Вы находитесь на странице: 1из 25

Rank-Order Tournaments as Optimum Labor Contracts Author(s): Edward P.

Lazear and Sherwin Rosen Source: Journal of Political Economy, Vol. 89, No. 5 (Oct., 1981), pp. 841-864 Published by: The University of Chicago Press Stable URL: http://www.jstor.org/stable/1830810 . Accessed: 16/09/2013 11:04
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

The University of Chicago Press is collaborating with JSTOR to digitize, preserve and extend access to Journal of Political Economy.

http://www.jstor.org

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

Rank-Order Tournamentsas Optimum Labor Contracts

EdwardP. Lazear and SherwinRosen


Research University of Chicagoand NationalBureau of Economic

This paper analyzes compensation schemes which pay according to an individual's ordinal rank in an organization ratherthan his output level. When workersare riskneutral,itis shown thatwages based allocation of resources as an upon rank induce the same efficient incentivereward scheme based on individual output levels. Under some circumstances, risk-averseworkersactually prefer to be paid on the basis of rank. In addition, if workersare heterogeneous in ability, low-quality workers attempt to contaminate high-quality firms, resultingin adverse selection. However, if abilityis known in advance, a competitive handicapping structure existswhichallows all in the same organization. workersto compete efficiently

I.

Introduction

It is a familiar proposition that under competitive conditions workers are paid the value of their marginal products. In this paper we show that competitive lotteries are often efficient and sometimes superior to more familiar compensation schemes. For example, the large salaries of executives may provide incentives for all individuals in the firm who, with hard labor, may win one of the coveted top positions. This paper addresses the relation between compensation and incentives in the presence of costly monitoring of workers' efforts and We are indebted to Jerry Green,MertonMiller, JamesMirrlees, GeorgeStigler, and EarlThompson Joseph Stiglitz, for helpful comments. The research wassupported
in part by the National Science Foundation. The research reported here is part of the NBER's research program in Labor Studies. Any opinions expressed are those of the authors and not those of the National Bureau of Economic Research. ?
Journalof PoliticalEconomy, 1981, vol. 89, no. 5] 1981 by The University of Chicago. 0022-3808/81/8905-0006$01.50

841

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

842

JOURNAL

OF POLITICAL

ECONOMY

output. A wide varietyof incentive payment schemes are used in practice. Simple piece rates, which have been extensivelyanalyzed (see, e.g., Cheung 1969; Stiglitz1975; Mirrlees 1976), gear payment to output. We consider a rank-order payment scheme which has not been analyzed but which seems to be prevalent in many labor contracts.This scheme pays prizes to the winnersand losers of labor market contests.The main differencebetween prizes and other incentiveschemesis thatin a contestearningsdepend on the rank order of contestants and not on "distance." That is, salaries are not contingent upon the output level of a particular game, because prizes are fixed in advance. Performanceincentivesare set by attemptsto win the contest.We argue that in many circumstancesit is optimal to set up executivecompensation along these lines and thatcertainpuzzling features of that market are easily explained in these terms. Central to this discussion are the conditions under which mecha(Alchian and Demsetz 1972). nismsexist for monitoringproductivity are available, then the If inexpensive and reliable monitorsof effort best compensation scheme is a periodic wage based on input. Howso that workers can alter their ever, when monitoringis difficult, input with less than perfect detection, input-wage schemes invite shirking.The situation often can be improved if compensation is related to a more easily measured output level. In general, inputbased pay is preferablebecause it changes the riskborne by workers in a favorableway. But when monitoring costs are so high that moral fromusing outputhazard is a serious problem,the gain in efficiency losses. Paying workers on based pay may outweigh the risk-sharing the basis of rank order alters costs of measurement as well as the nature of the risk borne by workers.It is for these reasons that it is incentivestrucsometimesa superior way to bring about an efficient ture. In the development below we start with the simplestcase of risk the basic issues. Then the more general case of to illustrate neutrality risk aversion is treated in Section III. Section IV considers issues of when workersare heterogeneous. sortingand self-selection II. Piece Rates and Tournaments with Risk Neutrality

To keep thingssimpleand to avoid sequential and dynamicaspects of the problem,we confineattentionto a single period in all thatfollows. Therefore, the reader should thinkof the incentiveproblem in terms of career development and lifetime productivityof workers. The worker's(lifetime)output is a random variable whose distribution is controlledby the workerhimself.In particular,the workeris allowed to control the mean of the distributionby investingin costly skills

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

RANK-ORDER

TOURNAMENTS

843

prior to enteringthe market. However, a given productivity realization also depends on a random factor which is beyond anyone's control. Employers may observe output but cannot ascertain the extentto whichit is due to investment expenditure or to good fortune or to both,thoughworkersknowtheirinputas well as output. Worker j produces lifetimeoutput qj according to
qj =/ tj + Ej,

(1)

where pj is the level of investment,a measure of skill or average output,chosen by the workerwhen youngand priorto a realizationof the random or luck component, Ej. Average skill, ,j, is produced at cost C(,), withC', C" > 0. The random variable E, is drawn out of a knowndistribution withzero mean and varianceC2.1 Here E is lifetime luck such as life-persistent person-effects or an abilityfactor,whichis revealed veryslowlyover the worker'slifetime.The crucial assumption is thatproductivity riskis nondiversifiable by the workerhimself. That is another reason forchoosing a long period forthe analysis.For example, if the period were very short and the random factorwas independentlydistributedacross periods, the workercould diversify per period risk by repetitionand a savings account to balance off good and bad years. Evidentlya persistentperson or abilityeffect cannot be so diversified when it is undiscoverablequickly,as appears true of managerial talent,for example. It is assumed, however,thatE is i.i.d. across individuals, so that owners of firmscan diversify risk either by pooling workers together in one firm or by holding a portfolio. To concentrate on incentive aspects of various contractual arrangements,we adopt the simplesttechnologyfor firms.Production requires only labor and is additively separable across workers. By virtue of the independence assumptions, managers act as expected value maximizersor as if they were risk neutral. Free entryand a competitive output marketset the value of the product at V per unit. Again, these assumptionsare adopted to illustratebasic issues in the simplestway. The analysis also applies when there are complemenbut more tarities among workersin production,whichis more realistic to exposit. difficult Piece Rates The piece rate is verysimple to analyze when workersare riskneutral. It involves paying the worker the value of his product. Let r be the
1 In thispaper the workerhas no choice over o-.This does not affect the risk-neutral solution but does have an effectif workers are risk averse, since they tend to favor overlycautious strategies.Also, virtually all the resultsof this paper hold true if the error structureis multiplicative rather than additive.

