i. The reason is that managers assign with a certain probability the higher wage to the less-skilled worker.

for information levels 3 and 4. That is. the estimated coefficients become larger in absolute terms with increased uncertainty. of course). In regressions 2M to 4M the dummies for each information environment compared to the first with perfect information are also negative. wH is the wage assigned to the perceived high-ability worker and wL is the wage assigned to the perceived low-ability worker. We did not 29 We get this result by combining Eqs. the difference between wH and wL is strictly decreasing in uncertainty about the true type of the worker (this simply follows by Jensen’s inequality).5. and seeing that e(wL ) > e(wH ) for a concave effort function.33 We have also included the actual test scores of the managers.5 is the probability the manager perceives correctly who the more productive worker is and  i is the productivity of a worker of type i ∈ {L. the hypothesis is rejected at the 10% level. Any p < 0. That is. if the high-ability worker only withdraws effort when her wage is strictly lower than her co-worker’s. if we can observe a higher degree of wage compression in periods with less information compared to those with more information. wH > wL . and 2 and 3 are equal to each other are both rejected at the 1% level. large. To show wage compression. Taking derivatives with respect to these wages yields: pH e(wH ) + (1 − p)L e(wH ) = 1 (2) pL e(wL ) + (1 − p)H e(wL ) = 1.5 is of course equivalent to assigning the higher wage to the worker who appears to be of lower skill – we can simply redefine the probability of getting it right then as pˆ = 1 − p. then the wage compression due to equity considerations should be differenced out. the greater the degree of wage compression. When there is complete uncertainty. p = 1 − p. large. making use of the result that the perceived high-ability worker is always assigned a higher wage for p > 0.5.32 Including manager fixed effects in regression 4M does not change the magnitude or significance of the estimates much (the goodness of fit increases. 31 In the knife-edge case where p = 0. wH .5 < p˜ ≤ 1 is defined implicitly by ∗ ˜ ˜ H  = E[(w∗ .2. 33 A potential problem is that the pre-estimate could be correlated with other factors that would lead the manager to pay more unequal wages.5. = dp −e (wL )(pL + (1 − p)H ) (5) The numerator is positive in the first case and negative in the second case. managers should pay one worker a slightly higher wage than the other so that there is a 50% chance that it is not violated. H}. i. we asked participants how many questions they think they answered correctly. The difference between wH and wL is thus a monotonically-increasing function of p which is zero when p = 1 − p = 0. we thought that the wage disparity should positively depend on the confidence a manager has in his ability to discern the high. while we fail to reject the hypothesis that the coefficients for information levels 4 and 5 are equal. First. 30 . the hypotheses that coefficients for information levels 1 and 2. instead. rather than simply reflecting the perceived ability to better discern high. As a measure of wage disparity we use the absolute value of the difference between the two wages the manager assigned. it would actually be optimal not to pay equal wages.and low-ability workers. We also include either some control variables (in regressions 1M to 3M) or fixed effects for the manager (in regression 4M). Now we introduce a term in the firm’s expected payoff equation which reflects the merit pay hypothesis. The larger the negative coefficient. In regression 2M. ˜ E[(wL∗ . Gross et al. then a negative coefficient for an information level implies that the wage differential is decreasing in size as compared to perfect information. and highly significant. since such considerations should be separate from the information condition. We regress this on a linear term for information in regression 1M and a dummy for each information environment in regressions 2M to 4M. (6) This does not affect any of the first-order conditions and wage compression is thus independent of merit pay considerations. pi)]. moreover. For the control variables in regressions 1M to 3M.29 Second. 32 If the dependent variable is the absolute value of the wage differential |wH − wL |. and highly significant. Empirical test of wage compression We can identify this hypothesis independently of equity motives.e. where 0. This is the variable pre-estimate. providing evidence of wage compression. In regression 1M the linear information term is negative. w∗ . pi)] − (1 − pi) where wL∗ . At the end of the aptitude test in the beginning of the experiment.5)30 : dwH (H − L )e(wH ) >0 = dp −e (wH )(pH + (1 − p)L ) (4) dwL (L − H )e(wL ) < 0.244 T. Our prior is confirmed in the sense that the estimate for it is positive and significant. (2) and (3). as e (w) < 0 by concavity. since this would result in a certain violation of merit pay. we take the total derivative of the first-order conditions and rearrange (this obviously only holds for p > 0. wages assigned to more productive workers are higher than for their less productive colleagues.and low-ability workers. The denominator is positive in both cases. which were – as previously mentioned – not known to them.31 6. then the model would result in a jump. In Table 11 we observe that there is a strong wage-compression effect. (3) This has two implications. with probability 1 − p merit pay is violated and the high-ability worker will reduce effort by some amount  > 0: E(MP ) = p{H e(wH ) + L e(wL )} + (1 − p){H [e(wL ) − ] + L e(wH )} − wL − wH . / Journal of Economic Behavior & Organization 117 (2015) 233–247 where p > 0. then equal wages should be assigned for all p ≤ p˜ .2. If we consider the modified merit-pay hypothesis as in regression (i) in Section 5. wH∗ = argmaxE[] and w∗ = argmaxE[] subject to wL = wH . the wage assignments are completely random guesses and wH = wL . since e(w) > 0 and  H >  L . Thus.

