<<

ARE CEOS REWARDED FORLUCK?THEONES WITHOUTPRINCIPALS ARE*

MARIANNE BERTRAND AND

Thecontracting viewof CEO pay assumesthat pay isusedby shareholdersto solvean agencyproblem. Simple models of thecontracting viewpredict thatpay shouldnot be tied to luck, whereluck isdeŽ ned as observable shocks to perfor- mancebeyond the CEO’ s control.Using several measures of luck, weŽ ndthat CEO pay infact respondsas much to a lucky dollaras to a generaldollar. A skimmingmodel, where the CEO hascaptured the pay-setting process, is consis- tentwith this fact. Becausesome complications to the contracting viewcould also generatepay forluck, wetest for skimming directly by examining the effect of governance.Consistent with skimming, we Ž ndthat better governed Ž rmspay theirCEO lessfor luck.

I. INTRODUCTION CEO pay is usually viewedthrough the lens of principal agentmodels. Under this contractingview, pay is usedto reduce themoral hazard problemthat arisesbecause CEOs oftenown verylittle of theŽ rmsthey control. Shareholders (perhaps acting throughthe board orthe compensation committee) optimally design thepay packagein orderto increase the CEO’ s incentiveto maximizeŽ rmvalue. 1 Simplemodels of the contracting view generateone important prediction. Shareholders will notreward CEOs for observable luck.By luck,we mean changes in Žrm performancethat arebeyond the CEO’ s control.Tying pay to

*Theresults in this paper were previously circulated as part ofa larger working paperentitled “ DoCEOs SetTheir Own Pay? TheOnes Without Princi- palsDo.” We areextremely grateful to , Rajesh Aggarwal, George Baker,Patrick Bolton,Peter Diamond, Robert Gibbons, Denis Gromb, Brian Hall, BengtHolmstrom, Caroline Hoxby, Glenn Hubbard, Lawrence Katz, Jo¨rn-Steffen Pischke,Nancy Rose,David Scharfstein, Robert Shimer, , ,and seminar participants at the University of California at Berkeley, ColumbiaUniversity, the , , the Mas- sachusettsInstitute of Technology, , and the National Bu- reauof Economic Research Corporate Finance Summer Institute 1999 for very helpfulcomments. We thank Kenneth Ayotte and Michael Mitton for excellent researchassistance, Michael Haid for giving us access to his data set of oil companies,and David Yermack for giving us access to his data on executive compensation.Financial support was provided by the Russell Sage Foundation, thePrinceton Industrial Relations Section, and the Princeton Center for Economic PolicyStudies. e-mail: [email protected]; [email protected]. 1.Murphy [1985,1986] is a forerunnerof thevast empirical literature on the contracting view.Murphy [1999]and Abowd and Kaplan [1999] summarize the CEO pay literature.Formal tests of thecontracting viewcan be found in Gibbons andMurphy [1990,1992], Garen [1994], Hubbard and Palia [1994], Bertrand and Mullainathan[1999], and Aggarwal andSamwick [1999a,1999b].

© 2001by thePresident and Fellowsof Harvard Collegeand theMassachusetts Institute of Technology. The Quarterly Journal ofEconomics, August 2001

901 902 QUARTERLY JOURNALOF luck,therefore, cannot provide better incentives and will only makethe contract riskier [Holmstrom 1979]. 2 Thispaper starts by examiningwhether or not CEOs arein fact paid forluck using threemeasures of luck. 3 First,we perform acasestudy ofthe oil industry wherelarge movements in oil pricestend toaffect Žrmperformanceon aregularbasis. Second, weuse changes in industry-speciŽc exchangerate for Ž rmsin the traded goodssector. Third, we use year-to-year differences in meanindustry performanceto proxy for the overall economic fortuneof asector. 4 Forall threemeasures, we Ž nd that CEO pay respondssigniŽ cantly toluck. In fact, weŽ nd that CEO pay is as sensitiveto a luckydollar as toa generaldollar. Moreover, these resultshold as wellfor discretionary components of pay— salary and bonus—as theydo foroptions grants. Theseresults are inconsistent with a simple contracting view.Motivated by practitionerssuch as Crystal [1991], wepro- posean alternative,skimming, which can explain theseresults [Bertrand and Mullainathan 2000a]. Theskimming view also beginswith theseparation of ownership and control,but itargues that this separationallows CEOs togain effectivecontrol of the pay-setting processitself. Both because of entrenchment,such as packing theboard with supporters,and becauseof thecomplexity ofthe pay process,many CEOs defacto set their own pay with littleoversight from shareholders. Their pay levelthen becomes constrainedby an unwillingnessto draw shareholders’attention. Payfor performance arises in theskimming view because good performancemay ease these constraints, in essencecreating

2.Note our emphasis on observable luck. Inany model, given the random- nessof theworld, CEOs (andalmost everybody else) will end up being rewarded forunobservable luck. Notealso our emphasis on the fact thatthis prediction holdsin simple agency models. As we will discuss shortly, complications to the agencymodel can in principlealter this result. 3.Blanchard, Lopez-de-Silanes, and Shleifer [1994] present suggestive evi- denceon pay forluck byshowing that windfall gains from court rulingsraise the payof CEOs. Itis only suggestive since court rulingsmay not be luck butrather aresultof the CEO’ s work. Inother domains, Shea [1999] independently performs anexercise similar to ours for baseball players. 4.This last test very muchresembles the approach followed in the relative performanceevaluation (RPE) literature[Gibbons and Murphy 1990;Janakira- man,Lambert, and Larcker 1992;Aggarwal andSamwick 1999a].Problems can arisewith RPE asa specialcase of luck. Filteringthis speciŽ c kindof luck maynot beoptimal from an agency theoretical point of view. As Gibbons and Murphy [1990]note, relative performance evaluation can distort CEO incentivesif they can“ takeactions that affect the average output of thereference group.” Aggarwal andSamwick [1999b]develop a formalmodel along these lines. By usingother shocksto performance that are even more objectively beyond managerial in u- ence,we circumvent these problems. ARECEOS REWARDED FORLUCK? 903 slackfor the CEO. In otherwords, when the Ž rmis doing well, shareholdersare less likely to notice a largepay package.To the extentthat luckydollars createslack as readily as generaldol- lars,pay forluck arises. Finding pay forluck, however, does not necessarily single out theskimming model. Complications to the agency model can makeit suchthat paying forluck is in fact optimal.For example, supposethat thevalue of a CEO’s humancapital risesand falls with industry fortunes.One would thenŽ nd that pay correlates with luckbecause the CEO’ s outsidewage moves with luck. Anotherpossibility is that boards maytie pay toluck in orderto motivateCEOs toforecast or respond to luck shocks. Subsection II.D discusseswhether arguments such as thesecan really ex- plain thepay forluck relationship. Tofurther differentiate skimming from these explanations, weempirically examine a directimplication of the skimming model.Skimming should beless prevalent in bettergoverned Žrms.Well-governed Ž rms,such as thosewith alargeshare- holderpresent on theboard, limit the CEO’ s ability tocapturethe pay process.We test this hypothesisusing severalmeasures of governance:presence of large shareholders (on the board and overall),CEO tenure(interacted with thepresence of largeshare- holdersto better proxy for entrenchment), board size,and frac- tionof directors that areinsiders. Consistent with skimming,we generallyŽ nd that thebetter governed Ž rmspay lessfor luck. 5 Theseeffects are strongest for the presence of large shareholders onthe board. An additional largeshareholder on the board re- ducespay forluck by between23 and 33 percent.Large share- holdersare especially important as CEO tenureincreases, con- sistentwith theidea that uncheckedCEOs canentrench themselvesover time. If pay forluck were optimal, we would have expectedwell-governed Ž rmsto pay forluck as muchas (if not morethan) poorlygoverned Ž rms.For example, whether or nota largeshareholder is present,the CEO would haveto be rewarded fora risein thevalue of his humancapital. TheseŽ ndings suggestthat at leastsome of the pay forluck in poorlygoverned Žrmsis dueto skimming by CEOs.

5.Whenever we refer to “ lesspay forluck” we mean that there is lesspay for luck relativeto the amount of pay for performance. Thus, these results would not bedriven by well-governedŽ rmssimply giving less overall pay forperformance. Infact, weŽ ndthat governance correlates very littlewith pay forperformance, onlywith pay forluck. 904 QUARTERLY JOURNALOF ECONOMICS

II. PAY FOR LUCK TEST

II.A.Theoretical Background Asimpletheoretical model will makemore precise what agencytheory says aboutthe reward for observable luck. Con- sidera standard agencysetup whererisk-neutral shareholders tryto induce a risk-aversetop managerto maximizeŽ rmperfor- mance.Since the actions of the CEO canbe hard toobserve, shareholderswill beunable to sign acontractthat speciŽes these actions.Instead, shareholderswill offera contractto the CEO whereher compensation level is madeto depend onthe Ž rm’s performance.Let p representŽ rmperformance and a the CEO’s actions,which by assumptionare unobservable to the sharehold- ers.Firm performance depends ontheactions of theCEO and on randomfactors. We split therandom factors into two components: thosethat can beobserved by shareholdersand thosethat cannot. Foran oilŽ rm,the price of crude oil would bean observable randomfactor. Letting o bethe observable factor and u be the unobservablenoise term, we assume that performancecan be written as p = a + d o + u. Undersome technical conditions (CARA utility and Brown- ian motionfor the performance process), Holmstrom and Milgrom [1987] calculatethe optimal incentivescheme for this model.Let s denotethis incentivescheme. Since shareholders can only ob- servetwo variables, p and o,theincentive scheme could at most depend onthese two variables. In fact, shareholderswill only rewardCEOs forperformance netof the observable factor: (1) s 5 a 1 b ( p 2 d o) 5 a 1 b (a 1 u). In otherwords, the optimal incentive scheme Žlters the ob- servableluck from performance. This is becauseleaving o in the incentivescheme provides no added beneŽt tothe principal as,by deŽnition, the agent has nocontrol over o.Motivating heron o has noincentive effects. Beyond providing nobeneŽt, tying pay to luckactually coststhe principal becausethe variance of the incentivescheme is higher,and theprincipal mustincrease mean pay tocompensate the risk-averse CEO. In practice,explicit incentivecontracts, such as options, rarelyŽ lter.For example, options are rarely if everindexed against marketperformance. This need not be inconsistent with alackof Ž ltering,however. It maybe that thediscretionary componentsof pay, suchas salary and bonus,are the ones used to ARECEOS REWARDED FORLUCK? 905

