tournaments without prizes - evidence from personnel records

MANAGEMENT SCIENCEVol. 57, No. 10, October 2011, pp. 1721–1736issn 0025-1909 �eissn 1526-5501 �11 �5710 �1721 http://dx.doi.org/10.1287/mnsc.1110.1383

© 2011 INFORMS

Tournaments Without Prizes:Evidence from Personnel Records

Jordi Blanes i Vidal, Mareike NossolManagerial Economics and Strategy Group, London School of Economics, London WC2A 2AE,

United Kingdom {[email protected], [email protected]}

We use a quasi-experimental research design to study the effect of giving workers feedback on their relativeperformance. The setting is a firm in which workers are paid piece rates and where, for exogenous reasons,

management begins to reveal to workers their relative position in the distribution of pay and productivity. Wefind that merely providing this information leads to a large and long-lasting increase in productivity that iscostless to the firm. Our findings are consistent with the interpretation that workers’ incipient concerns abouttheir relative standing are activated by information about how they are performing relative to others.

Key words : tournaments; relative concerns; status concerns; relative performance feedback; relativeperformance evaluation

History : Received November 4, 2009; accepted April 1, 2011, by Olav Sorenson, organizations. Publishedonline in Articles in Advance August 4, 2011.

1. IntroductionDoes receiving feedback on relative performancemake workers strive for success, or does it leadinstead to dissatisfaction and loss of motivation? Doesthis response depend on whether the comparisonwith other workers is favorable or unfavorable? Inthis paper we use a unique data set from a firm-levelquasi experiment to investigate these issues.

An important body of work has studied the use ofrelative performance from the perspective of explicitincentive theory (Lazear and Rosen 1981). Accord-ing to tournament theory, an agent’s performancecan provide a useful benchmark upon which otheragents’ performance can be evaluated and rewarded.The notion that competing with others for mone-tary rewards can have motivation effects has receivedempirical support in various economic settings (e.g.,Knoeber and Thurman 1994, Eriksson 1999, Casas-Arce and Martínez-Jerez 2009).

Alternatively, it may be that receiving feedback onrelative performance has in itself substantial conse-quences for employee satisfaction, motivation, andproductivity (Kluger and DeNisi 1996). Consider, forinstance, a setting in which workers are compen-sated using piece rates, so that better performers alsoreceive higher compensation. If workers display rela-tive concerns (Festinger 1954, Frank 1985, Buunk andGibbons 2007), information on relative performancewill affect well-being and effort decisions even whenpay is independent of relative performance per se.

Understanding this direct motivation effect isimportant for several reasons. First, at a general level

it can contribute to our knowledge of the interlinksbetween preferences, behavior, and the informationavailable to individuals. Second, it can shed lighton the additional, probably unintended, effects ofexplicit incentive schemes, like tournaments, that relyon comparisons across workers. Third, it has directimplications for optimal communication policies. Inorganizations where information on relative standingcan be computed easily, there is clearly potential fora judicious communication policy to increase produc-tivity at little or no cost to the firm.

In this paper we use a field quasi experiment tostudy the direct effect of informing workers abouttheir own position in the productivity and pay dis-tribution. There is a scarcity of evidence using realworkers in real economic conditions to study theeffects of relative performance feedback. In additionto the usual difficulties in obtaining data from orga-nizations, there is the additional difficulty that, whenrelative performance is communicated in firms, it typ-ically has explicit or implicit monetary consequences.For instance, a worker being informed that he hasperformed better than his colleagues may reasonablyconclude that he has a chance of receiving a promo-tion to a better paid position, which could translateinto higher effort. This makes it difficult to disentan-gle the direct effect of relative performance feedbackfrom the effect of the monetary reward to which suchfeedback is often linked.

In this paper we take advantage of a uniquemicrolevel data set, constructed from a German

1721

Blanes i Vidal and Nossol: Tournaments Without Prizes: Evidence from Personnel Records1722 Management Science 57(10), pp. 1721–1736, © 2011 INFORMS

wholesale and retail organization. In the main ware-house of this organization, workers were being paidpiece rates. From some point onward, however, theybegan to receive private information on their ownposition in the distribution of pay and productivity.For reasons that we describe in detail below, this rel-ative performance information exercise (henceforth,RPIE) did not have (and was known to not have)any type of compensation or career consequences. Itsintroduction was triggered by exogenous reasons andwas not part of any wider managerial policy.

Although not a randomized trial, a key feature ofour data set, namely, the availability of informationon individual daily performance, provides the basisfor a quasi-experimental research design. We makeuse of this rich information to estimate the produc-tivity increase following the introduction of the RPIE.We find the immediate average increase to be 608%and find no evidence that this effect vanishes overtime. The increase is consistent across the distribu-tion of productivity. We also show that this long-termincrease in productivity (i.e., in “quantity”) did notoccur at the expense of either a long-term decrease inthe quality of the work produced or an increase in thequit rate among the workforce.

To understand the economic mechanism behindthis increase in productivity, we exploit the differen-tial timing in the two parts that constitute the RPIE.The information was communicated on the first dayof the month and referred to the relative perfor-mance during the month before the previous one. Forinstance, on November 1, workers learned about theirperformance during the month of September. As aresult of this timing, we can distinguish the “kick-off moment” (the date in which workers performancefirst started to count toward the information that theywould receive) from the “revelation moment” (thedate in which workers first learned their relative per-formance). We find that productivity increases notonly at the revelation moment, but also at the kickoffmoment. The fact that productivity starts increasingeven before workers actually learn any new informa-tion casts doubt on most potential explanations of ourfindings, including those based on pure worker learn-ing about absolute ability or the distribution of wages.

Our preferred interpretation suggests that theobserved increase in productivity was caused by anincrease in concerns about relative standing. As wediscuss in detail below, our results are consistent withinformation about relative performance and the antic-ipation of such information leading to an increase inrelative concerns that shifted the distribution of pro-ductivity to the right, because all types of workerstried to improve their relative position in terms ofpay and/or productivity. Our findings therefore addto previous studies arguing that workers are affected

by relative concerns, whereby increases in others’ per-formance always has a negative effect on own utility(Luttmer 2005, Charness and Kuhn 2007, Clark et al.2008). It has been argued that the existence of suchconcerns leads to excessive levels of effort and con-sumption, creating a role for welfare increasing publicpolicy (Oswald 1997). We interpret the findings of ourpaper as providing evidence on the existence of “con-tests for relative position” using field data.

1.1. Related LiteraturePsychologists and management scholars have longrecognized that workers are motivated by more thanmonetary rewards (Etzioni 1971, Deci 1975). Individ-uals exert effort at work partly because of interestor enjoyment in the task itself or because completingthe task provides a feeling of accomplishment or self-worth. Typically, the presence of extrinsic rewards isbelieved to have the effect of “crowding out” intrin-sic motivation (Kohn 1993, Deci et al. 1999). At thesame time, Deci (1972) and Anderson et al. (1976) findthat (positive) feedback increases intrinsic motivation.In our setting, we argue that relative performancefeedback does indeed increase relative concerns andtherefore intrinsic motivation. The existence of perfor-mance pay most likely reinforced the strengtheningof relative concerns following the RPIE and thereforeserved to enhance the increase in intrinsic motivation.

Feedback interventions can sometimes lead to adecrease in performance (Kluger and DeNisi 1996).A substantial body of work in human resourcemanagement and applied psychology has thereforeanalyzed the circumstances that make job feedbackenhance satisfaction and performance (Nadler 1979,Latham and Wexley 1981, Moore and Klein 2008).Because in most organizations evaluations are mainlysubjective, research efforts have often focused onwhich organizational actors should contribute to theevaluation, what format the rating should take, andhow to best present it to the evaluated worker (for anoverview, see, for instance, London 2003). In our firm,performance was easily quantifiable, and thereforeit was possible to be both objective and detailed inthe information presented to workers.1 Furthermore,the objective nature of performance and the clarityof the goal implied that workers dissatisfied withtheir relative performance could easily understandthat a reduction in the gap between expectations andfeedback was possible through the exertion of higher

1 “Constructive” feedback is regarded as more likely to lead toincreased performance (Baron 1988). London (2003, pp. 15–16)argues that feedback is constructive when “it offers concrete infor-mation that can be used,” “it is clear and easily understood,” and“it is interpreted similarly by the source and recipient.” All thesecharacteristics clearly apply to the RPIE that we study.