91*** 0.03*** 0. respectively.06 0.34 −1. However.05 0. the test score coefficient is not significant in the wage compression regression. 5%.43 The dependent variable is the absolute value of the wage difference.05 0. it is not statistically significant. Regression 1M Information Good information Normal information Poor information No information Preestimate Test score Rel.and low-ability workers. 35 As further robustness checks for all specifications.08*** 0.24 No 0. This depends on the variation of the managers’ test scores.45*** 0.25 0.34 In regression 3M we have also included an interaction term of this dummy with the information environment. while regressions 2M and 3M contain a dummy for whether or not the response was 2 (matters a lot). 2. to which the others compare to.3 and the standard deviation 2. when we run regressions on how well the test score predicts mistakenly assigning the “wrong” wage (i. Including a dummy variable for the cycle did not change the results either. and 0 otherwise.02 0. / Journal of Economic Behavior & Organization 117 (2015) 233–247 245 Table 11 Wage compression under uncertainty of worker ability.03 −0. Including a separate dummy for the one manager who answered 0 does not change the results.16 0.02 −0.27 −1.40 −0.84*** 0.74*** 0.03 0.26 −0.26 −1. In regression 1M the variable Information takes the value 1 for perfect.08*** 0. there were 1. . since equality concerns could be dependent on the managers’ intentions and thus the information available.08*** 0. Adding this and any subset of control variables did not change the main coefficients of interest in any substantial way: all of the numerous specifications that we ran yielded approximately the same result for wage compression. including controls for the information level and obviously leaving out the perfect information setting.68** 0.T.13 Rel.20] and the observed minimum in our sample is 4. the interaction term for test scores in the normal-information environment is positive.35 One element of our experimental design is that managers differ in their ability to discern the high.05 0. and 1% level.12 No 0.04 0. As discussed above. which were virtually the same as robust errors. 34 We have thus combined the answers 0 (that it does not matter) and 1 (that it matters a little). While the testscore coefficient is positive in a regression on managers’ profits. The asterisks (1. It thus appears that concerns about relative wages did not matter for wage compression. and a lot is 2. The variable Rel. 2 for good. Rel.16 0. robust standard errors below the estimates. wage matters (dummy) Rel. The variable Q: Merit takes the value 1 when the manager positively answered the question if only high-ability workers respond negatively when paid less than their co-workers. While we had anticipated a negative coefficient. but when we regress manager profits on test scores interacting with the information level (plus controls for the information level). large. 3M.25 −1. though. have any prior on this and the coefficient is insignificant. or 3) refer to results significant at the 10%. It holds for a linear probability model.26 −1. large. a little is 1. wage matters*Info Fixed effects R2 No 0.02 −0. 3 for normal.51*** 0. In regressions 2M. wage matters*Info.02 −0.05 0.e.6.22 −1. depending on what the managers answered to the following question in the questionnaire at the end of the experiment: “Do you think that workers respond to their co-workers wages?” The answer no corresponds to a 0.03 −0.78*** 0.26 Yes 0. and 19 observations in these respective categories.53*** 0. wage matters takes a value of 0. Both results seem to support our idea of inept managers: they assign more often “wrong” wages and their profits suffer in consequence. This term is also insignificant. and highly significant at the 1% level. we cannot identify the clusters correctly. whereas the ability to correctly identify the high-ability worker does matter: this lends support to the notion that wage compression is not driven by equality concerns in our experiment. 13.28 −1. The test score for managers is in the interval [0. while being insignificant at the other information levels.88*** 0.49*** 0. and 4M the perfect information environment is the omitted dummy. a probit. we have also used errors clustered on managers. that the wage assigned to the perceived high-ability worker does not correspond to the wage received by the actual high-ability worker). or 2. and 5 for no information. 324 observations. In regression 1M we use this regressor linearly.25 −2.35 0. When we include manager fixed effects. and logit. Gross et al.27 0.23 −1. 1. the maximum 16.70*** 0. so there is substantial variation in employer accuracy. 4 for poor. and highly significant at the 1% level. then the coefficient is negative.99*** 0. wage matters (linear) Regression 2M Regression 3M Regression 4M −0. The average is 11. we report some further results on the relevance of test scores below. it turns out to be insignificant. All regressions include a constant.