Žlter.In theory,these other components could adjust enoughto undothe effect of the options value  uctuating with luck.Such adjustmentmay happen if aboard wereto monitor luck and alter eachyear’ s salary,bonus, and numberof newoptions granted so that theCEO’ s overallpay packageremains free of luck. II.B. Empirical Methodology Withinthe agency framework, most of the empirical litera- tureon CEO pay estimatesan equationof the form,

(2) yit 5 b p perfit 1 g i 1 x t 1 a X p Xit 1 e it, where yit is totalCEO compensationin Žrm i at time t, perfit is aperformancemeasure, g i areŽ rmŽxedeffects, x t aretime Ž xed effects,and Xit areŽ rm-and CEO-speciŽc variables suchas Žrm sizeand tenure.The coefŽ cient b capturesthe strength of thepay forperformance relationship. Performanceis typically measuredeither as changesin ac- countingreturns or stock market returns, and wewill useboth 6 measures. In measuringcompensation yit,muchof theliterature focuseson the  owof new compensation. Ideally, thecompensa- tionin agivenyear would alsoinclude changes in thevalue of unexercisedoptions granted in previousyears [Hall and Liebman 1998]. Sucha calculationrequires data ontheaccumulated stock of options held by theCEO eachyear, whereas existing data sets, including ours,contain only information on newoptions granted eachyear. Consequently, our compensation measure excludes this componentof the change in wealth.For our purposes, how- ever,this exclusiondoes not pose much of aproblem.The change in wealthdue to changing option values is mechanicallytied to lucksince options are not indexed. Thus, even if thesedata were available, focusingon the subjective components of pay would still bea natural strategy.We discuss this issueat greaterlength in subsectionII.D. Toestimatethe general sensitivity of pay toperformance,we will followthe literature and estimateequation (2) using astan- dard Ordinary LeastSquares (OLS) model.To estimate the sen- sitivity ofpay toluck, we need to use a two-stageprocedure. In theŽ rststage, we will predict performanceusing luckin orderto isolatechanges in performancethat arecaused by luck.In the

6.These are  owmeasures. In practice,given the Ž rm Žxedeffects, we will usemarket value and level of accountingreturns as measuresof perf i t. 906 QUARTERLY JOURNALOF ECONOMICS secondstage, we will seehow sensitive pay is tothese predictable changesin performance.This two-stage procedure is essentially an InstrumentalVariables (IV) estimationwhere the luck vari- able is theinstrument for performance. 7 Letting o beluck, the Ž rstequation we estimate is

perfit 5 b p oit 1 gi 1 ct 1 aX p Xit 1 eit, where oit representsthe luck measure (oil price,for example). Fromthis equationwe predict aŽrm’sperformanceusing only ˆ informationabout luck. Call this predicted value perfit. We then askhow pay respondsto these predictable changesin perfor- mancedue to luck: ˆ yit 5 b Luck p perfit 1 g i 1 x t 1 a X p Xit 1 e it.

Theestimated coefŽ cient b L uc k indicates howsensitive pay is to changesin performancethat comefrom luck. Since such changes should beŽ ltered,basic agencytheory predicts that b L u ck should equal 0.

II.C. Oil IndustryStudy Wenow turn to the oil industry as acasestudy ofpay for luck.As FiguresI and IIshow,the price of crude oil has uctu- ated dramatically overthe last 25 years.These large  uctuations havecaused large movements in industry proŽts. Moreover, these large uctuationsin crudeoil prices are likely to have been beyondthe control of a singleAmerican CEO. Forexample, the sharp declinein crudeoil price at theend of 1985 was causedby Saudi Arabia’s decisionto reform its petroleumpolicy and to increaseproduction, an actionhardly attributable (and never attributed) tothe CEOs ofAmericanoil Ž rms.Similarly, the large oilprice increase between 1979 and 1981 is usually attributed to an internalpolicy change by OPEC. Oil pricemovements there- foreprovide an ideal place totest for pay forluck: they affect

7.One might wonder why weshould use this procedure rather than simply include o directlyinto the pay forperformance equation (2) and run OLS to estimate yit 5 b p perf it 1 f p o 1 g i 1 x t 1 a x p Xit 1 e it. Thisequation is hard to interpret, however. Even if thereis nopay forluck, the coefŽcient f willnot equal 2 b but rather 2 b d ,aswecan see from equation (1). Sincewe do not estimate d ,theestimated coefŽ cient f canbe smalleither because thereis pay forluck orsimply because d issmall. The Ž rst equationin the IV procedurecircumvents this problem by scalingthe effect of luck onperformance. ARECEOS REWARDED FORLUCK? 907

FIGURE I RealPrice of aBarrelof Crude Oil performance,are measurable, and areplausibly beyondthe con- trolof the CEOs. Weuse a data seton the pay and performancefor the 51 largestAmerican oil companies between 1977 and 1994 toimple-

FIGURE II MeanRate of Accounting Return 908 QUARTERLY JOURNALOF ECONOMICS

FIGURE III OilIndustry CEO Payand Crude Oil Price mentthe methodology of theprevious section. 8 Beforemoving to regressionanalysis, it is usefulto look directly at howpay uc- tuatescompared with themovements in FiguresI and II. In FigureIII wehave graphed changesin oilprices for each year and changesin meanlog pay in theindustry. Two striking facts emerge.First, pay changesand oilprice changes correlate quite well.In twelveof theseventeen years they are of thesame sign: bothare up, orbothare down. This is suggestiveof pay forluck. Second,the remaining Ž veyearswhere pay and oilprices move in oppositedirections are all yearsin whichthe oil price drops. This hints at an asymmetry:while CEOs arealways rewardedfor good luck,they may not always bepunished forbad luck. Thisgraphical analysis doesnot quantify pay forluck. One cannotcompare it with thesize of the overall pay forperfor- mance.It alsodoes not control for other Ž rm-speciŽc variables that mightbe changing over time. Table I followsthe empirical

8.We areextremely grateful to Michael Haid for making the data set used in thissection available to us. See Haid [1997] for further detailsabout the construc- tionof the data set. Appendix 1 providesmean and standard deviations for the mainvariables of interest.While the original data set covers 51 companies over theperiod 1977 to 1994, information on CEO pay isavailablefor only 50 of these original51 companies. Moreover, CEO pay isavailablefrom 1977 on foronly 34 ofthese 50 companies. The Ž naldata set we use covers 827 company/ year observations. ARECEOS REWARDED FORLUCK? 909

TABLE I PAY FOR LUCK FOR OIL CEOS (LUCK MEASUREIS LOG PRICE OF CRUDE OIL) DEPENDENT VARIABLE: ln (TOTAL COMPENSATION )

SpeciŽcation: General Luck General Luck (1) (2) (3) (4)

Accountingrate .82 2.15 — — of return (.16) (1.04) Log(shareholder — — .38 .35 w ealth) (.03) (.17) Age .05 .07 .05 .05 (.02) (.03) (.02) (.02) Age2 p 100 2 .04 2 .05 2 .04 2 .04 (.02) (.02) (.02) (.02) Tenure .01 .01 .01 .01 (.01) (.01) (.01) (.01) Tenure2 p 100 2 .03 2 .03 2 .03 2 .03 (.02) (.01) (.02) (.02) FirmŽ xedeffects Yes Yes Yes Yes Yearquadratic Yes Yes Yes Yes Sample size 827 827 827 827 Adjusted R2 .70 — .75 —

a. Dependentvariable is thelogarithm of total compensation.Performance measure is accountingrate of returnin columns (1) and (2) and the logarithm of shareholder wealth in columns (3) and (4). All nominal variablesare expressed in 1977 dollars. b. Summarystatistics forthe sample of oil Ž rms areavailable in Appendix 1. c.The luck regression (columns (2) and (4)) instrument for performance with the logarithm of the price ofa barrelof crude oil in that year,expressed in 1977 dollars. d.Each regression includes Ž rmŽ xedeffects and a quadraticin year. e.Standard errors are in parentheses. methodologypresented in subsectionII.B, whichpermits a more systematicanalysis. All regressionsuse log (total compensation) as dependentvariable and includeŽ rmŽ xedeffects, age and tenurequadratics, and aperformancemeasure as dependent variables.9 Wealso include a yearquadratic toallow forthe fact that CEOpay has beentrending up during this period.Column (1) estimatesthe sensitivity of pay toa generalchange in account- ing performance.The coefŽ cient of .82 suggeststhat if an oilŽ rm increasesits accountingreturn by onepercentage point, total compensationrises by .82 p .01 = .0082 logpoint. Roughly, aonepercentage point increasein accountingreturns leads toa .8percent increase in pay. Notethat thesign and magnitudeof all theother covariates in theregression seem sensible. Pay in-

9.Total compensation in this table and all other tables includes salary, bonus,other incentive payments, and value of options granted inthat year. 910 QUARTERLY JOURNALOF ECONOMICS creaseswith ageand toa lesserextent with tenure.Both the age and tenureproŽ le are concave (the negative coefŽ cient on the quadratic term). Column(2) estimatesthe sensitivity of pay toluck. As de- scribedabove, we instrument for performance with thelog of oil price.10 ThecoefŽ cient in column(2) nowrises to 2.15. This suggeststhat aonepercentage point risein accountingreturns due to luck raisespay by 2.15 percent.Given the large standard errors,one cannot reject that thepay forluck coefŽ cient and pay forgeneral performance coefŽ cient are the same. One can, how- ever,strictly reject the hypothesis of completeŽ ltering:oil CEOs arepaid forluck that comesfrom oil price movements. Columns(3) and (4) performthe same exercise for a market measureof performance, shareholder wealth. The coefŽ cient of .38 oncolumn (3) suggeststhat a1percentincrease in share- holderwealth leads toroughly a .38 percentincrease in CEO pay. In column(4) weŽ nd that a1percentincrease in shareholder wealthdue to luck leads to.35 percentincrease in CEOpay. Again, pay forluck matches pay forgeneral performance.