Blanes i Vidal and Nossol: Tournaments Without Prizes: Evidence from Personnel RecordsManagement Science 57(10), pp. 1721–1736, © 2011 INFORMS 1723

effort (Kluger and DeNisi 1996, Bandura and Cervone1983). This fact may contribute to explain the positiveeffect on productivity that we document.

In economics, Eriksson et al. (2009) find no effectof providing relative performance feedback to labo-ratory subjects compensated via piece rates. Kuhnenand Tymula (2010) develop a theoretical model of self-esteem, study subjects in a laboratory flat wage setting,and find that relative performance feedback has bothex ante (i.e., for agents anticipating it) effects and expost (for agents receiving it) effort effects. Their resultechoes our finding that the RPIE affects performanceboth at the “kickoff stage” and at the “revelation stage.”

Two recent experimental papers try to separatelymeasure the importance of cardinal versus ordi-nal considerations. Clark et al. (2010) find thatagents’ rank in the income distribution affects effort,although, controlling for rank, relative income doesnot. Sunde et al. (2011) observe laboratory subjectsusing functional magnetic resonance imaging. Theyfind that both relative income and rank affect regionsof the brain associated with the processing of basicrewards. In our firm, the RPIE consisted of both cardi-nal and ordinal information, and unfortunately theirrelative importance cannot be separately identified.

Several papers study relative performance feedbackusing field data. In a context where workers are com-pensated according to team performance, Delfgaauwet al. (2009) find that information on relative perfor-mance leads to higher productivity, whereas Bandieraet al. (2009) find the opposite. In a schooling context,Azmat and Iriberri (2010) and Tran and Zeckhauser(2009) argue that learning about relative performanceleads to higher student effort.

Finally, this paper contributes to the literature ontournament schemes by documenting a new mar-gin through which competition operates, namely, the“pure contest for relative position,” which ariseswhen individuals derive utility directly from out-performing others in terms of performance or pay.Importantly, incentives along this margin are free tothe principal: the 608% average increase in produc-tivity that we observe in our setting was achievedat no cost to the firm, other than the negligible costof communicating with workers. Thus, even tourna-ments without prizes can have large incentive effects.

1.2. Outline of This PaperSection 2 discusses the institutional setting in whichthe RPIE occured. Section 3 presents and interpretsour main results. In §4 we examine other effects ofthe RPIE. Section 5 concludes.

2. Institutional Setting2.1. Workplace Conditions and CompensationWe study an intervention occuring in the second halfof 2001 in the main warehouse of a German wholesale

and retail organization. The company’s wholesaledivision handles a large number of food and nonfoodproducts, which are sold to independent supermar-kets as well as the company’s own retail division.

During our sample period, the firm employedaround 65 workers to perform the core task of thewarehouse. This task involves picking up customerorders, assembling the requested products, packingthem onto a trolley, and moving the trolley to thegoods-out area of the warehouse.2 Work is strictlyindividual, and opportunities to observe each other’swork are limited. Worker’s salary comes in threeparts: a fixed base salary, which is the same forevery worker,3 a “productivity” or “quantity” perfor-mance component, and a “quality” performance com-ponent. The structure of the compensation schemewas constant between August 1999 and August 2002.The levels of the piece rates have, however, changedover time.

The “quantity” performance component has twoparts, with their corresponding linear piece rates: thefirst part depends on the number of orders completed,and the second on the number of goods dispatched.The number of goods dispatched is obviously theproduct of the number of orders and the averagenumber of goods per order. In 2001 the “quantity”performance component represented around 25% ofthe average worker total compensation.

The “quality” performance component depends onthe number of wrongly dispatched goods and, inpractice, is much less important. When a customer issent the wrong good and he complains to the firm,this fact is recorded and tracked back to the individ-ual worker in charge of that particular good. A smalllinear discount is applied to a worker’s total compen-sation when his monthly mistake rate exceeds a cer-tain threshold. In 2001 the average discount was lessthan 1% of total compensation.

Both performance components are paid with a onemonth delay. Hence, in the pay slip received, say,on November 1, workers learn about performanceand wage per hour corresponding to the month ofSeptember, rather than October.

Our measure of productivity is the average numberof goods dispatched per hour by a given worker on agiven day. Because customer orders contain a varyingnumber of goods, we always control for the average

2 The allocation of orders to workers took place following a “cabrank system,” so that a worker finishing an order would bematched with the order at the top of the list, and so on. In regres-sions available upon request we found that a worker’s productivityin a month could not predict the characteristics of the orders that,on average, he would receive the following month.3 In this firm there is a one-to-one mapping between performanceand pay. For this reason, we will not be able to disentangle whetherconcerns about relative standing refer to performance or to pay.


number of goods per order. Otherwise we would beoverestimating the productivity of those workers thathandled a few large orders rather than many smallones. Once the number of goods per order is heldconstant, the number of goods per hour and the num-ber of orders per hour are equivalent measures ofproductivity.

Upon joining the firm, workers face a six monthprobationary period, and they may not be offereda regular contract at the end of this period. How-ever, once they survive the probationary period theyare very unlikely to be fired. After two years, therestrictions imposed by German law, the fact that lessproductive workers are cheaper to the firm, and thescreening performed during the probationary periodmake termination of contracts all but unheard of.

If the job that we study is unusually secure, it isalso a cul-de-sac, because the skills acquired in itdo not transfer easily to other positions in the firm.Among the 207 workers holding this job during thelast 10 years, only 2 workers have ever been pro-moted. The reason for this is that there is no naturalhigher-level position for which the skills gained inthis job provide either good training or a good signalof future productivity. Thus, any form of career con-cern is a truly unlikely source of motivation for theworkers in our study.

Last, there is a certain amount of learning and skillin this occupation. This is because being able to iden-tify the optimal route to gather the goods orderedby a customer leads to a significant gain in timeand productivity. This ability/skill component trans-lates into a strong serial correlation in workers’ rel-ative performance across time. During our sampleperiod, workers’ rank-order position in the productiv-ity distribution during a certain month had a correla-tion of 0.88 with their position in the previous month.

2.2. The Relative PerformanceInformation Exercise

In the summer of 2001, a few members of theworkforce—employees are not unionized—requestedfrom management access to information about thewage per hour being earned by the average worker.Given the low computational cost of creating anddistributing this information, the firm agreed tothis demand. All workers were then notified thattheir September performance report—that is, the oneincluded with the November 1, 2001, pay slip—wouldbe the first one for which they would receive thisinformation.

In addition, management decided to also commu-nicate to workers their own rank-order position inthe productivity/wage per hour distribution. The rea-son for this was very specific: two workers had beenconstantly complaining about the conditions of the

job and spreading discontent among the workforce.These two workers happened to be among the worstperformers, so management thought that privatelyrevealing to them this fact would help to alter theirbehavior.4

The introduction of the RPIE provides the basisof a quasi-experimental research design. Note thatthe RPIE was first triggered by the demand of asmall number of workers and then by a very spe-cific managerial motivation that was orthogonal tothe compensation scheme in place. Furthermore, com-pany insiders stated to us that the introduction ofthe RPIE did not coincide with any type of shock tothe dispatching technology. The identification strat-egy, which we discuss further below, uses both indi-vidual daily productivity and a short time windowaround the introduction of the RPIE to ensure thatpotential preexisting productivity trends are not con-founding our empirical findings.

The information finally communicated on Novem-ber 1 and subsequent months consisted of the mini-mum, maximum, and average wage per hour and aworker’s rank-order position in the wage per hourdistribution. Workers were notified prior to Septem-ber 1 that they would receive this information.

3. The Productivity Effectsof the Relative PerformanceInformation Exercise

Our discussion above suggests two moments at whichworkers’ behavior may have changed as a resultof the RPIE. First, on September 1—the “kickoffmoment”—workers’ daily productivity first started tocount toward the RPIE, and they may have alteredtheir effort in response to this fact. If workers antici-pate that they will derive utility tomorrow from learn-ing that they outperformed their colleagues today,then increasing effort today to obtain that futureutility seems like a reasonable response. Second,on November 1—the “revelation moment”—workersreceived information about their relative performancefor the first time. This may have led to a furtherbehavioral response if receiving such information fur-ther increased the salience and importance of relativeconcerns.

