The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
M AY / J U N E 1 9 9 9 W. Bentley MacLeod is professor of economics and law at the University of Southern California. Daniel Parent is researcher at the Center for Inter-university Research on the Analysis of Organizations (CIRANO), Montreal, Quebec, and an assistant professor at McGill University, Montreal, Quebec. The authors thank Joseph Ritter and Jim Rebitzer for helpful comments. They gratefully acknowledge the financial support of National Science Foundation, Grant SBR-9709333, the Federal Reserve Bank of St. Louis, and CIRANO. Daniel Parent's research is supported in part by Quebec's FCAR. Job Characteristics, Wages, and the Employment Contract W. Bentley MacLeod and Daniel Parent T his paper explores some of the determinants of compensation in the United States. We suggest that compensation systems should be viewed as an integral part of the production process. We also wish to highlight the diversity in observed systems of pay that is often overlooked when examining wage trends from a macroeconomic perspective.1 A goal of the work reviewed here is to introduce compensation models that make predictions based upon observed job characteristics, and illustrate how compensation form may respond to changes in both the nature of work and labor-market conditions. The extent to which we are able to relate compensation to job characteristics is very much limited by the data. Fortunately, available data sets do have some information that we can use. In this essay we use both the National Longitudinal Survey of Youth (NLSY) and the Panel Study on Income Dynamics (PSID) to explore these issues. These data are not perfect, but they do provide information on some quite distinctive compensation practices. Table 1 reports the incidence of pay method by occupation for the NLSY. Workers were asked if during the current year they received any of the following types of compensation: 1. Hourly: Pay that depends upon the number of hours worked. 2. Salary: Pay by fixed period, such as weekly, monthly or yearly. Hours of work may vary from pay period to pay period, with no corresponding change in salary. 3. Piece Rate: Payment based upon the number of pieces produced by the worker, typically a supplement to hourly pay. For the PSID, workers are also asked if they are paid a combination consisting of an hourly rate and a piece rate. 4. Commission: Pay based upon some dollar measure of output, such as sales in the last period, typically commissions supplement salary pay. For the PSID, workers are also asked if they are paid a combination consisting of a salary and commission. 5. Bonus: Pay above one’s salary or hourly pay that is not contractually linked to a measure of performance, and hence its level is at the discretion of the employer. 6. Promotion: Movement to a higher rank, usually, though not always, associated with greater pay.2 This list does not exhaust the types of pay that we observe in practice, though it does move beyond the types of pay that would be considered in most macroeconomic models. In the next section, we briefly review the standard agency model. This model, the starting point for the economic theory of contracts, helps us understand the conditions under which a firm should link measures of performance to pay. As Table 1 illustrates, however, explicit payfor-performance contracts are by no means ubiquitous. In a later section entitled “Opportunism and Contract Complexity,” we will explore the limitations of the agency model in the context of Williamson’s (1975) concept of opportunism. When the employment relation is complex, then pay-for-performance contracts are incomplete, and hence workers may engage in inefficient opportunistic behavior. A solution to this problem, discussed in a section entitled “Relational Contracts,” is to use a relational contract that delays specifying rewards and exact performance F E D E R A L R E S E R V E B A N K O F S T. L O U I S 13 1 However, there are a large number of possible models of compensation, as nicely outlined in the review of Ritter and Taylor (1997). 2 See the Data Appendix for the exact question pertaining to pay-for-performance in the NLSY. M AY / J U N E 1 9 9 9 Table 1 Pay Method by Occupation National Longitudinal Survey of Youth (NLSY) 1988-90 Occupation Managers and admin.except farm Writers, artists, etc. Sales workers Prof., tech, except eng. techn. Personal service workers Secretaries Engineering and science techn. Clerical and unskilled 1* Office machine operators Clerical and unskilled 2** Transport equip. operatives Food service workers Mechanics and repairmen Cleaning service workers Craftsmen and kindred 1*** Precision machine operatives Laborers, except farm Health service workers Textile operators Operatives exc. precis. machines and textile Hourly Salary Piece Rate Commission Bonus Promotion 19.98% 21.84% 25.07% 27.94% 36.81% 37.20% 42.37% 43.18% 43.88% 48.76% 50.48% 52.46% 53.16% 54.46% 60.32% 60.44% 60.71% 65.99% 66.67% 80.02% 78.16% 74.94% 72.06% 63.19% 62.80% 57.63% 56.83% 56.12% 51.24% 49.52% 47.55% 46.84% 45.55% 39.68% 39.56% 39.29% 34.01% 33.33% 0.68% 2.30% 0.78% 0.43% 1.84% 1.02% 0.00% 1.34% 0.84% 1.99% 3.38% 0.52% 4.54% 1.49% 2.67% 36.81% 6.02% 2.03% 9.76% 9.84% 9.20% 37.98% 1.99% 20.25% 1.37% 5.09% 3.12% 1.27% 1.74% 8.21% 1.29% 9.56% 0.50% 1.60% 1.10% 1.88% 0.51% 0.71% 28.46% 17.24% 25.58% 15.46% 9.20% 11.60% 9.32% 13.21% 13.50% 10.20% 13.53% 7.49% 9.89% 7.43% 10.68% 9.34% 10.34% 8.63% 11.43% 18.91% 14.94% 11.37% 13.76% 9.82% 13.99% 18.64% 16.32% 14.77% 14.93% 10.14% 11.37% 12.16% 9.90% 17.97% 10.44% 13.16% 9.65% 10.00% 68.93% 31.07% 8.75% 1.79% 10.54% 7.32% * From bank tellers to meter readers for utilities (Census 301 to 334) **From shipping clerks to ticket agents and other miscellaneous clerks (Census 374 to 395) ***From auto accessory installers to machinist apprentices (Census 401 to 462) owned by the principal.3 There are three basic ingredients in such a model: 1. The agent is risk averse. 2. The output of the agent is a stochastic function of effort. 3. The agent’s effort is imperfectly observable. For simplicity, assume that the principal is risk neutral, given that the agent is risk averse, this implies that the individual would prefer to receive a fixed income stream that is independent of the project’s fortunes. Given that effort is not easily observable, however, this may give rise to moral hazard: The agent may choose less than the efficient level of effort. The prin- expectations until after the worker has selected effort. Under the appropriate conditions, this provides a solution to the problem of opportunistic behavior. Moreover, it has the empirical prediction that firms are more likely to use bonus pay rather than efficiency wages when labor markets are tight. We test and find some support for this hypothesis. The final section of the paper contains concluding remarks. 3 See Hart and Holmstrom (1987) for a good overview of the agency model. See also Gibbons (1995) for a more upto-date review of this literature. AGENCY THEORY The agency model begins with a principal who wishes to hire an agent to carry out a task, usually involving the assets F E D E R A L R E S E R V E B A N K O F S T. L O U I S 14 M AY / J U N E 1 9 9 9 cipal can provide incentives for performance by making the agent’s pay conditional upon the available performance measures. More formally, suppose that the agent’s preferences are given by: Notice that even though the principal cannot directly observe the actions of the agent, the contract is designed so that in equilibrium the agent chooses to work hard. Assuming that the solution can be characterized by the first order conditions for the optimum, then the optimal contract solves the following equation: U (ω , e ) = u (ω ) − υ e , (1) where ωis income and e [{L,H} is low or high effort. The utility for income is assumed to be twice differentiable, and satisfy u ′( ω ) > 0, u ′′( ω ) < 0 for every ω> 0. The disutility for effort satisfies υH > υL > 0. The effort of the agent results in a stochastic output denominated in dollars, y e Y # R, as well as a vector of performance measures, m = {m1,...,mn} e M. Let fe (y,m) denote the joint distribution of y and m as a function of effort, where it is assumed that fe (y,m) > 0 for all (y,m) [Y ×M.4 Let us further suppose that it is efficient for the agent to produce a high level of effort (otherwise the problem is trivial), and that the principal offers a wage contract that is a function of the observable signals (y,m), given by w + c(y,m). In this case the principal agent problem is given by: ( (5) ) subject to: (4) {( )} E U c( y, m), H ≥ U , and {( )} { ( E U c( y, m), H ≥ E U c( y, m), L )} where {( E U c( y, m), e ( ) where µ, λ≥ 0 are the LaGrange multipliers associated with constraints 3 and 4, respectively. If there were no moral hazard problem, then constraint 4 would not be binding, and λ= 0 with the optimal contract given by a constant wage ω∗ satisfying υ′(ω∗)= 1/υ′(ω∗)= 1/µ. The interesting case is when moral hazard is a problem, and λ> 0. In that case, the sensitivity of the contract to y and m depends upon the behavior of the f L ( y ,m) likelihood ratio r ( y, m) = def . When f H ( y ,m) the likelihood ratio is a decreasing function of y, called the monotone likelihood ratio condition, then the optimal contract will be increasing in y. This condition implies that FH first-order stochastically dominates FL (though the converse is not true). As discussed in detail by Hart and Holmstrom (1987), the intuition is that a high y signals high effort, and the agent should receive a greater reward. In equilibrium the principal has correct expectations concerning worker effort, and the signaling effect is to provide ex ante incentives, and does not provide information to the principal per se. The signaling perspective does provide guidance about when additional measures of performance should be incorporated into the optimal contract, as shown in the following proposition.5 Proposition 1. Suppose that the solution to the principal agent problem satisfies the first-order condition 5, then the optimal contract c* (y,m) depends upon the signal mi if and only ∂r ( y, m) / ∂mi ≠ 0 for some value (y,m). ∫ y − c( y,m ) fH ( y, m) dyd m, (2) max c(.,.) (3) f L ( y, m) = µ + λ 1 − , f H ( y, m) u ′ c * ( y, m) 1 )} = ∫ υ (c (y, m)) fH ( y, m) dyd m − υ e . Constraint 3 is the individual rationality constraint that ensures the agent receives as much as his or her next best alternative, –denoted U . The next constraint, 4, is the incentive constraint that ensures that the agent prefers to work hard rather than to shirk. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 15 4 This is the so-called full-support assumption that is a necessary (though not sufficient) condition to use the first-order approach to characterize the optimum. Harris and Raviv (1979) show that if the support moves with effort then one can implement the first best. We also assume that the density is a differentiable function of y and m. 5 See Holmstrom (1982) for more details. M AY / J U N E 1 9 9 9 6 7 We use a linear probability model rather than a logit or probit because we can better control for selection effects and misclassification error. The main drawback of a linear probability model is that it is less efficient, but in general it is more robust to specification errors than a nonlinear model would be. Note also that the standard errors are adjusted for group effects (see e.g., Moulton, 1986) and that we take into account possible selection (into occupation) effects. See MacLeod and Parent (1997) for complete details. To correct for misclassification error, we borrow from Krueger and Summers (1988). For example, if mi represents the clothes of the agent or their hairstyle, and these provide no information concerning their effort, then they should not enter into the optimal contract. Any other measures, however, such as customer complaints, supervisor reports, etc., that provide additional information concerning performance above and beyond y should be included in the optimal contract, even if the contract already depends upon y. Consider for example a sales person who is paid on commission. Sales is a discrete variable that depends upon a number of factors, including price, buyer preferences, store location, etc. Hence a sale may be made even if a salesperson is rude (for example, the buyers had to purchase the good immediately and could not search further). Rudeness, however, is likely to affect the probability of a sale in many cases. Even if the sale is consummated, the optimal contract would entail a penalty if the customers report to the manager that the salesperson is rude. The model predicts that even a single report of rudeness should generate a negative financial consequence, and more generally, as Gibbons (1995) observes, agency theory generically predicts a sensitivity to available performance measures that we rarely observe in practice. were asked in 1979 and 1982, which we can use to carry out a preliminary investigation of the relationship between performance pay and job characteristics. The relevant question in those years was: “We would like to know what kind of opportunities this job offers you. How much opportunity does this job give you? A minimum amount, not too much, a moderate amount, quite a lot, or a maximum amount? 1. To do a number of things (variety). 2. Deal with people. 3. For independent thought or action (autonomy). 4. Friendships. 5. To do a job from beginning to end (probe if necessary: that is, the chance to do the whole job) (complete TASK).” Answers are re-coded to zero if respondents answer either “a minimum amount, not too much, or a moderate amount,” while they are re-coded to one if respondents answered either one of the last two possibilities. For each one of 20 occupation cells, we compute the average of the answers in both the 1979 and the 1982 surveys. We then merge these averages to each corresponding occupation category for the 1988-90 period. This, of course, is a crude way to proxy the different dimensions of the jobs, but we think that it is not too unreasonable to assume that jobs that are in the same occupation cell may share some common characteristics. In Table 2 we report the results from a linear probability model of different types of performance pay.6 Given that piece rate workers also are categorized as wage earners (notice that all workers are categorized as either wage or salary workers), then we can ask what job characteristics are associated with the use of piece rates. These results are reported in the first two columns, with the second column correcting for biases that may be introduced due to misclassification of worker occupation.7 Notice that requiring workers to perform complete tasks is negatively related to the use of piece rates. This may suggest that individuals on straight wages are more Some Evidence To understand why performance-pay contracts are not ubiquitous, we begin by looking at some of the determinants of performance pay. Even if agency theory is not a complete model, it still provides important insights into the necessary conditions for the use of a performance measure. In particular, jobs for which the cost of obtaining good measures are low should have a higher incidence of performance pay. As we can see from Table 1, we have data from the NLSY that describes certain types of performance pay during the 1988-90 period. Unfortunately, no questions pertaining to the characteristics of the jobs were asked in the NLSY during the 1988-90 period. But such questions F E D E R A L R E S E R V E B A N K O F S T. L O U I S 16 M AY / J U N E 1 9 9 9 Table 2 The Effect of Job Attribute on the Likelihood of a Compensation Characteristic Based upon the National Longitudinal Survey of Youth (NLSY) (1988-90) Is the following attribute important in your job? Piece Rate(1) vs. Hourly Wage (0) Commission (1) vs. Salary and/or Bonus Pay (0) Bonus + Salary (1) vs. Salary + Termin. Contract (0) Autonomy -0.1331 (0.5382) -0.1835 (0.3536) 1.5634 (0.4433) 2.2259 (0.5464) 0.982 -0.9165 1.1825 -0.5275 Complete Task -1.4971 (0.6352) -1.4102 (0.4173) -0.7975 (0.5231) -1.2647 (0.5960) 0.3077 (0.9044) -0.4598 (0.6226) Variety 0.9406 (0.4795) 0.6816 (0.3451) -1.1221 (0.3949) -1.156 (0.4429) -1.1146 (0.7175) -0.5263 (0.4700) Friendships -0.5213 (0.6105) -0.0419 (0.4029) -0.3344 (0.5052) -0.5861 (0.6794) -0.3302 (1.2908) -0.6134 (0.6012) Deal with People -0.0435 (0.1921) 0.0611 (0.1262) 0.2367 (0.1582) 0.1735 (0.3429) 0.1426 (0.2593) 0.4136 (0.1883) Correction for Misclassification? No Yes No Yes No Yes F-Test of No Selection (P-Value) Sample Size 0.0878 3927 0.2599 3927 4238 0.7084 4238 3832 3832 Notes: Standard errors are in parentheses, with 5 percent significance given in white, and 1 percent significance in grey. These are adjusted for structural group effects where applicable. Other covariates include tenure, labor market experience, and dummies for region, industry, year, residence in Standard Metropolitan Statistical Area (SMSA), unemployment rate, schooling, union status, and increase in responsibility. likely to be assigned specific tasks, with target completion dates, this is consistent with our view that a worker is paid a fixed hourly wage but does not imply a lack of incentive pay. Rather, the worker is paid for the time spent on the job, where he or she is required to achieve a satisfactory level of performance. Relative to piecerate contracts, tasks with less variety would be easier to monitor on a day-to-day basis, hence performance can be measured in terms of acceptable/unacceptable, with termination being the consequence if there is unacceptable performance. The Autonomy variable has positive sign in the Commission vs. Fixed-Salary regression, while the complete task variable is negative. Given that commission workers are rewarded based upon a measure of output, direct monitoring is less necessary and hence they have more autonomy. This also implies that those workers who are not paid commissions would be more closely monitored, an observation that is consistent with the negative coefficient for the Complete Task variable. Consistent with earlier results by Brown (1990), we find that Variety has a negative effect on the likelihood that commission contracts are used. This result does not follow directly from the agency theory that would predict the use of more, not less, performance pay. In the next section we outline a model based upon F E D E R A L R E S E R V E B A N K O F S T. L O U I S 17 M AY / J U N E 1 9 9 9 Williamson’s (1975) notion of opportunism, which may help explain this effect. It is also interesting to observe that job characteristics have little impact upon the choice of whether to use bonus pay. If bonus pay is not directly related to job characteristics, then what is its role? The use of bonus pay is not a prediction of the agency model because it is not an explicit function of a performance measure, rather it is the consequence of some subjective performance-evaluation system. More generally, the data also suggests that for many workers, contracted-performance pay (piece rate or commission) is not always an important ingredient of compensation, especially when Variety is important— even though agency theory predicts that even imperfect measures of performance should be incorporated into pay. In the next section we discuss how a model of contract incompleteness based upon a simple complexity argument can explain both the use of noncontingent pay and why the incidence of bonus pay may not depend upon job characteristics. view. Recall that in an agency model the optimal contract incorporates the incentives for shirking via the IncentiveCompatibility constraint, and thus, firms would never be surprised by worker behavior ex ante. Kerr’s observation of unexpected, dysfunctional behavior ex post is consistent with Williamson’s (1975) notion of opportunism: self-interest seeking with guile. In the context of an agency relationship, we define guile as behavior that takes advantage of the incentive system by increasing the agent’s payoff at the expense of the principals that is not anticipated via the Incentive Constraint. For example, consider a firm that rewards typists based upon the measured number of keystrokes per day. This is a clear pay-forperformance contract committing the firm to a pay method that is a simple function of “output.” The difficulty with this system, as was discovered when the system was implemented at one firm, is that one typist discovered that she could increase her income by pressing the same key repeatedly. Had the firm anticipated this behavior, it would have implemented additional monitoring to ensure the quality of output. The agency model explicitly assumes that all possible types of dysfunctional behavior are anticipated and controlled with the appropriate contract terms and conditions. Hence, the introduction of a negative behavior such as guile necessarily requires the relaxation of the complete-contracts assumption, which in turn requires a fundamental modification of the standard economic model of decision-making.8 The conceptual starting point is to view contract incompleteness as arising from the problem of exchanging complex goods, such as labor services. A distinguishing feature of a complex good, relative to an exchange of a simple good or commodity, is that quality is difficult to define, and therefore difficult to enforce using a contingent contract enforced by the threat of a court action. Secondly, both the creation of complex goods and the formation of contracts to govern their exchange are innovative activities that do not fit easily into the standard agency model. OPPORTUNISM AND CONTRACT COMPLEXITY 8 See MacLeod (1997) for a complete discussion of this point. What we learn from the agency model is that generically optimal contracts should incorporate all available performance measures. This implies that pay-for-performance should be the norm rather than the exception. There is a large body of evidence in the management literature that emphasizes the dysfunctional attributes of performance pay. For example, if we were to reward computer programmers based upon the number of lines of code that they produce, then the likely consequence is not necessarily high output, but many lines of inefficient and error-ridden code. An immediate response is that lines of code is not an appropriate measure of output. As the famous study by Kerr (1975) eloquently illustrates, many organizations and firms have implemented pay-for-performance systems, only later to discover that they result in dysfunctional behavior from the organization’s point of F E D E R A L R E S E R V E B A N K O F S T. L O U I S 18 M AY / J U N E 1 9 9 9 The problem can be illustrated formally with a simple model of employment based upon the multitasking model of Holmstrom and Milgrom (1991): 1. The principal and agent agree on compensation and expectations for performance (which may include the continuation of a previous agreement). 