33
168 JLEO, V17 N1 Empirical Effects of Performance Contracts: Evidence From China Mary M. Shirley The World Bank Lixin Colin Xu The World Bank Performance contracts (PCs)—contracts signed between the government and state enterprise managers—have been used widely in developing countries. China’s experience with such contracts was one of the largest experiments with contracting in the public sector, affecting hundreds of thousands of state firms, and offered a rare opportunity to explore how PCs work. On average, PCs did not improve performance and may have made it worse. But China’s PCs were not uniformly bad; in fact, PCs improved productivity in slightly more than half of the participants. PC effects were on average negative because of the large losses associated with poorly designed PCs. Successful PCs were those that featured sensible targets, stronger incentives, longer terms, man- agerial bonds, and were in more competitive industries. Selecting managers through bidding was not associated with performance improvement. Good PC features were more often observed in state-owned enterprises (SOEs) under the oversight of local governments, that faced more competition, that were smaller in size, and that had better previous performance. 1. Introduction Performance contracts (PCs) are widely used to reform state-owned enter- prises (SOEs). Since France pioneered their use in the 1970s, PCs have been tried in more than 50 countries (Ghosh, 1997). The World Bank (1995) found 565 PCs in 32 developing countries as of June 1994, where they were We are grateful for the constructive comments of two referees and of Pablo Spiller (the editor). We appreciate the helpful suggestions and comments of Thorsten Beck, George Clarke, Robert Cull, Deon Filmer, Hanan Jacoby, Ross Levine, Scott Masten, Taye Mengistae, Dilip Mookherjee, Itai Sened, Hongling Wang, Oliver White, Chunlin Zhang, and seminar participants of the World Bank. The dataset was kindly provided by the Chinese Academy of Social Science. We acknowledge the financial support of the Ford Foundation in developing this dataset. The sur- vey was carried out by the Institute of Economics under the Chinese Academy of Social Sciences in collaboration with economists from the University of Michigan, University of California, San Diego, and Oxford University. The findings, interpretations, and conclusions expressed in this article are entirely those of the authors. They do not necessarily represent the views of the World Bank, its executive directors, or the countries they represent. 2001 Oxford University Press

Empirical Effects of Performance Contracts: Evidence From China

  • Upload
    m-m

  • View
    214

  • Download
    0

Embed Size (px)

Citation preview

  • 168 JLEO, V17 N1

    Empirical Effects of Performance Contracts:Evidence From China

    Mary M. ShirleyThe World Bank

    Lixin Colin XuThe World Bank

    Performance contracts (PCs)contracts signed between the government andstate enterprise managershave been used widely in developing countries.Chinas experience with such contracts was one of the largest experimentswith contracting in the public sector, affecting hundreds of thousands of statefirms, and offered a rare opportunity to explore how PCs work. On average,PCs did not improve performance and may have made it worse. But ChinasPCs were not uniformly bad; in fact, PCs improved productivity in slightly morethan half of the participants. PC effects were on average negative because ofthe large losses associated with poorly designed PCs. Successful PCs werethose that featured sensible targets, stronger incentives, longer terms, man-agerial bonds, and were in more competitive industries. Selecting managersthrough bidding was not associated with performance improvement. Good PCfeatures were more often observed in state-owned enterprises (SOEs) underthe oversight of local governments, that faced more competition, that weresmaller in size, and that had better previous performance.

    1. IntroductionPerformance contracts (PCs) are widely used to reform state-owned enter-prises (SOEs). Since France pioneered their use in the 1970s, PCs havebeen tried in more than 50 countries (Ghosh, 1997). The World Bank (1995)found 565 PCs in 32 developing countries as of June 1994, where they were

    We are grateful for the constructive comments of two referees and of Pablo Spiller (theeditor). We appreciate the helpful suggestions and comments of Thorsten Beck, George Clarke,Robert Cull, Deon Filmer, Hanan Jacoby, Ross Levine, Scott Masten, Taye Mengistae, DilipMookherjee, Itai Sened, Hongling Wang, Oliver White, Chunlin Zhang, and seminar participantsof the World Bank. The dataset was kindly provided by the Chinese Academy of Social Science.We acknowledge the financial support of the Ford Foundation in developing this dataset. The sur-vey was carried out by the Institute of Economics under the Chinese Academy of Social Sciencesin collaboration with economists from the University of Michigan, University of California,San Diego, and Oxford University. The findings, interpretations, and conclusions expressed inthis article are entirely those of the authors. They do not necessarily represent the views of theWorld Bank, its executive directors, or the countries they represent.

    2001 Oxford University Press

  • Empirical Effects of Performance Contracts 169

    principally used for large utilities and other monopolies, and another 103,000in China, where they were also used for manufacturing SOEs.

    This article analyzes the experience with PCs in China, the country thatexperimented with this tool most extensively. We define PCs broadly as writ-ten agreements between SOE managers, who promise to achieve specifiedtargets in a given, usually short, time frame, and government, which usuallypromises to award achievement with a bonus or other incentive. PCs are thusa variant of pay-for-performance or incentive contracts, which are often usedto motivate managers in private firms. PCs have been suggested as a way toimprove central government agencies (Mookherjee, 1997)1 as well as stateenterprises (Jones, 1991; Ghosh, 1997).

    This article, using a panel dataset, analyzes the experience with PCs inroughly 500 Chinese SOEs. As a natural experiment, Chinas experienceoffered many advantages. First was the large number of contracts. TheChinese experiments with performance contracts could well be the largestexperiment in contracting in the public sector ever conducted. Second,Chinese PCs exhibited interesting variations, differing in term length, tar-gets, intensity of incentives, method of selecting managers, and whetherthe manager posted a bond as a pledge to improve performance. Third, theenterprises that signed PCs were in many different industries with largevariations in size, capital-labor ratio, precontract performance, and the levelof jurisdiction of the government that owned them. Firms in this dataset alsofaced widely different degrees of market competition (Li, 1997).

    This article addresses two questions. First, did PCs work? To what extentdid Chinese PCs enhance firm productivity on average? Second, can PCswork? How was firm performance affected by different PC provisions, inparticular, incentives, targets, bidding, contract length, managerial bonding,and the extent of product market competition? These are important ques-tions given the wide use of PCs and the interest in ways to improve SOEperformance when privatization is not an option.

    Neither economic theory nor empirical evidence provides clear-cut answersto these questions. Principal-agent theory (see Ross, 1973; Stiglitz, 1974;Sappington, 1991) suggests that PCs can be useful in reducing agency prob-lems as long as they can systematically lessen information asymmetry andimprove incentives. The principal (government officials in the case of stateenterprises) can only observe outcomes and cannot measure accurately theeffort expended by the agent (the SOE manager) or distinguish the effects ofeffort from other factors affecting performance (Laffont and Tirole, 1993).A negotiated incentive contract is viewed by its proponents as a device toreveal information and motivate managers to exert effort (Jones, 1991; Ghosh,1997). Proponents also argue that the contract can translate the multipleobjectives of the multiple principals who govern state-owned firms (differentministries, the president, and the legislature) into clear targets measured by

    1. New Zealand, for example, has used incentive contracts for ministries and other govern-ment bodies.

  • 170 The Journal of Law, Economics, & Organization, V17 N1

    specified criteria and weighted to reflect priorities. Moreover, targets canbe set to take into account circumstances where SOE managers have lesscontrol over their firms than comparable managers do in the private sec-tor. For example, performance might be judged against the firms past trendsrather than against an industry standard, to take into account situations wherethe firms performance is substandard because of government-imposed con-straints (such as prohibitions on layoffs, price controls, etc.). By specifyingtargets and evaluating results ex post, the PC is seen by its advocates as away to encourage governments to reduce ex ante controls, giving managersmore freedom and motivation to improve operating efficiency.

    However, several factors could reduce the positive effects on performancewhich PC proponents expect. The information problem may not be solvedby contracts, because of lack of information on SOEs whose shares are nottraded in a stock market and weak accounting and controls in developingcountries. Empirical studies (Jensen and Murphy, 1990; Crocker and Masten,1991) suggest that high-powered incentives are a problem even for privatefirms in developed countries. Owners may fear that because of informationasymmetry they will not be able to measure achievement well and waste theirbonus (Laffont and Tirole, 1993) or reward managerial self-dealing (Shleiferand Vishny, 1994). In SOEs there may also be political barriers to payingsuccessful managers considerably more than ministers or legislators.

    Another problem common to state-owned firms is that some of the multipleprincipals may derive benefits from objectives that run contrary to the PCsgoal of improving efficiency. Politicians may benefit when SOEs maximizethe employment of their constituents; bureaucrats might benefit from SOEactivities that increase their power, prestige, or perks (Shleifer and Vishny,1994). Many of these objectives are likely to be harder to contract on thanprofit maximization. Even if all parties agree, the full set of things they careabout are rarely quantifiable (multitask problem; Holmstrom and Milgrom,1991) and dont lend themselves to automatic or mechanistic types of con-tracts. But less formulaic contracts that leave grounds for interpretation expost, and rely on institutions, such as reputation, arbitration, or courts, toreduce opportunistic behavior (Crocker and Masten, 1991). These institu-tions are likely to be weak in developing countries, and perhaps nonexistentwhere one of the parties is the government.

    Commitment (Williamson, 1985) is thus an especially severe problem forPCs because one of the signatories is the government. There are likely to beno neutral third parties with the power to compel the government to meet itscommitments. Furthermore, in developing countries the institutions that curbarbitrary actions by governments and bind administrations to the promisesof their predecessors, such as checks and balances or reputation, are oftenweak (North and Weingast, 1989; Levy and Spiller, 1996). Managers maynot exert effort if they expect the government will renege on, for example,paying the promised incentive.