3.1. Descriptive AnalysisWe first provide graphical evidence that the RPIE wasassociated with a large and sudden shift in workers’

4 According to company sources, these two workers were indeedembarrased to learn that they ranked at the bottom of the distribu-tion and stopped complaining. Their contracts were not terminatednor were they disciplined or even threatened in any way, a fact thatshould alleviate any lingering concerns that workers’ reaction mayhave been due to the fear of job terminations.


Figure 1 Evolution of Productivity and Other Variables Over Time

RevelationKickoff

90

100

110

Goo

ds v

alue

Aug99 Sep01 Nov01 Aug02

TimeAug99 Sep01 Nov01 Aug02

Time

Panel C

RevelationKickoff

160

150

140

130

120

170

180

Pro

duct

ivity


TimeAug99 Sep01 Nov01 Aug02

Time

Panel A

0.2

0.4

0.6

0.8

1.0

Rat

io fr

esh

good

s

Panel B

5

6

7

Exp

erie

nce

Panel D

RevelationKickoff

Kickoff Revelation

Notes. Dependent variables are the number of goods dispatched per hour worked (panel A), the ratio of fresh goods over fresh and dry goods (panel B),the average value of a good (panel C), and worker experience (panel D). Each dot is the (averaged across workers) dependent variable of a different workingday. Vertical dashed lines display September 1, 2001 (kickoff date), and November 1, 2001 (revelation date). In panels A–C, solid lines display local linearregressions (with an eight-day bandwidth) computed separately before the kickoff date, after the revelation date, and in between.

productivity and not with similar discontinuities inobservable determinants of productivity.

Figure 1A shows the evolution of daily productiv-ity averaged across workers. It is clear from the fig-ure that the introduction of the RPIE coincided with asharp increase in productivity. Before September 2001,an average worker dispatched less than 145 goods perhour, whereas by November 2001 productivity hadincreased to around 157 goods per hour. Such increaseseems out of line both with the flat or mildly pos-itive preexisting trend and with preexisting fluctua-tions around the trend.

In Figure 1B and C, we display the evolution ofcharacteristics of the goods dispatched by workers.Goods can be either fresh or dry, and it is claimedby company insiders that fresh goods take longer todispatch because the aisles where they are placed arenarrower, and this creates jams in the warehouse. Inearly 2001, a weekend shift was added, leading to amore extreme concentration of fresh and dry goods in

different days of the week. However, it is clear fromFigure 1B that the RPIE did not coincide with any bigdiscontinuity in the ratio of fresh goods. Similarly, weplot the evolution of the average value of the goodsdispatched by workers over time in Figure 1C. Again,no large and permanent discontinuity seems apparentfrom the figure.

In Figure 1D we display the evolution of aver-age worker experience over time. Average experiencetrends upward if there are no changes to person-nel and instead decreases in periods, like the secondpart of 2000, in which the firm hires new workers.Such fluctuations are obvious in the figure, but, again,no large changes are apparent during our period ofinterest.

It seems difficult from Figure 1A to attribute thesharp increase during the September 2001–November2001 period to any preexisting trend in productivity.To ensure, however, that potential mispecificationof preexisting time trends does not confound our


Table 1 Descriptive Statistics

(1) (2) (3) (4)Mean SD Minimum Maximum

Panel A: Firm-level characteristicsGoods per day 11107024 152073 474078 11561079Goods per hour 149073 7085 126084 172014Average value of good 100 28057 306 122022Ratio of fresh goods 0056 0022 0022 1Number of observations 152

Panel B: Individual-level characteristicsMale 0063 0048 0 1Experience (in years) 3099 4006 0050 23088Age (in years) 38075 8035 21027 58048Left December 31, 2007? 0038 0049 0 1Ranking August 2001 34001 18035 3 66Number of observations 63

Notes. Descriptive statistics are based on the sample comprising July 1,2001, to December 28, 2001. The number of worker-day observationsincluded in the sample is 5,275. Panel A displays variables at firm-day levelby average worker. Average value of a good is normalized at 100. Panel B dis-plays variables at the worker level. The sample is restricted to those workerswith at least six months of experience in the firm.

estimates, we focus our main analysis on a short timewindow comprising six months in total. Our mainsample consists of three periods: (1) the “baselineperiod” includes the two months (July and August)just before the introduction of the RPIE; (2) the“kickoff period” includes the two months (Septem-ber and October) during which workers did nothave any new information but knew that their dailyproductivity would form the basis on which relativeperformance would be calculated; and (3) the “rev-elation period” includes the two months (Novemberand December) right after workers first learned theirrelative performance.

Using this sample, panel A of Table 1 displaysthe main input and output variables at firm level. Inpanel B of Table 1 we display a number of workercharacteristics. The average worker in our sample hasbeen at this job for four years and is 39 years of age.More than 60% of workers in our sample were still atthe job six and a half years later, which implies thatthe yearly separation rate was just 6%. We also dis-play information on the rank-order performance dur-ing August 2001, which has been constructed by usbut was not computed by the firm.5

5 We restrict our study to workers with more than six months ofexperience. We do this for several reasons. First, the returns to expe-rience are both very steep and very concave among workers in theirfirst six months. As a result, controlling for a linear trend may notadequately capture the evolution of productivity for these workers,even in such a short time horizon as the one used here. Second,these workers are in their probationary period and may have dif-ferent incentives from those faced by the rest of the workforce. Inparticular, we cannot rule out the possibility that they may perceive

3.2. Econometric FrameworkOur basic estimating equation takes the following lin-ear form:

yit = �kKt +�rRt + X′

it · �+�i + g4t5+ �it1 (1)

where yit is (the log of) worker i’s productivity, mea-sured in number of goods dispatched per hour on dayt; Kt = 1 in the post–September 1 period; and Rt = 1in the post–November 1 period. Our model also con-tains a vector of covariates X′

it, including, as discussedabove, the (log of the) average number of goods percustomer order by worker i on day t. We allow forindividual time-invariant unobserved effects, �i, toaccount for the possibility that individual workersmay differ across time periods.

Our identification strategy relies on the comparisonof the same workers over time. Relying on such varia-tion implies that we cannot be completely nonrestric-tive in accounting for time effects. We instead allowproductivity to evolve smoothly over time throughthe parametric function g4t5. We show however thatour results are robust to the specification of g4t5 aswell as to the period over which g4t5 is calculated. Thevariable �it captures independent and identically dis-tributed (iid) person-specific idiosyncratic shocks.

To account for correlation of the observations fromthe same date, we adjust the standard errors by allow-ing for arbitrary variance–covariance matrix withineach date across individuals. In practice, this does nothave much effect on the standard errors.

Our parameters of interest are �k and �r . The iden-tifying assumption in estimating these parametersfrom Equation (1) is that, without the introduction ofthe RPIE, productivity would not have deviated fromthe trend g4t5.

3.3. Baseline ResultsTable 2 displays our baseline results. Column (1)shows that productivity was 207% higher in the kick-off period and 509% higher in the revelation period,relative to the baseline period.

The fit of the model is very large even in this firstspecification with just three explanatory variables.The reason is that the number of goods per order is avery strong determinant of the number of goods dis-patched per hour, because workers can dispatch moregoods when they do not have to make constant tripsto the goods-out area of the warehouse and to pickup new orders. This fact underlines the importance ofcontrolling for this variable in every specification.

the information on their relative position as a signal on the likeli-hood that their contract will be renewed. By focusing on workersunlikely to be making those considerations, we will be able to arguethat pure concerns for relative position lie behind their reaction tothe RPIE. This restriction reduces our sample by just three workers.