Am. 2004. Q.doi. 1299–1324. Econ. 818–827. we have barely scratched the surface and invite other researchers to join us in exploring this important area. Who should be paid how much and who should be promoted? It is often the case that it is difficult for a superior to precisely assess the ability and productivity of a worker.. Appendix A.. Hole. 664–681. Does our experiment shed light on this real-world problem? We believe that the basic mechanism of hedging bets against paying unidentified high-performers a low wage should be the same in the real world as in our experiment.. J. 112 (3). Wage inequality and team production: an experimental analysis. . The degree of uncertainty changes from period to period. 2. G. Kamlani. we find that the more questions the manager thinks he answered correctly. 97 (3). at References Abeler.... S. Harvard University Press. J. Inequity in social exchange. J. In previous experiments where the high-ability worker is almost always paid more than the low-ability worker.org/10. S. M. O. Bewley. The reasons for wage rigidity: evidence from a survey of firms. 2011.. Psychol. Eur. Q. Bolton. Conclusion We explore wage setting and effort choices in a multi-worker setting when there is heterogeneity in worker productivity and managers’ perception of this productivity is imperfect. The managers’ intentions did not seem to play a role for the responses to violations of merit pay.009. von Siemens. Psychol. Mostly Harmless Econometrics: An Empiricist’s Companion.. Bolton. Rev. Altmann. Socioecon. Econ. Kube. E.246 T. J. This is consistent with our hypothesis.. Sørensen. 917–962... Angrist. 102 (6). Adams. A. 93–104. J. C. Sørensen. then actual managers should do so a forteriori.. J. lending credence to our assumption that this wage compression is independent of concerns for inequality. A. The merit hypothesis is strongly supported by our experimental data. 14 (3). Bardsley.2015. 2005. K. J. Department of Economics. Princeton University Press. Adv. Agell.. 26 (5). B.. violations of merit pay were still significant even for the entire group of high-ability workers. Saez. Since the proportion of high-ability workers who believe in merit pay was high and the effect so strong. whereas low-ability workers who are paid a lower wage will not respond negatively. We create a novel mechanism that induces an endogenous degree of uncertainty towards the true ability level of workers. J. Barankay. A.. 759–789. To the best of our knowledge. Why Wages Don’t Fall During a Recession. Gift exchange and workers’ fairness concerns: when equality is unfair.. 2010.D. Ockenfels. this article can be found. The pluralism of fairness ideals: an experimental approach. Are efficiency wages equality wages? Exogenously induced fairness norms in working environments. Bennmarker. in an innovative design. reciprocity. Inequality at work: the effect of peer salaries on job satisfaction. manager uncertainty concerning this ability is related to the manager’s own test performance.-S. A. allowing us to explore a related issue: if relative wages do not matter. Cappelen... Card. G... 32 (1). 54 (3). Econ. N... 33 (4). S. I. J. then why should firms engage in wage compression? We expected (and find) that managers would assign more equal wages with increased uncertainty about workers’ true ability. University of Cologne. R. P. Lit. Brown. While these are new results. Worker ability is assigned via an aptitude test and. Moretti. it is difficult to test for this effect. E. Rev. the more equal the wages that managers assign on average to their high. D. F. We view uncertainty about worker ability as a realistic and important part of management’s leadership role. T. Bandiera. Social preferences and the response to incentives: evidence from personnel data. J. and competition. as the previous results of the experimental literature suggest. 2008. Q. Rev. The higher the uncertainty concerning workers’ true ability. Q. 1245–1279.W.. Bekman. we are the first to find experimental evidence of optimal wage compression due to uncertainty. 