II.D.More General Tests Theoil industry casestudy, whileinstructive, raises the questionof howgeneralizable these results are. In this subsection wewill examineluck shocks that affect abroaderset of Žrms.We focuson twomeasures of luck:movements in exchangerates and meanindustry performance.By affecting theextent of import penetrationand henceforeign competition, exchange rate move- mentscan strongly affect aŽrm’sproŽtability. 11 Movementsin meanindustry performancealso proxy for luck to the extent that aCEOdoesnot in uence how the rest of her industry performs. As wementioned before, this last instrumentis morequestion- able.In practice,however, we Ž nd that meanindustry move- mentsoperate exactly like exchange rate or oilprice movements. Toimplement these tests, we use compensation data on792 largecorporations over the 1984 –1991 period.The data setwas graciouslymade available tous by David Yermackand Andrei Shleifer.It is extensivelydescribed in Yermack[1995]. Compen-

10.This table does not report the Ž rst-stageregressions of performanceon oil price.But asone would expect from Figures I andII, theseregressions show very signiŽcant coefŽcients on oil price ( p < .001). 11.Revenga [1992] uses exchange rates as an instrument for import pene- tration.Bertrand [1999]shows its effects on Ž rm proŽtability. ARECEOS REWARDED FORLUCK? 911 sationdata werecollected from the corporations’ SEC Proxy, 10-K, and 8-K Žllings.Other data weretranscribed from the Forbes magazineannual surveyof CEOcompensationas wellas fromSEC Registrationstatements, Ž rms’Annual Reports,direct correspondencewith Žrms,press reports of CEO hiresand depar- tures,and stockprices published by Standard &Poor’s. Firms wereselected into the sample on the basis oftheir Forbes rank- ings. Forbes magazinepublishes annual rankingsof the top 500 Žrmson four dimensions: sales, proŽ ts, assets and marketvalue. Toqualify forthe sample, a corporationmust appear in oneof these Forbes 500 rankingsat leastfour times between 1984 and 1991. In addition, thecorporation must have been publicly traded forfour consecutive years between 1984 and 1991. Yermack’s data areattractive in that theyprovide both gov- ernancevariables and informationon options granted, not just informationon optionsexercised. But theydo notinclude changes in thevalue of options held, which we must therefore exclude fromour compensation measure. If anything,this biases usto- ward understanding theamount of pay forluck. Since options are notindexed, changes in thevalue of options held will covary perfectlywith luck.Including thesechanges in thecompensation measurewould onlyincrease the measured pay forluck. This data limitation,therefore, is lessof a concernfor our purposes. TableII presentssummary statistics forthe main variables ofinterestin thefull Yermackdata. 12 All nominalvariables are expressedin 1991 dollars.The average CEO earns$900,000 in salary and bonus.His total compensation is nearlytwice that amountat $1,600,000. Thedifference indicates thelarge fraction ofaCEO’s pay that is dueto options grants. The average CEO is roughly57 yearsold and has beenCEO ofthe Ž rmfor nine years. As far as governancegoes, the average Ž rmin oursample has 1.12 largeshareholders, of whichless than afourthare sitting on theboard. There are on average thirteen directors on a board. Forty-twopercent of themare insiders. 13

12.In practice,depending on therequired regressors, the various tests in the followingsections will be performed on various subsamples of theoriginal data. Noneof these main variables of interest signiŽ cantly differ in any of these subsamples. 13.Technically, we deŽ ne insiders to be both inside and gray directors.An insidedirector is deŽ ned as a directorwho is a current orformer ofŽ cer ofthe company.A gray directoris a relativeof acorporateofŽ cer, orsomeone who has substantialbusiness relationships with the company outside the course of regular business. 912 QUARTERLY JOURNALOF ECONOMICS

TABLE II SUMMARY STATISTICS: FULL YERMACK CEO SAMPLE

Mean S. D.

Age of CEO 57.42 6.84 Tenureof CEO 9.10 8.08 Salaryand bonus 901.69 795.15 ln(Salary and bonus) 6.62 .60 Totalcompensation 1595.85 3488.32 ln(Total compensation) 6.98 .81 Numberof largeshareholders (All) 1.12 1.42 Numberof largeshareholders on board .24 .74 Board size 13.45 4.54 Fractionof insiders on board .42 .19

a. Sampleperiod is 1984–1991. b. All nominalvariables are expressed in thousands of 1991 dollars. c.Column 1 is themean for each variable, while column 2 is thestandard deviation. d.A largeshareholder is deŽned as someonewho owns more than 5 percentof the common shares in thecompany, excluding the CEO. Insiders here denote directors who are current or former ofŽ cers of the company,relatives of corporate ofŽ cers, or anyone who has asubstantial business relationshipwith thecompany. Relationships arising in the normal course of business wouldnot be called an insider. Our insiderdeŽ nition corresponds to insider + grayin the original Yermack data. e.Total compensationequals salary plusbonus plus total valueof options grants plusother compensation.

OurŽ rstgeneral measure of luck focuses on exchange rate movements.We exploitthe fact that exchangerates between the U.S.dollar and othercountry currencies  uctuategreatly over time.We alsoexploit the fact that different industriesare affected by different countries’exchange rates. For example, since the toy industry maybe more affected by Japanese importswhile the lumberindustry maybe more affected by Bolivia,these two industries mayexperience very different shocksin thesame year. Thisallows ustoconstruct industry-speciŽc exchangerate move- mentswhich are arguably beyonda speciŽc CEO’s controlsince theyare primarily determinedby macroeconomicvariables. The exchangerate shock measure is based onthe weighted average of thelog real exchange rates for importing countries by industry. Theweights are the share of eachforeign country’ s importin total industry importsin abaseyear (1981– 1982). Real exchangerates arenominal exchange rates (expressed in foreigncurrency per dollar) multiplied by U.S.CPIand divided by theforeign country CPI. Nominalexchange rates and foreignCPIs arefrom the InternationalFinancial Statistics ofthe International Monetary Fund. ARECEOS REWARDED FORLUCK? 913

TABLE III PAY FOR LUCK

ln (total ln (total Dep. var.: Cashcomp ln (cash) comp) ln (cash) comp) SpeciŽcation: GeneralLuck General Luck General Luck General Luck General Luck

PanelA:LuckMeasureis Exc hange RateSho ck

Income .17.35 — ——————— (.02) (.16) Income/assets ——2.132.94 2.36 4.39 — ——— (.16) (1.28) (.28) (2.17) ln (shareholder ——————.22.32 .31 .57 wealth) (.02) (.13) (.03) (.23) Sample size 17371737 1729 1729 1722 1722 1713 1713 1706 1706 Adjusted R2 .75— .75— .58— .75— .59—

PanelB:LuckMeasureis Mean IndustryPerformance

Income .21.34 — ——————— (.02) (.10) Income/assets ——2.184.02 2.07 4.00 — ——— (.12) (.53) (.21) (.86) ln (shareholder ——————.20.22 .25 .29 wealth) (.01) (.12) (.02) (.19) Sample size 46844684 4648 4648 4624 4624 4608 4608 4584 4584 Adjusted R2 .77— .81— .70— .82— .71—

a. Dependentvariable is thelevel of salary andbonus in columns 1 and2, thelogarithm of salary and bonusin columns 3, 4, 7, and8 andthe logarithm of total compensationin columns 5, 6, 9, and10. Performancemeasure is operatingincome before extraordinary items incolumns 1 and2 (inmillions), operatingincome to total assets incolumns 3 to6 andthe logarithm of shareholderwealth in columns 7 to 10. All nominalvariables are expressed in real dollars. b. Inthe luck regressions in Panel A, theperformance measure is instrumentedwith current and lagged appreciationand depreciation dummies and current and lagged exchange rate index growth. First-stage regressionsare presented in Appendix 2. c.In the luck regressions in Panel B, theperformance measure is instrumentedwith the total assets- weightedaverage performance measure in the Ž rm’s two-digitindustry (the Ž rmitself is excludedfrom the meancalculation). d.Each regression includes Ž rmŽxedeffects, year Ž xedeffect and demographic controls (quadratics in ageand tenure). e.Standard errors are in parentheses.

PanelA ofTable III examinesthis luckmeasure. Note that sincethe exchange rate measure can only be constructed for industrieswhere we have imports data, thesample size is much smallerhere than forour full sample.All regressionscontrol for Žrmand yearŽ xedeffects as wellas forquadratics in tenureand age.14 Column1 usesas dependentvariable thelevel of cash

14.We do not report the coefŽ cients on age and tenure to save space, but theyresemble the oil industry Ž ndings:positive but diminishing effects of tenure and age. 914 QUARTERLY JOURNALOF ECONOMICS compensation.Thus, relative to our standard speciŽcation, we do notrun this regressionin logsand donot include value of options granted.Since proŽ ts arereported in millionsand pay is reported in thousands,the coefŽ cient of .17 in column1 suggeststhat $1000 increasein proŽts leads toa 17 centincrease in perfor- mance.Column 2 performsthe same exercise for pay forluck: we instrumentfor performance using theexchange rate shocks. 15 As in theoil case, we Ž ndapay forluck coefŽ cient that is ofthe same orderof magnitudeas thepay forgeneral performance coefŽ cient. Columns3 through6 runthe more standard regression wherewe usethe logarithm of pay and an accountingmeasure of performance(operating income divided by totalassets). In col- umns3 and 4weuse only cash compensation, while in columns5 and 6weuse total compensation. In bothcases, we Ž nd the sensitivityof pay toluck to be about the same as thesensitivity of pay togeneral performance. When accounting performance rises by onepercentage point, compensation (either total or cash) rises by about2 percent,whether that risewas dueto luck— exchange ratemovements— or not. Columns7 through9 replicatethese four columns for market measuresof performance. Again, weŽ nd pay forluck that matchesthe pay sensitivityto a generalshock. A 1percent increasein shareholderwealth raises pay (again eithertotal or cash) ofabout .3 percent, irrespective of whether this risewas causedby luckor not. Twoimportant points should betaken away fromthis panel. First,the average Ž rmrewards its CEO as muchfor luck as it doesfor a generalmovement in performance.There seems to be verylittle if any Žlteringat all.Since we use a totally different shock,these Ž ndings address theoreticalconcerns about the use ofmean industry shocks(such as thoseraised in Gibbons and Murphy [1990] and Aggarwal and Samwick[1999b]) and show that thelack of Žlteringobserved in RPEŽndings generalizesto othersources of luck. Second,there is as muchpay forluck on discretionary com- ponentsof pay (salary and bonus) as thereis onothercomponents