2. The state of the world ω t ∈ Ω is revealed. 3. The agent divides a time endowment of Y among k different tasks: y t ∈ℜ K . 4. The principle pays the agent Wt. 5. Both principle and agent decide whether to continue the relationship or not. The date is denoted by the subscript t, and K is the number of possible tasks. The twist upon the previous literature concerns the interpretation of the state of nature. Suppose that both the costs and benefits of different actions are unknown ex ante; for example, a fireman may not know which house will catch fire; how difficult it will be to put out the fire; nor is he able to anticipate the set of actions that will need to be carried out upon entering the burning house. A state space that incorporates uncertain costs and benefits for each of the possible tasks can be defined as follows: (6) Ω= The quadratic term implies that the marginal cost of effort in a single task is increasing with effort, ensuring an interior optimum. The function δ(yit ) is 1 if yit is positive and zero otherwise, which implies that there is a fixed cost f of supplying a positive level of effort to a particular task. When there are a large number of tasks this implies that the individual will supply effort to only a subset of possible tasks. The benefits and costs have been modeled as functions, however it is explicitly assumed that a measurement system does not exist. Consider a secretary who carries out a variety of tasks including typing, answering the phone, filing, making travel reservations, etc. The costs and benefits for these different activities vary with the day-to-day demands of the office. For example, several people in the office may need to go to the same conference, raising the productivity of allocating time to travel plans, and resulting in a cutback in typing throughput. On the cost side, if the conference occurs during a busy period (for example college convocation), then one may have to call several hotels to find accommodations. Not only do these costs and benefits vary in an independent way from day-to-day, it is not clear (at least to me) how one would construct a measurement system to directly compare the costs and benefits of the different actions. Notice that in the principal agent model it assumed that all signals, m, are verifiable and can be used to construct an explicit contract; however the yit are assumed to not be measurable. Here, we suppose that the yit can be observed, but there exists no contractible m. For example, if one had a measure of individual contribution, m t = α T y t , this could be used to construct an efficient, explicit contract. For many, if not most jobs, it is very difficult to construct such a measure. The lack of a measurement system aggregating performance implies that the contract must explicitly describe each state and specify the appropriate associated action.9 This is common in many contracts. For example, the contract for a singer at a concert may explicitly list acceptable reasons, such as laryngitis, that excuse the {{α ,...,α } × {β ,..., β }} , 1 n { where α k ∈ α 1 ,..., α n 1 k m } denotes one of n levels of productivity for task k, while { β k ∈ β 1 ,..., β m } represents one of the m cost levels for task k. The total benefit from an effort choice yt is defined by αTyt (boldface represents a vector), while the total cost to the worker of producing this effort is (7) C( yt , β) ≡ ∑ K i=1 ( 2 i yit ) − δ ( yit ) f . F E D E R A L R E S E R V E B A N K O F S T. L O U I S 19 9 This assumption can be contrasted with the agency approach to compensation as outlined in Baker (1992) and Holmstrom and Milgrom (1991). This work examines the optimal way to incorporate imperfect signals of worker performance into the pay package. M AY / J U N E 1 9 9 9 Table 3 Cost of a Complete-State Contingent Contract Number of Tasks Number of Cost and Performance Levels 2 5 10 15 2 $0.16 $10 $10,000 $10 million 3 $0.81 $600 $35 million $2 trillion 4 $2.56 $10,000 $11 billion $11,000 trillion 5 $6.25 $100,000 $1,000 billion $10 million trillion Cost of a contract clause: 1 cent –where U is the one-period alternative utility for the worker. Following Townsend (1979) and Dye (1985), let us suppose that there is a cost for including additional contract contingencies, given by γ per contingency. For this multitasking model one has the following result. Proposition 2. The cost of implementing the complete contract procedure when all states occur with positive probability is nkmkγ. What is important to observe is that the cost of the contract is an exponential function of the number of tasks. The literature on computational complexity emphasizes the impossibility of implementing algorithms whose costs are exponential in the size of the problem (see Garey and Johnson, 1979). To see why this is the case, suppose that γ = 1 cent, and that the number of cost and performance levels are the same (n = m). Table 3 presents the costs of the complete contract as a function of the number of tasks and effort levels. As one can see, the use of a complete contract when there are more than say 10 tasks is impossible. Furthermore, given that these costs reflect the number of underlying states, dynamic programming is impossible because one could not compute the expected value of the relationship. Observe that the piece rate contracts correspond to basing compensation on one dimension of output. In this simple individual from providing the contracted upon services. Formally the contract is a function c : Ω → X = ℜ × ℜk , where for each state w e Ω, the ( ) c(ω ) = ω (ω ), y(ω ) ∈ X defines the wage payment and the output expected from the agent. This assumption differs from the incomplete contracts literature where it is assumed that such a contract is impossible, while maintaining the hypothesis that individuals understand all the possible outcomes and can recontract based on the ex post realization of the state. For this model an efficient complete contract, ( ) c * (ω ) = ω (ω ), y(ω ) , is the solution to the following program: (8) y (ω ) ∈ arg max a y′ − C( y′, β), y′ subject to: k (9) y ≡ ∑ yi′ = Y , i=1 and (10) ( ) w(ω ) = U + C y(ω ), β , F E D E R A L R E S E R V E B A N K O F S T. L O U I S 20 M AY / J U N E 1 9 9 9 setup complete contracts are very inexpensive; therefore, they should be observed when there is a small number of tasks to be measured. A solution to the problem of complexity is to use an ex post evaluation of the employee using supervisor reports. The subjective nature of these reports make third-party enforcement impossible. Hence, performance depends upon what MacNeil (1974) calls a relational contract, which is discussed in more detail in the next section. Given that direct supervision of the employee is an essential ingredient of the relational contract, then not only should workers in such contracts have less autonomy, but they also should have well-defined goals that are determined by their supervisors. contingent contract with no ex post evaluation).10 The theory developed in MacLeod and Malcomson (1989) makes some predictions concerning the effect of market alternatives for workers upon the incidence of bonus pay that we briefly outline here. Suppose the employment contract is given by c={w,b}, where w is a fixed wage that is paid at the end of the period regardless of performance, and b ≥ 0 is a discretionary bonus payment that depends on the firm’s subjective ex post evaluation of performance. Given this contract, individual utility and firm profits are given by: RELATIONAL CONTRACTS (11) U (c) = ω + b − υ e + δ U c , (12) Π (c) = θ e − w − b + δ Π c , where e [ {0,1} is a noncontractible effort choice taken by the worker, Uc and Πc are the utility and profit, respectively, from continuing the relationship, assumed to be discounted at the rate δ. The parameters υ and θare respectively the cost and benefit of one unit of effort. The implicit agreement between the firm and worker requires the firm to pay the bonus if and only if the worker selects the high level of effort.11 Should either party shirk, then the relationship is termi––nated immediately. Letting U and Π denote the market alternatives for the worker and the firm, then a contract is self-enforcing if and only if the following incentive conditions are satisfied: When an explicit contract is not possible, the firm must rely upon some form of ex post incentive to ensure performance. There are essentially three types of noncontracted ex post rewards that we observe in the NLSY: 1. Termination contracts—pay the worker a fixed salary, and fire the worker at the end of the period if performance is not satisfactory. 2. Bonus contract—pay the worker a discretionary bonus at the end of the period that depends on performance. 