    The few existing empirical assessments of PCs reach different conclusions.Song (1988) suggests positive outcomes based on experience in Korea, but

  • Empirical Effects of Performance Contracts 171

    he partly relies on employee and management opinions that could be biased.Trivedi (1990) finds that Indias variant of a PC (memorandum of understand-ing) improved the dialogue between SOE management and government, butdoes not rigorously analyze their impact on firm performance. Ghosh (1997)examines econometrically the experience of 12 Indian companies in imple-menting PCs and finds positive effects; but he does not control for simul-taneous reforms such as liberalization and government disinvestment. Nellis(1988) finds ambiguous PC effects in France and many African countries, inpart because at the time of the study the experience was still recent. Shirleyand Xu (1998) find that PCs did not improve total factor or labor productivityor profitability because they failed to reduce information asymmetry, providesufficiently high-powered incentives, and credibly commit both parties to thegoals of the contract. The sample, however, was small (12 company cases insix developing countries) and limited to natural monopolies.

    This study is, to the best of our knowledge, the first econometric studyto systematically evaluate the productivity effects of PCs in China. Moreimportantly, it is also the first study to relate PC effects to contract provisionsalong three dimensions suggested by the contracting literature: information,incentives, and commitment.

    Our primary findings are three. First, PCs on average were not signif-icantly correlated with improvements in productivity in a large sample ofSOEs in China. In fact, on average PCs were found to have a negativeand significant correlation with productivity when the endogenous natureof PC participation was taken into account. Our second finding is that PCscan improve productivity when they simultaneously specify sensible targets,offer strong incentives, and signal commitment, especially in a competitiveenvironment. Thus productivity was higher in more competitive firms whosePCs had higher wage incentives, longer terms, managerial bonds, and profit-oriented targets. In fact, slightly more than half of the PCs were associ-ated with productivity improvements. Bidding, in contrast, was not associatedwith productivity improvement, perhaps reflecting the design of the auctionsor weak enforcement of bidding contracts. Finally, good PC features weremore often observed in those SOEs that were under the oversight of localgovernment, faced more competition, were smaller in size, and had betterprevious performance.

    Our analysis differs from other studies of Chinese contracts. Byrd (1991)describes in rich detail the principal types of performance contracts and sug-gests some of the main advantages of PCs over the traditional mode ofgovernment oversight. He argues that the main problems with PCs are thestrong bargaining power of managers, not enough risk-bearing on the partof firms, ambiguous ownership type, and noncredibility of contracts. Byrd(1991), however, does not offer systematic evidence about the effectivenessof PCs. Groves et al. (1995) examine the contractual provisions that affectSOE managers (such as the length of the contract, management turnover, andchanging management pay sensitivity) and find that these provisions are con-sistent with a well-functioning managerial labor market. Groves et al. (1995)

  • 172 The Journal of Law, Economics, & Organization, V17 N1

    also analyze the determinants of many other provisions, but do not systemat-ically assess how they affect productivity, with the exception of the impact ofmanagement turnover. In contrast, our focus is on the quantitative importanceof alternative specifications of performance contractssuch as profit orien-tation, managerial bond, firm-level pay sensitivity, competition, the length ofthe contract termmost of which Groves et al. did not examine.

    The next section briefly describes the implementation of PCs in China.Section 3 presents our hypotheses about the effects of PCs. Section 4 inves-tigates the effects of PCs on performance of our sample and compares theeffects of alternative provisions. The final section draws policy implicationsfrom our findings.

    2. Performance Contracts in ChinaThe Chinese government began to experiment with PCs for SOEs in the mid-1980s.2 Not until 1986, however, did the government implement PCs on asignificant scale; in fact, PCs became the national policy for reforming SOEsfrom 1987 to 1994.3 In our dataset, the share of state enterprises under PCsgrew from 8% in 1986 to 42% in 1987; it then skyrocketed to 88% by 1989.

    We defined a firm as being under a PC if its response to the question-naire indicated the existence of a contract that the manager had signed withthe government. We captured the differences between the types of contractsby analyzing the impact of the PC provisions summarized in Table 1. Oneprovision was duration of contracts (LENGTH), which ranged from one toeight years. Seventy-one percent of the contracts explicitly specified a wageelasticity (WELASTICITY) that was set ex ante and remained constantthroughout the length of the contract; WELASTICITY is imputed from thequestionnaire responses as to the percentage by which total wages wouldincrease when profits increased by 1 percent.4 WELASTICITY varied from0.1 for the 5th percentile to 0.8 for the 95th percentile, with a median of0.6. The managers in charge of implementing the contracts could be selectedby a number of methods: by the government, by election, or by bidding.5

    2. About Chinese SOE reforms, see Johnson (1990), Jefferson and Xu (1991), Groves et al.(1994, 1995), Jefferson and Rawski (1994), Gordon and Li (1995), Li (1997), and Xu (2000).

    3. See Byrd (1991), and Ghosh (1997), Lin, Cai, and Li (1997) for more details about theimplementation of PCs in China.

    4. This incentive measure differs from a related measure of wage incentive used in recentstudies of Chinese SOEs: bonus/total wage bill (Groves et al., 1994 for instance). This bonusshare, observed both before and after PC adoption, was more susceptible to simultaneity bias.W.ELASTICITY, in contrast, was set ex ante and remained constant throughout the period, andthus is less susceptible to simultaneity bias. We view the bonus/wage bill ratio as a result ofthese reforms, and to avoid simultaneity bias, we do not control for the bonus ratio variable.Note that we have controlled for a more complete list of reforms than similar studies usingthe same data (Groves et al., 1994, 1995; Li, 1997): for instance, we control for another wageincentivethe managerial discretion to determine employee wagebut other studies do not.

    5. This category also includes those firms whose managers were directly appointed by thecontract signer.

  • Empirical Effects of Performance Contracts 173

    Table 1. Definitions and Descriptive Statistics of PC Features

    MeanDefinitions (Standard Error)

    PC Dummy variable: one if a firm was under a PC

    Conditioned on that a firm was under a PC:

    WELASTICITY Firm-level wage elasticity, the percentage 525increase of total wage bill of the firm whenthe profit increase by 1 percent 314

    BIDINCUMBENT Dummy variable that is one if the manager 149who won the bid for the PC was anincumbent manager. 356

    BIDNEWCEO Dummy variable that is one if the manager 020who won the bid for the PCwas a new manager. 136

    LENGTH The length of the contract (in years). 28581523

    PROFIT Dummy variable that is one if the primary 427target of a PC was profit. 495

    BOND Dummy variable that is one if the manager of 285a PC posted a bond as a guaranteefor performance targets. 452

    Bidding was used to select managers for 17% of the contracts signed. Themore local the government in general, the more likely it was that biddingwas employed to select the manager for the contract. Among firms whosemanagers were chosen through bidding, roughly 13% (11 firms) appointednew managers (BIDNEWCEO) and 87% appointed incumbent managers(BIDINCUMBENT).

    In some cases (16% of the companies) the manager posted a bond that wasforfeited if he failed to achieve the contracts goals (BOND). As with bidding,posting a bond was more likely, the more local the government authority.The amounts of the bonds were nontrivial, averaging about 32,400 yuan,several times a CEOs annual wages. On average, managers of SOEs underthe central government, some of the largest firms, posted the smallest amount,followed by managers of county firms, which are some of the smallest SOEs.

    The targets specified by the PCs also varied in the weight given to outputgoals, tax receipts, and profits.6 Almost 42% of firms in our sample reportedthat their PCs primary target was total before-tax profits (PROFIT hereafter),another 28% reported profits plus taxes remitted to the government as theprimary PC goal, and 26%, output quantity and value and labor productivity.The choice of primary target also varied by oversight authority. The PC

    6. We know from the questionnaires whether the primary target of the contracts focused onprofitability (before tax), taxes, or outputs. Although it might be desirable to let profit be profitafter tax, the questionnaire does not make this feasible.

  • 174 The Journal of Law, Economics, & Organization, V17 N1

    targets for firms governed by municipal and county governments tilted morefrequently toward profit than PCs for firms governed by the two upper levelsof government.

    3. PCs Effects, Information, Incentive, and CommitmentWe draw on agency theory (Ross, 1973; Grossman and Hart, 1983; Laffontand Tirole, 1993) and contracting literature (Williamson, 1985) to shed lighton how the PC features might affect firm performance. A PC can be under-stood as an outcome of a game between managers with disutility of effortand a government with imperfect information about the managers effort.Theory suggests that PCs will improve performance when they reduce theinformation advantage enjoyed by managers, increase managers incentivesto overcome their disutility of effort, and strengthen the governments andthe firms commitment to honor the contracts.

    Should PCs not reduce information asymmetry, we expect that managerswill exploit the opportunity to shirk, perhaps by negotiating lower targetsthan they could potentially achieve, and performance will not improve. Inour empirical analysis, the reduction of information asymmetry is representedby bidding and a competitive environment. First, we expect that bidding, byproviding government with more information about the firms and potentialmanagers [as well as by adjusting the managers incentive to the condi-tions of the firm, as shown by Nalebuff and Stiglitz (1983) and McAfeeand McMillan (1987)], will reduce information asymmetry, thus allowingthe government to use strong performance incentives and reduce shirking.Hence bidding should be associated with better performance. As suggestedby Groves et al. (1995), we also try to distinguish the effects when thebid was won by the incumbent or by a new manager. Incumbent managersinformation advantage allows them to cherry-pick firms with better prospects.Moreover, new managers may be at a disadvantage in building up a workingrelationship with labor. Consequently, we expect firms whose contract bidswere won by incumbent managers to perform better after PC adoption thanthose won by nonincumbents.