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

844

JOURNAL

OF POLITICAL

ECONOMY

the worker'snet income is rq - C(pu). piece rate. Ignoringdiscounting, Risk-neutralworkerschoose u to maximize expected net return
E[rq
-

C(g)] = r,

C(u).

The necessaryconditionis r = C'(tk) or the familiarrequirementthat investment equates marginalcost and return.On the other hand, the expected profitof a firmis
E(Vq
-

rq) = (V -r),

so free entryand competitionfor workersimplyr = V. Consequently


V
= C'(p.).

equals its social return,yieldingthe The marginalcost of investment standard result that piece rates are efficient. Tournaments Rank-Order We shall consider two-playertournamentsin which the rules of the game specify a fixedprize W1to the winnerand a fixedprize W2to the loser. All essential aspects of the problem readily generalize to any number of contestants.A worker's production follows (1), and the winner of the contestis determinedby the largestdrawing of q. The contestis rank order because the margin of winningdoes not affect earnings. Contestants precommit their investmentsearly in life, knowingthe prizes and the rules of the game, but do not communicate witheach otheror collude. Notice thateven though thereare two and not oligopolisplayersin a given matchthe marketis competitive tic,because investment is precommittedand a given player does not know who his opponent will be at the time all decisions are made. Each person plays against the "field." The We seek to determinethe competitive prize structure (W1,W2). method proceeds in two steps. First,the prizes W1 and W2 are fixed and workers' investmentstrategies are analyzed. Given arbitrarily these strategies,we then find the pair (W1,W2) that maximizes a constraintby firms. worker'sexpected utility, subject to a zero-profit It will be seen that a worker's incentivesto invest increase with the spread between winningand losing prizes, W1 - W2. Each wants to improve the probabilityof winning because the return to winning varies with the spread. The firmwould always like to increase the spread, ceteris paribus, to induce greater investment and higher productivity, because its output and revenue are increased. But as contestantsinvest more, their costs also rise. That is what limitsthe too large a spread induce excesspread in equilibrium:Firmsoffering A competing firmcan attractall of these workersby sive investment. decreasing the spread because investmentcosts fall by more than

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

RANK-ORDER

TOURNAMENTS

845

expected product, raising expected net earnings. Increasing marginal cost of skill implies a unique equilibrium spread between the prizes that maximizes expected utility. More precisely,consider the contestant'sproblem, assuming that C (,), so thattheirbehavior is both have the same costs of investment identical. A contestant'sexpected utility(wealth) is
(P)IW1 - C(y)] + (1 - P)[W2 - C(Z)] = PW1 + (1 - P)W2
-

(2) thatj wins is of winning.The probability where P is the probability (3) = prob (g j - t4k > G) G (g j - t), - g(e), G(-) is the cdf of A,E(f) = O, andE(f2) = where Ek Ej, 6 2cr2(because Ej and Ek are i.i.d.). Each player chooses pi to maximize (2). Assuming interiorsolutions, this implies
(W1-W2) a

P = prob (qj >

qk)

= prob (Uj

Ilk

>

Ek -

Ej)

OtFi)

= ?

and (W1 - W2)


P2 - C"()<Oi

=j,k.

(4)

We adopt the Nash-Cournot assumptionsthat each player optimizes against the optimum investmentof his opponent, since he plays against the marketover whichhe has no influence.Therefore,j takes fork. It then and conversely his investment /k as givenin determining followsfrom (3) that, for playerj
aP/luj = aG(gj /Lk)/tL =

g(/Lj

into (4) yieldsj's reaction function which upon substitution


(W1 - W2)gQ(
.k) -

C'

0.

(5)

with (5). Player k's reaction functionis symmetrical impliesthatwhen the Nash solutionexists,Pj = /k and P Symmetry G (0) = 1/2, so the outcome is purelyrandom in equilibrium.Ex ante, each player affectshis probabilityof winning by investing.2
2 However, it is not necessarilytrue that there is a solution because with arbitrary densityfunctionsthe objectivefunctionmay not be concave in the relevantrange. It is solutionexistsprovided that.j2 is sufficiently large: possible to show thata pure strategy factor.This resultaccords with Contestsare feasible only when chance is a significant differenceof opinion is and is in the spiritof the old sayingthata (sufficient) intuition = g(lij - AOk)and g( ) is a pdf, necessaryfora horse race. Stated otherwise,sinceOP/0,uj = = of second-orderconditionsin (4) and fulfillment g'; (jIlk) may be positive, implies sharp breaks in the reaction function.If U2 is small enough the breaks occur at and a Nash equilibrium in pure strategieswill not exist. verylow levels of investment, Existence of an equilibrium is assumed in all that follows.

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

846
to

JOURNAL

OF POLITICAL

ECONOMY

Substituting ,j

,tk

at the Nash equilibrium,equation (5) reduces


i =j,k,

C'(p~i) = (W1 - W2)g(0),

(6)

verifying the point above that players' investmentsdepend on the spread between winningand losing prizes. Levels of the prizes only influencethe decision to enter the game, which requires nonnegativityof expected wealth. The risk-neutral firm'srealized gross receiptsare (qj + qk) V, and its costs are the totalprize money offered,W1 + W2.Competitionfor labor bids up the purse to the point where expected total receipts
equal costs W1 + W2
=

condition reduces to equilibrium,the zero-profit V11= (W1 + W2)/2.

(9,i

9.k)

V. But since

pj

9k

g in

(7)

The expected value of product equals the expected prize in equilibrium. Substitute(7) into the worker'sutility function(2). Noting that P = 1/2 in equilibrium,the worker'sexpected utility at the optimum investmentstrategyis
Vpu- C(g). (8)

The equilibriumprize structure selectsW1and W2to maximize(8), or [V - C'(,W)](au/(Wi)


=
0

1, 2.