15.First-stage regressions are reported in Appendix 2. In practice,we use as instrumentscontinuous variables for exchange rate appreciation (current and lagged)as well as dummies to allow for nonlinear effects of appreciation (also current andlagged). The dummies are formed at the 2 and4 percentcutoffs both forappreciation and depreciation. The instruments are jointly highly signiŽ cant, withŽ rst-stage p-valuesof lessthan 1 percentin all cases. ARECEOS REWARDED FORLUCK? 915 suchas optionsgranted. This rules out the notion that pay for luckmechanically arises because Ž rmscommit (implicitly orex- plicitly) tomultiyear stock option plans wherethe number of optionsgrants is Žxedahead oftime. Under such plans, as Žrm valuerises, so does the value of precommitted options grants [Hall 1999]. Becausesalary and bonusare the most subjective componentsof pay, Žnding pay forluck on thesevariables is very suggestive.Boards arerewarding CEOs forluck even when they couldŽ lterit. In PanelB ofTableIII, wereplicatePanel A exceptthat our measureof luck becomes mean performance of the industry, whichis meantto capture external shocks that areexperienced by all theŽ rmsin theindustry. More speciŽ cally, as an instru- mentfor Ž rm-levelrate of accounting return in agivenyear, we usethe weighted average rate of accountingreturn in that yearin thetwo-digit industry that Žrmbelongs to, excluding the Žrm itself fromthe calculation. 16 Theweight of agivenŽ rmin agiven yearis theshare of its totalassets in theaggregate “ totalassets” ofthe two-digit industry theŽ rmbelongs to. Similarly, as an instrumentfor Ž rm-levellogarithm of shareholder wealth in a givenyear, we use the weighted average of the log values of shareholderwealth in thetwo-digit industry in that year,again excludingthe Ž rmitself fromthe calculation and using total assetsto weight each individual Žrm. 17 As in PanelA, all regressionsinclude Ž rmŽ xed effectsand yearŽ xedeffects. We alsocontrol for a quadratic in CEO ageand aquadratic in CEO tenure.The regressions include more than twicethe data points ofPanelA becausewe cannow use all Žrms, notonly those in thetraded goodssector. Panel B showsa pattern quitesimilar to Panel A. Thepay forluck relationship in all speciŽcations again roughlymatches the pay forgeneral perfor- mance.Besides reinforcingthe Ž ndings ofPanel A, theselatest Žndings suggestthat previousRPE resultsarose probably not becauseof mismeasurement of the reference industry orof the industry shockbut becauseof true pay forluck.

16.We alsoinvestigated the use of one-digitand three-digit industry means asinstruments and found qualitatively similar results. 17.These mean industry performance measures are constructed from COMPUSTAT. Tomaximize consistency of the performance measures between Yermack’s Žrmsand the rest of industry,we also compute shareholder wealth and incometo assets ratios from COMPUSTAT forthe Ž rmsin Yermack’ s. Because notall Ž rmsin Yermack’ s dataare present in COMPUSTAT inevery year, we lose about800 Ž rm-yearobservations. 916 QUARTERLY JOURNALOF ECONOMICS

III. WHY IS THERE PAY FOR LUCK? Theresults so far clearlyestablish pay forluck. There are severalreactions possible to this evidence.First, one could take it as evidenceof skimming. To understand howthe skimming model predicts pay forluck, consider a CEO whohas captured thepay process.His primary worry in settingpay will bethat outrageous skimmingmay cause otherwise passive investorsto stand upand notice.Good performance, however, provides the CEO with extra slack.For example, shareholders may scrutinize a Žrmmore closelyduring bad times.This allows higherpay whenperfor- manceis goodand producesa positivelink between pay and performance,but fordifferent reasonsthan in thecontracting view.If goodperformance creates slack irrespective of whetherit was lucky,pay forluck will result. 18 Alternatively,one could argue that pay forluck is in fact optimaland that theevidence so far is consistentwith thecon- tractingview. One reason why pay forluck might be optimal is that theCEO’ s outsideoption may in fact depend onluck.When theoil industry enjoysa goodfortune, the ofoil CEOs maysimply becomemore valuable. Firms then pay theiroil CEOs moresimply tomatch their increased outside options. Thus,pay forluck is optimal herenot as an incentivedevice, but merelybecause the optimal level ofpay increaseswith luck. 19 Objectionscan be raised against this view.First, our sugges- tiveevidence of asymmetryin pay forluck may be hard torecon- cilewith this view.Average CEO compensationin theoil industry always goesup whenthe price of crude oil goes up but doesnot always godown whenthe price of crudeoil goes down. In ourfull sample,we performed a similartest using theindustry luck shock.For accounting measures, we again found that pay re- sponds moreto positive industry shocksthan tonegative ones (with noasymmetry on generalpay forperformance). For market

18.The scrutiny ofotherwisepassive investors may be triggered byabsolute performancefor several reasons. The very natureof decidingwhere to pay atten- tionrequires focusing on variables immediately at hand. Passive investors may useaccounting returns, earnings growth, orstock return indeciding when to act ona Žrm andall these variables are clearly not preŽ ltered. The idea that outra- geouspay actuallyproduces political intervention of someform has been pointed outin Jensenand Murphy [1990].Empirical work onthiscan be found in Joskow, Rose,and Wolfram [1996]. 19.While the argument made here is rather imprecise, Himmelberg and Hubbard[2000] provide a formalmodel as wellas empirical results that interpret payfor luck inthis light. ARECEOS REWARDED FORLUCK? 917 measures,we could not reject symmetry. Second, it is unclear why aCEO’s humancapital should becomemore valuable as industry fortunesrise. In fact, it maybe exactly in bad timesthat having theright CEO is mostvaluable. A priori,either relation- ship seemsplausible. Totestthis assumption,we examined turn- overin theCEO market.We found nostatistically signiŽcant relationshipbetween a CEO’s turnoverand industry returns (after controllingfor the Ž rm’sreturns)and apoint estimatethat was negative.This suggests that, if anything,turnover is coun- tercyclical.Third, we tested the effect of the industry’ s average CEO turnoverrate on pay forluck. If pay forluck were caused by marketcompetition for CEOs, thenindustries with higherturn- overshould exhibitthe greatest pay forluck. For accounting measures,we found that industrieswith thehighest turnover in fact showedthe least pay forluck. For market measures of per- formance,we found norelationship between industry turnover and pay forluck. Of course,for the last twoŽ ndings, onecould always arguethat competitivepressures operate through the threatof turnover rather than throughactual turnover.As a whole,though, we have been unable to Ž nd positiveevidence that outsidebidding upofCEO wagescould explain ourresults. Anotherreason why pay forluck may be optimal is that one maywant toprovide incentives to the CEO toforecast or respond to luck.20 Thiskind ofargumentcan be most readily evaluated in ouroil industry application. Suppose that aparticularly talented CEO in theoil industry understoodthe political subtletiesof the Arab countriesand forecastthe coming of thepositive oil shock at thebeginning of the 1980s. By increasingoutput fromexisting oil wells,increasing inventories, or intensifying searchfor new wells, hecould have increased his Žrm’sproŽts whenthe shock did come.Shouldn’ t shareholdersreward this farsighted CEO? The importantpoint hereis that thoseCEOs whowere exceptional in having forecastshould indeed berewarded. But this is notwhat wetest for. We use none ofthebetween Ž rmvariation in response tothe oil shock. We merely test whether the average Žrm expe- riencesa rise(or fall, forthe negative shocks) in pay. Putanother

20.An argumentsimilar to hedging has been made by Diamond [1998]. Tying pay toluck maygenerate incentives for the CEO tochange his correlation withthe luck variable.In practice,diversiŽ cation seems to be more in theinterest ofmanagement than shareholders. Tufano [1996], for example, demonstrates that managerialcharacteristics, such as share or option ownership, are quite predic- tiveof risk managementin a sampleof goldŽ rms. 918 QUARTERLY JOURNALOF ECONOMICS way,our results suggest that aCEO whoresponds to the shock exactlythe same way as everyother oil CEO is rewarded.This cannotbe a rewardfor having forecastwell. Again, onemay want toreward CEOs forexceptional responsiveness to shocks, but thereis littlereason to reward them for just average responsiveness. AŽnal setof responses to pay forluck would beto abandon theliteral contracting view and arguethat Žlteringout luck is simply impossible.This might be becauseof cognitivecomplexity in understanding what is luckand what is notluck. Part of this cognitivecomplexity may be a pureinformation issue if thereare notenough data available toŽ gureout the appropriate effectof luck.For example, estimating the coefŽ cient d in equation(1) may simply notbe possible. Part of the cognitive complexity may be psychologicalas in theevidence on the fundamental attribution error[Durell 1999]. 21 Noneof theevidence so far directlyrefutes this argument.