    Second, we expect that PCs signed with firms facing more competition areassociated with less information asymmetry and hence have higher produc-tivity. In competitive sectors there is more information about the circum-stances in which the manager operates competitive markets provide aricher information base on which to write contracts (Holmstrom and Tirole,1989:96). As a result, shirking will be more evident in competitive firmsrelative to firms with more market power, and performance measures will bemore meaningful there. Moreover, contracts with competitive firms may bemore formulaic and easier to enforce, thus less likely to be exposed to rene-gotiations (Crocker and Masten, 1991). In the empirical analysis, competitionwas measured by the markup ratio (based on Li, 1997; see appendix).

    Not surprisingly, we expect incentives to raise productivity. We also expectthat the strength of the incentive is likely to be constrained by information

  • Empirical Effects of Performance Contracts 175

    asymmetry. Since incentives to SOEs have a cost to societyfor instance,government revenue would dropgovernments will try to ensure that theincentive payment does not exceed the social gain from increased firm effi-ciency. Hence, when information asymmetry is severe, government will tendto set the incentive too low to motivate much improvement in performance.Thus, without solving the information problem, introducing incentives canbring about only limited change (Laffont and Tirole, 1993). In our analy-sis, the incentive we investigate is firm-level wage elasticity. This incentivedevice has limitations. First, it may be distributed equally among employeesin a context where there are limited ways to punish shirking (e.g., there areoften restrictions on firing SOE workers), which may reduce its incentiveeffects. Second, if the incentives for investment are relatively weak, a strongpush for profit reduces funds available for long-term investment and thus thelong-term growth rate. Finally, since it is aimed only at workers, it will nothave a sustained impact unless the manager is also motivated to improvemanagement practices and take other steps to enhance productivity.

    A second way that incentive design may affect productivity is through set-ting PC targets. Profit targets align firms incentives better toward improvingproductivity than other targets such as taxes and outputs. This hypothesis isimplied by the multitasking perspective. When the primary target is output,the firm may overproduce low-quality or high-cost products without increas-ing profits or efficiency; when it is tax, the firm may sacrifice investment orother expenditures important to long-term growth to be able to pay the tax.Profit targets, in contrast, reward with equal weights both output increase andcost reduction.7 We thus expect that PCs focusing on profitability are morelikely to improve performance than those focusing on tax or output goals.8

    In a related vein, we expect profit targets in PCs will be more produc-tivity enhancing for firms facing less competition if they are constrained inincreasing prices to meet the profit target. Again, this could be best under-stood from the multitasking point of view. Since managers of firms facingless competition can afford to spend more effort in pursuing nonproductivity-related activitiesas Hicks famously put it, the most important monopolyprofit is a quiet lifetheir effort devoted to raising productivity is lowerthan that of managers of competitive firms. With an exogenous increase ofthe weight on profits (as a proxy for productivity), firms with more marketpower but limited scope for raising prices have more room for improvingproductivity by shifting effort from private-benefit activities to productivity-

    7. This idea of balanced incentives when there are multiple tasks can be traced to the seminalpaper of multitask analysis of Holmstrom and Milgrom (1991).

    8. Many economists emphasize the importance of a profit orientation in reforming SOEsin China. Lin, Cai, and Li (1997), for instance, suggest that profitability could be a sufficientstatistic for performance in competitive industries without soft budget constraints. Implicit in theideas of market socialism was also the belief that SOEs could perform well when they pursuedprofit goals.

  • 176 The Journal of Law, Economics, & Organization, V17 N1

    enhancing activities.9 Firms with monopoly power in China faced price con-trols on much of their production, and we expect that they would have alreadyincreased uncontrolled prices as much as possible before the profit target wasimposed.

    Our third hypothesis is that PCs will improve performance if they elicitboth the governments and the firms commitment. When managers are notcommittedperhaps because they expect to use their information advantageto bargain down the targets ex postthen ex ante they will exert only enougheffort to fulfill the anticipated bargained-down targets. In this case a PC willfail to improve the managers incentives and fail to turn around the firmsperformance. Alternatively, when government is not committed, it will failto enforce the contract and/or will renege on paying the promised incentives.Since government is both a signatory and the enforcer of the contract, itis especially important that its commitment be credible. If the manager andemployees do not believe that the government will honor the contract and paythe incentive if they meet their targets, ex ante they will hesitate to work hardor put in optimal investment for fear of ex post expropriation. Unfortunatelyour data do not permit us to test this hypothesis by assessing measures ofcommitment such as contract enforcement, reputation, or the like. Insteadwe proxy commitment by the length of the contract and by bonding. Longer-term contracts signal to managers that the government is more committed;managers then may invest with a longer time horizon and therefore expandthe production possibility frontier. As for bonding, it has been suggested as afirst-best solution to agency problems (Becker and Stigler, 1974). Managers,concerned with the possibility of losing their bonds if performance tends tofall under the targets, will naturally work harder. We thus expect that longercontracts and the use of bonds will be associated with greater improvementsin performance.

    4. Effects of PC Participation and ProvisionsWe examined how PCs affect the productivity of SOEs using a panel datasetconsisting of 769 firms from 1980 to 1989 located in four provinces ofChina from A Survey of Chinese SOEs: 19801989. (See the data appendixfor more details.) Most but not all of the firms in our sample signed a PC.

    9. A slightly more rigorous approach might be more convincing. Suppose the objectivefunction of a manager is e1+Be2Ce1 + e2, where is the weight given to profit, is profit, e1 is effort devoted to raising profits, B is private benefits to the managers for effortsspent on objectives other than raising profits (e2 is the weight given to private benefits,and C is the total psychic costs of efforts. Both types of efforts are assumed to be perfectlysubstitutable. Profit targeting is represented by a higher . Let M be market power. Then should increase with M : competitive firms experience a larger cost or a smaller benefit frompursuing private-benefit activities, since they are more likely to become insolvent, or they have tocut wages with lower profits. With the standard assumptions of concave B, and convex C (forinstance, C = ce1 + e22 = e1/21 B = be1/22 it is fairly easy to establish that e1/M > 0from solving the two first-order conditions of e1 and e2.

  • Empirical Effects of Performance Contracts 177

    We assume the following CobbDouglas production function (the use of atranslog production function produced quite similar results about PC effects):

    yit = j0 +L lnLit +k ln kit +ZZit +RRit +PCPCit +i+it (1)

    where i indexes firm, j indexes industry, and t indicates year. To take intoaccount the potential differences in technology for firms in different indus-tries, we decomposed the firms into four industries: chemical, light, machine,and materials.10 The variables are defined as follows (see the data appendixfor details of variable construction and the associated deflators):

    yit: log (value added per worker) for firm i at year t.j : industry-specific total factor productivity (TFP) level.11

    lnLit: the number of employees, excluding those absent for more than halfa year.

    ln kit: capital per worker, with capital constructed by the perpetual inven-tory method.

    ZZit: captures the effects of other controls such as provincial-specificgrowth rates, and the industry-year dummies (to control for industry-wide shocks such as overall credit cycle effects and other businesscycle-related effects, and industry-wide technological progress).12

    Rit: other major reforms and changes in the market environment thatwere applied to both PC and non-PC participants, including

    marginal profit retention rates (MARGINAL RATE), managerial wage discretion (i.e., letting managers determine

    employee wages) (WDISCRETION), delegating production autonomy to managers (AUTONOMY), reducing the markup ratio (MARKUP), and the presence of new managers (NEW CEO). [A manager is

    defined to be a new manager if he was still incumbent at thetime of interview (1990) and appointed after 1980.]

    PC: either the PC dummy or a vector of PC variables interacted with thePC dummy.

    PC : The corresponding coefficient(s) of the PC variable(s).i: fixed effects for firm i, to capture all firm-specific time invariant

    factors that account for productivity.it: time-varying error term for firm i at year t.

    10. We follow Li (1997), which uses the same dataset. The data contain firms in 36 two-digit industries, but many industries contain too few observations to justify more disaggregatedtreatment.

    11. We have also tried allowing industry-specific coefficients for capital and labor as well asindustry-specific TFP levels. The results about the PC effectsboth for PC status and for PCprovisionsremain quite similar.

    12. Statistical testsfor both our FE model and our OLS modelsshow that the year effectsacross different industries are different.

  • 178 The Journal of Law, Economics, & Organization, V17 N1

    This institutionalized production function assumes that, besides physicalinputs, a firms output also depends on how production is organized, itsinternal incentives, and the economic environment (See Jensen and Meckling,1979; Coase, 1991; McMillan et al., 1989; and Holmstrom and Tirole, 1989).Note that we include other reforms in the production function. It is impor-tant to bear in mind that these reforms were not part of the contents of PCs.They had different timing and cross-sectional incidence than the PC vari-ables, and were regarded as different reform measures by the Chinese gov-ernment. Since other studies using the same data as this article have foundsuch reforms to be important in explaining performance (Groves et al., 1994;Li, 1997; Xu, 2000), one has to control for these reforms to isolate the PCeffects.13

    4.1 Empirical IssuesPerhaps the most important issue is the potential endogeneity of the PCvariables (and the other reforms). Since the selection of firms to participate inPCs was not random, it could be that PC participants are systematically better(worse) performers than non-PC participants for reasons that have nothing todo with PC participation. The most likely bias is upward, since managers ofbetter SOEs would have a reason to push for inclusion. It could also be thatboth PC participation and other reforms are associated with some excludedvariable that is determining changes in productivity. In other words, PCit andRit might be correlated with i and it .