3. Deferred compensation—reward the worker with a promotion or permanent wage increase. Bonus pay and deferred compensation are not perfect substitutes since a promotion entails a permanent increase in income. Given that we are using only indicators rather than levels, however, we have coded bonuses and deferred compensation into the same category. This reduces the error associated with imputing the true value of the promotion. Between 10 percent to 14 percent of the individuals in our data set receive some form of bonus pay (as opposed to piece rates or commissions, which are forms of complete ( ) (13) δ U c − U ≥ υ − b, (14) δ Π c − Π ≥ b. ( ) Notice that it is necessary to pay a bonus – only if d(Uc-U ) < v. For example. if unemployment rates for the worker were to –increase, this would lower U and increase – the likelihood that d(Uc-U) ≥ v. In this case, the threat of termination alone provides sufficient incentives for the worker not to shirk. Conversely, with a tight local labor F E D E R A L R E S E R V E B A N K O F S T. L O U I S 21 10Some individuals in the NLSY data receive both piece rates and bonuses. They are a small fraction of our sample, however, and so we do not explicitly consider this case. 11MacLeod and Malcomson (1989) prove that there is no loss of generality when contracts are restricted to take this form. M AY / J U N E 1 9 9 9 Table 4 Tobit Analysis of Determinants of Bonus Pay Panel Study of Income Dynamics (1984-91) $1979 (Standard Errors in Parenthesis) Variable All Observations SMSA Workers Only Local Unemployment Rate -360.47 (76.97) -357.2 (76.83) Industry Unemployment Rate (one-digit) -91.36 (328.33) -125.58 (77.96) Schooling 186.98 (66.65) 200.69 (66.63) Union -1920.59 -2059.63 -1869.24 (554.94) (548.56) (559.72) -2165.24 -2408.64 -2258.56 (982.56) (970.28) (995.49) Potential Experience -10.73 (20.80) -11.6 (20.52) -52.79 (20.21) 38.8 (35.39) 39.57 (35.46) -9.38 (34.29) Tenure 14.63 (26.12) 15.63 (25.69) 30.94 (26.20) 26.38 (44.00) 30.67 (43.00) 40.91 (44.16) Live in a SMSA 571.09 657.22 347.53 (345.38) (344.84) (103.79) Yes No Yes* -7724 -7733.5 -7721.6 5119 5119 5119 Industry Dummies Yes Log Likelihood -14116 N 10217 No -525.37 (75.41) -893.73 (152.61) -396.76 -36.86 (598.02) (133.36) -47.69 (56.34) Yes* -14124.2 -14113.6 10217 -570.49 -542.78 (158.22) (157.94) 10217 242.25 277.29 (107.46) (107.66) -6.64 (92.18) Note: Workers paid commissions are excluded from the analysis. Additional regressors include time and occupation dummies, as well as a dummy for being married. *A full set of Year X Industry (one-digit) dummies. market, when the worker can always find alternative work easily, the incentive constraints imply that some form of end-of-the-period bonus must be paid. Therefore, we expect the incidence of bonus pay to be a decreasing function of the local unemployment rate. In Table 4 we present some evidence of this effect using the Panel Study on Income Dynamics. We also explore the effect of both the local and industry unemployment rates upon the amount of bonus pay. Table 5 shows the same relationship regarding the incidence of bonuses/promotions in the NLSY. One explanation for the incidence/amount-of-bonus pay is as a form of profit sharing between the firm and the worker. Most firm’s profits are correlated with industry rather than local unemployment rates. When this is the case, it implies that bonus pay incidence will increase with a decrease in the industry unemployment rate, while the local rate would be unimportant. The selfenforcing contract model makes the opposite prediction. As we can see from the regression results, the industry rate is not significant, while the local unemployment rate has a negative impact upon the amount and the incidence of bonus pay. Also, as we would expect, this effect is stronger when we restrict analysis to urban areas where workers would have better market alterna- F E D E R A L R E S E R V E B A N K O F S T. L O U I S 22 M AY / J U N E 1 9 9 9 Table 5 tives. More surprising for us, is the fact that the local labor market effect increases in the PSID data set when we add controls for time-varying industry effects. If bonus pay were the result of profit sharing, then the addition of such controls would make the effect of local unemployment either small or less precise, whereas we observe exactly the opposite. In this model we have assumed that the supervisor can perfectly observe performance ex post. We could add imperfect observability, as in Shapiro and Stiglitz (1984), and obtain the same result. It is sometimes believed that it is imperfect observability that generates an efficiency wage. As the results of Holmstrom (1982) demonstrate, however, an imperfect but contractible measure of output would completely eliminate the equilibrium unemployment result for a standard efficiency wage model. Hence, the use of bonus pay and/or efficiency wages are a consequence of increases in job complexity that make it impossible to fully specify ex ante an employer’s performance expectations. Therefore, our results provide more support for efficiency-wage type models. In the absence of bonus pay, an efficiencywage model implies that the wage must be above market clearing, and if unemployment falls this may lead to an increase in inflation. Recently, the economy has appeared to have both low inflation and low unemployment. This could occur if firms move towards a system of bonus pay, rather than raise wages. In Figure 1 we illustrate the trend in the incidence of bonus pay, inflation, and unemployment from 1976 until 1991. While this is not a test, it does show a definite upward trend in the use of bonus pay during this period. Fixed-Effect Results—NLSY 1988-90 Bonus/Promotion vs. Termination Contract (Bonus=1) (Salaried Workers Only) Bonus/Promotion vs. Termination Contract (Bonus=1) (All Noncommission Workers) Autonomy 0.982 (0.9165) 0.5743 (0.8136) Complete Task 0.3077 (0.9044) 0.0397 (0.8621) Variety -1.1146 (0.7175) -0.793 (0.6839) Friendships -0.3302 (1.2908) 0.4767 (1.2377) Deal with People 0.1426 (0.2593) 0.4306 (0.2472) Unemployment Rate in Local Labor Market -0.0774 (0.0161) -0.0321 (0.0159) Unemployment Rate in Industry -0.0299 (0.0225) 0.0123 (0.0213) Schooling -0.0104 (0.0316) 0.0134 (0.0206) Union -0.0807 (0.0336) 0.0092 (0.0264) 3832 7682 Variable Sample Size Note: Standard errors are in parentheses, with 5 percent significance given in white, and 1 percent significance in grey. These are adjusted for structural group effects where applicable. Other covariates include tenure, labor market experience, and dummies for region, industry, year, residence in Standard Metropolitan Statistical Area, and increase in responsibility. explain the data. Rather, the data suggests that compensation systems depend on explicit performance measures when these accurately measure the contribution of work. In complex environments, firms must depend upon subjective measures of performance associated with ex post rewards to the worker. We have also presented evidence showing that the amount of bonus pay is dependent upon the state of the local labor market. One benefit of bonus pay is that its level can be adjusted easily from year-to-year in response to business cycle fluctuations, CONCLUSIONS In this essay we have reviewed some preliminary evidence relating job characteristics to the form of compensation. Our main message is that we observe a variety of compensation systems used in practice, the form of which depends upon job characteristics. There is no single economic model of contract formation that can F E D E R A L R E S E R V E B A N K O F S T. L O U I S 23 M AY / J U N E 1 9 9 9 which as Weitzman (1985) has argued, can result in both low unemployment and low inflation. The recent trend increase in the use of bonus pay may be one reason why inflation has not increased, even though the United States also is experiencing low unemployment. Currently, we do not know if this trend is the consequence of secular changes in the nature of work, or the result of innovative activity on the part of the firm. Given that the form of compensation is likely to affect the responsiveness of incomes to inflation and business cycle fluctuations, it is important to better understand the reasons for these changes. We can conclude that it is an oversimplification to view wage formation as the simple consequence of supply and demand forces, and that better understanding the source of variation in pay systems may have important implications for the nature of monetary policy, a question we hope to explore in future work. Kerr, Steven. “On the Folly of Rewarding A, While Hoping for B,” Academy of Management Journal (December 1975), pp. 769-83. Krueger, Alan B., and Lawrence H. Summers. “Efficiency Wages and the Inter-Industry Wage Structure,” Econometrica (March 1988), pp. 259-94. MacLeod, W. Bentley. “Complexity, Bounded Rationality and Heuristic Search,” mimeo, University of Southern California, (August 1997). _______, and James M. Malcomson. “Implicit Contracts, Incentive Compatibility, and Involuntary Unemployment,” Econometrica (March 1989), pp. 447-80. _______, and Daniel Parent. “Jobs Characteristics and the Form of Compensation,” mimeo, University of Southern California (1997). MacNeil, Ian R. “The Many Futures of Contracts,” Southern California Law Review (1974), pp. 691-816. Moulton, Brent R. “Random Group Effects and the Precision of Regression Estimates,” Journal of Econometrics (August 1986), pp. 385-97. Ritter, Joseph A. and Lowell J. Taylor. “Economic Models of Employee Motivation,” this Review 79 (September/October, 1997), pp. 3-21. REFERENCES Baker, George. “Incentive Contracts and Performance Measurement,” Journal of Political Economy (June 1992), pp. 598-614. Shapiro, Carl, and Joseph E. Stiglitz. “Equilibrium Unemployment as a Worker Discipline Device,” American Economic Review (June 1984), pp. 433-44. Brown, Charles. “Firm’s Choice of Method of Pay,” Industrial and Labour Relations Review (Special Issue 1990), pp. 165S-82S. Dye, Ronald A. “Costly Contract Contingencies,” International Economic Review (February 1985), pp. 233-50. Townsend, Robert. “Optimal Contracts and Competitive Markets with Costly State Verification,” Journal of Economic Theory (October 1979), pp. 265-2. Garey, Michael R. and David S. Johnson. Computers and Intractability, W. H. Freeman and Co., 1979. Weitzman, Martin L. “Profit Sharing as Macroeconomic Policy,” American Economic Review (May 1985), pp. 41-45. Gibbons, Robert. “The Employment Relationship,” mimeo, survey paper prepared for the Seventh World Congress of the Econometric Society, Tokyo, August 1995. Williamson, Oliver. Markets and Hierarchies: Analysis and Antitrust Implications, The Free Press (1975). Hart, Oliver D., and Bengt Holmström. “The Theory of Contracts,” Advances in Economic Theory, Fifth World Congress., Truman Bewley, ed., Cambridge University Press, 1987, pp. 71-155. Holmstrom, Bengt. “Moral Hazard in Teams,” Bell Journal of Economics (Autumn 1982), pp. 324-40. _______, and Paul Milgrom. “Multi-Task Principal-Agent Analyses: Incentive Contracts, Asset Ownership, and Job Design,” Journal of Law, Economics, and Organization (1991), pp. 24-52. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 24 M AY / J U N E 1 9 9 9 DATA APPENDIX any of these types of compensation. Please do not include profit sharing or employee stock purchase plans. National Longitudinal Survey of Youth (1988-90) The National Longitudinal Survey of Youth (NLSY) data set surveyed 12,686 young males and females who were between the ages of 14 and 21 in 1979. In 1988, 1989, and 1990, respondents were asked whether all or part of their earnings were based on job performance. They were also asked a few questions on their work environment. For instance, we know if the respondents were supervising other employees and whether they had received a promotion since their last interview. Unfortunately, we do not know the precise dollar amounts of incentive pay received by workers nor do we know the proportion of their earnings which is due to pay-for-performance. We asked the following question pertaining to pay-for-performance: “The earnings on some jobs are based all or in part on how a person performs the job (hand card D). On this card are some examples of earnings that are based on job performance. Please tell me if any of the earnings on your job (are/were) based on 1. Piece rates. 2. Commissions. 3. Bonuses (based on job performance). 4. Stock options. 5. Tips. 6. Other.” They also were asked whether they had received a promotion on their current/ most recent job since the last interview. We restricted the sample to individuals who were in the labor market on a fulltime basis. The people who were considered as meeting that criterion were those: 1. Whose primary activity was either working full-time, on a temporary lay-off or looking actively for a job, 2. Who had worked at least half the year since the last interview and who were working at least 20 hours per week. Table 6 Average Real Wage Change, NLSY 1988-90 No Promotion No Bonus Bonus Only Promotion Only Bonus and Promotion All Jobs 6.7% 7.6% 12.0% 11.6% Within Existing Employment Relationships Only 6.2% 7.2% 11.7% 3.8% Incidence of Different Combinations, NLSY 1988-90 No Promotion No Bonus Bonus Only Promotion Only Bonus and Promotion All Jobs 72.3% 10.5% 13.5% 3.7% First Time Observed With Employer 72.7% 9.8% 13.7% 3.8% F E D E R A L R E S E R V E B A N K O F S T. L O U I S 25 M AY / J U N E 1 9 9 9 Figure 1 Computation of Bonuses from PSID Data Evolution of Bonus Incidence in the United States. Variables V5285, V5784, V6393, V6983, V7575, V8267, V8875, V10258, V11399, V12798, V13900, V14915, V16415, V17831, V19131, and V20431: “Head’s income from bonuses, overtime, and/or commissions.” Note that starting with interview year 1986, the codebook specifies that the values for this variable represented any extra bonus, overtime and commissions income not included in heads of household's income from wages and salaries during the preceding calendar year. Therefore, it is possible that some workers who actually received a bonus from their employer did not report it separately from their usual. income. Variables V5419, V5906, V6517, V7120, V7743, V8405, V9036, and V10563: “Did you work any overtime which isn’t reported in [average hours per week worked last year]?” Variables V11142, V12541, V13741, V14831, V16331, V17740, V19044, and V20340: “The values for this variable [...] represent the annual overtime hours worked on all main jobs, if reported separately from regular work hours.” Variables V4515, V5426, V5913, V6524, V7127, V7720, V8388, V9019, V10468, V11659, V13062, V14162, V15170, V16671, V18109, V19409: “How is that? Neither salaried nor paid hourly.” This question refers to the method of pay where the respondent was paid neither a straight salary nor an hourly rate. From this question, we were able to identify those workers paid commissions or a base salary plus commissions. Variables V10465, V11656, V13059, V14159, V15167, V16668, V18106, V19406: This is the overtime hourly rate for salaried workers. Variables V10467, V11658, V13061, V14161, V15169, V16670, V18108, V19408: This is the overtime hourly rate for hourly paid workers. Variables V10469, V11660, V13063, V14163, V15171, V16672, V18110, PSID 1976-91 16 14 % Paid Bonuses 12 10 8 Unemployment Rate 6 Inflation Rate 4 2 0 1976 77 78 79 80 81 82 83 84 85 86 87 88 89 90 1991 Individuals excluded from the sample were those who have been in the military at any time, the self-employed, and all public-sector employees. These restrictions left us with an unbalanced sample of 8,165 observations (3,847 workers), of which 3,832 were paid either a salary or a salary and a bonus. The Panel Study of Income Dynamics (PSID), (1976-91) 12In the PSID, data on hours worked during year t, as well as on total labor earnings, bonuses/commissions/overtime income, and overtime hours, are asked at the year t+1 interview. Thus, we actually use data covering interview years 1976-92. 13Since we cannot separately identify the amount of income derived exclusively par from commissions, we have to remove these workers from the calculations. Note that removing all negative estimates of the bonuses probably biases the par mean bonus paid upward. The sample consisted of white male heads of households aged 18 to 64 with positive earnings for the period spanning the years 1976-91.12 Individuals in the public sector and those who worked less than 500 hours were excluded from the analysis. We know whether each worker was paid a piece rate, a commission, an hourly rate, or a salary. One interesting feature of the PSID for the 1976-91 period is the fact that we were able to determine whether a worker received a bonus during the last year. In the PSID questionnaire, workers were asked the amount of money they received from either working overtime, or from commissions, or from bonuses paid by the employer. Given that workers reported either their number of overtime hours worked (or simply that they worked overtime) as well as the hourly rate for overtime, we were able to compute an estimate of the amounts paid in bonuses.13 F E D E R A L R E S E R V E B A N K O F S T. L O U I S 26 M AY / J U N E 1 9 9 9 V19410: This is the overtime hourly rate for workers not paid either a salary or an hourly rate. Since no information on overtime hours is available before 1984, we could not compute an estimate of overtime income for the years 1976-83. Thus, we simply deleted from the sample all workers who report working overtime between 1976 and 1983 and those who report positive hours of overtime work between 1984 and 1991.14 We also deleted commission workers. It is worth repeating that we may have a noisy measure of bonuses paid. The reason is that the questions on overtime are not clear cut because workers were NOT asked to report any overtime activity during the previous calendar year. Instead, they were asked to report all overtime work not already included in the usual hours per-week worked. Measures of Local Labor Market Conditions. From the beginning of the PSID to interview year 1989, questionnaires were sent to state employment offices asking about current labor market conditions in these counties. Specifically, the unemployment rate measure refers to a specific period during the corresponding interview year. For interview year 1976, the reference month is August; for interview years 197779, it is November; for interview years 1981 and 1983, it is December, while for interview years, 1982, 1984-88, it is September. Starting with interview year 1990, they replaced the variables about the availability of unskilled jobs and unemployment rates with the average annual unemployment rates for the respondents’ counties for the calendar year prior to the interview. These figures come from the U.S. Bureau of Labor Statistics (BLS) Local Area Unemployment Statistics Program. The industry (1 digit) level unemployment rate series also comes from the BLS. 14Restricting the sample to 1984- 91 and using the amount earned in overtime to compute bonuses does not change the results, apart from the standard errors. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 27 M AY / J U N E 1 9 9 9 F E D E R A L R E S E R V E B A N K O F S T. L O U I S 28 M AY / J U N E 1 9 9 9 James B. Rebitzer is the Frank Tracy Carleton Professor of Economics at Case Western Reserve University. Commentary incentive contract should pay attention to all informative indirect measures of output (e.g. rudeness by a sales person) as well as output (sales). M & P correctly observe that compensation is not nearly as responsive to these direct and indirect performance measures as our model would suggest. The question then becomes what’s wrong with the principal agent model? What follows are some answers. James B. Rebitzer A new consensus is emerging among labor economists: Wage formation is the result of something more than the interaction of supply and demand. Organizations structure compensation as part of the larger task of structuring incentives and other features of employment relationships. Competitive labor market forces play a role in this design process, but so do many other things. The core empirical problem posed by this new view of wage determination is understanding the enormous diversity we observe in the structure of compensation. In my view, understanding this diversity is as important to organizational economics as understanding the diversity of species was to natural history in the 19th century. Unlike the natural historians, however, our theoretical prowess far outstrips our empirical knowledge. As a result, our theories are at grave risk of diverting our attention towards the wrong phenomena. The paper by Bentley Macleod and Daniel Parent is important because it builds a bridge between theory and empirics. I think that it goes about as far as you can with the conventional labor economics data sets. As a result of Macleod and Parent’s (henceforth referred to as M & P) efforts, I found myself having to confront directly the issue of how organizationally minded economists should undertake the empirical study of compensation. Complex Tasks The most novel feature of M & P’s analysis of principal agent models is their analysis of how difficult it is to write a complete contract when tasks are even a little bit complex. In their framework, the cost of implementing a complete contract is nkmkg where k is the number of tasks, g is the cost of writing a contract provision, and n and m are the number of productivity and cost levels associated with each task. Because costs increase exponentially with the number of tasks, an optimal contract can’t be written when there is even a small amount of complexity. The clear implication is that “constrained optimal” incentive schemes (i.e. those designed while taking into account the costs of writing complex contracts) will typically focus on a subset of informative actions and outcomes. If complexity causes firms to operate with incentives that do not include all informative measures of input and output, then multi-task issues must be very commonplace. Multi-task