Table 2 Baseline Results

(1) (2) (3)Unconditional Time trend Other covariates

Kickoff 00027∗∗∗ 00033∗∗∗ 00028∗∗∗

4000055 4000105 4000065Revelation 00032∗∗∗ 00038∗∗∗ 00040∗∗∗

4000065 4000105 4000075(log)Goods per order 00308∗∗∗ 00308∗∗∗ 00336∗∗∗

4000115 4000115 4000065Time trend −000001 −000001∗

40000015 400000075(log)Goods piece rate −00045

4001045Ratio fresh goods 00013

4000105Average value of goods −00174∗∗∗

(in euros) 4000225Day of week dummies No No YesWorker experience dummies No No YesWorker fixed effects No No YesAdjusted R2 0022 0022 0077Number of observations 5,275 5,275 5,275

Notes. Dependent variable is (log of) number of goods dispatched per hourworked. Data comprise July 1, 2001, to December 28, 2001.Kickoff is a dummyvariable taking value 1 in the post–September 1, 2001, period. Revelation is adummy variable taking value 1 in the post–November 1, 2001, period. Workerexperience dummies capture the quarter of experience of the worker. Sampleis restricted to those workers with at least six months of experience in the firm.Robust standard errors allow for cluster at the day level.

∗, ∗∗, and ∗∗∗ denote estimates significant at 10%, 5%, and 1%,respectively.

Column (2) adds a linear time trend to the specifi-cation. The estimated coefficients of interest are verysimilar in magnitude to those of column (1), andthe estimated linear trend is not statistically differentfrom zero. Thus, unobserved time-varying factors cor-related with productivity are unlikely to be biasingour estimates.

In column (3) we add extra covariates, day of weekdummies, experience dummies, and worker fixedeffects. We find the average effect of kickstarting theRPIE to be 208%, whereas the average effect of actu-ally revealing the relative performance information is400%. The cumulative effect is 608%.

The coefficients on the covariates are interesting intheir own right. Consider the level of the goods piecerate, which the firm altered in every month of oursample. These changes were by very small percent-ages (usually 1% or 2%), so it is plausible that theworkforce barely (if at all) reacted to them. The neg-ative and statistically insignificant coefficient on thepiece rate indicates indeed that workers did not seemto work harder in months in which the piece rate wasmarginally higher.

We find, however, that workers were more pro-ductive when the goods dispatched were cheaper.Company insiders suggested that this may be either

because cheaper goods are smaller in size or becausethey are more common and workers are better awareof their location.

3.4. Placebo RPIE and Robustness ChecksBecause the RPIE was introduced at the same time forall workers, identification of its effect on productiv-ity arises from a comparison over time of the sameworker. The estimated effects are therefore biasedupward to the extent that they capture factors thatcause productivity to rise independently of the RPIEand that are not captured by the linear trend. Table 3presents a number of robustness checks to addressthis and other concerns.

To judge whether our findings are due to season-ality, we repeat the estimation of (1) substituting ourbaseline sample from 2001 with an equivalent samplefrom the previous year. This new “placebo” samplespans July 1, 2000, to December 31, 2000, and con-sists of 3,836 observations. We define a placebo kickoffdummy to be equal to one after September 1, 2000,and a placebo revelation dummy to be equal to one afterNovember 1, 2000. We include worker fixed effects, atime trend, experience dummies, and the other con-trols. Column (2) of Table 3 suggests that our mainresults were not due to seasonality. The estimatedcoefficients are much smaller in magnitude and areeither not statistically significant or weakly significantbut with a negative sign.

We next examine the robustness of our estimates tothe sample width by estimating (1) on a sample com-prising 12 months in total. As column (3) of Table 3shows, the estimated coefficients are still statisticallysignificant and only slightly different in magnitude.

Second, we study whether our results are robustto allowing the parametric time trend g4t5 to takedifferent formats. We first allow g4t5 to take aquadratic form. Because a problem of severe collinear-ity between t and t2 arises in our baseline time win-dow of 6 months, we use a window of 12 months.Column (4) of Table 3 shows that our main findingsare indeed robust to allowing g4t5 to take a more flex-ible form.

In our benchmark specification from column (3) inTable 2, the linear trend is estimated using both pre-treatment and treatment observations. Alternatively,we may want to estimate the trend exclusively usingpretreatment information. This allows us to evaluatethe effect of the RPIE above or below what we wouldexpect productivity to be given its pretreatment trend.To do this, we estimate the empirical model

yit = X′

it · �+�i + �41 −Kt5 · t + �it (2)

and use the estimated coefficients to project theexpected value of productivity yit given the estimatedpretreatment trend �. We then compute the difference


Table 3 Placebo RPIE and Robustness Checks

(1) (2) (3) (4) (5) (6)Baseline 12 months Quadratic Pretreatment Pretreatmentresults Placebo window trend trend effects

Kickoff 00028∗∗∗ 00020∗∗∗ 00029∗∗∗ 00027∗∗∗ 00032∗∗∗

4000065 4000055 4000065 4000055 4000055Placebo kickoff 00008

4000065Revelation 00040∗∗∗ 00035∗∗∗ 00049∗∗∗ 00027∗∗∗ 00040∗∗∗

4000065 4000035 4000045 4000055 4000065Placebo revelation −00011∗

4000065Table 2 covariates Yes Yes Yes Yes Yes Yes

Standard errors Day Day Day Day Boot- Boot-cluster cluster cluster cluster strapped strapped

Adjusted R2 0074 0071 0074 0074 0074 0074Window months 6 6 12 12 6 6Observations 5,275 3,836 10,176 10,176 5,275 5,275

Notes. Dependent variable is (log of) number of goods dispatched per hour worked. Column (1) is baseline regression(i.e., column (3) from Table 2). Column (2) data comprise July 1, 2000, to December 31, 2000, and the placebo kickoffand placebo revelation dummies take a value of 1 in the post–September 1, 2000, and post–November 1, 2000, periods,respectively. In columns (3) and (4) data comprise April 1, 2001, to March 31, 2002. Column (4) includes a quadratictime trend. In column (5), the linear time trend is estimated exclusively on the baseline period (July and August 2001).In column (6), the linear time trend, the worker fixed effects, and the effects of the covariates are estimated exclusivelyon the baseline period (July and August 2001). Robust standard errors allow for cluster at the day level unless specifiedotherwise.

∗, ∗∗, and ∗∗∗ denote estimates significant at 10%, 5%, and 1%, respectively.

between actual and predicted productivity �it = yit −yit and estimate

�it = �kKt +�rRt + �it0 (3)

The estimated coefficients �k and �r are displayed incolumn (5) of Table 3. The coefficients are very similarto those of our benchmark specification from Table 2,suggesting that our results are robust to the form ofthe parametric time trend.

In column (6) we use only pretreatment informa-tion to estimate the individual fixed effects, �i, andthe effects of other covariates, �, as well as the lineartrend, �. Again, we find that the estimated coefficientsare statistically significant and similar in magnitude.6

3.5. Heterogeneous EffectsTable 2 shows that the average effect of the RPIE is608%. The effect may however be different for differ-ent types of workers, perhaps even negative for some.

6 One important limitation of our study is the absence of a con-trol group. To partially remedy this drawback, we examined theperformance of a comparable German food trader, Metro Group.Productivity measures at the warehouse worker level were unfor-tunately unavailable. Instead, we studied the evolution over timeof a quarterly measure of operating profits, earnings before interest,taxes, depreciation, and amortization. We found that this measureof performance was not characterized by any significant disconti-nuity in the second part of 2001 (details are available upon requestfrom the authors).

In Table 4 we split the workforce on the basisof observable characteristics and run our benchmarkspecification (1) separately for different groups ofworkers. Panel A of Table 4 shows that the reactionsof males and females to the RPIE are very similar.Although the estimated coefficients are slightly differ-ent in magnitude, these differences are not statisticallysignificant. In panel B we find that only workers inthe middle and top thirds in terms of experience reactto the kickoff of the RPIE. On the other hand, we findno differences in the revelation coefficient for workerswith different levels of experience. In panel C we usethe (undisclosed) ranking from August 2001 to splitthe sample by preexisting level of performance. Wefind weak evidence that workers in the bottom thirdincrease their productivity more at the kickoff stage,whereas workers in the top third increase in produc-tivity more at the revelation stage. The overall effectof the RPIE is very similar for workers with differentpreexisting levels of performance.

Instead of aggregating workers on the basis oftheir observable characteristics, we can study sepa-rately the reaction of every individual worker to theRPIE. To do this, we ran 58 separate worker-levelregressions equivalent to (1) and computed the sumsof the estimated kickoff and revelation coefficients.Although there was substantial heterogeneity acrossworkers, the reaction to the RPIE was positive forthe overwhelming majority of workers (i.e., 51 of 58).