    We use a three-pronged strategy to deal with this issue. First, since endo-geneity stems largely from missing information (i.e., omitted-variable bias),we explore as many reasonable determinants of productivity as possible. Asspelled out earlier, we have included province-specific TFP growth, province-year dummies, and (in sensitivity checks) province-year-specific time dum-mies; these variables should filter out a significant portion of productivityimprovements unrelated to reforms and PCs. Second, in all specifications,we use firm dummies to control for endogeneity due to firm heterogeneity.Third, we try multiple techniques, specified below, to deal with the possiblecorrelation between RitPCit , and it .

    We primarily employ the fixed-effects two-stage least squares (FE2SLS)method to deal with any remaining endogeneity problem. Since a large por-tion of the variation in PC provisions across firms was due to local govern-ments discretion in implementing PCs, our instruments include (but are notlimited to) the provincial means of PC variables. The assumption is that theregional mean of PC provisions will be correlated with the firms PC provi-sions, but not with firm-level productivity. (To avoid superficial correlation,in computing provincial means we excluded the firm under consideration.)Table 2 presents the R2s for regressions in which a PC variable for firm i is

    13. This is the first study about PCsas far as we knowthat controls for other reformsin disentangling the PC effects. See, for instance, Nellis (1988), Song (1988), Trivedi (1990),Ghosh (1997), and Shirley and Xu (1998).

  • Empirical Effects of Performance Contracts 179

    Table 2. Explanatory Power of Provincial Patterns ofPC Variables on Firm-Level PC Features

    PCj R2

    PC 0667WELASTICITY 0415BIDINCUMBENT 0127BIDNEWCEO 0009a

    TERM 0234PROFIT 0473BOND 0318

    The regression is PCjit= 0 + 1 (provincial mean of PCjit ) where

    j means PC status or a PC feature.aThe coefficient is low because only 11 SOEs had this feature at theend of the period.

    regressed on the provincial mean of that PC variable. In general the R2s arelarge. These instruments thus satisfy at least one of the crucial requirementsfor good instrumental variables: explanatory power.14

    Besides the regional means of PC provisions, our instruments also include(i) year times oversight by central, provincial, municipal, and county gov-ernments, and times large or small size15 (since the trend in implementingreforms and PCs across government jurisdictions and sizes should affect thePC provisions of a firm), (ii) the once-lagged other reforms and PC pro-visions (Ri t1 and PCi t1) as indicators of the existing incentive structuresof the firm, and (iii) the once-lagged share of engineers in total numberof employees as an indicator of technology which might affect contractualchoices. The use of these instruments is based on the assumption that afirms oversight status, historically determined size classification, previousreforms, PC status, and technology correlate with contractual choices but,after controlling for capital, labor, industry-year dummies, and firm dummies,are not correlated with current productivity shocks. Since it is not obviousthat these variables would be uncorrelated with productivity, we conduct anover identifying restrictions test. As will be shown later, the overidentify-ing restrictions test suggests that these are valid instruments, that is, theyare uncorrelated with productivity residuals. As another way to ensure thatour identifying restrictions are valid, and in light of the fact that we relylargely on the regional means of PCit as additional instruments, we haveexamined whether the PC effects are robust even if we explicitly controlfor province-year-specific productivity effects. The estimates of PC effects

    14. Staiger and Stock (1994) have emphasized the need to check if the instruments aresignificant predictors for the endogenous variables. If not, the bias of instrumental estimatesmay also be quite large.

    15. Here large and small sizes are historical classification of SOEs. Large-size firms onaverage had 3600 employees, while small-size firms had 560. These dummies should not captureeconomies of scale because we have already controlled for ln(number of employees).

  • 180 The Journal of Law, Economics, & Organization, V17 N1

    when we control for province-year dummies are very close to that (or those)we estimate when not controlling for these dummies, and hence we do notreport the specifications that control for these dummies. The finding shouldrender more credence to the use of provincial means of Rit and PCit as animportant part of the instrumental variables for identification. Finally, sincethe presumption of endogenous PCs could be costly if wrong, we conduct theHausmanWu test for the exogeneity of the PC variable(s), as we describein more detail below.

    Besides the FE2SLS method, we also employ two other methods to checkthe sensitivity of the results with respect to the way we deal with endogeneity.This may be important because the large number of endogenous variablesmay make the results sensitive to the way we deal with endogeneity. Our twoother methods are dynamic panel estimation a la Arellano and Bond (1991),and the split-sample instrumental variable approach (Angrist and Krueger,1995), the details of which will be explained later.16

    4.2 Effects of PC Participation: Benchmark ResultsFor the empirical application, we deleted any observation that was miss-ing the dependent variable, or was missing the capital-labor ratio, or hadan unbalanced panel.17 This leaves us with a balanced panel of 560 SOEs,roughly 60 of which did not participate in PCs by the end of the sampleperiod. Table 3 reports the benchmark results of how PCs on average affectedproductivity.

    As mentioned, since the participation in performance contracts could beselective and thus endogenous, we conduct the HausmanWu test for theexogeneity of PCit to see if the PC dummy might be endogenous.

    18 In imple-menting the test, PCit and other reforms are all treated as endogenous. Themaintained instruments include the once-lagged other reform variables, themean percentage of firms adopting PCs among other firms in the province,a once-lagged share of technical workers, and year times several dummies(dummies for large size, and small size, and oversight by central, provincial,

    16. An interesting article by Athey and Stern (1998) may be useful as an alternative approachfor dealing with endogeneity. They suggest a structural approach to estimate a simultaneousequation system describing productivity and the demand for organizational design practices.This method is more ambitious in data requirements and aims for richer strucuture (such as thecomplementarity of organization practices) than this article wants. An implementation of thisapproach is therefore beyond the scope of this article.

    17. Roughly 9% of observations dropped because of an unbalanced panel. They involvelarger measurement errors because the deflator for the output and capital stock were imputedbased on industry-year cells. In addition, markup ratio, an important feature with which PCstatus will interact, could not be reliably estimated. The estimation of markup ratio requires thedata about each years price information on outputs and inputs (see the data appendix).

    18. Under the null hypothesis that the PC dummy is uncorrelated with it , both the FE andthe FE2SLS will be consistent, though the FE will be efficient but the FE2SLS will not. Underthe alternative hypothesis, the FE is inconsistent but the FE2SLS is consistent.

  • Empirical Effects of Performance Contracts 181

    Table 3. The Effects of PC: Basic Results

    (1) FE2SLSa (2) FE (3) FE

    lnK/L 0133 0132 0132510 528 529

    lnL 0137 0100 0099206 141 139

    Markup 0512 0496 0496709 668 668

    Marginal rate 0308 0086 0087233 143 145

    Autonomy 0082 0063 0063180 252 251

    Wdiscretion year 0030 0026 0026230 225 226

    New CEO year 0041 0045 0046431 525 527

    PC 0196 0013 0007241 038 022

    Terminating PC 0039062

    N 5094 5660 5660R2 046 047 047

    , and present statistical significance at the 10%, 5%, and 1% levels. The parentheses containt-statistics. The standard errors are White adjusted. Other controlled variables: industry dummies, industry-year dummies, province dummies year, and missing indicators for marginal profit retention rates.aPC (and other reforms when applicable) are all treated as endogenous. The instruments include (theonce-lagged other reform variables when other reforms are included and treated as endogenous), themean percentage of adopting PCs among other firms in the province, a once-lagged share of technicalworkers, year (the dummies of large size, the dummy of small size, the dummies indicating the oversightstatus being central, or being provincial, or being municipal, or being county). The HausmanWu test rejectsthe exogeneity of PC (with a p-value of .008). The overidentifying restrictions cannot be rejected (with ap-value of .973). The first-stage regressions almost all have high R2 (more than 0.40), and F statistics forexcluded instruments all have p-value of .0000.

    municipal, or county authority). The HausmanWu test rejects the exogene-ity of the PC dummy (with a p-value of .008). Since we have more instru-ments than endogenous variables, we test the validity of our instruments bythe overidentifying restrictions test (Davidson and MacKinnon, 1993:236).The p-value of the test is .973, so we cannot reject the null hypothesis thatthe instruments are valid. Moreover, the first-stage regressions all have highR2 values (e.g., almost all are more than 0.40) and p-values of 0.000 for theF statistics associated with the excluded instruments.

    Since the exogeneity of PCit is rejected, we rely on the FE2SLS results incolumn 1. On average, the PCs were significantly associated with a produc-tivity disadvantage of roughly 20 percentage points. The rest of the resultsare consistent with previous literature: more competitive firms had higherproductivity; higher marginal profit retention rates had a positive correla-tion with productivity; and firms with production autonomy, with managerial

  • 182 The Journal of Law, Economics, & Organization, V17 N1

    wage discretion, and new CEOs had higher productivity (Groves et al., 1994,1995; Li, 1997; Xu, 2000).19

    The comparison of FE2SLS with the FE specifications (in column 2 ofTable 3) is illuminating. In the FE estimation, PCs were not significantlyassociated with productivity: the coefficient is 0.013 (t = 038). Since thePC effect in FE2SLS was negative and significant, this suggests that PCparticipation was positively correlated with the productivity residual (it).Thus it appears that SOEs facing favorable productivity shocks were morelikely to take advantage of the incentive provision of PCs.

    Our finding that PCs on average are not associated with higher productiv-ity could be due to a delay in restructuring. Many of the PCs did not expirewithin our sample period, and if we assume that CEOs use the early yearsof the PC to restructure, the positive PC effects may come in the later periodof the PC. To check this possibility, we create a dummy variable that is onewhen the PC expired before or in the last year of our sample period. Inour dataset, slightly less than 100 SOEs had terminal PCs, that is, contractsthat expired by 1989. Adding an additional term of PC times the terminalPC dummy in column 3, and using the FE specification, we find that thedifference in PC effects for terminating and nonterminating PCs are statisti-cally indistinguishable. Thus the delay in the restructuring hypothesis doesnot appear to be plausible.