III.A. TheEffect ofGovernance Whilewe haveargued against someof thevarious extensions ofthesimple agency model, in theend we still believethat they meritserious consideration. They suggest to us that thepay for luckŽ nding doesnot per se rule out agency models. The results arealso consistent with theidea that Žlteringout luck is justnot feasible.Therefore, we nowturn to testinga speciŽc predictionof theskimming view rather than arguingagainst theother views. Sincethe skimming view emphasizes the CEO’ s ability togain controlof the pay process, should play an importantrole in skimming.It is exactlyin thepoorly governed Žrmswhere we expect CEOs tomost easily gain controlof thepay process.This suggests that weshould expectmore pay forluck in thepoorly governed Ž rms. 22

21.Another possibility is thatluck isnotcontractible. In practice,we do not believethis is important for most of our Ž ndings.First, it is difŽ cult to believe that noncontractingissues can explain our results in the oil industry case study: the priceof crude oilcaneasily be measuredand written into a contract. Second,even inthe presence of noncontractibility, subjective performance evaluation should effectivelyŽ lter(see Baker, Gibbons, and Murphy [1994]). 22.Would the alternative explanations above predict aneffectof governance? First,one might argue that better governance increases the efŽ cacy ofmonitoring mechanismsand hence reduces the need to pay forperformance. This argument wouldpredict achangein theoverall level of pay forperformance, not onlyin pay forluck. Wecircumvent this problem by looking at the change in pay forluck relativeto the change in generalpay forperformance. As anaside,general pay for ARECEOS REWARDED FORLUCK? 919

Toexamine how pay forluck differs betweenwell-governed and poorlygoverned Ž rms,we estimate two equations. First, in orderto provide a baseline,we ask how pay forgeneral perfor- mance(not luck) differs betweenwell-governed and poorlygov- ernedŽ rms.We estimatean OLSequationsimilar to equation (2) exceptthat weallow thepay forperformance coefŽ cient to depend ongovernance:

(3) yit 5 b p perfit 1 u p (Govit p perfit)

1 g i 1 x t 1 a X p Xit 1 a G p Govit 1 e it, where Govit is ameasureof governance. To understand this equation,differentiate both sides with respectto performance to get ­ yit /­ perfit = b + u p Govit.In words,this speciŽcation allows thepay forperformance sensitivity to be a functionof the governancevariable. A positivevalue for u would imply that bettergoverned Ž rmsshow greater pay forperformance. Equation (3) ofcourse tells us nothing about pay forluck, merelyabout pay forperformance. To get at pay forluck, we reestimatethis equationusing ourtwo-stage instrumental vari- ables procedure. 23 Wethen compute an estimateof the effect of governanceon pay forgeneral performance, uˆ and anestimateof ˆ theeffect of governance on pay forluck, u L u ck . ˆ ˆ Ourtest then consists in comparing u and u L uc k . We will speak of morepay forluck in poorlygoverned Ž rmswhen poorly governedŽ rmsdisplay morepay forluck relative to pay for generalperformance. If poorlygoverned Ž rmssimply gavemore pay forperformance and pay forluck rose as aconsequence,we would notrefer to this as morepay forluck. In practice,we will seethat it is pay forluck that changeswith governance,while pay forperformance hardly changes.We will alsoverify that these performancedoes not systematically correlate with governance. Second, for the- oriesthat rely on changesin the value of theCEO’ s humancapital, it is unclear why thesechanges would happen more in thepoorly governed Ž rms.Finally, the presumptionthat Ž lteringis somehow cognitively impossible would clearly be refutedif some Ž rmscould Ž lter. 23.An extremelyimportant caveat here: our approach allows for the possi- bilitythat better governed Ž rmsmay have a differentresponsiveness of perfor- mance toluck. So,for example, if well-governedŽ rmswere merely less responsive intheir performance (not their pay) toluck, thiswould not create a spuriouspay forluck differencebetween well-governed and poorly governed Ž rms.Technically, performance perf it ,theendogenous variable we need to instrument, appears both directlyand indirectly (the term Govi t p perf it )inthis equation. When we instrument,we perform two Žrst stages,one for the direct effect perf i t and one for theinteraction term Govi t p perf it .Thisprocedure is crucialbecause it allows the effectof luck onperformance to depend on governance. 920 QUARTERLY JOURNALOF ECONOMICS

TABLE IV LARGE SHAREHOLDERSAND PAY FOR LUCK (LUCK MEASURE IS MEAN INDUSTRY PERFORMANCE) DEPENDENT VARIABLE: ln(Total Compensation)

Governance measure: Largeshareholders Largeshareholders on board General Luck General Luck General Luck General Luck SpeciŽcation: (1) (2) (3) (4) (5) (6) (7) (8)

Income Assets 2.184.59 — —2.144.49 — — (.238) (.912) (.217) (.882) Governance* 2 .094 2 .416 — — 2 .181 2 1.48 — — Income/assets (.094) (.204) (.176) (.396) ln (shareholder ——.249.383 — — .258 .318 wealth) (.018) (.219) (.017) (.199) Governance* — — .001 2 .066 — — 2 .019 2 .076 Log(shareholder (.007) (.036) (.016) (.053) wealth) Governance 2 .009 .018 2 .017 .411 2 .006.084 .100 .480 (.011) (.018) (.049) (.240) (.021) (.033) (.108) (.356) Sample size 46104610 4570 4570 4621 4621 4581 4581 Adjusted R2 .695 .706 .694 .706

a. Dependentvariable is thelogarithm of total compensation.performance measure is operatingincome tototal assets. All nominalvariables are expressed in real dollars. b. Inall theluck regressions, both the performance measure and the interaction of the performance measurewith the governance measure are instrumented. The instruments are the asset-weighted average performancein the two-digit industry and the interactions of the industry performance with that governance measure. c.“ Largeshareholders” indicates the number of blocks ofat least 5percentof the Ž rm’s commonshares, whetherthe block holder is oris nota director.“ Largeshareholders on board”indicates the number of blocks ofat least 5percentof the Ž rm’s commonshares that areheld by directors of the board. d.Each regression includes Ž rmŽ xedeffects, year Ž xedeffects, a quadraticin age, and a quadraticin tenure. e.Standard errors are in parentheses.

resultsare robust to allowing pay forluck to vary by Žrmsize. Otherwise,one might simply worrythat governanceis aproxyfor size.24 In TableIV, weimplement this frameworkfor the case of largeshareholders. We ask whether the presence of largeshare- holdersaffects pay forluck. Shleifer and Vishny [1986], among others,argue that largeshareholders improve governance in a Žrm.A singleinvestor who holds alargeblock of sharesin aŽrm

24.One must also be careful in interpreting the results from this exercise. Theyare merely suggestive of the cross-sectional relationship between gover- nanceand the extent of skimming.They do not necessarily imply that a policyof changinga speciŽc governancevariable will necessarily lead to a changein the extentof skimming. To make strong policy suggestions such as this, one would needmore exogenous variation in governance and see its effects on CEO pay,such asin Bertrand andMullainathan [1999]. The governance results here are useful becausethey demonstrate the relevance of the skimming view and not because theyisolate policy mechanisms to reduce skimming. ARECEOS REWARDED FORLUCK? 921 will havegreater incentives to watch overthe Ž rmthan adis- persedgroup of smallshareholders. 25 In ourcontext, the idea of largeshareholders Ž tsmostnaturally as this matchesthe intui- tionof “ having aprincipal around.”Yermack data containa variable that countsthe number of individuals whoown blocks of at least5 percentof the Ž rm’s commonshares. When the CEO happens toown such a block,we exclude this blockfrom the count.We furtherknow whether these large shareholders are on theboard ornot.A priori,one might expect that largesharehold- erson the board havethe greatest impact. They can exert their controlnot just through implicit pressureor voting,but alsowith adirectvoice on theboard. Since the information is available, we will considerthe effect both of all largeshareholders and ofonly thoseon the board. TheŽ rstfour columns of TableIV useall largeshareholders as ourmeasure of governance. All regressionsinclude the usual controls.Column (1) estimateshow the sensitivity of pay toper- formancedepends ongovernancefor accounting measures of per- formance.The Ž rstrow tells us that aŽrmwith nolarge share- holdersshows a sensitivityof log compensation to accounting returnof 2.18.An increasein accountingreturn of onepercentage point leads toan increasein pay ofabout 2 percent.The second rowtells us that adding alargeshareholder only weakly de- creasesthe sensitivity of pay togeneral performance, and this effectis notstatistically signiŽcant. For example, a onepercent- agepoint increasein accountingreturn now leads toa 2.09 percentincrease in pay whenthe Ž rmhas onelarge shareholder (comparedwith 2.18 in theabsence of any largeshareholder). Column(2) estimateshow large shareholders affect pay for luck.26 Asbefore,the Ž rstrow tells us that thereis signiŽcant pay forluck. The second row here, however, tells us that this pay for luckdiminishes signiŽ cantly in thepresence of a largeshare- holder.A onepercentage point increasein accountingreturns due toluck leads toroughly a 4.6 percentincrease in pay whenthere is nolarge shareholder but onlya 4.2 percentincrease in pay

25.They also point out a possibleopposing effect: very largeshareholders mayhave a greaterability to expropriate rents for themselves. This effect is likely tobe greatest in othercountries where investor protection is weakest.Empirical evidenceon the efŽ cacy oflarge shareholders can be found in Zeckhauser and Pound[1990], Shivdasani [1993], and Denis and Serrano [1996]. 26.In all that follows, we will use mean industry performance as our mea- sureof luck sincethis produces the most powerful Ž rst stagesin the IV framework. 922 QUARTERLY JOURNALOF ECONOMICS whenthere is onemore large shareholder. Each additional large shareholderdecreases this effectby .4percent. This is a10 percentdrop in thepay forluck coefŽ cient for each additional largeshareholder. 27 Columns(3) and (4) estimatethe same regressions using marketmeasures of performance.In this case,the pay forgeneral performancedoes not depend at all onthe existence of a large shareholder(a coefŽcient of .001 with astandard errorof .007). Weagain Žnd, however,that pay forluck diminishes with the presenceof a largeshareholder. While the result is onlysigniŽ - cant at the10 percentlevel, the economic magnitude is larger. Thepay forluck coefŽ cient now drops .066/.383 ’ 17 percentfor eachlarge shareholder. In columns(5) through(8) werepeat the above exercise but alterthe governance measure. We nowfocus only on largeshare- holderson theboard. Comparing columns(6) and (2), weseethat thegovernance effect strengthens signiŽ cantly with respectto theŽ lteringof accounting performance. We see that thepay for luckdrops by 33 percentfor each additional largeshareholder. Theresults are very statistically signiŽcant. On marketperfor- mancemeasures, we Ž nd theeffect also rises but lessdramati- cally.In column(8) thepay forluck drops 23 percentwith each largeshareholder on the board. Moreover, this last resultis insigniŽcant. In summary,our Ž ndings in TableIV highlight how largeshareholders (especially thoseon the board) affect theex- tentof pay forluck. Firms with morelarge shareholders show for lesspay forluck. Theresults in TableIV simply compareŽ rmswith large shareholderswith Žrmswithout. This ignores the effects of CEO tenure,another important determinant of governance.A common beliefis that CEOs whohave been with theŽ rmlongerhave had achanceto become entrenched, perhaps by appointing friends on theboard. In this case,we would expecthigh tenure CEOs to