Table 4 Decomposition by Gender Experience and Performance (August Ranking)

Panel A: Decomposition by gender

(1) (2) (1)–(2)Males Females p-value

Kickoff 00023∗∗∗ 00038∗∗∗ 00284000085 4000115

Revelation 00047∗∗∗ 00026∗∗ 00174000085 4000125

Worker FE Yes YesCovariates Yes YesObservations 3,307 1,968

Panel B: Decomposition by experience

(1) (2) (3) (1)–(2) (1)–(3) (2)–(3)Least experience Middle experience Most experience p-value p-value p-value

Kickoff −00010 00029∗∗ 00037∗∗∗ 0002 0000 00414000125 4000105 4000105

Revelation 00045∗∗∗ 00046∗∗∗ 00036∗∗∗ 0098 0055 00564000135 4000125 4000115

Worker FE Yes Yes YesCovariates Yes Yes YesObservations 1,484 1,833 1,907

Panel C: Decomposition by performance (August ranking)

(1) (2) (3) (1)–(2) (1)–(3) (2)–(3)Bottom ranked Medium ranked Top ranked p-value p-value p-value

Kickoff 00052∗∗∗ 00015 00019∗ 0002 0004 00784000125 4000095 4000115

Revelation 00035∗∗ 00028∗∗ 00053∗∗∗ 0072 0025 00114000145 4000115 4000115

Worker FE Yes Yes YesCovariates Yes Yes YesObservations 1,553 1,860 1,862

Notes. The dependent variable is (log of) number of goods produced per hour worked. Data comprise July 1, 2001, to December 28,2001. The sample is restricted to those workers with at least six months of experience in the firm. Covariates are as in column (3)of Table 2, except in panel B, which excludes experience quarter dummies. Robust standard errors allow for cluster at the day level.FE, fixed effects.

∗, ∗∗, and ∗∗∗ denote estimates significant at 10%, 5%, and 1%, respectively.

Furthermore, of the 19 workers for whom the com-puted overall effect was statistically different fromzero at the 10% level, 18 had a positive sign, whereasonly 1 had a negative sign.

We argue in §3.7 that two findings from this sec-tion are unlikely to be compatible with basic modelsof learning about ability or the distribution of perfor-mance: (a) very experienced workers do not react lessat the revelation stage than less experienced workers,and (b) the effect of the RPIE can be regarded as pos-itive for almost all workers.

3.6. Effects on InequalityWe now study whether the RPIE affected the amountof inequality across workers. Figure 2 plots the kerneldistribution of productivity for August, September,and November 2001. It seems that the RPIE shifted thedistribution to the right, without noticeable effects on

the dispersion of productivity. To confirm this findingformally, we computed the coefficient of variation forthe productivity variable for each day of our baselinesample. We found that this measure of dispersion wasnot different on average following the introduction ofthe RPIE (details are available upon request from theauthors). Thus, static inequality did not change.

We can also study inequality from a dynamic per-spective by examining the likelihood that a worker ata position in the performance distribution in a partic-ular month repeats the position the following month.In Figure 3 we plot workers’ rank position in termsof wage per hour across consecutive months for eachmonth in our baseline sample. There is significant per-sistence.7 Overall, it does not seem as if persistence

7 We do not find the existence of strong serial correlation in rankincompatible with the notion that agents exert more effort when


Figure 2 Distribution of Productivity Over Time

200150

August 2001 (baseline)September 2001 (kickoff)November 2001 (revelation)

100

0

0.005

0.010

0.015

Den

sity

0.020

0.025

Productivity (number of goods per hour)

Notes. This figure displays the kernel densities of the productivity variablefor three selected months: August 2001 (the last baseline month), Septem-ber 2001 (the first kickoff month), and November 2001 (the first revelationmonth). Productivity is defined in terms of goods dispatched per hour.

in rank position varied following the introduction ofthe RPIE. We confirmed this by regressing a worker’srank position in a month on his position in the pre-vious month and interacting this variable with thekickoff and revelation variables. We found that thecoefficients associated with these interactions werenot statistically different from zero (details are avail-able upon request from the authors).

To summarize, the RPIE led to an increase in aver-age productivity and wages while leaving both staticand dynamic disparities between workers essentiallyunchanged.

3.7. Interpretation and Alternative ExplanationsTo summarize the main findings: (a) the effect of theRPIE is on average positive both at the kickoff stageand at the revelation stage (Figure 1A and Table 2);(b) individual regressions show that the effect is pos-itive for a large proportion of workers and negativefor just one worker (§3.5); (c) the increase in produc-tivity at the kickoff stage is larger for workers withmore experience (panel B of Table 4) and for workersat the bottom of the performance distribution (panelC of Table 4); and (d) the overall increase in produc-tivity is, however, statistically equal for workers withdifferent observable characteristics (Table 4).

We now discuss the potential explanations of ourfindings.

their relative concerns have stronger. This is because (a) relativeconcerns could be cardinal as well as ordinal, (b) rank mobilityover time can be high in the presence of strong serial correlationprovided that the number of workers is large, and (c) increasedrelative concerns could be affecting the marginal return to efforteven while having no effect on equilibrium rank positions.

3.7.1. Preferred Interpretation: Increase in Rel-ative Concerns. We regard the evidence presentedhere as consistent with the notion that workers’ incip-ient concerns about relative standing were activatedby the RPIE. This interpretation is consistent withthe documented productivity increase at the kickoffstage. By increasing effort at that stage, a workerwould have increased his chances of outperforminghis colleagues, a fact that, on being learned later,would generate a utility increase. This explanation isalso consistent with the revelation of relative perfor-mance further increasing the salience of these rela-tive concerns and leading to the additional increasein productivity documented in Table 2. Last, the exis-tence of concerns for relative pay is perfectly com-patible with more experienced workers reacting morestrongly at the kickoff stage, because it is natural tothink that these concerns should be stronger for indi-viduals that have been members of the workforce fora longer period of time (Frank 1985).

What type of relative concerns can explain our find-ings? Consider first relative concerns of the “inequityaversion” type (Fehr and Schmidt 1999, Fehr andGächter 2000). These preferences would be consistentwith workers at the bottom of the distribution increas-ing their effort by a large amount, and with workersat the top not working harder and even decreasingtheir effort to allow the others to catch up. Yet, this isthe opposite of what we find in §3.5.

We instead regard relative concerns whereby in-creases in others’ performance always have a nega-tive effect on own utility (Frank 1984, Fershtman et al.2005, Moldovanu et al. 2007) as consistent with ourfindings. As Frank (1984), Fershtman et al. (2005), andMoldovanu et al. (2007) discuss, an increase in theseconcerns translates into everybody increasing theirlevel of effort and nobody decreasing it, as we findin §3.5.

We now discuss alternative explanations.

3.7.2. Alternative Explanation: Career Concerns.Could concerns about either termination of employ-ment or a potential promotion present an explanationof our findings? As discussed in §2, the likelihoodof both terminations and promotions in our firm isclose to zero. Working conditions and base pay arefixed and identical for every worker, regardless of per-formance. Furthermore, workers were explicitly toldthat the relative performance information would notbe used in any way by management. Thus, we do notregard career concerns as a valid explanation.

Using career concern interpretations, it is also diffi-cult to explain why productivity did not decrease foralmost any worker in the firm. Consider for instancethe incentives of a weak performer who, before theintroduction of the RPIE, is unaware of this fact


Figure 3 Relation Between Ranks in Two Consecutive Months

0

0.5

1.0

0

0.5

1.0

0 0.5 1.0 0 0.5 1.0 0 0.5 1.0

Baseline 1 (July 2001) Baseline 2 (August 2001) Kickoff 1 (September 2001)

Kickoff 2 (October 2001) Revelation 1 (November 2001) Revelation 2 (December 2001)

Nor

mal

ized

ran

k cu

rren

t mon

th

Normalized rank previous month

Notes. This figure presents a scatter plot capturing the relation between the rank (in terms of the wage per hour) obtained by an individual in a month and therank obtained by the same individual in the previous month. The ranks are all normalized to one by dividing by the total number of workers working in the firmon that month. Six different scatter plots are presented for each of the months comprising our baseline sample.

and exerts some extra effort with the hope of rank-ing among the top and meriting a promotion. Wewould expect that upon learning his mediocrity, hewould regard a promotion as virtually impossible anddecrease his efforts to attain it. Yet we do not find thisto be the case for almost any worker. Similarly, veryable workers fearing a potential termination shouldhave put less effort, not more, after learning that theyranked at the top of the distribution. Again, this is theopposite of what we find.