    Some observers have been concerned that PCs might have provisions sim-ilar to the other reforms and thus by controlling for other reforms we mightreduce the possibility that PCs have a positive correlation with productivity.To check this possibility, we rerun the column 13 specifications excludingthe other reforms. The resultsunreportedare indeed different. When theendogeneity issue is ignored, PCs on average had a positive and statisticallysignificant effect of 6 to 8 percentage points. However, it is worth repeatingthat it is important to isolate the effects of other reforms, which had distincttiming and content from the PCs, if we are to judge the PCs themselves.Moreover, once endogeneity is considered, PCs had a negative though statis-tically insignificant effect of roughly 11 percentage points even when otherreforms are excluded from the regression.

    19. We also experimented with the dynamic panel method in the spirit of Arellano and Bond(1991). First, we take a first difference to filter out the firm-specific fixed effects. Then theresidual becomes (it i t1), the endogenous regressors (say Yit) become (Yit Yi t1), andpotentially valid instruments include Yi t2 Yi t3, and so on. Then we make use of the momentconditions E it i t1yi t1s " s = 12 t 2 and the generalized method of momentsto estimate the system. With this method, the coefficient of PCit is negative (0132), butstatistically insignificant (t = 088). The insignificance is not very surprising, since with manydummy variables, first-differencing may yield too little variation and result in a weak instrumentproblem. Nevertheless, it is comforting to know that completely different identifying restrictionsshow similar negative PC effects as with the FE2SLS specification. We did not use the dynamicpanel methods to identify the effects of PC provisions because of weak instruments. All thePC provisions had identical timings. After first-differencing, the PC provisions possess fewvariations and moreover become quite collinear, and dynamic panel methods we tried provedunable to identify the effects of each PC provision.

  • Empirical Effects of Performance Contracts 183

    In sum, our estimates suggested that PCs on average did not improveproductivity. This finding is similar to that of Shirley and Xu (1998) forPCs in market economies. However, we still wish to know if PCs can workwhen properly designed. We next test whether some PC designs did improveproductivity, and whether the ways different PCs affected productivity areconsistent with our previous conjectures.

    4.3 Effects of PC ProvisionsTo assess the effects of the specific provisions, we now allow each PC vari-able to affect productivity. We shall only report the specifications that controlfor the other reforms since the results excluding the other reforms asregressors give quite similar results for a variety of specifications. To havea parsimonious specification, we experimented with allowing the PC vari-ables to affect both the productivity level and growth rate, and only reportthe more significant result.20 In the end, most features entered affecting lev-els only, with only the dummy for posting managerial bonds and two of theother reforms (wage discretion and new CEO) entering with a growth rateeffect (i.e., we allow the dummy to interact with the year).

    Since the PC dummy was found to be endogenous in the last subsec-tion, we need to consider the possibility that the PC provisions are alsoendogenous. (Note that we are only considering the potential correlationof the PC features and other reforms with it since we have already con-trolled for the firm fixed effects.) The HausmanWu test indeed finds bothother reforms and PC features to be endogenous. Thus again we rely onFE2SLS; the instruments include once-lagged other reforms and PC fea-tures, provincial means of PC features, and once-lagged share of engineersin total number of employees, and year times several dummies (central gov-ernance, provincial governance, municipal governance, county governance,large size, small size). Because we have more instruments than endoge-nous variables, we can also explicitly test our overidentifying restrictions(Davidson and MacKinnon, 1993:236). The null hypothesis is that the set ofinstruments are valid instruments (i.e., they are not correlated with it).

    21 Thep-value of the test statistic is .346, hence, we cannot reject the null hypothesisthat our instruments are valid. Moreover, since it is important that instrumentvariables have predictive power for the endogenous variables (Bound, Jaeger,and Baker, 1993; Stager and Stock, 1994), we also checked the R2s of first-stage regressions. They are uniformly high: ranging from 0.32 to 0.96, with

    20. Where only the level or rate effect was found to be statistically significant, we kept onlythe significant term in the final specification. If neither was significant we kept at least one typeof effect for each PC provision, choosing whichever was closer to statistical significance. Thereare enough PC-participants in our sample with sufficient post-PC histories to identify the rateeffects. There were 15 firms with 6 or more years of experience, 16 firms with 5 years, 27 firmswith 4 years, 265 firms with 3 years, and 313 firms with 2 years.

    21. By the test, nR2, with R2 being generated by regressing the residual of the 2SLS tothe set of instruments, follows a chi-squired distribution with a degree of freedom equal to thenumber of extra instruments.

  • 184 The Journal of Law, Economics, & Organization, V17 N1

    the p-values of the F -statistics for the excluded instruments being 0.0000.Using these instruments, we find that almost every individual PC feature isshown to be endogenous, with BIDINCUMBENT and BOND YEAR atthe margin. To be conservative, we thus treat all PC features (along with theother reforms) as endogenous.

    The fixed-effects 2SLS results using the aforementioned instruments arereported in column 1 of Table 4. The results favor our theoretical predictions.

    Consistent with our hypothesis, firms with a higher wage elasticity per-formed better: the coefficient of employee wage elasticity is positive andsignificant. A 1 SD increase (0.31) would raise productivity by 28 percent-age points (FE2SLS).

    Bidding won by the incumbent manager is not significantly correlated withproductivity. This implies that, holding constant other PC features, new PCswon by incumbent managers would largely be indistinguishable from otherPCs without bidding. Consistent with the hypothesis that asymmetric infor-mation led incumbent managers to cherry-pick firms with good prospects,PCs with new CEOs winning the bid had much lower productivity, 39 per-centage points lower than PCs with an incumbent CEO winning the bid.This finding is consistent with the previous finding of Groves et al. (1995).22

    The failure of bidding to raise productivity may be due to the poor designof auctions. For instance, the process may not be transparent. This expla-nation is supported by the fact that incumbent managers won 83 of the 94bids. Alternatively, bidding may only succeed when contract enforcement iseffective, a condition hard to achieve when government is one of the parties.

    Targeting profits in PCs is positively correlated with productivity. Theimplied effect is 38 percentage points. This result is also consistent witha vast literature on the effects of alternative compensation schemes, sum-marized by Prendergast (1999). He states that, In summary, firms getwhat they pay for; by emphasizing certain outcomes in pay, they make thoseoutcomes more likely to occur.

    The length of PCs is positively and significantly correlated with product-ivity. A 1 SD increase of LENGTH (1.5 years) would increase productivityby 23 percentage points. Equally interesting, the posting of a managerialbond is positively associated with productivity, though not in a statisticallysignificant way (t = 146), implying a large gain in the productivity growthrate of 5.6 percentage points. These results support our hypothesis that PCfeatures signaling commitment should improve performance.

    PCs appeared to work far better where competition reins, and the effectis strong. For firms that signed PCs without profit targets, a 1 SD dropin the markup ratio23 implies an increase in productivity of more than 200

    22. The slight difference between our results and those of Groves et al. (1995) can probablybe attributed to differences in measuring the outcome and the specifications. They used relativeperformance of a firm within an industry, and they did not control for capital, labor, and otherreforms.

    23. Note that the markup ratio is standardized.

  • Empirical Effects of Performance Contracts 185

    Table 4. The Effects of PC Provisions: Dependent Variable = log(Per-Employee ValueAdded)

    (1) FE2SLSa (2) FE (3) SSIV (4) FE

    lnK/L 0151 0129 0118 0130403 516 320 520

    lnL 0334 0102 0426 0092264 144 363 131

    Markup 0371 0484 0328 0487424 661 251 661

    Marginal rate 0227 0071 0561 0077133 117 406 126

    Autonomy 0079 0067 0040 0066127 267 060 263

    Wdiscretion year 0011 0021 0033 0022050 180 149 183

    New CEO year 0030 0043 0038 0044219 495 282 512

    PC 1599 0125 0674 0151377 218 202 248

    PC Welasticity 0897 0203 0797 0187431 328 449 285

    PC bidincumbent 0009 0089 0029 0091008 193 027 189

    PC bidnewCEO 0397 0393 0757 0398194 438 455 444

    PC Length 0151 0012 0090 0018340 105 242 148

    PC Profit target 0384 0072 0582 0079332 196 391 215

    PC 1(bond) year 0056 0006 0087 0003146 027 207 011

    PC markup 2049 0087 0847 0108313 247 174 295

    PC profit target markup 1179 0106 3016 0110165 297 357 299

    PC terminating PC 0192109

    Welasticity terminating PC 0115061

    bidincumbent terminating PC 0064037

    Length terminating PC 0073140

    Continued

  • 186 The Journal of Law, Economics, & Organization, V17 N1

    Table 4. Continued

    (1) FE2SLSa (2) FE (3) SSIV (4) FE

    1(bond) year terminating PC 0064091

    PC markup terminating PC 0130190

    Profit target markup 0004 terminating PC 003

    Observations 5094 5660 2539 5660R2 0020 0477 0490 0488

    present statistical significance at the 5% and 1% levels. The parentheses contain t-statistics. The standard errorsare White-adjusted. Other controlled variables: industry dummies, industry-year dummies, province dummies year,and missing indicators for marginal profit retention rates.aPC and other reforms are all treated as endogenous. The instruments include the once-lagged other reformvariables, the provincial means of PC variables among other firms in the province, a once-lagged share of technicalworkers, the dummies of large size, of small size, the dummies indicating the oversight status of the firm (central,provincial, municipal, and county). The HausmanWu test rejects the exogeneity of all PC variables except that ofbid incumbent and of 1(bond) year (with a cutoff of p-value of .10). The overidentifying restrictions cannot berejected (with a p-value of .346). The R2s of the first stage range from 0.32 to 0.96, with the p-value of the F -statisticsfor the excluded instruments being 0.0000.

    percentage points. This finding is again consonant with Shirley and Xus(1998) finding that PCs did not improve productivity in natural monopolies(for which the markup ratio should be relatively high). With as many endoge-nous variables as we have, FE2SLS would probably struggle to distinguishthe effects of each component. Nevertheless, it is perhaps safe to concludethat competition must be the most important ingredient in making PCs work.