2981–3003.. Working Paper Series in Economics 56. Gross et al. ERC: a theory of equity. Psychol. Supplementary data Supplementary data associated with http://dx.. Econ. Pischke. Machin. / Journal of Economic Behavior & Organization 117 (2015) 233–247 7. Edited by Kenneth Arrow..E. Rev. 2012.. 1998. in the online version. Our paper establishes that: (1) when higher productivity is the result of ability rather than random assignment. When in Rome: conformity and the provision of public goods.. 2001. Sausgruber. B. Econ.jebo. In addition. if students are able to consistently recognize the problem and assign wages accordingly... 2010. and (2) that wage compression is more common with increasing uncertainty concerning which is the more productive worker.. Am. Econ. the higher the inequality of wages he assigns. Econ. Indeed. Eur. J. Responsibility for what? Fairness and individual responsibility. 1997. 113 (4). Conformity and reciprocity in public good provision. Samuel Bowles. 2007. J. Labour Econ.. H. Am. C. since answering more questions correctly reduces the uncertainty regarding the true type of a worker. 1990. 120 (3). Implications of skill-biased technological change: international evidence. 429–441. J. A.. Cappelen.and low-ability workers. and Steven Durlaf. 347–369. Tungodden. Campbell III. Bartling.. Assoc.. 1–16. 395–408. B. J.D. Yellen. Our merit-pay hypothesis asserts that a worker’s ability level (and the corresponding productivity difference) is critical in anticipating an effort response to coworker wages: a high-ability worker who believes in merit pay will be disgruntled if paid a lower wage and will punish the manager who assigned her the wage by withholding effort. thereby generating a relatively large number of observations that violate merit pay.. Soc. The fair wage-effort hypothesis and unemployment. 2012.1016/j.W. 1999. 255–283. 39 (1). Ø. Carpenter. 16. Numerous robustness checks confirm the validity of our result. Mas. 2007. Ø. Wibral. 6–193. a highproductivity worker typically reduces her output when she receives a lower wage than a less productive worker.E.. 2005. 267–299. Exp. Econ. Werner. G. 2000. Moreover. Bound. Wage incentives and wage rigidity: a representative view from within. 105 (2). Econ. J. Econ. I.. We consider asymmetric effects of wage comparisons on worker effort. 8. Rasul. Tungodden.. managers’ evaluations of whether relative pay mattered for the provision of worker effort was insignificant for wage compression.06. “Review of” Meritocracy and economic inequality. Econ. Akerlof... 1966. J.F. J. Econ. E.

D. G. Weichbold.. 82–101. G. 437–459. Fehr. Königstein. Rivas. Charness. R. 2001. Econ. S. 74 (4).. 573–594. Leonard. Financ. 29 (2). Ind.. Rev. The impact of pay comparisons on effort behavior. P. Top executive pay: tournament or teamwork? J. 2007. Satisfaction and comparison income... M. Cowherd. G. Does pay inequality affect worker effort? Experimental evidence. Econ. Roberts. Maximiano.. D. Procedural Fairness and the Tolerance for Income Inequality..01. C. J. E. 1981. Gantner. 2013. P. J. Behav. 783–821. 1072–1091. 47 (2).. D.. Q. 262–280.14). Gross et al. Gächter. J. 49 (3). Kirchsteiger. Knell. Violante.... 2007. 1–1302. J. Kay. J. P. Frank. A.. Goerg. Executive compensation and tournament theory: empirical tests on Danish data. 1999.. S. Scand. B. J. J.. Labor Econ. Q. 324. Salmon. Wage gaps destroy employee morale. Charness. Econ. Fairness within firms: the case of one principal and multiple agents. Treating equals unequally: incentives in teams. 2012.. 1982. J. Equitable choices in bargaining games with joint production. Milgrom.-V. Zultan. Rosen. 69 (2). G. Adm. F. Bell J.. reference points. 117 (522). J. 2006. D. and production technology. 1025–1050.. Soc. 2005. D. 2001. 46 (2). G. and police performance.. Econ. T. 2002.T.. 1993.. Camerer. Policy 5 (2).. 89 (5). J. Q. Hibbs Jr. 2–407... Organ. Scand. 1999.. S. J. Public Econ. Rev.. 124 (1). Econ. Oswald.. / Journal of Economic Behavior & Organization 117 (2015) 233–247 247 Carpenter. 764–773. Econ. 417–435. H. Rev. vol. Econ.. J.. J. E. C. Rev. 1993. 1–49.. J. 32. Pers. 1346–1367.N. Who compares to whom? The anatomy of income comparisons in Europe. A. competition. J. J. Choosing the Joneses: endogenous goals and reference standards.. 359–381. A. Zweimüller. J. Bazerman. 2007.. 