3.7.3. Alternative Explanation: Learning AboutOwn Absolute Ability. One potential explanation ofour findings is that the RPIE may have helped work-ers to learn about their own ability for the job, and thismay have led to a readjustment in the chosen level ofeffort. One potential benchmark for this argument isprovided by Ertac (2006), who presents a theoreticalmodel in which effort and ability are strategic comple-ments in a worker’s production function, and workersare uncertain about their own ability. Past productioncan help workers learn about their own ability but issubject to an unobserved firmwide productivity shockthat makes learning imperfect. In this setting, Ertac(2006) shows that information on how much otherworkers produced can help a worker to learn abouttheir own ability. For instance, a worker with low pro-ductivity in a month will revise expectations about hisown ability upward after observing that others also

did badly, because his low productivity is likely tobe the result of an unfavorable firmwide productivityshock rather than the result of low ability.

We see two problems with this explanation.The main problem is that any interpretation basedexclusively on learning cannot explain the existenceof a positive productivity effect at the kickoff stage,in which workers do not learn anything.

An additional problem is that the predictions of themodel outlined above do not seem consistent withour empirical findings. In the model outlined abovethe arrival of new information makes some workersrevise their expected ability upward and exert moreeffort, whereas other workers revise it downward andexert less effort.8 This is not what we found in thedata. We found in §3.5 that there is almost no hetero-geneity in the sign of the reaction to the RPIE: almostall workers react by increasing their effort.9

8 To gain intuition, consider a situation in which all workers have thesame prior expected ability and learn the average level of produc-tion of their coworkers. Roughly speaking, workers producing more(respectively, less) than the average will update their ability upward(downward) and exert more (less) effort upon learning this fact.9 A second prediction of Ertac (2006) is that workers less uncertainabout their own ability should react less to the information con-veyed by the RPIE. Again, this is inconsistent with our findings.We find in panel B of Table 4 that highly experienced workers (pre-sumably well informed about their own ability) do not react less atthe revelation stage that highly inexperienced workers.


Nevertheless, it may be that a different model oflearning about absolute ability can explain the reve-lation stage effect.

3.7.4. Alternative Explanation: Learning Aboutthe Distribution of Performance. One potential ex-planation of our findings is that the RPIE may havehelped workers to learn about a productivity param-eter that is common to all workers, and, again, thismay have led to a readjustment in the chosen level ofeffort. For instance, a worker with bad luck in a par-ticular month may, in the absence of RPIE, wronglybelieve that the returns to effort in the company arenot very high. However, by observing the higher pro-duction of his coworkers, he will update his beliefabout the returns to effort upward, which may leadto more effort.

We believe that this explanation is unlikely. Anylearning explanation is inconsistent with the docu-mented positive reaction at the kickoff stage, in whichthere can be no learning. Second, we would expectthat some workers should react to the arrival of newinformation by increasing their effort, whereas othersshould decrease it. We find instead in §3.5 that the signof the reaction to the RPIE is positive for almost allworkers.

Again, it may be that a different model of learningcan explain the revelation stage effect.

3.7.5. Alternative Explanation: Certification toOutside Employers. It could be argued that the rela-tive performance information included in the pay slipcould have been used by workers as a verifiable signalof high productivity to prospective outside employ-ers. Under this hypothesis, the interest in improvingthis signal could have generated the increase in effortand productivity that we have documented.

Although we cannot completely rule out thisnotion, we find it unlikely for two reasons. First, wewould not expect workers at the bottom third of thedistribution to plan on disclosing such a signal to out-side employers, because little can be gained by admit-ting one’s relative incompetence.10 As a result, wewould expect their effort to remain unchanged, con-trary to our findings. Second, the availability of suchcertification should have led to an increase in the quitrate among the workforce (and especially among thebest performers), as information about workers’ abil-ity becomes less asymmetric. We show in §4 that thisis not the case.

10 This is of course provided that (a) the most that such workerscan aspire to is relative mediocrity and that (b) outside employersare unaware that our firm provides this information, two notionsthat we find highly plausible.

4. Other Effects of the RelativePerformance Information Exercise:Mistake Rate and Separation Rate

In this section we briefly investigate whether the pro-ductivity increase documented in the previous sectionwas accompanied by a permanent change in (a) thequality of the work done, and (b) the separation rate.

4.1. Effect on the Mistake RateThe “quality” component represents a tiny determi-nant of total compensation (on average, the discountin compensation due to mistakes is less than 1%). Yet,it is worth studying whether the RPIE was accom-panied by a permanent increase (or occured at theexpense of a permanent decrease) in the quality ofwork done. Unfortunately, in our firm informationon mistakes is recorded only at monthly intervals,rather than at daily frequencies. Because this makes itmore difficult to study the instantaneous behavioralresponse to the introduction of the RPIE, our find-ings in this Section should be regarded as suggestive,rather than conclusive.

We display in Figure 4 the evolution of the averagemistake rate over time. The introduction of the RPIEdid not coincide with a dramatic discontinuity in theevolution of the mistake rate. The mistake rate is, onaverage, higher during the 12 months previous to theRPIE than during the months after it, but this declin-ing trend is apparent well before the kickoff date. Themonth of September 2001 (the first month after thekickoff date) is associated with a mistake rate that ismuch higher than that of the two previous months(July and August). However, in October, November,and December the mistake rate was back to normalpretreatment levels.

Figure 4 Evolution of Mistake Rate Over Time

Kickoff Revelation

0.0015

0.0020

0.0025

0.0030


Month

Mis

take

rat

e (m

ista

kes/

good

s)

Notes. The dependent variable is (log of) number of mistakes per good dis-patched. Vertical dashed lines display September 1, 2001 (kickoff date), andNovember 1, 2001 (revelation date). Each dot is the (averaged across work-ers) mistake rate of a different month.


We now use our baseline six month time windowto estimate the following panel data specification:

Mit = dt + X′

it · �+�i + �it1 (4)

where Mit is the (log of) mistake rate of worker i onmonth t, dt is a dummy variable capturing month t, �i

is an individual time-invariant unobserved effect, X′it

is a vector of covariates similar to that of column (3)of Table 2, and �it are iid person-specific idiosyncraticshocks.

The main parameters of interest are dSeptember,dOctober, dNovember, and dDecember, where the first twoparameters capture the kickoff period, and the secondtwo parameters capture the revelation period. Theparameter dJuly is omitted, and the parameter dAugustcaptures the pretreatment time trend in the mistakerate.

Table 5 reports ordinary least squares estimates of(4). In column (1) we do not include worker fixedeffects or extra covariates. We find that in Augustthe mistake rate is smaller than in July, whereas inSeptember it is larger. From October onward we donot find any statistical difference in the mistake raterelative to the omitted month of July. Column (2)shows that this finding is robust to conditioning onindividual fixed effects and a vector of covariates sim-ilar to that of column (3) in Table 2.

Thus, the results in this section are consistent with atransitory (i.e., in the month of September) increase inthe mistake rate following the kickoff date. There is,however, no evidence to suggest that the RPIE led toa permanent change in the quality of work produced.

4.2. Effect on Workers’ Separation RateIn this section we investigate whether the separa-tion rate can be regarded as being different beforeand after the introduction of the RPIE. We also studywhether workers of different ability reacted differ-ently to this introduction. To do this, we estimatea Cox semiparametric proportional hazard model,which allows for a fully flexible, nonparametric base-line hazard. In our model, the covariates shift thebaseline hazard proportionally, through the function

�it =�Rt +�Ai +�4Rt ·Ai5+ X′

i · �1 (5)

where Rt = 1 if the first date of the quarter of tenure toccurs after the revelation moment, Ai is a measureof individual ability,11 and X′

i is a vector of workercharacteristics.