    Consistent with our conjectureprofit targeting in PCs would lead toa higher performance gain in firms with more market powerthe interac-tion term, PC PROFIT MARKUP, is positive and significant. Thusprofit targeting was more effective in firms with more market power. Indeed,a 1 SD increase of MARKUP raises the effect of profit targeting by roughly118 percentage points.

    The comparison of the FE2SLS results with those of FE (column 2) isagain interesting. Not surprisingly, the FE2SLS results are more consistentwith our hypotheses than the FE estimates; the magnitudes in the FE2SLS aremore pronounced than the FE estimates. The qualitative conclusions about theeffects of PC provisions are similar except in the following cases. First, whilethe FE2SLS results suggest no productivity disadvantage for PCs signed bythe incumbent manager, the FE suggests otherwise. Second, while contractduration had no statistically significant relationship with productivity in theFE specification, the relationship is statistically and economically signifi-cant in FE2SLS. This may suggest that the government allowed managers ofpoorly performing SOEs a longer time to turn them around; and the FE2SLSresults suggest that the intention was achieved. Finally, the FE2SLS speci-fication suggests a positive and a far larger effect of managerial bond thanthe FE specification. This is consistent with the notion that managers of

  • Empirical Effects of Performance Contracts 187

    poorly performing SOEs were asked to post bonds for new contracts, andsuch incentives improved productivity as a result.

    Even though the instruments appear to be valid, there may still be small-sample bias associated with the conventional 2SLS method (Nelson andStartz, 1990; Bound, Jaeger, and Baker, 1993). To check for this we usedanother instrumental variables-based estimator, the split-sample instrumentalvariables (SSIV) estimator (Angrist and Krueger, 1995), which is consistentwith a matrix attenuation bias (this does not imply that each individual coef-ficient is attenuated).24 This way we can examine how robust the qualitativeresults are with respect to the method of identification. The results of SSIVare reported in column 3 of Table 4. The effects of PC provisions remainremarkably similar to the FE2SLS results in column 2. One noticeable changeis that the coefficient of managerial bond becomes significant, which suggeststhat small sample bias might play a role in the statistical insignificance of thisvariable in the pooled FE2SLS specification. Since these two methods havedistinctive directions of bias, the similarity of the estimates should providemore confidence that the qualitative results about the PC effects are aboutright.

    Again we ask, did the effects of PC provisions differ by whether the PCsended during the period under examination? Column 4 of Table 4 suggeststhat most of the effects of PC features did not hinge on whether PCs expiredor notwith one exception. Terminating PCs are significantly correlated withproductivity gains for firms with more market power. This may be explainedby restructuring. If firms with more market power restructured more, but thebenefits of this restructuring showed up only in the later phases of the con-tracts, then terminating contracts signed by more monopolistic firms wouldshow a positive and large PC effect, and nonterminating contracts would not.

    4.4 Comparison of Alternative PC ProvisionsWe wish to know the effects of PCs with different combinations of provi-sions. Using the FE2SLS estimates in column 1 of Table 4, we calculated thecombined effects, reported in Table 5. To simplify the presentation, for eachdimension of PC provisions and market competition (markup ratio), PCj , weclassify firms either as having PCgoodj or PC

    badj .

    25

    24. The SSIV estimation proceeds as follows: (1) randomly split the sample into two halves,say, S1 and S2; (2) use S1 to estimate parameters of the first-stage equation; (3) use the aboveequation to obtain fitted values for S2; then use the fitted values as regressors to obtain the 2SLSestimators in S2.

    25. In particular, for the dummy variables, good PC features were no bidding, targets witha profit orientation and managerial bonding, and bad was the opposite. (The bidding dummyis based only on BID.INCUMBENT; we ignore BID.NEWCEO since it would yield similar,albeit somewhat more negative effects on productivity and had only 11 firms). For the con-tinuous variables, a good PC feature was above the mean for that provision or the markupratio below its mean. The evaluation resulted in the following means for the two subsam-ples: WELASTICITYbad = 016, WELASTICITYgood = 066; TERMbad = 208, TERMgood =428; MARKUPbad = 055, MARKUPgood = 081 (note that MARKUP is standardized withmean 0). The values are the means for the good or the bad subsamples.

  • 188 The Journal of Law, Economics, & Organization, V17 N1

    Tabl

    e5.

    Dis

    tribu

    tion

    ofIm

    pute

    dPC

    Effe

    cts

    byC

    ontra

    ctTy

    pe(T

    otal

    PCPa

    rtic

    ipan

    ts=

    499a

    )

    LEN

    GTH

    :low

    LEN

    GTH

    :low

    LEN

    GTH

    :low

    LEN

    GTH

    :low

    LEN

    GTH

    :hig

    hLE

    NG

    TH:h

    igh

    LEN

    GTH

    :hig

    hLE

    NG

    TH:h

    igh

    W.E

    LAS:

    low

    W.E

    LAS:

    low

    W.E

    LAS:

    high

    W.E

    LAS:

    high

    W.E

    LAS:

    low

    W.E

    LAS:

    low

    W.E

    LAS:

    high

    W.E

    LAS:

    high

    mar

    kup:

    high

    mar

    kup:

    low

    mar

    kup:

    high

    mar

    kup:

    low

    mar

    kup:

    high

    mar

    kup:

    low

    mar

    kup:

    high

    mar

    kup:

    low

    BID

    :ye

    s2

    28

    0.52

    18

    10.

    991

    95

    0.85

    14

    71.

    33

    PRO

    FIT:

    no3.

    531.

    523

    16

    2.52

    33

    42.

    282

    87

    3.03

    BO

    ND

    :no

    BID

    :ye

    s2

    19

    0.61

    17

    21.

    081

    80

    1.00

    13

    21.

    47

    PRO

    FIT:

    no3

    55

    1.73

    31

    52.

    643

    31

    2.48

    27

    83.

    11

    BO

    ND

    :ye

    s

    BID

    :ye

    s1

    25

    00

    60

    77

    0.42

    09

    10.

    280

    44

    0.75

    PRO

    FIT:

    yes

    16

    50

    12

    10

    90.

    851

    29

    0.56

    06

    61.

    40

    BO

    ND

    :no

    BID

    :ye

    s1

    16

    0.36

    06

    90.

    500

    76

    0.43

    02

    90.

    90

    PRO

    FIT:

    yes

    15

    90.

    711

    01

    0.98

    11

    50.

    800

    46

    1.54

    BO

    ND

    :ye

    s

    Con

    tinue

    d

  • Empirical Effects of Performance Contracts 189

    Tabl

    e5.

    Con

    tinue

    d LEN

    GTH

    :low

    LEN

    GTH

    :low

    LEN

    GTH

    :low

    LEN

    GTH

    :low

    LEN

    GTH

    :hig

    hLE

    NG

    TH:h

    igh

    LEN

    GTH

    :hig

    hLE

    NG

    TH:h

    igh

    W.E

    LAS:

    low

    W.E

    LAS:

    low

    W.E

    LAS:

    high

    W.E

    LAS:

    high

    W.E

    LAS:

    low

    W.E

    LAS:

    low

    W.E

    LAS:

    high

    W.E

    LAS:

    high

    mar

    kup:

    high

    mar

    kup:

    low

    mar

    kup:

    high

    mar

    kup:

    low

    mar

    kup:

    high

    mar

    kup:

    low

    mar

    kup:

    high

    mar

    kup:

    low

    BID

    :no

    22

    70.

    531

    80

    1.00

    19

    40.

    861

    46

    1.33

    PRO

    FIT:

    no3

    51

    1.55

    31

    52.

    563

    32

    2.33

    28

    83.

    10

    BO

    ND

    :no

    BID

    :no

    21

    80.

    611

    71

    1.09

    17

    91.

    011

    31

    1.48

    PRO

    FIT:

    no3

    51

    1.75

    31

    32.

    673

    29

    2.51

    27

    73.

    16

    BO

    ND

    :ye

    s

    BID

    :no

    12

    40

    05

    07

    60.

    420

    90

    0.28

    04

    30.

    76

    PRO

    FIT:

    yes

    16

    20

    11

    10

    70.

    901

    27

    0.60

    06

    41.

    47

    BO

    ND

    :no

    BID

    :no

    11

    50.

    040

    68

    0.51

    07

    50.

    430

    28

    0.91

    PRO

    FIT:

    yes

    15

    60.

    080

    98

    1.03

    11

    20.

    830

    44

    1.60

    BO

    ND

    :ye

    s

    For

    each

    cell,

    the

    first

    row

    (und

    erlin

    ed)

    refe

    rsto

    the

    incr

    ease

    inpr

    oduc

    tivity

    leve

    lass

    ocia

    ted

    with

    asp

    ecifi

    cPC

    prov

    isio

    n,w

    ithth

    ese

    cond

    row

    bein

    gits

    t-st

    atis

    tic.T

    hees

    timat

    esth

    atth

    ista

    ble

    isba

    sed

    onis

    the

    FE2S

    LSsp

    ecifi

    catio

    n(c

    olum

    n2

    ofTa

    ble

    5).