(9)

The marginalcost of investment equals itsmarginalsocial return,V = C'(g), in the tournamentas well as the piece rate. Therefore,competitive tournaments,like piece rates, are efficient and both result in exactlythe same allocation of resources. Some furthermanipulationof the equilibriumconditionsyieldsan interesting interpretation in terms of the theoryof agency (see Ross 1973; Becker and Stigler 1974; Harris and Raviv 1978; and Lazear 1979):
W, = VpL + C'(,u)/2g(O)
W2 = Vg - C'U(j)2g(0)
= =

Vt + V/2g(O) - VI2g(O). (10)

The second equality followsfrom V = C'(/i). Now thinkof the term C'(g)/2g(O) = V/2g(0) in (10) as an entrancefee or bond thatis posted by each player. The winningand losing prizes pay off the expected marginal value product plus or minus the entrance fee. That is, the players receive theirexpected product combined witha fair winnertake-allgamble over the totalentrancefeesor bonds. The appropriate social investment incentivesare given by each contestant'sattemptto win the gamble. This contrastswiththe main agency result,where the bond is returnedto each workeraftera satisfactory performancehas

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

RANK-ORDER

TOURNAMENTS

847

been observed. There the incentive mechanism works through the employee's attemptsto work hard enough to recoup his own bond. Here it works through the attemptsto win the gamble. Comparative staticsfor this problem all follow from (9) and (10) once a distributionis specified. For example, if E is normal with variance _2, theng(0) = 2uV'cr. It followsfrom(10) thatthe optimal spread varies directlywith V and oC2. While several other interesting observationscan be made of this sort,we note a somewhatdifferent but important practical implication of this general scheme. Even though the optimal prize structuredetermines expected marginal product throughits effecton workerchoice of g and the zero-profit condition (7) implies that expected prizes equal expected producdo not equal tivity, neverthelessactual realized earnings definitely productivity in either an ex ante or ex post sense. Consider ex ante first. Since gj = 9k= a, expected productsare equal. Since W1> W2is the paymentthat receives never required to induce any investment, equals the payment that k receives. It is impossible that the prize is equal to ex ante product,because ex ante productsare equal. Nor do wages equal ex post products. Actual product is Vq rather than Vtk. But q is a random variable,the value of whichis not knownuntilafter the game is played, while W1and W2are fixedin advance. Only under the rarest coincidence would W1 = Vqj and W2 = Vqk. Consider the salary structurefor executives. It appears as though the salary of, say, the vice-presidentof a particular corporation is below thatof the presidentof the same corporation. Yet substantially presidentsare oftenchosen fromthe ranksof vice-presidents. On the to presiday that a given individual is promoted fromvice-president dent, his salary may triple. It is difficult to argue that his skillshave tripled in that 1-day period, presenting difficultiesfor standard theorywhere supply factorsshould keep wages in those two occupations approximatelyequal. It is not a puzzle, however, when interpreted in the context of a prize. The president of a corporation is viewed as the winner of a contest in which he receives the higher prize, W1. His wage is settledon not necessarily because it reflects his currentproductivity as president,but rather because it induces that individual and all other individuals to performappropriatelywhen they are in more junior positions. This interpretation suggests that presidentsof large corporations do not necessarilyearn high wages because they are more productive as presidents but because this particular type of payment structuremakes them more productive over theirentireworkinglives. A contestprovides the proper incentives for skill acquisition prior to coming into the position.3
3 If E is a fixed effect, there is additional informationfromknowing the identity of of a winner is g + E(Ej I qj > qk), while winnersand losers. The expected productivity that of a loser is g + E(Ej I qj < qk). In a one-period contest there is no possibility of

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

848
Comparisons

JOURNAL

OF POLITICAL

ECONOMY

indifferent Though tournamentsand piece rates are substantially for creatingincentives,we have demonstratedthe surprisstitutions ing resultthatboth achieve the Pareto optimal allocation of resources when workersare riskneutral. In factother schemes also achieve this allocation. For example, instead of playing against an opponent, a workermightbe compared witha fixedstandard q, withone payment and another, higher, payawarded if output falls anywhere below _q mentawarded ifoutput fallsanywhereabove standard. Attempting to as attempting to beat beat the standard has the same incentiveeffects another player. Using the same methodsas above, it is not difficult to show thatthere are spread-standardcombinationsthatinduce Pareto optimum investments.Since all these schemes involve the same investmentpolicy,and since average payout by the firmequals average product forall of them,theyall yieldthe same expected rewardsand, therefore,the same expected utilityto workers.4 In spite of the apparent equality of these schemes in termsof the preferences of risk-neutralworkers, considerations of differential and measurementmayserve to break thesetiesin costsof information practical situations.The essential point follows from the theory of measurement (Stevens 1968) that a cardinal scale is based on an underlyingordering of objects or an ordinal scale. In that sense, an ordinal scale is "weaker" and has fewerrequirementsthan a cardinal scale. If it is less costlyto observe rank than an individual's level of output, then tournamentsdominate piece rates and standards. On the other hand, occupations for which output is easily observed save resources by using the piece rate or standard, or some combination, and avoid the necessityof making directcomparisons withothers as the tournamentrequires. Salesmen, whose output level is easily obare paid by piece rates,whereas corporate executives, served,typically whose output is more difficult to observe, engage in contests. In a modern, complex business organization, a person's productivityas chief executive officer is measured by his effect on the profitability of the whole enterprise.Yet the costsof measurementfor
taking advantage of this information.However, in a sequential contest with no firmspecificcapital, the informationwould be valuable and would constrain subsequent wage paymentsin successive rounds through competitionfrom other firms.It is not difficultto show that this does not affect the general nature of the bond-gamble elements or firmsadopt if the investmenthas firm-specific solution. Alternatively, do not necessarily policies thatbind workersto it (as in Lazear 1979), these restrictions apply. since for anyq a correspondingspread 4The level of the standard is indeterminate, This is also true of contestsamong can be chosen to achieve the optimal investment. the prizes of N - 2 of them are indetermimore than two players.WithN contestants, is dropped, the indeterminacyvanishes in both cases. nate. When risk neutrality

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

RANK-ORDER

TOURNAMENTS

849

expensive. Instead, it each conceivable candidate are prohibitively mightbe said thatthose in the runningare "tested"by assessmentsof performanceat lower positions.Realizations fromsuch testsare sample statistics in these assessments,in much the same way that grades are assigned in a college classroom and IQ scores are determined. The point is that such tests are inherentlyordinal in nature, even though the profitability of the enterpriseis meteredby a well-defined, cardinal ratio scale. It is in situationssuch as this that the conditions seem ripe for tournaments to be the dominant incentive contract institution. Notice in this connection that the basic prize and piece-rate structuressurvivea broad class of revenue functionsotherthan summable ones. Even if the production functionof the firmincludes complior substitution among cated interactionsinvolvingcomplementarity individual outputs, there exists the possibilityof paying workers either on the basis of individual performanceor by rank order. The revenue functionitselfcan even involve rank-orderconsiderations, stillexist.For example, spectatorsat a horse race and both possibilities generally are interestedin the speed of the winning horse and the closeness of the contest. Then the firm's(track) revenue function few order statistics; yetthe horses could be paid depends on the first on the basis of theirspeed ratherthan on the basis of win,place, and show positions. Both methods would induce them to run fast.5 of the problem of tournament There has been verylittletreatment prize structureand incentives in the literature.Little else but the well-knownpaper by Friedman (1953) based on Friedman-Savage literature preferencesforlotteriesexistsin economics. In the statistics there is an early paper by Galton (1902) that is worthyof brief and second-place discussion. Galton inquired into the ratio of firstprize money in a race of n contestants,assuming the prizes were divided in the followingratio:
W11W2= (Q1 - Q3)/(Q2 - Q3).