11 To compute this measure, we run a regression where productivity(i.e., the number of goods per hour) is the dependent variable, andday fixed effects, worker fixed effects, and quarter of experiencedummies are the explanatory variables. Our measure of individualability, Ai , is the estimated worker fixed effect from this regression.The estimated standard errors are bootstrapped in Equation (5) toaccount for the fact that Ai is generated, rather than observed.

Table 5 Evidence on the Mistake Rate

(1) (2)Unconditional Controls

July (baseline month 1) Omitted OmittedAugust (baseline month 2) −00167∗∗∗ −00224∗

4000765 4001125September (kickoff month 1) 00192∗∗∗ 00194∗∗

4000695 4000815October (kickoff month 2) 00017 00114

4000725 4000835November (revelation month 1) 00036 00143

4000635 4001185December (revelation month 2) −00056 −00016

4000805 4001125Worker fixed effects No YesTable 2 covariates No YesAdjusted R2 0003 0070Number of observations 332 332

Notes. Dependent variable is (log of) number of mistakes per good dis-patched. Data comprise July 1, 2001, to December 28, 2001. July and Augustare the two months previous to the introduction of the RPIE. September andOctober are the two months comprising the kickoff period. November andDecember are the first two months of the revelation period. The sample isrestricted to those workers with at least six months of experience in the firm.Robust standard errors allow for cluster at the worker level. Covariates areas in column (3) of Table 2 (except for the linear time trend and the (log of)goods piece rate). The covariates are monthly averages.


We estimate this model exclusively on workerswho joined the firm after the end of the probation-ary period (i.e., for whom the quarter of tenure t isat least 3). By focusing on workers past their proba-tionary period we can interpret our findings as refer-ring to the workers’ quit rate.

The small size of the workforce, together with thefact that separations are relatively rare events, makesit unfeasible to circumscribe the analysis to a narrowtime window centered around the introduction of theRPIE. We instead use eight years of data, betweenFebruary 1999 and December 2007. Because of thelarge width of the sample time window, the evidencefrom this section should again be regarded as sugges-tive, rather than conclusive.

We include in our analysis only workers who joinedthe firm after August 1999 (the first month for whichwe have observations) to avoid problems of left cen-soring. In total, our sample contains 127 workers, outof which 41 had left the firm by December 2007.

Table 6 reports the results from the proportionalCox hazard model (5). In column (1) we observe thatthe likelihood of a separation is lower in the periodafter the introduction of the RPIE than in the periodbefore. The hazard ratio is 00181, indicating that thelikelihood of a separation in the period after the RPIEis just 18% of the likelihood in the period before. In


Table 6 Likelihood of a Separation Before and After the Revelation ofRelative Performance

Dependent variable: equal to 1 if worker i left the firm in quarter t

(1) (2) (3)Cox regression Cox regression Cox regressionunconditional ability other covariates

Revelation −1070∗∗∗ −1072∗∗∗ −1064∗∗∗

4003595 4003675 40053756001817 6001787 6001937

Worker ability −6024∗∗∗ −7088∗∗∗

410195 4104256000017 60000037

Revelation × −00008 00216Worker ability 4004915 4004895

6004867 610247Worker covariates No No Yes

Number of observations 719 719 719Number of subjects 127 127 127Number of separations 41 41 41

Notes. Revelation takes value 1 if the first date of quarter t occurs afterNovember 1, 2001. Worker ability is the estimated worker fixed effect of aregression that has productivity as a dependent variable and includes expe-rience dummies and day fixed effects, as well as worker fixed effects, on theright-hand side. Worker covariates include gender, marital status, age entryin the firm, and a linear cohort effect (the calendar year in which the workerjoined the firm). Data comprise August 1, 1999, to December 31, 2007. Thesample is restricted to workers joining the firm after August 1999. The sam-ple is restricted to workers surviving the probationary period (i.e., the first sixmonths of employment). Robust standard errors are in parentheses in col-umn (1). Robust bootstrapped standard errors are in parentheses in columns(2) and (3). Hazard ratios are in square brackets.


column (2) we introduce our measure of worker’sability both in levels and interacted with Rt . Wefind that workers with a higher value of our abilitymeasure are less likely to leave the firm, suggestingthat our measure does indeed capture an importantdimension of worker heterogeneity. However, the dif-ference in separation rates across workers with differ-ent abilities does not change after the introduction ofthe RPIE. Last, we find in column (3) that our find-ings are robust to the introduction of a number ofworker characteristics, including gender, marital sta-tus, age at entry in the firm and a linear cohort effect(captured by the calendar year in which the workerjoined the firm).

To summarize, we do not find any evidence indicat-ing that workers are more likely to quit the firm afterthe introduction of the RPIE. Subject to the caveatsmentioned above, our evidence in fact suggests theopposite.

4.3. Relation with Monetary Incentives andEffect on Profits

We now relate the effects of the RPIE to the effectsof monetary incentives. To estimate how workers

respond to an increase in the goods piece rate, we usedata between January 2002 and December 2007.12 Weregress the (log of) individual productivity on the (logof) the goods piece rate, the (log of) the orders piecerate (which the firm changed during this time period),a cubic time trend, individual fixed effects, and thetime-varying controls from Table 2. Our estimate ofthe effect of increasing the goods piece rate implies anelasticity of 43%. This elasticity implies that the 608%achieved by the introduction of the RPIE would beequivalent to an increase in the goods piece rate ofaround 15.8%, equal to 608%/0043. Thus, the nonpecu-niary incentives provided by the RPIE are reasonablylarge not only in absolute terms but also relative tothe effects achieved by monetary incentives.

We can also calculate the potential reduction inlabor costs induced by the RPIE. Because the existing65 workers started producing on average 608% moregoods per hour, the firm was in principle able to pro-duce the same amount of goods in the same amountof time with 4.20 (6.5%) fewer workers. How wouldthis have translated into labor cost savings? To calcu-late this, we need to take into account the fact thatthe performance component of worker compensationis independent of the number of workers in the firm.The performance component is given by the prod-uct of the piece rate(s) and total production, whichis exogenously given by the demand for the firmproducts. The firm could, however, have saved onthe fixed component of worker compensation, whichrepresented around 75% of the average worker payduring this period. Producing the same amount with6.5% fewer workers therefore implies potentially sav-ing 409% = 608% × 75% in terms of labor costs. Inour opinion, this represents a economically signifi-cant, although not huge, amount.13

5. Concluding RemarksWe have provided evidence that social comparisonsaffect human behavior independently of the mone-tary incentives to which these comparisons are usu-ally linked.

12 We use this longer time period because the fact that the firmchanged the level of the piece rate on repeated occasions and bysignificant amounts allows us to obtain good estimates of the effectsof monetary incentives.13 We refer to these savings as potential savings rather than actualsavings, because no worker was actually fired as a result of theincrease in productivity. The firm does not typically fire employ-ees past the probationary period, and during this time period thedemand for the firm products and the size of the workforce wasgrowing over time. The firm therefore adjusted by failing to expandthe workforce at the same rate as it would have in the absenceof the RPIE. Note, finally, that the decrease in the separation ratefound in §4.2 will also reduce search costs and therefore increaseprofits.


In our firm, simply receiving information about rel-ative standing—and anticipating such future informa-tion—leads to an increase in individual productivitythat is statistically and economically significant. Wehave found no evidence of this increase being purelytransitory. Apart from a transitory (i.e., one month)slight decrease in the quality of work done, we havefound no evidence of workers reacting along otherdimensions in ways that are detrimental to the firm.Taken together, our results suggest that making infor-mation about relative performance available is unam-biguously beneficial to the firm’s profitability.

We believe that the notion that workers display rel-ative concerns provides the best possible explanationof our findings. Other interpretations fall short in oneway or another, notably in explaining why workersalready increase their effort at the kickoff stage.

We have been able to identify the causal effectsof social comparisons thanks to unique access to thedetailed personnel records of a firm that, for largelyexogenous reasons, started to provide information onrelative performance to workers. Our findings arebased on a single firm study, and may not necessar-ily generalize to other settings. To understand theirexternal validity, there are two features of our firmand work environment that we regard as particularlysignificant:

(1) The effects of providing information on rela-tive productivity and pay may depend on the tech-nology of production, the incentive scheme in place,and the nature of workers affected. In our firm, theexistence of absolute performance pay implies a one-to-one mapping between performance and compen-sation.14 Furthermore, work is individual, with littlescope for cooperation among workers. As a result,workers dissatisfied with their relative position haveone single natural channel through which they canremedy their situation: the exertion of more effort.The effects of information on social comparisons mayobviously be different in fixed-wage settings or in set-tings where industrial politics are potentially impor-tant (Lazear 1989). Similarly, workers’ attitudes maybe critical in determining whether social comparisonsare morale boosting or have the opposite effect. Inshort, although we may find similar effects among,say, door-to-door salesmen, our results do not neces-sarily generalize to, for instance, corporate lawyers,because for these team-work is important and there islarge scope for mutual sabotage.