    Her

    ear

    eth

    eva

    lues

    used

    for

    impu

    ting

    the

    PCef

    fect

    sin

    this

    tabl

    e:(1

    )B

    ID:y

    es,1

    ;no,

    0.(2

    )PR

    OFI

    T:ye

    s,1;

    no,0

    .(3)

    BO

    ND

    :yes

    ,1;n

    o,0;

    (4)

    LEN

    GTH

    :low

    ,2.0

    8ye

    ars;

    high

    ,4.2

    8ye

    ars.

    (5)

    W.E

    LAST

    ICIT

    Y:lo

    w,0

    .16;

    high

    ,0.6

    6.(6

    )(th

    est

    anda

    dize

    d)m

    arku

    p:hi

    gh,0

    .55,

    low

    ,0.

    81.

    The

    grow

    thra

    teef

    fect

    sof

    BO

    ND

    istra

    nsla

    ted

    into

    leve

    leffe

    cts

    byth

    efo

    llow

    ing

    form

    ula:

    bo

    nd

    1+L

    ENG

    TH/

    2,w

    hich

    isth

    eav

    erag

    ele

    vele

    ffect

    sof

    BO

    ND

    over

    the

    perio

    d.W

    eex

    clud

    edth

    ose

    firm

    sw

    ithbi

    ddin

    gw

    onby

    new

    man

    ager

    s(in

    tota

    l11

    firm

    s).

    Itsef

    fect

    sw

    illbe

    sim

    ilar

    toth

    ose

    bidd

    ing

    firm

    sw

    onby

    incu

    mbe

    ntm

    anag

    ers

    exce

    ptth

    atth

    ere

    isan

    addi

    tiona

    lpro

    duct

    ivity

    disa

    dvan

    tage

    .

  • 190 The Journal of Law, Economics, & Organization, V17 N1

    The range of PC effects appears to be enormous. A PC that used no bid-ding, with no profit orientation, no managerial bond, short LENGTH, lowWELASTICITY, and high markup ratio (see cell [5, 1]) would reduce pro-ductivity by as much as 227%. In stark contrast, a PC with long TERM,high WELASTICITY, and low markup ratio (see the last columns) wouldincrease productivity by 75 to 148 percentage points. The elements thatseem to be crucial to the success of PCs are a low markup ratio, a highWELASTICITY, a long duration, and managerial bonds. However, even ifall other good features are absent, implementing PCs in competitive sec-tors and asking for managerial bonds (see cell [2, 2] and cell [6, 2]) stillobtains a positive effect of 61% (with a t-statistic of around 1.73). Glancingat the entire table one cannot help but be impressed by the importance ofcompetition.

    Finally, we projected the effects of PCs on TFP using the coefficient foreach provision in column 1 of table 4 (FE2SLS), and the actual combinationsof PC provisions used in our sample.26 We projected that PCs on averagewould reduce TFP by roughly 39%. PCs had positive effects in about 57%of the PC participants, and for these firms the mean gain in TFP attributableto PCs was 67%, with a standard deviation of 41%. For the rest of the SOEsin our sample, the mean TFP loss from negative PC effects was 179%,with a standard deviation of 162%. Thus the average negative PC effectslargely reflect the poor designs of PCs for a minority of the sample.27 Sincethese findings are based on the available sample, which is not necessarily

    26. The rate effect of managerial bonding is transformed into a level effect by 0056 BOND 1 + TERM/2, where 0.056 is the FE2SLS estimates of 1BOND year sinceposting managerial bond).

    27. A referee points out another interesting possibility that might account for the negativePC effects for a significant portion of the sample: renegotiation concerns. Between 1988 and1990 China experienced its first high inflation (overall retail price index being 7.3 in 1987,18.5 in 1988, and 17.8 in 1989) in the past 30 years. Since profit targets specified in nomi-nal terms would become much easier to fulfill in an inflationary enviornment, the governmentmight want to renegotiate the terms of the contracts. Anticipating this, firms might deliberatelyreport bad news in order to obtain lower targets in the future. If the expectation of renego-tiation does lead to underreporting rather than underperforming, we have an estimate withdownward bias. However, the performance of nonterminating PCs should be less underreportedthan that of terminating PCs. This is because when terminating contracts need to be re-signed,underreporting or underperforming is more likely to reduce profit targets for the next contract,whereas for nonterminating PCs, wage-profit link and other current incentives would be morelikely to induce truereporting and hard work. But in our empirical examination we do not findsubstantial differences in the effects of these two types of PCs. We have also discussed thiscriticism with two Chinese experts who are more familiar with Chinese SOEs than we are, Dr.Chunlin Zhang (enterprise specialist of Resident Mission of The World Bank, formerly of forthe economic committee of the Chinese government), and Professor Hongling Wang (of themicroeconomics division of the Economics Institute of CASS). They suggest that the typicalsituation that provoked renegotiation was when managers were going to receive extremely highbonuses for obviously unjustified reasons, then government officials launched renegotiation onlywhen their own interests, and not necessarily that of the state, were at stake. It is unlikely thatthe government was so rational and efficient as to start renegotiation in anticipation of relativelyhigh inflation. We appreciate their discussions.

  • Empirical Effects of Performance Contracts 191

    representative of the universe, caution should be used in inferring PCs effectson all Chinese SOEs based on our findings.

    4.5 When Did PCs Perform Better?The disparity of the effects of good and bad PCs begs the question: Underwhat circumstances did PCs perform better? A complete investigation of thisquestion would demand a new article on the political economy of adopt-ing PCs. Nevertheless, we believe the simple regressions we present belowpartially address this question. The dependent variable in these regressionsis the total imputed PC effects (based on column 1 of Table 4) from theactual PC provisions of an SOE, which should be a good summary statis-tic of the soundness of the design of PCs. In fact, this measure can bestbe viewed as an index of PC effectiveness. Since the PC provisions did notchange for the complete duration of the contract, our sample consists of thefirm-years that an SOE adopted a PC. The independent variables includebasic characteristics of the SOE and its environment. These were lagged per-formance, lagged capital-labor ratio, lagged size of the labor force, markupratio, province dummies, industry dummies, year dummies, oversight dum-mies, and other reforms. We include a subset of these variables in columns1 and 2 of Table 6, and all of them in column 3. Since the results incolumn 3 are consistent with those in columns 1 and 2, we shall only discusscolumn 3.

    Not surprisingly, PCs show better effects in SOEs facing more competition.Surprisingly there was not much regional, industry, or year effects. Studentsof the Chinese economy might conjecture that there might be a JiangSu(province) effect, given that JiangSu was one of the richest provinces andone that implemented the reforms more dramatically. But in our results, noneof the province (or industry or year) dummies are statistically significant.

    Nevertheless, government oversight status did matter for PCs effects. Ceterisparibus, SOEs under the oversight of municipal or county governments had sig-nificantly better results from their PCs relative to SOEs under the oversight ofthe central and provincial governments. This is a strong result, for we have con-trolled for market competition, firm characteristics, other reforms, and laggedperformance. Thus local governments appeared to be able to design better PCs,perhaps because local officials had better information or faced harder budgetconstraints prompting them to design incentive systems that give more weightto efficiency.

    Capital intensity (lagged by one period) did not seem to affect the results ofPCs. But larger firms (lagged by one period and measured by the number ofemployees) had worse PC effects. This could be because it was more difficultto motivate a larger team, or there may be political economy reasons, suchas fear of layoffs, that lead governments to design less efficient PCs for largeemployers.

    SOEs with better lagged performance, measured by both labor productivityand profitability, appeared to have superior PC effects. To the extent thatbetter contracts are those with higher-powered incentives, this piece of

  • 192 The Journal of Law, Economics, & Organization, V17 N1

    Table 6. What Types of SOEs Had Good PC Designs?

    (1) (2) (3)

    Markup 1432 1487 1493447 463 464

    lnK/Lt1 0084 0132 0110073 113 094

    lnLt1 0564 0664 0664225 259 258

    JiLin Province 0148 0149 0116155 156 119

    JiangSu Province 0028 0018 0004030 020 004

    ShanXi Province 0104 0101 0097109 106 101

    Chemical industry 0291 0125 0117(the default category is light industries) 040 017 016

    Machine industry 0250 0041 0032031 005 004

    Material industry 0216 0054 0023034 008 004

    year 1987 0044 0062 0057033 047 043

    year 1988 0079 0095 0067058 068 048

    year 1989 0089 0060 0059049 032 032

    Oversight under provicial government 0039 0025 0032(the default category is oversight undercentral government) 026 016 021

    Oversight under municipal government 0231 0238 0226186 193 182

    Oversight under county government 0205 0234 0211123 140 125

    lnvt1 0139 0139212 212

    Profits/salest1 0681 0641209 195

    Managerial wage discretion 0130185

    New CEO 0010012

    Marginal rate 0038031

    Production autonomy dummy 0095138

    Constant 0880 0837 0782157 149 139

    Observations 489 484 484R2 084 084 084

    Dependent variable = the imputed PC effects using actual PC provisions and the coefficients of column 1 of Table 4. , and present statistical significance at the 10%, 5%, and 1% levels, respectively. In parentheses aret-statistics. The estimation method is OLS.

  • Empirical Effects of Performance Contracts 193

    evidence is consistent with the notion that agents of a better type choosehigher-power incentives (Laffont and Tirole, 1993).

    Another issue is whether PC designs were complementary to the otherreforms already in place. That is, did SOEs with more advanced statusin other reforms also have better PC designs? All the other reformsexcept managerial discretion are insignificant, thus there seems to be nostrong complementarity between other reforms and PC designs.

    5. ConclusionsOur analysis suggests that PCs on average did not improve the productivityof state enterprises in China. We also find that PC provisions mattered agreat deal: PCs can improve productivity when they provide high-poweredincentives, use sensible targets, and signal commitment through longer termsand managerial bondand especially when they are implemented in a morecompetitive environment. In the absence of these good features, PCs can hurtproductivity. In our analysis of the circumstances under which good PCswere implemented, we find that good PC features were more often observedin SOEs under the oversight of local governments, facing more competition,smaller in size, and with better previous performance.