etc. (fastest)order statistic, Here Qi is the expected value of the firstWhile a moment'sreflection suggeststhis criterionto be roughlyrea priori lated to marginal productivity, Galton proposed it on strictly grounds.He wenton to show the remarkableresultthatthe ratioabove
5The reader is reminded that throughout this section and the next workers are identicala prioriand differ onlyex post throughthe realizationof E. In the real world, marketparticipantsare sorted into different where there is population heterogeneity, contests.There players (and horses, for that matter)who are known to be of higher quality ex ante may play in games with higher stakes. If it can be accomplished, the sorting is by anticipated marginal products. In that sense, pay differencesamong contestantsof known quality resemble the effectof a "piece rate." These issues are more thoroughlydiscussed below.

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

850

JOURNAL

OF POLITICAL

ECONOMY

is approximately3 when the parent distribution of speed is normal. Hence, thiscriterionresultsin a highlyskewed prize structure.From what we know today about the characteristic skew of extreme value distributions, a skewed reward structurebased on order statistics is less surprising for virtuallyany parent distribution.In the more modern statistical literature,the method of paired comparisons has tournament-likefeatures. Samples from differentpopulations are compared pairwise, and the object is to choose the one with the largest mean. Comparing all samples to each other is like a roundrobin tournament. An alternativedesign is a knockout tournament withsingle or double elimination.The latterrequires fewersamples and is thereforecheaper, but does not generate as much information as the round robin (David 1963; Gibbons, Olkin, and Sobel 1977). Galton's original work and the more modern developments it has given rise to are not helpful to us; theydeal withsamples fromfixed populations, so the reward structureis irrelevantfor resource allocation.The problemwe have treatedhere is thatof choosing the reward structure to provide the proper incentiveand elicitthe sociallyproper distributions. III. Optimal Compensation with Risk Aversion

All compensation systems can be viewed as schemes which transform the distribution of productivity to a distribution of earnings. A piece rate is a linear transformation of output,so the distribution of income is the same apart froma change in locationand scale. A tournamentis a highly nonlinear transformation:It converts the continuous disinto a discrete,binomial distribution of intributionof productivity come. When workers are risk neutral, both schemes yield identical investments and expected utility because their first momentsare the same. In thissection,it is shown thatwithriskaversionone method or the other usually yields higher expected utility, because the interaction between insurance and action implies substantially different first in the two cases.6 and second moments of the income distribution We have been unable to completely characterize the conditions under which piece rates dominate rank-ordertournamentsand vice versa,but we show some examples here. Truncation offeredby prizes implies more control of extreme values than piece rates but less control of the middle of the distribution.Differentutilityfunctions
6 One might think that risks could be pooled among groups of workers through sharing agreements,but that is false because of moral hazard. A worker would never agree to share prizes since doing so would resultin ,u = 0, and consequentlyE(qj + qk) = 0 and bankruptcy forthe firm.As a result,firms tournamentsor piece rates offering in the pure sense yield higher expected utilitythan the sharing arrangement.

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

RANK-ORDER

TOURNAMENTS

851

weight one aspect more than the other so that tournaments can actually dominate piece rates. Linear Piece Rate7 Optimum The piece-ratescheme analyzed pays workersa guarantee,I, plus an incentive, rq,wherer is the piece rate per unit of output. The problem for the firmis to pick an r, I combination that maximizes workers' expected utility max [E(U) = maxf U(y)6(y)dy], where y =I + rq - C()
=

(11)

(12)
-

I + rp.+

rE

C(k)

and 0(y) is the pdf of y. The worker'sproblem is to choose tkto maximize expected utility given I and r. If E f(E), the worker's problem is maxE(U) = f U[I + rp. + The first-order condition is 9E(U) = f [U'(y)][r- C'(It)]f(E)dE ~49 which convenientlyfactorsso that r = C'(Q).
=

rE

-C

(g)]f (E)dE.

0,

(13)

Condition (13) is identical to the risk-neutral case, because E is independent of investmenteffort, g. Assuming risk-neutral employers,Vgtis expected revenue from a workerand I + rg is expected wage payments.Therefore, the zeroprofitmarketconstraintis = I + rgu. ( 14) V,(A into (12), the optimum contract Solving (14) for I and substituting maximizes f U{Vg(r) + rE- C[g(r)]}f(E)dE the withrespectto r, where ,t = ,u(r)satisfies (13). Aftersimplification
7 The following is similarto a problemanalyzed byStiglitz(1975). A linear piece-rate structure is a simplification. A more general structurewould allow for nonlinear piece rates (see Mirrlees 1976).

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

852 marginal condition is


[V
-

JOURNAL

OF POLITICAL

ECONOMY

C'()]

d_ EU'

dr

+ EEU'=

0.

(15)

Since riskaversion impliesEEU' < 0, (15) shows thatV > C'(pu) in the is optimum contractfor risk-averseworkers.This underinvestment the moral hazard resultingfrominsuranceI > 0 and r < V implied by (15). function Using familiarTaylor series approximationsto the utility and a normal densityfor E, the optimum is approximated by

c'-1 ( 1 + sC"o-2)
and
0_2 202 (1 + sC"o-2)2

(16)

(17) (

-U"/U' evaluated at mean income is the measure of absowhere s lute riskaversion. Investmentincreases (see [16]) in V and decreases in s, C", and 0_2, because all these changes implysimilarchanges in the marginal piece rate r which influencesinvestment throughcondition on (13). The same changes in V, s, and C" have correspondingeffects the variance of income (see [17]), but an increase in 0-2 actually reduces variance,if 0-2 is large, because it reduces r and increasesJ*8 Prize Structure Optimum in a two-playergame is The worker'sexpected utility E(U) = P{U[W1
-

C(tk*)]} + (1

P){U[W2

C(tk*)]},

(18)

where * denotes the outcome of the contestratherthan the piece-rate scheme. The optimum prize structureis the solution to max (E(U*)
W1 sW2

= max{P
*

U[W1

C(Q*)]

+ (1 -P)

U[W2 -C(W)]})
(19)

constraint subject to the zero-profit


V11* = PW1 + (1 - P)W2.