models apply to settings where only a subset of value creating activities can be metered and incented. In a multi-task world, high powered incentives are costly because they divert effort and attention away from valuable, but hard to meter, activities.1 M & P go on to suggest that the problem of contract complexity can be resolved by relational contracts (i.e. incentive arrangements that rely on noncontracted, ex post rewards based on the subjective evaluations of the THEORY The authors begin with a very nice exposition of a principal agent model of incentive design. The key points of the model are: (i) incentives are costly because agents are risk averse; (ii) the optimal F E D E R A L R E S E R V E B A N K O F S T. L O U I S 29 1 Holmstrom and Milgrom (1991) argue that some multitask problems can be reduced by the design of jobs. M AY / J U N E 1 9 9 9 supervisor). This is an interesting claim. I don’t think, however, that M & P adequately address the question of how subjective, ex post evaluations solve the complexity problem. I see two classes of answers to this question. The first answer focuses on subjective evaluations of complex behaviors. Human beings are very good at parsing information that is very complex and messy. We can form accurate judgements about things that we couldn’t begin to write down in an explicit contract or algorithm. Think how easily young children solve the problem of recognizing faces or understanding a sentence, both problems that are notoriously difficult to write down explicitly. Subjective impressions about the contributions of individual employees may, therefore, be a “good enough” foundation on which to anchor incentive pay. The second approach focuses on timing. By relying on ex post assessments, managers can reduce the complexity of the assessment task because they do not need to consider states of nature that didn’t actually occur. Accepting for the moment that incentive schemes based on subjective, ex post assessments and rewards solve the complexity problem, it seems clear that relational contracts raise other problems. In particular, subjective incentive schemes require that supervised employees trust supervisors to fairly assess and evaluate performance.2 Arranging things so that employees can trust the subjective assessments of managers is no doubt difficult and costly. I would have liked to see M & P analyze the benefits and costs of relational contracts relative to feasible principal agent contracts (i.e., a contract that relies on a small number of indicators) and to then identify conditions under which one incentive scheme outperforms the other. 2 See Baker, Gibbons and Murphy’s (1994) analysis of incentive compatible implicit contracts. 3 See Levy and Murnane (1996) and Ichniowski, Shaw, and Prennushi (1997). incentives that ignores job design is probably going to get important things wrong. Casual examination of the literature on high-performance work organizations suggests what organizations want from their employees is problem solving. Specifically, employers increasingly want to combine the tacit information available to front-line employees with their problem-solving skills to achieve higher levels of quality and service. Car manufacturers and steel companies want production workers to use local, tacit information to sustain a process of continuous quality improvement. Insurance companies want their front-line employees (those having direct contact with customers) to have the information, knowledge and communication skills to offer full-service, one-stop shopping to all customers.3 Problem-solving responsibilities (and the associated “soft tasks” like communication and team work) are extremely complex and, following M & P’s logic, are probably impossible to fully specify ex ante. My impression is that high performance organizations spend at least as much time and effort on job design as incentives. Experts, for example, are divided on the importance of incentives for modern manufacturing, but they are unanimous on the importance of such job-design issues as problem-solving teams and giving employees responsibility for their own quality control. Economics offers only the most rudimentary theory of job design and this lacunae poses a real problem for developing economic theories of incentive design. The Sociology of Groups Another limitation of principal agent models (one that M & P are aware of but didn’t write about in this paper) has to do with the individualistic nature of the production process in the principal agent model. A large amount of human productive activity takes place in small groups and the sociology of groups must matter for the design of incentives. In standard principal agent models, incentive pay is costly because employees Job Design and Problem Solving The principal agent model takes jobs, or tasks, as given. This view is certainly too narrow. Organizations design both jobs and incentives to elicit desirable behaviors from employees. A theory of F E D E R A L R E S E R V E B A N K O F S T. L O U I S 30 M AY / J U N E 1 9 9 9 are risk averse. Thus, risk aversion causes employers to operate with lower-powered incentives than would otherwise be the case. Small groups, however, create additional incentive costs. If, for example, individuals care about their income relative to others in the work group, firms may eschew high-powered incentives to avoid morale-lowering invidious comparisons. Concerns about morale may cause supervisors with responsibility for performance evaluation to alter their evaluations so that all individuals perform “above the average.” In this way, group sociology will alter the functioning of the relational contracts described by M & P. Group sociology also can make problematic incentives based on objective performance measures. If individuals in groups can engage in mutual help activities, high-powered incentives based on individual output can discourage valuable mutual help activities in groups.4 The sociology of small groups does more than create incentive costs. Informal interactions among members of the group can also create additional incentive instruments—notably peer pressure. If individuals in a work group have superior information about local conditions than managers, mobilizing peer pressure can significantly improve performance. they are normally conceived nor expectations of future consequences enter directly into the calculus…Rule following is grounded in a logic of appropriateness… The process is not random, arbitrary or trivial. It is systematic reasoning, and often quite complicated. In those respects, the logic of appropriateness is quite comparable to the logic of consequences. But rule-based decision-making proceeds in a way different from rational decision-making. The reasoning process is one of establishing identities and matching rules to recognized situations. (March 1994, p. 57-58) In a similar vein, some sociologists and a few economists have written about the detrimental effect that high-powered incentives can have on intrinsic motivation. Consider, for example, Akerlof’s (1982) gift-exchange models. In Akerlof’s models individuals are powerfully motivated by a desire to take actions that are appropriate to the relationship they have with their employer. By offering high and stable wages, employers can reduce opportunism by inducing individuals to adopt actions appropriate to a gift exchange. Close monitoring and/or high-powered incentives, in contrast, cause individuals to adopt actions appropriate with a low-trust situation. Why might high-powered incentives undermine intrinsic motivation? I haven’t found a convincing answer to this question.5 One plausible possibility is that incentives work differently depending on the beliefs employees attribute to management’s use of incentives. Close monitoring may send the message that the employer doesn’t believe employees are reliable people. Perhaps highpowered incentives send a similar signal. Two provocative implications flow from the social psychological view of incentive problems: The Social Psychology of Incentives The principal agent model requires that individuals be rational opportunists. This means that individuals take actions now with expectations about how these actions will influence their future economic welfare. This view of decision-making is ubiquitous in economics but not elsewhere. Consider, for example, the following quote from James March’s How Decisions Happen. March argues that rule-following can be more important than rationality in the operation of incentives. 4 These issues are analyzed in • Individuals may be less opportunistic than conventional models suppose; and • Treating people as if they were opportunistic can create more opportunism. When individuals and organizations fulfill identities, they follow rules or procedures that they see as appropriate to the situation in which they find themselves. Neither preferences as F E D E R A L R E S E R V E B A N K O F S T. L O U I S 31 Encinosa, Gaynor, and Rebitzer (1998). 5 See Kreps (1997) for an intriguing statement of the problem. M AY / J U N E 1 9 9 9 These claims are inherently difficult to investigate empirically because the smartest opportunists will shirk only where it is hardest to catch them. If, however, they are correct, the principal agent model profoundly misunderstands the ways in which incentives motivate behavior. offers me a great deal of autonomy, variety and task completion. It is not hard to imagine, however, that a manufacturer’s representative might report that she also enjoys lots of autonomy, variety and task completion. Does this mean that autonomy, variety and task completion mean the same thing in these two occupations? I don’t think so. Using cross-occupational variation in a small number of job characteristics to explain individual compensation strikes me as especially tricky because of the plethora of omitted variables that may be correlated with occupation-level job characteristics. Piece rates make good sense for manufacturers’ representatives but not for university professors, because the product of the former occupation is easy to measure and the output of the latter is complex, multi-dimensional and hard to measure. As this example (and our previous discussion) indicates, some of the most interesting omitted variables are likely to relate to job-design issues because job design and incentives are likely to be jointly determined. The third set of empirical results in the paper concern the relationship between unemployment rates and the likelihood of bonus pay. M & P ask us to imagine relational contracts in which there are essentially two types of ex post sanctions/rewards, dismissal or the payment of a bonus. They argue that as local unemployment rates fall, dismissal threats become less effective, so