27.One may recall from subsection II.D thatwe are excluding the cumulated optionsfrom our measure of compensation as thesewill mechanically covary with stock priceand, hence luck. However,if Ž rmswith large shareholders provided moreoptions, they may effectively provide more pay forluck becausea bigger fractionof compensation(and hence for pay forluck) comesfrom the mechanical portion.To test at least this presumption, we compared the fraction of total compensationthat were options grants betweenwell-governed and poorly gov- ernedŽ rms.We found no consistent economically or statisticallysigniŽ cant dif- ferencefor our governance measures. ARECEOS REWARDED FORLUCK? 923

TABLE V TENURE, LARGE SHAREHOLDERS , AND PAY FOR LUCK (LUCK MEASURE IS MEAN INDUSTRY PERFORMANCE) DEPENDENT VARIABLE: ln (TOTAL COMPENSATION )

Any largeshare holderontheb oard? No Yes No Yes General Luck General Luck General Luck General Luck SpeciŽcation: (1) (2) (3) (4) (5) (6) (7) (8)

Income Assets 2.143.35 1.28 2.47 — — — — (.30) (.96) (.65) (2.60) CEO tenure p .00 .13 .063 2 .006 — — — — Income/assets (.02) (.05) (.045) (.131) Log(shareholder — — — —.24.26 .27 .53 wealth) (.02) (.24) (.05) (.32) CEO tenure p — — — — .003 .009 2 .005 2 .013 Log(sh. wealth ) (.001) (.005) (.003) (.010) CEO tenure .01 2 .00.010 .016 2 .002 2 .045.044 .084 (.00) (.01) (.011) (.016) (.01) (.04) (.020) (.059) Sample size 38843884 740 740 3841 3841 743 743 Adjusted R2 .7030 .757 .715 .700

a. Dependentvariable is thelogarithm of total compensation.All nominalvariables are expressed in real dollars. b. Inall theluck regressions, both the performance measure and the interaction of the performance measurewith the CEO tenure are instrumented. The instruments are the asset-weighted average perfor- mancein the two-digit industry and the interactions of theindustry performance with the CEO tenure. c.Sample in columns (1), (2), (5), and (6) is theset ofŽrm-yearobservations for which there is nolarge shareholdersitting onthe board of directors; sample in columns (3), (4), (7), and (8) is theset ofŽ rm-year observationsfor which there is at least onelarge shareholder sitting ontheboard of directors. d.Each regression includes Ž rmŽ xedeffects, year Ž xedeffects, a quadraticin age, and a quadraticin tenure. e.Standard errors are in parentheses.

showthe greatest pay forluck. Moreover, we would expectthis effectto be strongestin thoseŽ rmswheregovernance is weakand thereis nolarge shareholder present to limit the increased en- trenchment.Hence, in theabsence of large shareholders, we expectfairly stronggovernance early in aCEO’s tenure,but this governanceshould weakenover time as heentrenches himself. In thepresence of large shareholders, we not only expect stronger governancebut alsothat this strongergovernance should last throughoutthe CEO’ s tenure.It is harderfor a CEO tobegin stackingthe board whenthere is alargeshareholder around. Thus,we expect a risein pay forluck with tenurein theabsence ofa largeshareholder, but lessof a rise(or even no rise) in the presenceof alargeshareholder. TableV teststhis idea.We Ž rstsort Ž rmsinto two groups based onwhether they have a largeshareholder present on the 924 QUARTERLY JOURNALOF ECONOMICS board.28 Thisproduces 740 orso data points forŽ rmswith large shareholdersand 3880 orso data points forŽ rmswithout large shareholders.For each set, we now separately estimateregression 3forthese two groups with tenureas ourgovernance measure. Columns(1) and (2) focuson accounting measures of perfor- mancein Žrmswithout a largeshareholder. The second row tells usthat whiletenure does not affect pay forperformance, it greatlyincreases pay forluck. In fact, aCEOwith (roughly)the mediantenure of nine years shows about .13 p 9/3.35 ’ 35 percentgreater pay forluck than onewho just began at theŽ rm. Letus contrast this with columns(3) and (4) whichestimate the sameeffect for Ž rmswith alargeshareholder present. Here we Žnd that tenuredoes not affect pay forluck at all, while,if anything,it seemsto raise pay forperformance slightly [Gibbons and Murphy 1992]. Thus,pay forluck increases with tenurein theabsence of a largeshareholder but doesnot change with tenurein thepresence of alargeshareholder. Columns(5) through(8) repeatthe tests of columns (1) through(4) formarket measures of performance. Here the results areless stark but still verysuggestive. Comparing columns (6) and (5), weseethat bothpay forperformance and pay forluck rise with tenure,but pay forluck rises three times as fast (.003 versus .009). ThecoefŽ cient on thepay forluck, however, is onlysigniŽ - cant at the10 percentlevel. The economic signiŽ cance, however, stays largeas aCEOwith atenureof nine years shows an increasein pay forluck of .009 p 9/.26 ’ 31percent,but arisein pay forperformance of only10 percent.In columns(8) and (7) we seethat, if anything,pay forluck and pay forperformance both diminish with tenure. Whilelarge shareholders correspond most closely to the idea ofaprincipal, othergovernance measures could also be used. Our data containtwo variables that wereshown to be important governancemeasures in thepast: thesize of the board and the fractionof board memberswho are insiders in theŽ rm.Small boards arethought to be more effective at governingŽ rms.Yer- mack[1996], forexample, shows that smallerboards correlate with larger q valuesfor Ž rms.Insiders ontheboard aregenerally

28.We focusonly on largeshareholders on theboard because these provided thestrongest results in Table IV. Wehaveused all large shareholders and found similar,although statistically weaker, results. ARECEOS REWARDED FORLUCK? 925 thoughtto weaken governance. 29 TheŽ rstfour columns in Table VI estimatethe effect of board sizeon pay forluck. Columns (1) and (2) showthat, for accounting measures, the direction of the effect is theopposite of what wepostulated but thecoefŽ cient is statistically insigniŽcant. Note that theactual sizeof the coefŽ cient is tiny:even ahugeincrease in board sizeof ten board membersleads onlyto a 2percentdrop in pay forluck. Columns (3) and (4), however,show that thereare signiŽ cant effectsfor market measures of perfor- manceand theseare of the expected sign. Consider the difference betweentwo boards, one of which has tenboard membersand oneof whichhas six.The big board Žrmshows a pay forperformance coefŽcient of .240 and apay forluck coefŽ cient of .229. The smaller board continuesto show a pay forperformance coefŽ cient of .228. But,it showsa pay forluck coefŽ cient of .177. This represents a drop of (.229 2 .177)/.177 ’ 30 percentin thepay forluck coefŽ cient. Theother measure we examine is ameasureof insider presence on theboard. This variable is measuredas fractionof board memberswho areŽ rminsiders or gray directors .Columns(5) and (6) showthat on accountingmeasures,insider presence dramatic ally increasesthe pay forluck coefŽ cient (signiŽ cant at the10 percentlevel). In aboard with tendirectors ,turningone of the outside directors from an outsiderto an insiderincreases pay forluck by 4.51 p .1/2.27 ’ 20 percent.The effect onpay forperformanc eis negativeand small.Columns (7) and (8) show that onmarket performan cemeasures, insider board presenceagain increasespay forluck, but whilethe coefŽ cient continues to be eco- nomicallylarge, it is statistically quiteinsigniŽ cant. Weturn to our last governancemeasure in columns(9) through(12), wherewe construct an index that aggregatesall the governancemeasures used so far: number of largeshareholders, numberof large shareholders on the board, board size,and in- siderpresence on board. 30 Toformthe index, we demean each of

29.Empirical evidence on theeffect of insiders on the board can be found in Baysingerand Butler [1985], Weisbach [1988], Rosenstein and Wyatt [1990], Hermalinand Weisbach [1991], Byrd andHickman [1992], and Brickley, Coles andTerry [1994].We have also examined CEO ownershipand whether the founderis present. We do not report these for space reasons but both produce generallysigniŽ cant effects.Founders and CEOs withhigh insider ownership bothshow greater pay forluck. 30.Market valuation of a Žrm mayprovide another index of governance.We haveexamined how a market-to-bookmeasure correlates with the extent of skimmingon accounting returns. We found,using a procedureidentical to the one usedfor the governance variables, that Ž rmswith higher market-to-book showed lowerlevels of pay forluck onaccounting returns. Clearly it wouldbe conceptually awkward todo a similarexercise for market returns. 926 QUARTERLY JOURNALOF ECONOMICS

TABLE VI CORPORATE GOVERNANCEAND PAY FOR LUCK (LUCK MEASURE IS MEAN INDUSTRY PERFORMANCE) DEPENDENT VARIABLE: ln (TOTAL COMPENSATION )

Governance measure: Board size Fractioninsiders General Luck General Luck General Luck General Luck SpeciŽcation: (1) (2) (3) (4) (5) (6) (7) (8)

Income Assets 2.615.19 — —2.302.27 — — (.558) (1.62) (.453) (1.24) Governance* 2 .045 2 .093 — — 2 .482 4.51 — — Income/assets (.043) (.094) (.853) (2.69) Log(shareholder ——.216.099 — —.241.241 wealth) (.034) (.210) (.029) (.215) Governance* ——.002.013 — —.027.126 Log(sh. wealth) (.002) (.006) (.05) (.190) Governance .012 .015 2 .013 2 .080 .158 2 .315 2 .066 2 .742 (.005) (.007) (.016) (.041) (.129) (.271) (.407) (1.29) Sample size 46244624 4584 4584 4624 4624 4584 4584 Adjusted R2 .695 .706 .695 .706