(20)

The workerselectsg* to satisfy oE(U)/h,* = 0. Since cost functions are the same and Ej and Ek are i.i.d., the Nash solutionimpliespi = gk
8 Furthermore, r-V/(1 + sC"o-2) and I-sV2o-2/(1 + sC"o2)2,so thatr = V andI = 0 in the case of riskneutrality (s = 0). All these approximationsuse first-order expansions for termsin U'( ) and second-orderexpansions fortermsin U( ). The same is true of the approximations below for the tournament.

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

RANK-ORDER

TOURNAMENTS

853

and P = plifiesto

'/2

as before. Then the worker's investmentbehavior simC' (u* (21) (1 = 1, 2. (22) (23)

2[U(l) - U(2)]g(0) U'( 1) + U'(2) where U(T) UIWT - C(p*)] and U'(r)- U'[WT C(,*)]forT Equation (21) implies
/I= 1*(W,,W2),

and the optimum contract(W1,W2)maximizes E(U*) =


12 U[W1 - C(,*)]

1/2U[W2 -C(9*)]

subject to (20), withP = '/2,and (22). Increasing marginal cost of investment and risk aversion guarantees a unique maximum to (23) when a Nash solution exists. Again, assuming a normal densityforE, second-order approximations yield - Cr-l (1 and
2

(24)

V 7rVOfr;i (1 + ITCrsC o-2)2'

(25) (5

where

= W, - C (/i*) ifq j ye = W2 - C (,*) ifq j

> <

qk qk

and E j- N(Oo2),Ek- N(0,o-2), and cov (Ej,Ek) = 0. The comparative staticsof (24) and (25) are similarto the piece rate (16) and (17) and need not be repeated.
Comparisons

Equations (16) and (24) indicate that investmentand expected incomel0are lower forthe contestthan forthe piece rate at givenvalues of s. Moreover, for values of o.2 in excess of 1/sC"\/, the variance of income in the tournament is smaller than for the piece rate. This would seem to suggest that contests provide a crude formof insurance when the variance of chance is large enough, but the problem is significantly more complicated than that because there is no separation between tastesand opportunitiesin this problem: The optimum
9 Futhermore,C'(p.*) - g(O)(W1 - W2), so the spread is stillcrucial for investment case. incentives,as in the risk-neutral 10Since y = Vu level of u when CQ(g),and since ,u is below the wealth-maximizing workersare riskaverse, lower,uimplieslowerybecause revenue fallsbymore than cost.

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

854

JOURNAL OF POLITICAL ECONOMY

mean and variance themselvesdepend on utility-function parameters. Thus, for example, for the constant, absolute risk-aversionutility functionU = -e-slYs, the insurance provided by the contest is insufficient to compensate foritssmallermean: It can be shown thatthe of the optimal piece rate exceeds that of the expected indirectutility optimal tournamentfor all values of o.2, at least withnormal distributions and quadratic investment-cost functions.However, when there is declining absolute risk aversion, we have examples where the contest dominates the piece rate. Illustrativecalculations are shown in table 1 using the utility function U = aya, which exhibitsconstantrelativebut declining absolute riskaversion,s(y) = (1 - a)Iy. Again quadratic costsand normal errors are assumed. However, this utilityfunction is defined for positive incomes only,so an amount of nonlabor income yo is assigned to the worker to avoid a major approximation error of the normal, which admits negative incomes (i.e., the possibility of losses). Table 1 shows that when yo = 100 so thats = .005, the contestis preferred until o.2 ? 3. However, if yo = 25 so that s = .020, the contestis only preferredfor a2 < .2. The intuitionis that piece rates concentratethe mass of the income distribution near the mean, while contestsplace 50 percentof the weightat one value significantly below the mean and the other value significantly above. Strongly risk-averse workersseem to dislike the binomial nature of thisdistribution when too much of the mass at low levels of 0J2 is high because it concentrates utility. However, when o-2 is small,the contestwhichtruncatesthe tails
TABLE 1
CONSTANT
'2

RELATIVE

RISK AVERSION

yA*

E (U)
Yo = 100; s(y0)= .005

E(U*)

.1 .5 1 3 6 12

.9995 .9975 .9950 .9852 .9710 .9436

.9984 .9922 .9846 .9552 .9142 .8420 Yo = 25; s(y0)= .020

5.012155 5.012150 5.012100 5.011940 5.011800 5.011420

5.012465 5.012445 5.012295 5.011925 5.011415 5.010515

.1 .2 1 12
No] p/2:

.9980 .9960 .9807 .8094


=

.9938 .9878 .9419 .5741

2.524665 2.524616 2.524237 2.519930

2.524725 2.524575 2.523437 2.514282

F.-U
a2.

ay';y

ye

+I +rq -C(y)forpiecerate;y

=yo + W -C(yA) forcontest(i = 1, 2);a = .5, V = 1,C (?)

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

RANK-ORDER

TOURNAMENTS

855

of the income distributionassociated with a linear piece rate has higher value. IncomeDistributions While it is not possible to make a general argument based on an example, table 1 suggests that persons with more endowed income and smaller absolute risk aversion are more likelyto prefercontests, and those withlow levels of endowed wealth and larger absolute risk aversion are more likelyto preferpiece rates. Consider a situationin which all persons have the same utility function,such as the one in table 1, and face the same costs and luck distribution, the only difference being the fact that some workers have larger endowed incomes than others.If thisdifference is large enough, it can be optimal to pay piece ratesto those withsmall values of endowed income and to pay prizes to those with large values. Individuals will self-selectthe paymentscheme in accordance withtheirwealth. The distribution of earnings among those selecting the piece-ratejobs is normal with mean Vg and variance r2o-2.It is binomial with mean Vp)* and variance (AW)2/4for those who enter tournaments.Note that , and A* depend upon s(y), which is smaller for workerswho select contests, and it can turn out as it does in table 1 thatexpected income is larger in the contestthan in the piece rate; for example, if o-2 = 1 then the rich prefercontests(5.012295 > 5.012100) and the poor preferpiece rates (2.524237 > 2.523437), but /* = .9846 exceeds p. .9807. This situationis shown in figure 1. The overall distribution is the sum of a binomial and a normal with lower mean, weighted by the number of individuals in each occupation (see fig. 1). It is positively skewed because Vp.* > Vp.. Note also that the distribution of wage income will be less skewed than that of total income. The reason is thatyo and mean-wage income are posicorrelatedbecause the likelihoodof choosing a contestincreases tively withyo. These implicationsconform to the standard findingson the distribution of income in an economy. This example is interesting because it is verycloselyrelated to some early resultsof Friedman (1953), who studied how alternativesocial thatcater to workers' arrangementscan produce income distributions risk preferences. He showed that the Friedman-Savage utility function leads to a two-classdistribution. Persons in the risk-averse region are assigned to occupations in which income follows productivity, while persons in the risk-preferring ticketsin very region buy lottery riskyoccupations in which few win very large prizes. The overall is the sum of these two and exhibitscharacteristic distribution skew. The Friedman-Savage utilityfunction implies that a person's risk

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

856
f(y)

JOURNAL

OF POLITICAL

ECONOMY

.5.