firms come to rely increasingly on bonuses as an alternative incentive mechanism. This logic implies that the incidence of bonus pay will be a decreasing function of the worker’s unemployment rate. If local unemployment rates are a good measure of the worker’s alternative job opportunities, we would then expect the incidence of bonus pay to decline as local unemployment increases. An alternative explanation is that bonuses are a form of profit sharing, and their incidence should, therefore, be positively correlated to firm profits. If industry unemployment rates are a good proxy for profits, we would then expect the incidence EMPIRICS M & P’s empirical excursions use data from conventional microeconomic surveys (NLSY 88-90 and the PSID) to examine the incidence of different types of compensation. In the first set of results (Table 1), M & P present cross-occupational variation in the form of compensation. From these results, it is clear that the form of compensation varies in intriguing ways with broad occupational categories. Piece rates, for example, are used widely for precision machine operatives (36.81%) and not for textile operators (9.76%). Not surprisingly, sales workers are frequently paid by commission (37.98%), but so are personal service workers (20.25%). These results are intriguing, but relying on such broad occupational averages may conceal as much as they reveal. My own research on incentives in medical groups has convinced me that there is enormous heterogeneity of compensation arrangements even within narrowly defined physician specialties (Encinosa, Gaynor and Rebitzer, 1998). I would be interested, therefore, in seeing a breakdown of the variation within and between the broad occupation group listed in Table 1. The second set of empirical results (M & P’s Table 2) indicate that occupational characteristics do matter for the incidence of piece rates and commissions. The authors regress the type of compensation an individual receives against occupationlevel averages of perceived job autonomy, task completion and variety. The finding that autonomy, task completion and variety matter for the form of compensation is intriguing, but difficult to interpret because it relies upon the interpretations that respondents give to these concepts. My job as an economics professor, for instance, F E D E R A L R E S E R V E B A N K O F S T. L O U I S 32 M AY / J U N E 1 9 9 9 of bonus pay to rise as industry unemployment rates fall. The results in M & P’s Table 3 indicate that the industry unemployment rate has little effect on the incidence of bonuses while the local unemployment rate has a negative impact. I am not convinced by this interpretation of the results. It seems to me that local and industry unemployment rates are poor proxies for the underlying variables of interest. Profits for firms offering geographically restricted services (e.g., hotels, restaurants, hospitals, cabs, etc.) are more likely to be affected by local rather than industry unemployment rates. Profits at manufacturing enterprises also may respond more to local than industrylevel unemployment if firms are heavily dependent on a local customer. M & P’s model hypothesis is that relational contracts are not incentive-compatible where the worker faces high unemployment. They investigate this idea using annual unemployment rate variation. Relational contracts, however, depend critically on trust and firms should therefore be loathe to alter them in response to short-term spikes in the unemployment rate. Thus, investigation of M and P’s hypothesis should rely on relatively long-term shifts in unemployment. Indeed, an interesting extension of M & P’s empirical work would be to examine the effect of short- and long-term movements in the unemployment rate. If short-term movements have more influence on compensation than long-term movements, this would argue against the incentivecompatible, relational contracts. If I am right that conventional data are not up to task of understanding the determinants of compensation, then what would the right data look like? I think the ideal data set would have the following features: • Data on a large number of firms in a specific industry and employees in specific jobs; • Data on the form of compensation contract as well as compensation outcomes; • Qualitative interviews about what managers and employees see as critical tasks of the job; • Information on how employee behaviors varied with variation in the form of compensation, intensity of supervision, job design and work setting. It is clear to me that the demanding data-collection efforts entailed by this list would only be manageable for narrowly focused case studies. For the same reason, it is unlikely that any single study would succeed in collecting data along all four dimensions at the same time. If large numbers of empirical organizational economists joined their colleagues in sociology and went about the business of collecting and analyzing this sort of data, we would be left with a rich assortment of case studies. Each of these studies would be limited and inadequate for developing a general theory of compensation. Taken together, however, these studies could (like the natural histories of 19th century biology) form patterns that we could then use to construct a suitable and general theory of compensation. WHAT WOULD THE RIGHT DATA SET LOOK LIKE? M & P’s paper is important because it takes good and reasonable economic models of compensation and compares their predictions to the patterns visible in the data. The problems they encounter in their empirics stem from the fact that the information collected in conventional microeconomic data sets are simply too distant from the phenomena of interest (i.e., jobs and incentives in organizations). REFERENCES Akerlof, George A. “Labor Contracts as Partial Gift Exchanges.” Quarterly Journal of Economics, (November 1982, 97:4), pp. 543-69. Baker, George, Robert Gibbons, and Kevin J. Murphy. “Subjective Performance Measures in Optimal Incentive Contracts.” Quarterly Journal of Economics, (November 1994), pp. 1125-56. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 33 M AY / J U N E 1 9 9 9 Encinosa, William E., Martin Gaynor, and James B. Rebitzer. “The Sociology of Groups and the Economics of Incentives: Theory and Evidence On Incentives in Medical Groups,” (unpublished, May 1998). Holmstrom, Bengt and Paul Milgrom. “Multi-task Principal-Agent Analyses: Incentive Contracts, Asset Ownership, and Job Design.” Journal of Law, Economics and Organization, (Spring 1991), pp. 24-52. Ichniowski, Casey, Katheryn Shaw, and Giovanna Prennushi. “The Effects of Human Resource Management Practices on Productivity: A Study of Steel Finishing Lines.” American Economic Review, (June 1997), pp. 291-313. Kreps, David M. “Intrinsic Motivation and Extrinsic Incentives.” American Economic Review, (May 1997), pp. 359-64. March, James G., A Primer on Decision Making: How Decisions Happen, New York: The Free Press, (1994). Murnane, Richard J. and Frank Levy, Teaching The New Basic Skills, New York: The Free Press, (1996). F E D E R A L R E S E R V E B A N K O F S T. L O U I S 34 M AY / J U N E 19 9 9 Truman Bewley is an Alfred Cowles Professor of Economics at Yale University. He is deeply grateful to Joseph Ritter for insightful comments on various drafts of this paper. Work Motivation I strove to avoid sample bias by holding interviews in a large and diverse set of companies as well as by using many distinct avenues of approach to gain access to them. Using these methods, I avoided talking to people from only a few circles of friends. The companies represented a broad spectrum of industries and a full range of sizes and financial conditions. Some were bankrupt, many were shrinking and experiencing heavy layoffs, and some were growing rapidly. Some had been founded only recently, while most were well-established. Some were unionized, whereas many had no union presence. Some were public corporations and others were closely held or family-owned. I made a point of finding businesses that had cut or frozen pay during the recession. There were few such; most firms continued to grant regular raises. My method did not yield a valid opinion survey nor reliable statistics on the incidence of various business practices. I believe, however, that I gathered valuable information about what happens in the labor market during a recession and what business people and labor leaders think about layoffs and pay cuts. The explanation of wage rigidity given by more than 275 business people and labor leaders I interviewed was based on views of worker motivation that deviate from the standard model. In this paper, I formulate a somewhat speculative model of work motivation stimulated by what I heard. The model incorporates ideas from psychology into the utility-maximizing framework of economics. Truman Bewley D uring 1992 and 1993, I undertook a field study in the Northeast of the United States to learn why wages and salaries seldom fall during recessions.1 I interviewed more than 330 business people, labor leaders, counselors of unemployed workers, labor market intermediaries (headhunters), labor lawyers, and management consultants. The purpose of the study was exploratory; much of my effort went into the search for hypotheses rather than tests of specific ones. For this reason, I did not require informants to answer a fixed list of questions, but informed them of the purpose of the study and invited them to tell me what they thought was relevant. I intervened only occasionally to seek clarification, show interest, or nudge the discussion in new directions. Only after informants had spoken at length did I ask specific questions to cover points that interested me. I usually avoided asking about economic theories until the end of interviews. Such questions sometimes stopped conversation, because the theories seemed naive and the questions led respondents to try to think like an economist, rather than to explain their world concretely in their own terms. Some business people refused such open-ended interviews, probably because they feared that while talking loosely they might say something that would embarrass them or hurt their company. I concluded that low response rates might make a random sample unrepresentative. (I had much less difficulty gaining the cooperation of the other types of respondents.) Most interviews with business people were obtained through personal contacts or by telephoning people and persuading them to cooperate. Often, people I interviewed arranged further interviews. WAGE RIGIDITY AND MORALE In this section, I will summarize what I heard in interviews, giving the reasons for wage rigidity explained to me by business people and labor leaders. In the first instance, the resistance to