Governance measure: Governance index General Luck General Luck SpeciŽcation: (9) (10) (11) (12)

Income Assets 2.07 4.23 — — (.210) (.865) Governance* .007 2 .216 — — Income/assets (.057) (.134) Log(shareholder wealth) — — .249 .252 (.016) (.232) Governance* — — 2 .003 2 .033 Log(sh. wealth) (.004) (.015) Governance 2 .016 .000 .010 .210 (.007) (.011) (.027) (.103) Sample size 4610 4610 4551 4551 Adjusted R2 .695 .705

a. Dependentvariable is thelogarithm of total compensation.All nominalvariables are de ated. Each regressionincludes Ž rmŽ xedeffects, year Ž xedeffects, a quadraticin age, and a quadraticin tenure. Standarderrors are in parentheses. b. “Boardsize” indicates the number of members of the board of directors,as listed inthe proxy statement nearthe start oftheŽ scal year.“ Fractioninsiders” is thefraction of inside and “ gray”directors on the board ofdirectors. “ Governanceindex” is theunweighted average of four standardized governance variables (numberof large shareholders, number of large shareholders on board, minus board size, andone minus fractioninsiders). c.In all theluck regressions, both the performance measure and the interaction of the performance measure withthe governance measure are instrumented. The instruments are the asset-weighted average performance in thetwo-digit industry and the interactions of the industry performance with that governancemeasure. thefour governance variables, divide it by its standard deviation, and thentake the sum of thesestandardized variables.For board size,we use negative of board sizein this procedure.For fraction ARECEOS REWARDED FORLUCK? 927 ofinsiders on the board, we use one minus that fraction.This guaranteesthat theresulting governance index has largervalues wheneverthe Ž rmis bettergoverned. 31 Foraccounting measures of performance (columns (9) and (10)), weagain Žndthat pay forluck diminishes with thegover- nance,while pay forperformance does not change. The coefŽ - cient,however, is onlysigniŽ cant at the10 percentlevel. To gaugethe magnitude of these effects, consider a one-standard- deviationincrease in thegovernance index, about 2. Such an increaseleads toa 2 .216 p 2/4.23 ’ 10 percentfall in thepay for luckcoefŽ cient. When we usemarket measures (columns (11) and (12)), increasesin thegovernance index greatlyreduce pay for luckbut hardly affect pay forperformance. In this case,the coefŽcients are signiŽ cant at the5 percentlevel. Moreover, their magnitudeis bigger.A one-standard-deviation increasein gover- nancedecreases pay forluck by 26 percent. Finally,we investigate the robustness of our Ž ndings. The primaryconcern one might have is that wehave not adequately controlledfor Ž rmsize. One might worry that largeŽ rmshave quitedifferent pay forperformance sensitivities than small Žrms [Bakerand Hall 1999]. If this alsotranslates into different pay for lucksensitivities, the estimates above might confuse this size effectfor a governance“ effect.” In TableVII weaddress this problemby controllingfor size interactedwith performance.We reestimate equation (3) but this timeinclude a term Size p perfit.Ourmeasure of size in these regressionsis averagelog real assets of theŽ rmover the period. Wereport the results for two governance measures, large share- holderson theboard and thegovernance index, although we have reestimatedall theprevious tables with thesecontrols and found similarresults. Columns (1) through(4) areto be compared with columns(5) through(8) ofTable IV. Wesee that theeffect of governanceon the Ž lteringof accounting rates of return in fact strengthenswhen these controls are added ( 2 2.23 versus 2 1.48).

31.This particular way ofproceedingwill tend to count large shareholders on the board twice, onceas on theboard and once as general large shareholders. This isa crude way ofincorporating our prior belief (supported by the Ž ndingsin Table IV)that large shareholders on the board matter more. When we use either measurein the index alone, we Ž ndqualitatively similar results. We have also estimateda regressionin which we include all four governance measures (and theirinteractions) together. These regressions showed all the governance mea- suresentering with the same sign and only the large shareholder variables being statisticallysigniŽ cant. 928 QUARTERLY JOURNALOF ECONOMICS

TABLE VII CORPORATE GOVERNANCEAND PAY FOR LUCK ROBUSTNESS CHECKS (LUCK MEASURE IS MEAN INDUSTRY PERFORMANCE) DEPENDENT VARIABLE: ln (TOTAL COMPENSATION )

Governance measure: Largeshareholders on board Governance index General Luck General Luck General Luck General Luck SpeciŽcation: (1) (2) (3) (4) (5) (6) (7) (8)

Income Assets .288 36.9 — — 2 .570 12.3 — — (1.14) (14.6) (1.31) (6.17) Governance* 2 .118 2 2.23 — — .056 2 .197 — — Income/assets (.18) (.52) (.061) (.126) Log(shareholder — — .223 2 .136 — — .251 2 .194 wealth) (.086) (.334) (.098) (.345) Governance* — — 2 .018 2 .059 — — 2 .003 2 .027 Log(shareholder (.016) (.053) (.004) (.012) wealth) Governance 2 .010.127 .094 .365 2 .019 2 .002.007 .170 (.021) (.038) (.109) (.357) (.007) (.010) (.029) (.085) Firm Size p YesYes Yes Yes Yes Yes Yes Yes Performance? Sample size 46214621 4581 4581 4610 4610 4551 4551 Adjusted R2 .695— .706— .695— .705—

a. Dependentvariable is thelogarithm of total compensation.Performance measure is operatingincome tototal assets. All nominalvariables are expressed in real dollars. Each regression includes Ž rmŽ xedeffects, yearŽ xedeffects, a quadraticin age, and a quadraticin tenure. b. Inall theluck regressions, both the performance measure and the interaction of the performance measure withthe governance measure are instrumented. The instruments are the asset-weighted average performance in thetwo-digit industry and the interactions of the industry performance with that governancemeasure. c.“ Largeshareholders on board” indicates the number of blocks ofat least 5percentof the Ž rm’s common shares that areheld by directors of the board. “ Governanceindex” is theunweighted average of four standardized governancevariables (number of large shareholders, number of large shareholders on board, minus board size, andone minus fraction insiders). “ Firm size”indicates average log assets overthe same period. d.Standard errors are in parentheses.

Theeffects on market performance measures, however, weaken slightly (.059 versus.076). Columns(5) through(8) usethe gov- ernanceindex and areto be compared with columns(9) through (12) in TableVI. Thiscomparison shows that theresults weaken veryslightly ( 2 .197 versus 2 .216) foraccounting measures as wellas formarket measures ( 2 .027 versus 2 .033). Forboth governancemeasures, however, the results remain economically signiŽcant. In twoof the cases (column (4) and tosome extent column(6)), theyremain statistically insigniŽcant. In theother two,they remain statistically signiŽcant. Wehavealso attempted other robustness checks. We checked whetherŽ lteringmay happen overlonger time horizons by ag- gregatingour data overseveral years as wellas lookingat lags. Wealsoallowed forinteractions between performance and yearin ARECEOS REWARDED FORLUCK? 929 ourregressions since there are known to be changesin thepay for performancesensitivity over this timeperiod. These modiŽ ca- tionsdid notalter our qualitative Žndings.

IV. CONCLUSION CEOs arerewarded for luck. Moreover, pay forluck is as large as pay forgeneral pay forperformance. Pay for luck also appears on themost discretionary components of compensation, salary and bonus.Looking closer, we found that pay forluck is strongestamong poorlygoverned Ž rms.Adding alargeshareholder on the board, for example,decreased the pay forluck by 23 to33 percent.This Ž nding weakenstwo prominent explanations of pay forluck: “ Payingfor luckis optimal”and “Filteringout luck is impossible.” Morebroadly, these results encourage a revisionof ourviews onCEO pay. Poorlygoverned Ž rmsŽ tthepredictions of the skimmingview. Well-governed Ž rmsŽ tthepredictions of the contractingview better. They are to remove some luck in setting pay. Thissuggests that bothviews hold somesway. Other empiri- cal facts support this idea.In Bertrand and Mullainathan [2000b] weshowed that well-governedŽ rmscharged CEOs morefor the optionsthey were granted. Options containa gift component becauseeven if theCEO doesnothing they have value from the intrinsicvolatility ofthe stock (their Black-Scholes value). We showthat Žrmswith largeshareholders, smaller boards, and so on,are better able tocharge their CEOs and betterable toremove this gift componentby reducingthe other components of pay. In otherwords, principal agentmodels work best when there are in fact individuals aroundto act as principals. Severalunanswered questions remain. First, it is unclear what effectsthe reward for luck has onoverallCEO utility.Does competitionin themarket for CEOs forcethe mean level of pay at initial hireto adjust sothat thereare no ex anterents to behad? Oris thehiring process sufŽ ciently closed or captured by insiders that suchadjustments are small? Second, while formal models of thecontracting view abound, there is nocareful analysis ofthe skimmingmodel. Without such an analysis, ourunderstanding of skimmingwill necessarilyremain vague. What arethe exact mechanismsby whichskimming is constrained?How speciŽ cally doesbetter governance translate into better pay packages?The resultsin this paper suggestthat moreenergy should bedevoted toclarifying theskimming alternative. 930 QUARTERLY JOURNALOF ECONOMICS

APPENDIX 1: SUMMARY STATISTICS—50 LARGEST U. S. OIL COMPANIES CEOS

Mean S. D.

Age of CEO 58.562 7.892 Tenureof CEO 10.181 9.781 Totalcompensation 608.269 597.194 ln(Total Compensation) 6.125 .722

a. Data set is 50 ofthe 51 largestU. S.oilcompanies over the period 1977– 1994. b. Total Compensationis deŽned as thesum of salary andbonus (cash and stock bonus),company contributionsto thrift plans, otherannual income, and the value of the options granted to the CEO during that year,in thousands of 1977 dollars.