+ Vp Y,

Z+Y( W
FIG. 1

yO+VtL*

W1+yO

Income

preferences depend on the part of his wealth that is not at risk. Therefore, Friedman's assignment of people to jobs really follows endowed wealth(yo),justas in our example. However, our framework offerstwo improvements.First,the problem of incentivesis directly incorporated into the formulationof the optimum policy. Second, workersin this model are risk averse for all values of incomes, but even so gambles can be the optimal policy.
Error Structure

Relative costs of measurementare stillimportantin choosing among incentiveschemes,but the error structure plays additional roles when workersare riskaverse. Suppose the output estimatorforworkeri in activity T is qi= qir + pT + PiT, where viT is random error and p, is an i- but common to all workerswithin errorthatis specificto activity that activity.In the piece rate the common error p adds noise which risk-averseworkersdislike, while the common noise drops out of a rank-ordercomparison because it affectsboth contestantssimilarly. That is, the relevantvariance for the contestis 2o-2, while thatfor the piece rate is o-2 + o-2. It is evident thatthiscan tip the balance in favor of tournamentsif o-2 is large enough and/or workersare sufficiently risk averse. The common error p bears two interesting interpretations. One is activity-specific measurement error. For example, j and k may have the same supervisorwhose biased assessmentsaffectall workerssimilarly.This is similarto monitoringall workersby a mechanical count-

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

RANK-ORDER

TOURNAMENTS

857

ing device that mightrun too fastor too slow in any given trial.The other interpretation of p is true random variation that affectsthe enterpriseas a whole. For example, suppose all firmsproduce with the same technology, but that in a given period some firms do better or worse than others. Then risk-averseworkers prefer not to have their incomes vary with conditions facing the firmas a whole, and wages based on a contesteliminatethis kind of variation.Withoutits eliminationthere would be excessive losses due to moral hazard. It must be pointed out that,in the absence of measurementerror, using a contestagainst a fixed standard q discussed above has lower variance than playing against an opponent. As shown in Section II, the relevant variance in a contest is that of ( = Ek - Ej, which has variance 20-2 against an opponent and only o-2 against a standard (since the standard is invariant, Ek 0). Consequently, we might expect risk-averseworkers to prefer absolute standards."1Again, however,the crucial issue is the costs of measurementand the error structure.For the complex attributesrequired for managerial positions,it is difficult to observe output and thereforedifficult to compare to an absolute standard. Insofar as samples and testsare necessary,it bears repeatingthatthese are inherently ordinal in nature. But thisleads us back to the problem of common error,where it is often impossible to know whethera person's output is satisfactory without comparisons to other persons. Further, when there are changing in the firmas a whole, it is difficult productioncircumstances to know whetherthe person failed to meet the standard because of insufficient investment or because the firmwas generallyexperiencingbad times, a problem of measuring "value added." Risk-averseworkersincrease utility by competingagainst an opponent and eliminating thiskind of firmeffect. IV. Heterogeneous Contestants

but ratherseem to Workersare not sprinkledrandomlyamong firms be sorted by abilitylevels. One explanation for this has to do with complementaritiesin production. But even in the absence of complementarities,sorting may be an integral part of optimal laborcontract arrangements. Informational considerations imply that
11 Playingagainst a standard is like Mirrlees's(1976) notion of an "instruction." It is clear thatusingstandardsas well as piece ratesmustbe superiorto using one alone. That to be paid I ifq < q and lo + rq forq -q. This is important schemewould allow workers because it truncatesthe possibilitieswhen Vq < 0. Given the technology,it is possible thatverylarge negativevalues of output can occur, and since it is impossibleto always tax workersthe fullextentof thisloss, some formof truncation is desirable. A contestis an alternativeway to control the tails of this distribution.

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

858

JOURNAL

OF POLITICAL

ECONOMY

compensation methods may affectthe allocation of worker types to firms. Therefore, thissectionreturnsto the case of riskneutrality and analyzes tournamentstructureswhen investment costs differamong persons. Two typesof persons are assumed, a's and b's, withmarginal < CQ4) forall costsof thea's being smallerthan those of the b's: CQ(pu) of disturbancesf(E) is assumed to be the same for pt.The distribution both groups. Many of the following results continue to hold, with usually obvious modification of the arguments,if the a's and b's draw from differentdistributions.The following section addresses the question of self-selection when workersknow theiridentities but firms do not. The next section discusses handicapping schemes when all cost-function differencescan be observed by all parties. Adverse Selection Suppose that each person knows to which class he belongs but that thisinformation is not available to anyone else. The principalresultis that the a's and b's do not self-sort into theirown "leagues." Instead, all workersprefer to work in firmswith the best workers(the major leagues). Furthermore,there is no pure price-rationing mechanism that induces Pareto optimal self-selection. But mixed play is inefficient because it cannot sustain the proper investment strategies. Therefore, tournamentstructuresnaturallyrequire credentials and other nonprice signals to differentiate people and assign them to the appropriate contest. Firms select their employees based on such informationas past performances, and some are not permittedto compete. The proof of adverse selection consistsof two parts. Firstwe show that players do not self-sort into a leagues and b leagues. Second, we show that the resultingmixed leagues are inefficient. 1. Playersdo notself-sort. -Assume leagues are separated and consider the expected revenue Ri generated by playingin league i = a, b with an arbitrary investmentlevel g. Then (26) Ri(pu)= Wi + (WI,- W)Pi, i = a, b, where (Wi ,W12) is the prize money,and Pi is the probability of winning in league i. Recall thatPi depends on the individual's level of investment and thatof his rivals.Therefore,pa = G (IL -E*) and pb = G (a ,a*), where ua* is the existingplayers' investments in the a league, where V = C'(/u4*), and similarly forpu. Recalling from(6) and (9) that = V/g(O)and from (10) that W' = VI* - V12g(0),equation I(26) becomes Ri(,) = Vg* V0)[1/2-G (
-

g*)].