315–333. J. 2. Ríos-Rull. A theory of sequential reciprocity. A. 2006. J. Charness. Wage dispersion and productive efficiency: evidence for Sweden. M. Sloof. Am. Am. 209–225. Kirchler. J.B. R. 1992. 1993. 28 (4). Economics. M. Fehr. 2009.. 2009. Econ. C. 2004. 60. 114 (3).... Organization and Management. Performance pay and wage inequality. Incorporating fairness into game theory and economics. Pfeffer. 1999. M. J. Behav. T. Econ. A. productivity. Q. G. 693–723.. Sci. C. Sonnemans. Gift exchange in a multi-worker firm. USA Today (accessed 09.. Behav. Econ.. Promoting employees: how to get it right. Ku.. Organ. 549–571. Harris. Gächter. 2000.. M. J. Post (accessed 09. Polit. Moral hazard in teams. 755–782. 90 (4).. Dufwenberg. Psychol. 2010. Lab. 2010. 25 (4). Sci. 30. R. A. 2010.. B. 788–794.. BE J.. J. 121 (3)... workers’ motivation. Rank-order tournaments as optimum labor contracts. Rev.. Stone. Winter. 1051–1072... M. L. The Centre for Decision Research and Experimental Economics. A theory of fairness.. and working collaboratively: evidence from college and university faculty. Incentives and discrimination. J. S. 531–543. 7. Econ. Eriksson. and Effort: Theory and Evidence. 94 (3). Kuhn. Econ. A. Econ. P.. 101 (1). Levine. L. Pay Inequality. Wade.. 2004. 1993. J. 1996. Seki. Deveau. Gächter. MacLeod. Dispersion in wage premiums and firm performance. Zala-Mezø. Q.. Mas. Thompson. Econ. 1995.. Falk. 61 (3).. Rabin. A. At work: employees’ biggest frustration? ‘Idiot’ bosses.. Charness. 817–868.. Holmstrom. Gill.. 426. 18 (4). 346–364. Holmstrom. 6–628. 57 (3). 108 (2). 1989.. Labor Econ. 106 (3). Senik. Adm. 16 (2).. Does fairness prevent market clearing? An experimental investigation. R. K.. Wourio. W.. T. G. Charness.. D.E. 117 (522).. NBER Working Papers 11786. 1992.... 53 (2). Lett..L. Econ. E. A theory of wage dynamics. Kirchsteiger. 2–320.14). Econ. Behav. 117 (3). Locking. 4–340. Schmidt. Intra-firm wage dispersion and firm performance. Kovács. M. W. 747–772. Fehr. Econ. Pay Secrecy. Competitive work environments and social preferences: field experimental evidence from a Japanese fishing community. Wage dispersion and workers effort. (accessed 09. 43.D. Thöni. Nosenzo. The impact of social comparisons on reciprocity. Am. Games Econ.. Prentice-Hall Englewood Cliffs... Are workers paid their marginal products? Am. Konow. Q. Games Econ. 2004.. 665–688. Fair shares: accountability and cognitive dissonance in allocation decisions. 2011. Econ. Microsoft Bus. E..01. Labor Econ.J. Econ. 76 (3). M. E. 2013.. D. Q. Krusell. N. J. 817–869. Social comparison and performance: experimental evidence on the fair wage-effort hypothesis. Labor Econ. Kyklos 52 (4). Güth. 114. A. NJ.. 1029–1054. Inc. 555–572. Fairness and desert in tournaments.. B. Levine. J... Langton. The effect of wage dispersion on satisfaction. Lazear. 2008. Clark. 10 (1).. J.. Knez. 2000. 2000. Winter-Ebmer. Intention and stochastic outcomes: an experimental study. S. Attribution and reciprocity in an experimental labor market. 841–864. Rev. In: Discussion paper. Labor Econ. Executive pay and firm performance. Martins. R. 38.. J. Pay.. E. Capital-skill complementarity and inequality: a macroeconomic analysis. P. Ohanian. J. 1982.01. University of Nottingham. Kube. 22 (3). Games Econ. Rel. Labor Econ. Florida State University. Rabin.. 2012. W. Schmalenbach Bus. 2010. . Econometrica 68 (5). S... Econ. and cooperation. D.. 2004. E. Anal. Econ. Loewenstein.. Clark. Riedl. Parent. Understanding social preferences with simple tests. Product quality and pay equity between lower-level employees and top management: an investigation of distributive justice theory.. In: Discussion paper. Stud. 128. School of Economics. 65–94. J. Kuhn. 1984. E. 1998. Behav. Social utility and decision making in interpersonal contexts. M.. 1989. When social norms overpower competition: gift exchange in experimental labor markets. Bull. 120 (544)..14). Lemieux. 2010.. Main. O’Reilly III. National Bureau of Economic Research. 63–65. Sefton. 268–298. Outside options and social comparison in three-player ultimatum game experiments. S. 17 (2). H. productivity. Konigstein. G. Labor Econ. M. Rev. 13-S. J. G. Güth. Econ..