APPENDIX 2: PAY FOR LUCK: FIRST-STAGE REGRESSIONS (LUCK MEASURE IS EXCHANGE RATE SHOCKS)A

Dep. Var.: IncomeInc. to Assets Ln (Sh.Wealth)

(1) (2) (3) 2% < Appr. < 4% 2 56.588 2 .006 2 .039 (Current) (26.408) (.004) (.047) 2% < Appr. < 4% 2 15.428 .004 2 .027 (Lagged) (24.271) (.004) (.048) Appr. > 4% 2 68.903 2 .013 2 .034 (Current) (32.039) (.005) (.058) Appr. > 4% 2 12.045 .006 .053 (Lagged) (30.646) (.005) (.055) 2% < Depr. < 4% 76.642 2 .000 .153 (Current) (24.647) (.004) (.045) 2% < Depr. < 4% 85.858 .010 .114 (Lagged) (25.942) (.004) (.047) Depr. > 4% 45.482 .007 .094 (Current) (27.761) (.005) (.050) Depr. > 4% 76.345 .017 .046 (Lagged) (29.791) (.005) (.054) Exch. rateindex growth 2 19.273 2 .000 2 .077 (Current) (167.134) (.030) (.302) Exch. rateindex growth 216.140 .038 .237 (Lagged) (175.302) (.031) (.316) Sample size 1737 1729 1713 Adjusted R2 .622 .700 .873 F-stat 3.48 2.6 2.47 (prob > F 2 stat) (.000) (.004) (.006)

a. Dependentvariable is thelevel of income in column (1), the ratio of operating income to total assets incolumn (2) and the log value of shareholder wealth in column (3). Income and shareholder wealth are expressedin millions of 1977 dollars.2% < Appr. < 4% is dummyvariable that equals1 ifthe industry- speciŽc exchangerate index appreciated by more than 2% andless than4% sincethe previous year. All the otherappreciation and depreciation dummies are deŽ ned in a similar way. b. Eachregression includes Ž rmŽ xedeffects and year Ž xedeffects. All regressionsalso includea quadraticin CEO age and a quadraticin CEO tenure. c.The 3 regressionsare the Ž rst-stage regressionsassociated withcolumns (2), (4), and (8) in Panel A of Table 3. d.Standard errors are in parentheses. ARECEOS REWARDED FORLUCK? 931

PRINCETON UNIVERSITY, NATIONAL BUREAU OF ECONOMIC RESEARCH, AND CENTER FOR ECONOMIC POLICY RESEARCH MASSACHUSETTS INSTITUTE OF TECHNOLOGYAND NATIONAL BUREAU OF ECONOMIC RESEARCH

REFERENCES Abowd,John M., and David S. Kaplan,“ ExecutiveCompensation; Six QuestionsThat NeedAnswering,” Journalof Economic Perspectives, XIII(1999), 145– 168. Aggarwal, RajeshK., and Andrew A.Samwick, “TheOther Sideof theTrade-Off: TheImpact of Risk on Executive Compensation,” Journalof Political Econ- omy, CVII(1999a), 65– 105. Aggarwal, RajeshK., andAndrew A.Samwick, “ExecutiveCompensation, Stra- tegicCompetition, and Relative Performance Evaluation: Theory and Evi- dence,” Journalof Finance, LIV(1999b),1999 – 2043. Baker,George, Robert Gibbons, and Kevin J. Murphy, “SubjectivePerformance Measuresin Optimal Incentive Contracts,” Quarterly Journalof Economics, CIX(1994),1125– 1156. Baker,George, and Brian Hall, “ UnderstandingTop Management Incentives: FirmSize, Risk and CEO Effort,”Harvard Business School, 1998. Baysinger,Robert D., andHenry N. Butler,“ CorporateGovernance and the Board ofDirectors:Performance Effects of Changesin BoardComposition,” Journal ofLaw, Economicsand Organization, I(1985),101– 124. Bertrand,Marianne, “ Fromthe Invisible Handshake to the Invisible Hand? How ImportCompetition Changes the Employment Relationship,” National Bu- reauof EconomicResearch Working Paper No. 6900, 1999. Bertrand,Marianne, and Sendhil Mullainathan, “ CorporateGovernance and Ex- ecutiveCompensation: Evidence from Takeover Legislation,” Princeton Uni- versity,1999. Bertrand,Marianne, and Sendhil Mullainathan, “ Agentswith and without Prin- cipals,” American EconomicReview, XC(2000a),203– 208. Bertrand,Marianne, and Sendhil Mullainathan, “ DoCEOs SetTheir Own Pay? TheOnes without Principals Do,” National Bureau of Economic Research WorkingPaper No. 7604, 2000b. Blanchard,Olivier J., FlorencioLopez-de-Silanes, and Andrei Shleifer, “ WhatDo FirmsDo with Cash Windfalls?” Journalof Financial Economics, XXXVI (1994),337– 360. Brickley, JamesA., Jeffrey L. Coles, and Rory L.Terry, “OutsideDirectors and the Adoptionof Poison Pills,” Journalof Financial Economics, XXXV(1994), 371– 390. Byrd, JohnW., andKent A. Hickman,“ DoOutside Directors Monitor Managers?” Journalof Financial Economics, XXXII(1992), 195– 221. Crystal,Graef, InSearch ofExcess: The Overcompensation ofAmerican Executives (NewYork, NY: W.W.NortonCo., 1991). Denis,David, and Jan Serrano, “ Active Investorsand Management Turnover FollowingUnsuccessful Control Contests,” Journalof Financial Economics, XL(1996),239 –266. Diamond,Peter, “ ManagerialIncentives: On theNear Linearity of Optimal Com- pensation,” Journalof Political Economy, CVI(1998),931– 957. Durell,Alan, “ Attributionin PerformanceEvaluation,” Dartmouth College, 1999. Garen,John E., “ExecutiveCompensation and Principal-Agent Theory,” Journal ofPolitical Economy, CII(1994), 1175– 1199. Gibbons,Robert, and Kevin J. Murphy, “RelativePerformance Evaluation for ChiefExecutive OfŽ cers,” Industrial andLabor Relations Review, XLIII (1990),S30 –S51. Gibbons,Robert, and Kevin J. Murphy, “OptimalIncentives in the Presence of CareerConcerns: Theory and Evidence,” Journalof Political Economy, C (1992),468 – 505. Haid,Michael, Incentive Compensationand the Market forCorporate Control: Substitutive Forces toDiscipline Management ofPubliclyHeld Organizations inthe U.S.:Empirical Evidence Fromthe Oil Industry 1977–1994 (Bern, Switzerland:Haupt, 1997). 932 QUARTERLY JOURNALOF ECONOMICS

Hall,Brian J., “ TheDesign of Multi-YearStock OptionPlans,” Journalof Applied Corporate Finance, XII(1999),97– 106. Hall,Brian J., and Jeffrey B. Liebman,“ Are CEOs ReallyPaid Like Bureaucrats?” Quarterly Journalof Economics, CXIII(1998), 653– 691. Hermalin,Benjamin, and Michael Weisbach, “ TheEffects of Board Composition andDirect Incentiveson Firm Performance,” Financial Management, XX (1991),101– 112. Himmelberg,Charles, and R. GlennHubbard, “ IncentivePay and the Market for CEOs: An Analysisof Pay-For-Performance Sensitivity,” Columbia Univer- sity, 2000. Holmstrom,Bengt, “ MoralHazard and Observability,” Bell Journalof Economics, X(1979),74 – 91. Holmstrom,Bengt, and Paul Milgrom, “ Aggregationand Linearity in the Provi- sionof IntertemporalIncentives,” Econometrica, LV(1987),303– 328. Hubbard,R. Glenn,and Darius Palia, “ ExecutivePay and Performance: Evidence fromthe U. S.BankingIndustry,” Journalof Financial Economics, XXXIX (1994),105– 130. Janakiraman,Surya N.,Richard A.Lambert,and David F. Larcker, “An Em- piricalInvestigation of the Relative Performance Evaluation Hypothesis,” Journalof Accounting Research, XXX(1992),53– 69. Jensen,Michael, and Kevin J. Murphy, “PerformancePay and Top-Management Incentives,” Journalof Political Economy, XCVIII(1990), 225– 264. Joskow,Paul L., Nancy L.Rose,and Catherine D. Wolfram,“ PoliticalConstraints onExecutive Compensation: Evidence from the Electric UtilityIndustry,” RandJournal of Economics, XXVII(1996), 165– 182. ——,“CorporatePerformance and Managerial Remuneration: An EmpiricalIn- vestigationof Managerial Labor Contracts,” RandJournal of Economics, XVII(1986),59 – 76. ——,“ExecutiveCompensation,” Handbookof Labor Economics, O.Ashenfelter andD. Card, eds.(Amsterdam, The Netherlands: North-Holland, 1999). Revenga,Ana L., “Exporting Jobs?The Impact of Import Competition on Employ- mentand Wages in U. S.Manufacturing,” Quarterly Journalof Economics, CVII(1992), 255– 284. Rosenstein,Stuart, andJeffrey G. Wyatt,“ OutsideDirectors, Board Indepen- dence,and Shareholder Wealth,” Journalof Financial Economics, XXVI (1990),175– 191. Shea,John, “ NominalIllusion: Evidence from Major League Baseball,” University ofMaryland,1999. Shivdasani,Anil, “ BoardComposition, Ownership Structure, andHostile Take- overs,” Journalof Accounting andEconomics, XVI(1993), 167– 198. Shleifer,Andrei, and Robert Vishny, “ LargeShareholders and Corporate Control,” Journalof Political Economy, XCIV(1986),461– 488. Tufano,Peter, “ WhoManages Risk? An EmpiricalExamination of RiskManage- mentPractices in theGold Mining Industry,” Journalof Finance, LI (1996), 1097–1137. Weisbach,Michael, “ OutsideDirectors and CEO Turnover,” Journalof Financial Economics, XX(1988),431– 460. Yermack,David, “ DoCorporations Award Stock OptionsEffectively?” Journal of Financial Economics, XXXIX(1995), 237– 269. ——,“HigherMarket Valuation of Companieswith a SmallBoard of Directors,” Journalof Financial Economics, XL(1996),185– 211. Zeckhauser,Richard J.,andJohn Pound, “ Are LargeShareholders Effective Monitors?An Investigationof ShareOwnership and Corporate Performance,” Asymmetric Information,Corporate Finance andInvestment, R.GlennHub- bard,ed. (Chicago, IL: Universityof Chicago Press, 1990).