(27)

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

RANK-ORDER

TOURNAMENTS

859

Note thatRi(p) = VAlu when , = and thatdRi/dpu R2'(Q) = Vg(p. = V and C'(,u*) = V, thentz* < a* so that - u*)/g(O) > 0. Since CQ(*) > Ri(Ipc). Furthermore, = R'(pi). Therefore, Rj(1_0) Rb[1, -(qa* -(l)] V fork = p* and Rb(pu) is a pure displacementofRa4L). Since R=*) Ri'(t) < V elsewhere,and sinceRi(,) is increasing,the revenue functions never cross. So Rb(,U) lies to the southwestof Ra(pu)(see fig. 2). Therefore,independent of cost curves,it is alwaysbetterto play in the a league than the b league: Workers will not self-select. are inefficient. 2. Mixedcontests -Suppose the proportionsof a's and If pairingsamong b's in the population are a and (1 - a), respectively. of a player of typei is a's and b's are random, then expected utility
W2 + [aPr + (1 - a)P'](W1 - W2)
-

Qbi),

where (W1,W2) is the prize money in mixed play and Pj7is the probability thata player of typei defeats a player of typeJ.The first-order condition for investmentof type i in this game is
a

+ a+(l-a) a

a)

- ( WI ] *

W2)

Ci'

1i)

A development similar to Section II implies equilibrium reaction functions


[ag(0) + (1
-

a)g(ia

-,b)](WI

W2) Ca(ja)

Rj(p)

Vfl

ib

2 FIG. 2

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

86o for a's and


[ag (jib -')

JOURNAL

OF POLITICAL

ECONOMY

+ (1 - a)g(O)](Wi

W2)

Cb'Vb) =

for b's. If the solution is efficient, then Cb'(Qb) = V implies


ag(O) + (1 - a)g(ji;a
-/b)

Ca(,ua), which

ag(~i2b -a)

+ (1 -a)g(O).

Since g is symmetric and nonuniform, thisconditioncan hold onlyifa = 1/2.Therefore, except in thatveryspecial case, mixed contestsyield inefficient investment:One type of player overinvestsand the other underinvestsdepending upon whether or not a ] l2. We conclude that a pure price systemcannot sustain an efficient competitiveequilibrium in the presence of population heterogeneity withasymmetric information. Marketscan be separated, but only at a cost. Consider, for example, the case where a's want to prevent b's from contaminatingtheir league. By making the spread, W7 - V2, sufficiently large, Ra(4) becomes steeper than Rb(4) in figure2 and crosses it so that the envelope covers Rb(Q) at low values of p and Ra(IL) at high values. Then, for some high levels of [L, it is more profitable to play in the a league and, forlow levels of p., the b league is preferable.Individuals may self-sort, but the cost is thata's overinvest. The result is akin to that of Akerlof (1976) and to those of Spence (1973), Riley (1975), Rothschildand Stiglitz(1976), and Wilson (1977). As they show, a separating equilibrium need not exist, but, even if it does, thatequilibriummay be inferiorto a nonseparatThe obvious practical resolution of these difficulties is the use of nonprice rationing and certification to sort people into the appropriate leagues based on past performance.Similarly, firmsuse nonprice factorsto allocatejobs among applicants. The rules for allocating thosejobs may be importantfor at least two reasons thatwe can describe here. only briefly First,sortingworkersof different skilllevels into appropriate positionswithina hierarchy maybe beneficial.In thispaper, productionis additive, so it does not matterwho works withwhom. To the extent that the production technologyis somewhat more complicated, sorting may well be crucial. A series of pairwise,sequential contestsmay efficiently perform that function. Suppose that qit = pi + 8i + qit, where 6i is an unobserved ability component forplayeri and q is white noise. Suppose it is efficient forthe individual withthe highest8 to be the chief executive. There will be a tendency to have winners play winners because
E(8j I qjl
> qkl)

ing equilibrium.

> E(5k

qjil >

qkl)

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

RANK-ORDER

TOURNAMENTS

86i

in the firstround. A sequential elimination tournament may be a way to select the best person. cost-efficient Second, workersmay not know preciselytheirown abilitiesor cost functions.A worker who is ignorant about his cost functionvalues information before selecting a level of investment expenditure. Therefore, firmsmay offer "tryouts"to provide informationabout strategies.In fact,one can imagine the existence optimal investment of firms which specialize in running contests among young to be used workers-the minor leagues-which provide information when and if the workersopt to increase the stakesand enter a bigger league. These issues point up an importantdifferencebetween piece rates and contests. In the pure heterogeneous case, where informationis and workersare riskneutral,a piece rate alwaysyieldsan asymmetric However, once slotefficient solution, namely,V = CA( = G(b). in productingof workersis importantbecause of complementarities about their tion, or if it is desirable for workersto gain information type,it is no longer obvious that a series of sequential contestsdoes not result in a superior allocation of resources. Handicap Systems This sectionmoves to the opposite extremeof the previous discussion and assumes that the identitiesof each type of player are known to mixed contests. everyone. Competitivehandicaps yield efficient Consider again two types a and b now known to everyone. Prize (11) and (12) are structuresin a-a and b-b tournaments satisfying but those conditions are not optimal in mixed a-b play. efficient, by /ua*and /4*,their Denote the sociallyoptimal levels of investment difference by AA, and the prizes in a mixed league by W1and W2.Let h be the handicap awarded to the inferiorplayer b. Then the Nash solution in the a-b tournamentsatisfies
-~a
-ab

h)AW = QA.)
h)i\W=

(28)

and
g
(-a

lib-

CG(b)-

of g[~I) Since (The second condition in [28] followsfrom symmetry = criterionis V C'(,u*) = C'(,u*), independent investment the efficient of pairings, the optimum spread in a mixed match must be
= V/g(A/,u - h). A JW

(29)

From (28), condition (29) insures the proper investmentsby both contestants. The spread is larger in mixed than pure contestsunless a

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

862

JOURNAL

OF POLITICAL

ECONOMY

gives b the full handicap h = a* - ib*. Otherwise, the appropriate spread is a decreasing functionof h. Prizes W, and W2mustalso satisfy the zero-profit constraint W, + W2 = V * (qu* + g4) independent of h since the spread is alwaysadjusted to induce investments and /1t*. gua* The gain to an a from playing a b with handicap h, rather than another a with no handicap, is the differencein expected prizes: ya(h)= PW1 + (1
-P)W2
-

C(La*)

-2[(Wy + Wa)/2 - Ct(a*)]

(30)

= pW1 + (1 -P)1/V2

- (Wal + Wa2)/2,

thata whereYa(h) is the gain to a and P = G (A - h) is the probability wins the mixed match. The corresponding expression for b is
Yb(h) =

(1 - P)W1 + PW2 - (Wbl+ Wb2)/2.