    Our findings send a mixed message. On the one hand, the fact that onaverage PCs had negative effects, coupled with our similar earlier findingsin six other countries, urges caution in using PCs as a tool to reform SOEs,and may explain why China stopped signing PCs after 1994. On the otherhand, PCs, when properly designed and implemented, can indeed improveproductivity. These more effective PCs are more likely in smaller, relativelywell-running, competitive firms. However, these same firms are also thosemost feasible to be privatized, and empirical studies find that competitivefirms perform better under private ownership than they did as SOEs (for asurvey of the literature see Shirley and Walsh, 2000).

    An important finding in light of Shirley and Xu (1998) is that PCs hadpositive effects in fully 57% of the participants. Without further study of thepolitical economy of incentive contracts in government settings, we can onlyconjecture about why some contracts were poorly specified but the majoritywere not. Lack of knowledge about how to specify an efficient contract is nota very plausible explanation under the circumstances. Some of the reasons wediscussed earlier may provide better explanations. One is that politicians orbureaucrats structured PCs to maximize their political benefits or rents ratherthan productivity (Shleifer and Vishny, 1994; World Bank, 1995). This expla-nation seems especially feasible given that PCs failed across many differentinstitutional settings and contract designs. Successful PCs might arise whenpoliticians or bureaucrats are constrained by a hard budget constraint, whichin China was more likely the more local the level of government.

    A related explanation for poor PCs is that governments multiple princi-pals may have targeted many goals (profitability, investment, worker bene-fits) such that the PCs targets deviated from productivity, or the contracts

  • 194 The Journal of Law, Economics, & Organization, V17 N1

    incentives were lowered to avoid maximizing one goal at the expense of theothers. The fact that SOE shares are not traded even where stock marketsexist, and the absence of good accounting practices, may have given SOEmanagers an information advantage and bargaining power that PCs could notcircumvent. Thus our earlier research found evidence of strong managerialbargaining power in decisions about performance targets in six developingmarket economies (Shirley and Xu, 1998). This information asymmetry mayalso explain the generally low power of incentives since government wouldbe reluctant to risk wasting its bonus if achievements cannot be measured(Laffont and Tirole, 1993).

    To what extent are our results applicable in other countries? We believe themessage is general. First, since we are interested in contracts between govern-ment and state enterprises, the differences between China and other countriesare less important than they would be if we were drawing conclusions forprivate firms. Studies of SOEs (such as World Bank, 1995) suggest that thesituation of SOEs in developing market economies closely resembles thatof state enterprises in China (although not township and village enterprises)and other transitional economies. In most developing countries governmentsintervene widely in SOE operations, extend them protection from competi-tion and bankruptcy, and provide subsidies and debt relief. Second, our studyof PCs in China produced results strikingly similar to our earlier analysis ofPCs in six market economies (Ghana, India, Korea, Mexico, the Philippines,and Senegal). Finally, we postulated that the productivity effects of PCs are afunction of how well the contracts addressed problems of information asym-metry, incentives, and commitment, contractual features which, judging fromthe literature on information economics, are the most important generic ele-ments in characterizing contracts and country circumstances. At the sametime we attempted to control for as many aspects of the unobservable as wecould, such as other reforms, the competitive environment, etc., all of whichshould reduce the influence on our results of factors special to China.

    Appendix A: The DatasetThe dataset we use is A Survey of Chinese State Enterprises: 19801989.It covers 769 SOEs in 21 cities in four provinces (Shanxi, Jilin, Jiangsu, andSichuan). The 769 firms constitute a stratified random sample of all SOEsin manufacturing. There was substantial variation in the size of these SOEs:the median SOE had 930 employees, the SOE at the 10th size percentile had304, and that at the 90th percentile had 3175.

    The dataset has two parts. Part 1 consists of quantitative tables filled outby the accountants of each enterprise. Each table includes 321 variablescovering details about products, costs, wages and labor utilization, invest-ment, financing, fixed assets, profit distribution, taxes, prices, and materialinputs. Part 2 consists of questionnaires answered by the managers of eachenterprise. The managers answered questions about performance contractssigned with the government, the relationship between the enterprise and the

  • Empirical Effects of Performance Contracts 195

    government, production autonomy, the characteristics of the management,and so on.

    Appendix B: Construction of Key Variables for the 19801989 DatasetIn constructing these variables, we have followed other users of this dataset,especially Li (1997), and Gordon and Li (1995). All quantities (value added,capital stock) are expressed in 1989 market values. We assume that the 1989prices reflected best the opportunity costs of the resources.

    B.1 Capital Price Indexes and Capital StockThe survey contains answers to questions about the inflation rate of the mixedprice of equipment between the periods 19651975, 19751980, 19801984,and for each year between 1985 and 1988. Based on these answers we com-puted average inflation rates for equipment. For 19801984, we assumedequal yearly inflation rates. For 1989, since we did not observe equipmentinflation, we used the output inflation rate in the machine industry as a proxy.

    Since the survey did not provide information on prices of buildings orplant, for that inflation measure we used the percentage increase in aggregateconstruction costs compiled by the State Statistical Bureau. This series hasalso been used by Li (1997).

    We computed the composite price index for capital goods by averaging theequipment price index and the buildings and plant price index, the weightsbeing the investment expenditures on equipment and plant.

    We based our measure of capital stock on capital assets for productiveuse, which includes plant and equipment for industrial production. (In con-trast, capital assets for nonproductive use are mainly buildings and expen-ditures on dormitories, cafeterias, employee housing, and other social welfarefunctions.) Following Li (1997) and Gordon and Li (1995), we did not usethe net value of capital stock as the base to compute capital stock because ittends to exaggerate the increase in enterprise capital stock during the sampleperiod in which the inflation rate was high, because the accounting rate ofdepreciation was artificially low and the depreciation was based on historicalcosts (Gordon and Li, 1995).

    Realized investment at year t is imputed by subtracting the nominal valueof productive capital assets at the end of year t 1 from that at the end ofyear t. The reported investment, usually different from our imputed figures,is not used because it measures the value of capital expenditure (rather thancapital formation) in a given year. It includes, for example, expenditures onongoing construction projects; while it excludes prior investment projectscompleted in the year.

    Assuming that investment occurs smoothly over the course of a year, wecan compute the capital stock in 1980 (Ki80), the initial year, as

    Ki80 = 05Ki79 +Ki80PK89/PK80

  • 196 The Journal of Law, Economics, & Organization, V17 N1

    where Kit is the productive capital asset in year t, and PKt is the cumula-

    tive price index for the composite capital goods.28 The capital stock for thefollowing years is then constructed by the following formula:

    Kit = Ki t1 +05Ii t1PK89PKi t1

    +05Ii tPK89PKi t

    t = 81 89

    where I is the imputed realized investment.With this procedure there are still a little more than 700 missing Kit . Their

    values are imputed as the industry-year averages for 36 industries.

    B.2 Price Index for Value AddedThe price index for value added is based on the price indexes of output andmaterial inputs. Let Pvt be the price index of value added in year t, PQt bethat of output, and PMt be that of intermediate inputs. Let Qt denote outputunits and Mt be input units. By definition, the Laspeyres price index of valueadded is computed as follows:

    PVtPVt1

    = PQtQt1 PMtMt1PQt1Qt1 PMt1Mt1

    Tyler expansion along (PQt1PMt1) gives the following formula for thepercentage price increase of value added based on those of output and ofintermediate inputs:

    lnPVtPVt1

    = Qt1Vt1

    PQt PQt1Mt1Vt1

    PMt PMt1

    [Below we discuss the construction of the output price index (PQt) and inter-mediate input price index (PMt). In the empirical implementation, we valuethe value added for each year at the 1989 price of value added.]

    B.3 The Output Price IndexThe survey reports the mixed (plan and market) price index for the firmsmain product. While most firms reported cumulative price indexes, somereported year-to-year price inflation. We checked carefully and correctedthose obvious coding errors. When in doubt we treated them as missing.Consequently we have around 500 firms reporting a reasonable mixed priceindex. For the rest of firms we computed the average year-to-year mixed priceinflation rates for their industry-year sample, then assigned that value as theimputed mixed price inflation rate. Then we converted them to a cumulativemixed price index.

    28. Ki79, unobserved in the dataset, is extrapolated as in Li (1994):

    beginning-of-year total capital80end-of-year productive capital80

    end-of-year total capital80

  • Empirical Effects of Performance Contracts 197

    We then estimated the market output price index. The survey has infor-mation about the sales under the state plan and to the market, and theirrespective prices. Based on this information, we constructed the market priceindex for output. Again, firms with missing values for the market price indexwere assigned their industry-year averages.

    These price indexes were then used to compute the gross value of output(GVO). The survey reports GVO in current mixed prices. We first obtainedGVO in current market prices by multiplying the reported GVO by the ratioof market output prices to mixed output prices in year t. That number wasthen translated into GVO in 1989 market prices by multiplying it by the ratioof the market price index in 1989 to the market price index in year t.

    B.4 Price Index of Intermediate InputsThe dataset has detailed information about the plan and the market pricesof the two primary materials but it does not provide information aboutenergy and other intermediate inputs. We therefore computed price indexesfor intermediate inputs based on the assumption that the inflation rate forintermediate inputs was the same as that of materials. This is reasonablesince materials accounted for the vast majority of intermediate inputs. A sig-nificant portion of the reported material price variables was missing: roughly40% of the answers were useful.

    We first computed the mixed price of each material input using the physicalshares of the plan and the m