(2) Perhaps more importantly, in our firm the infor-mation was disclosed in a purely private manner to the

14 Obviously this one-to-one mapping also implies that our resultsmay reflect workers caring about relative productivity per se, aswell as about relative pay. Disentangling these two possibilities fallsoutside the scope of this paper.

workers (although, of course, some may have sharedthat information among themselves). We are unfor-tunately unable to judge the extent of informationsharing among workers. However, to the extent thatsuch information sharing was not complete, we mayobserve different, possibly stronger, results under apublic disclosure of relative standing. Falk and Ichino(2006), for instance, find that the ability to observeother workers’ productivity leads to an increase ineffort in a flat wage setting. Although the distinctionbetween the impact of private and public informationis clearly beyond the scope of this paper, we regard itas an interesting direction for future work.

AcknowledgmentsThe authors thank David de Meza, Daniel Ferreira, ThomasKittsteiner, Ignacio Palacios-Huerta, and Yona Rubinsteinfor insightful comments. They are grateful to all thoseinvolved in providing the data. Joy Blundell and VeraRadyugina provided excellent research assistance. Thispaper has been screened to ensure that no confidentialinformation is revealed. All errors remain the authors’responsibility.

ReferencesAnderson, R., S. Manoogian, S. Reznick. 1976. The undermining

and enhancing of intrinsic motivation in preschool children.J. Personality Soc. Psych. 34(5) 915–992.

Azmat, G., N. Iriberri. 2010. The importance of relative perfor-mance feedback information: Evidence from a natural exper-iment using high school students. J. Public Econom. 94(7–8)435–452.

Bandiera, O., I. Barankay, I. Rasul. 2009. Team incentives: Evidencefrom a field experiment. Working paper, London School of Eco-nomics, London.

Bandura, A., D. Cervone. 1983. Self-evaluative and self-efficacymechanisms governing the motivational effects of goal sys-tems. J. Personality Soc. Psych. 45(5) 1017–1028.

Baron, R. 1988. Negative effects of destructive criticism: Impact onconflict, self-efficacy and task performance. J. Appl. Psych. 73(2)199–207.

Buunk, A., F. Gibbons. 2007. Social comparison: The end of a theoryand the emergence of a field. Organ. Behav. Human DecisionProcesses 102(1) 3–21.

Casas-Arce, P., F. A. Martínez-Jerez. 2009. Relative performancecompensation, contests, and dynamic incentives. ManagementSci. 55(8) 1306–1320.

Charness, G., P. Kuhn. 2007. Does pay inequality affect workereffort? Experimental evidence. J. Labor Econom. 25(4) 693–725.

Clark, A., P. Frijters, M. Shields. 2008. Relative income, happinessand utility: An explanation for the easterlin paradox and otherpuzzles. J. Econom. Literature 46(1) 95–145.

Clark, A., D. Masclet, M. Villeval. 2010. Effort and comparisonincome: Experimental and survey evidence. Indust. Labor Rela-tions Rev. 63(3) 407–426.

Deci, E. 1972. Effects of contingent and noncontingent rewards andcontrols on intrinsic motivation. Organ. Behav. Human Perfor-mance 8(2) 217–229.

Deci, E. 1975. Intrinsic Motivation. Plenum Press, New York.Deci, E., R. Koestner, R. Ryan. 1999. A meta-analytic review of

experiments examining the effects of extrinsic rewards onintrinsic motivation. Psych. Bull. 125(6) 627–668.


Delfgaauw, J., R. Dur, J. Sol, W. Verbeke. 2009. Tournament incen-tives in the field: Gender differences in the workplace. Workingpaper, Erasmus University, Rotterdam, The Netherlands.

Eriksson, T. 1999. Executive compensation and tournament theory:Empirical test on danish data. J. Labor Econom. 17(2) 262–280.

Eriksson, T., A. Poulsen, M. C. Villeval. 2009. Feedback and incen-tives: Experimental evidence. Labour Econom. 16(6) 679–688.

Ertac, S. 2006. Social comparisons and optimal revelation: The-ory and experiments. Working paper, Koç University, Istanbul,Turkey.

Etzioni, A. 1971. Modern Organizations. Prentice-Hall, EnglewoodCliffs, NJ.

Falk, A., A. Ichino. 2006. Clean evidence on peer effects. J. LaborEconom. 24(1) 39–57.

Fehr, E., S. Gächter. 2000. Fairness and retaliation: The economicsof reciprocity. J. Econom. Perspect. 14(2) 159–181.

Fehr, E., K. Schmidt. 1999. A theory of fairness, competition andcooperation. Quart. J. Econom. 114(3) 769–817.

Fershtman, C., H. Hvide, Y. Weiss. 2005. Cultural diversity, statusconcerns and the organization of work. Res. Labour Econom.24(1) 3–38.

Festinger, L. 1954. A theory of social comparison processes. HumanRelations 1(7) 117–140.

Frank, R. 1984. Are workers paid their marginal products? Amer.Econom. Rev. 74(4) 549–571.

Frank, R. 1985. Choosing the Right Pond: Human Behavior and theQuest for Status. Oxford University Press, New York.

Kluger, A., A. DeNisi. 1996. The effects of information interventionson performance: A historical review, a meta-analysis and apreliminary information intervention theory. Psych. Bull. 119(2)254–284.

Knoeber, C., W. Thurman. 1994. Testing the theory of tournaments:An empirical analysis of broiler production. J. Labor Econom.12(2) 155–179.

Kohn, A. 1993. Punished by Rewards. Houghton-Mifflin, Boston.Kuhnen, C., A. Tymula. 2010. Feedback, self-esteem and perfor-

mance in organizations. Working paper, Northwestern Univer-sity, Evanston, IL.

Latham, G., K. Wexley. 1981. Increasing Productivity Through Perfor-mance Appraisal. Addison-Wesley, Reading, MA.

Lavy, V. 2008. Gender differences in market competitiveness in areal workplace: Evidence from performance-based pay tourna-ments among teachers. NBER Working Paper 14338, NationalBureau of Economic Research, Cambridge, MA.

Lazear, E. 1989. Pay equality and industrial politics. J. PoliticalEconom. 97(3) 561–580.

Lazear, E., S. Rosen. 1981. Rank-order tournaments as optimumlabor contracts. J. Political Econom. 89(5) 841–864.

London, M. 2003. Job Feedback: Giving, Seeking and Using Feedbackfor Performance Improvement. Lawrence Erlbaum Associates,Mahway, NJ.

Luttmer, E. 2005. Neighbors as negatives: Relative earnings andwell-being. Quart. J. Econom. 120(3) 963–1002.

Moldovanu, B., A. Sela, X. Shi. 2007. Contests for Status. J. PoliticalEconom. 115(2) 338–363.

Moore, D., W. Klein. 2008. Use of absolute and comparative perfor-mance feedback in absolute and comparative judgments anddecisions. Organ. Behav. Human Decision Processes 107(1) 60–74.

Nadler, D. 1979. The effects of feedback on task group behavior:A review of the experimental research. Organ. Behav. HumanPerformance 23(3) 309–338.

Oswald, A. 1997. Happiness and economic performance. Econom. J.107(6) 1815–1831.

Sunde U., T. Dohmen, A. Falk, B. Weber, K. Fliessbach. 2011. Rela-tive versus absolute income, joy of winning, and gender: Brainimaging evidence. J. Public Econom. 95(3) 279–285.

Tran, A., R. Zeckhauser. 2009. Rank as an incentive. Working paper,Indiana University, Bloomington.

tournaments without prizes - evidence from personnel records

Documents

chosen level

denote estimates

denote estimates

experimental

linear cohort

linear time

robust standard

individual