(31)

The zero-profit constraintsin a-a, a-b, and b-b require that ya(h) + yb(h)= 0 forall admissibleh. The gain of playingmixed matchesto a is completelyoffsetby the loss to b and vice versa. If Ca(11) is not greatlydifferent fromCb(U), then LA.t = /4*- lit is - h). This approximation and small and P l/2 + [g(4g - h)](A constraintreduce (30) to the zero-profit
Ya(h) V . / -) h

(32)

The expression foryb(h)is the same, except its sign is reversed,so the gain to a decreases in h, and the gain to b increases in h. Therefore,h* = gI/2is the competitivehandicap, since it impliesya(h*) = yb(h*)= 0. If the actual handicap is less than h*, then Ya is positive and a's preferto play in mixed contestsratherthan withtheirown type,while b's prefer to play withb's only. The opposite is true if h > h*. A two-playergame is said to be fair when the players are handicapped to equalize the medians. The competitivehandicap does not result in a fair game, since h* = LAtI2 < Ag. The a's are given a competitive edge in equilibrium,because theycontributemore to total output in mixed matches than the b's do. This same result holds if Ea variance than Eb, but it may be sensitiveto the assumphas a different tion of statistical independence and output additivity. h can be constrainedto be zero. In thiscase, different Alternatively, wage schedules would clear the market. Since Ya(O) = -Yb(O) 3, to b's, paying W1 - 3 and W2 - 3 to a's, while paying W1 + /3,W2 + /3 the investments unaltered. It is easy leaves the spread and, therefore, between mixed and pure to verify thata's and b's are stillindifferent contests,because expected returnsare equal between segregated and integratedcontests for each type of player. With no handicaps, the market-clearing prizes available to a's in the mixed contestare lower

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

RANK-ORDER

TOURNAMENTS

863

than those faced byb's. Still,expected wages are higherfora's than b's in the mixed contest,because their probabilityof winningis larger. The b's are given a superior schedule in the mixed contest as an equalizing differencefor having to compete against superior opponents. This yields the surprisingconclusion that reverse discrimination, where the less able are given a head start or rewarded more lucrativelyif they happen to accomplish the unlikely and win the contest, can be consistentwith efficientincentive mechanisms and mightbe observed in a competitivelabor market. V. Summary and Conclusions

This paper analyzes an alternativeto compensationbased on the level of individual output. Under certain conditions,a scheme which rewards rank yields an allocation of resources identical to that generated by the efficient piece rate. Compensatingworkerson the basis of their relative position in the firmcan produce the same incentive workers as does the optimal piece rate. It structurefor risk-neutral might be less costly,however, to observe relative position than to measure the level of each worker's output directly.This results in paying salaries which resemble prizes: wages which differ from realized marginal products. When riskaversionis introduced,the prize salaryscheme no longer duplicates the allocation of resources induced by the optimal piece rate. Depending on the utilityfunctionand on the amount of luck involved, one scheme is preferred to the other. An advantage of a contestis thatit eliminatesincome variationwhichis caused by factors common to workersof a given firm. Finally,we allow workers to be heterogeneous. This complication adds an importantresult: Competitivecontestsdo not automatically sort workers in ways that yield an efficient allocation of resources when informationis asymmetric.In particular,low-qualityworkers attemptto contaminatefirms composed of high-quality workers,even if there are no complementarities in production. Contaminationresults in a general breakdown of the efficient solution if low-quality workers are not prevented from entering. However, when player types are known to all, there exists a competitive handicapping schemewhichallows all typesto workefficiently within the same firm.
References Akerlof,George. "The Economics of Caste and of the Rat Race and Other Woeful Tales." Q.J.E. 90 (November 1976): 599-617. Alchian, Armen A., and Demsetz, Harold. "Production, InformationCosts, and Economic Organization." A.E.R. 62 (December 1972): 777-95.

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

864

JOURNAL

OF POLITICAL

ECONOMY

Becker, Gary S., and Stigler,George J. "Law Enforcement,Malfeasance, and Compensation of Enforcers."J. Legal Studies3, no. 1 (1974): 1-18. Cheung, Steven N. S. The Theory to of Share Tenancy:WithSpecial Application Asian Agriculture and theFirstPhase of Taiwan Land Reform. Chicago: Univ. Chicago Press, 1969. David, Herbert A. TheMethod ofPaired Comparisons. London: Charles Griffin, 1963. Friedman, Milton. "Choice, Chance, and the Personal Distributionof Income."J.P.E. 61, no. 4 (August 1953): 277-90. Galton, Francis. "The Most Suitable Proportion between the Values of First and Second Prizes." Biometrika 1 (1901-2): 385-90. and Ordering Gibbons, Jean D.; Olkin, Ingram; and Sobel, Milton. Selecting Populations: A New Statistical Methodology. New York: Wiley, 1977. Harris, Milton,and Raviv,Artur."Some Resultson IncentiveContractswith Applications to Education and Employment,Health Insurance, and Law Enforcement."A.E.R. 68 (March 1978): 20-30. Lazear, Edward P. "Why Is There Mandatory Retirement?"J.P.E. 87, no. 6 (December 1979): 1261-84. Mirrlees,James A. "The Optimal Structure of Incentives and Authority withinan Organization." BellJ. Econ. 7 (Spring 1976): 105-31. Riley,John G. "Competitive Signalling."J. Econ. Theory10 (April 1975): 174-86. Ross, Stephen A. "The Economic Theory of Agency: The Principal's Problem." A.E.R. Papers and Proc. 63 (May 1973): 134-39. Rothschild, Michael, and Stiglitz,Joseph E. "Equilibrium in Competitive Insurance Markets:An Essay on the Economics of ImperfectInformation." Q.J.E. 90 (November 1976): 630-49. Spence, A. Michael. "Job Market Signaling." Q.J.E. 87 (August 1973): 355-74. and the Schemapiric View." Science Stevens, S. S. "Measurement, Statistics, 161 (August 30, 1968): 849-56. Stiglitz, Joseph E. "Incentives,Riskand Information:Notes towardsa Theory of Hierarchy."Bell J.Econ. and Management Sci. 6 (Autumn 1975): 552-79. Wilson, Charles A. "A Model of Insurance Markets with Incomplete Information."J. Econ. Theory16 (December 1977): 167-207.

This content downloaded from 128.91.116.8 on Mon, 16 Sep 2013 11:04:40 AM All use subject to JSTOR Terms and Conditions

Похожие интересы