pay reduction came from managers, not from workers, though it was anticipated employee discontent that motivated management to oppose pay cuts. The discontent, usually described as poor F E D E R A L R E S E R V E B A N K O F S T. L O U I S 35 1 The results of the study are reported in a book, entitled Why Wages Don’t Fall During a Recession, to be published by Harvard University Press in 1999. M AY / J U N E 19 9 9 morale, would not necessarily be expressed openly. Nonetheless, business people believed it could be harmful enough to cause monetary losses that would exceed the savings from a pay cut. The downward rigidity of the pay of existing and of newly hired employees have separate explanations. The reason almost all managers gave for not cutting the pay of existing employees was concern about morale. New employees would probably object little if, before they applied for their jobs, pay rates for new hires had fallen by no more than the pay of existing employees in the same jobs. New employees, in contrast, feel it is inequitable to be paid according to a scale lower than the one that applied to colleagues that were hired earlier. For this reason, downward pay rigidity for new hires exists only because the pay of existing employees is rigid. The pay of new hires is usually downwardly flexible when co-workers do not have enough contact with each other to know each other’s pay. This circumstance arises typically when labor turnover is high and when a large fraction of the employees work part-time on schedules that seldom overlap. Typical examples are floor crews in fast-food restaurants and in supermarkets. Good morale means many things in industry: a willingness to cooperate with company objectives, a sense of common purpose consistent with the firm’s goals, enthusiasm for the job, happiness, toleration of unpleasantness, moral behavior, and mutual trust. Business people value good morale because it reduces labor turnover, makes it easier to recruit good workers, and increases the productivity of the existing work force. The increase in productivity arises not so much because employees work harder at assigned tasks that are monitored, but because workers do the right thing even when no one is watching. They do extra things without instruction, make suggestions for improvements, help each other, and share information with each other and with superiors. Good morale is thought to be especially important for productivity in jobs where it is difficult to monitor performance, where good performance requires imagination and creativity, and where workers must deal with customers. Morale is important in the latter case, because employees handle customers better when cheerful. The morale of existing employees is hurt by pay cuts because of what may be called the insult effect and the standard of living effect. The latter occurs because lower living standards distract and aggravate workers and cause them to blame the company for the difficult adaptation to lower incomes. The insult effect occurs because workers associate pay with self-worth and recognition of their value to the company. Many workers receive regular increases, grow used to them, and interpret them as recognition of loyalty and good performance. Hence, a pay cut is interpreted as a signal of dissatisfaction with employees, even if everyone’s pay is reduced. These effects apply to both real and nominal pay reduction, though the effects of an abrupt nominal cut are stronger than those of a slow decline in the purchasing power of pay. Another reason a pay cut is interpreted as an affront is that it is viewed as unfair, because the company takes something away while giving nothing in return. A pay cut is not felt to be insulting if management can convince workers that the reduction is justified; that is, if it prevents a large number of layoffs. Pay cuts typically occur when a business is in danger of closing or has trouble competing in product markets. In these circumstances, workers usually accept cuts. Such circumstances are rare, however. A central fact of life for most businesses is that pay rates have little impact on total employment. That is, in most firms the elasticity of demand for labor is small. Pay cuts are more common among firms where this elasticity is high. Business people and labor leaders were confident they usually could convince employees, with some effort, that pay cuts were justified, if indeed they were. I was told that workers refuse to believe what they are told about company difficulties only when management has a reputation for duplicity, when relations between management and union representatives are bad, F E D E R A L R E S E R V E B A N K O F S T. L O U I S 36 M AY / J U N E 19 9 9 or when workers recoil from facing reality. I found little support for the many theories of wage rigidity based on information asymmetries, and, in particular, theories based on the assumption that management cannot persuade workers that low profits or competitive conditions require pay reduction. The general thrust of what was said was that normally information flows freely enough within businesses so that most employees know when their company is in trouble. In some small- and medium-sized companies, the workers may know this before management does, because it is low-level employees who take orders and keep accounts, and gossip spreads quickly. Nevertheless, asymmetries of information underlie the explanation of wage rigidity. Morale is important for productivity because management finds it prohibitively expensive to monitor employees closely. For this reason, companies rely on workers to do what they are supposed to do without being told, even when supervisors are unlikely to check up on them or never do so. Workers are likely to be so cooperative only if they have good morale. Though employees expect to share in company success through pay increases, they do not expect to share in its losses by having their pay reduced. Adjusting to lower income is too painful for workers to endure relative to the sacrifices made by company owners, and pay cuts raise the awkward issue of the disparity between the incomes of workers and owners. Morale is fragile and can be destroyed quickly by matters more minor than pay cuts. It can be hurt by any form of unfair or inequitable treatment by management, where the standards of fairness, especially regarding pay differentials, are often determined by company or industry traditions rather than by absolute standards of justice. Good morale normally takes a long time to build. It is fostered by: • Recognition and reward of contributions to the company; • Good explanations of the social contribution of the company’s products; and • Good explanations of a worker’s role in the production process. Collective activities within the company, such as charity drives and company picnics, also improve morale, as does almost anything that encourages workers to think of people other than themselves. Morale is hurt by threats, such as threats of being fired if performance is substandard. Though companies fire some workers, it is thought to be a bad business practice to have people work in a negative, menacing atmosphere. Just this style of management is sometimes used, however, with low-level and low-paid labor doing short-term jobs that are easily monitored. Firing is most useful for ridding an organization of scoundrels and ne’er-do-wells rather than as a way of motivating ordinary workers to perform. Positive incentives and an optimistic atmosphere encourage performance more effectively than do threats. Most workers want to do well and will do so, if given the opportunity, and if they understand what they are supposed to do. Furthermore, many people enjoy their work.2 A sense of pride, duty, and accomplishment can make even disagreeable jobs bearable. Nevertheless, strict discipline is necessary for good morale, for if some workers are allowed to get away with slacking, those who work hard feel they are being treated inequitably. What has been said is a fair summary, I believe, of the dominant views of business people and labor leaders. I now turn to the problem of formulating these ideas in ways that may be useful for economic theory. INTERPRETATION OF MORALE • Frank, but good, relations between subordinates and superiors; Good morale has three components: • Identification with the organization or internalization of its objectives, • Prospects for economic security and progress within the company; F E D E R A L R E S E R V E B A N K O F S T. L O U I S 37 2 Juster (1985) found in a survey that ordinary people preferred work to activities associated with leisure. M AY / J U N E 19 9 9 • Good moods, and unconsciously or semi-consciously. Conscious mental operations are slow, though adaptable. Unconscious ones are rapid, though restricted to learned routines. Thus, it is hard to learn to play the piano or to speak a foreign language, but, once those skills are learned, they occur smoothly and with only general conscious direction. Identity includes many such mental subroutines. For instance, it would be nearly impossible to function socially if we constantly weighed self-interest against collective advantage. These calculations are replaced, in most cases, by rules about how we should behave, who are our friends or foes, and what we should expect of them. Those rules are all part of identity. In summary, the function of identity is to make mental activity more efficient, and identity’s mental mechanism is a set of unconscious goals and mental subroutines. An additional advantage to an individual of group identification is that it contributes to his or her sense of being powerful, valuable, important, and wanted. A sense of self-worth is needed by people because it gives them a reason to survive and promote their own interests. Without a sense of worth and the power to shape their own lives, people can be incapacitated by what psychologists call learned helplessness (Gleitman, 1995, pp. 133-35). An obvious benefit of group identity is that it makes it easier to work with other people, which is important because most productive human activities require cooperation. This benefit does not explain morale’s fragility, however. Its function may be to protect individual self-interest. Though commitment to a group helps overcome prisoner’s dilemma or free-rider problems arising in cooperative activity, the same sense of responsibility exposes individuals to exploitation. It is useful to have a system that balances private and group advantage, and conventional standards of fairness offer an orderly way of accomplishing this. These establish rules of reciprocation among group members and between them and the organization, and the duties specified by these rules are accepted by members when they agree emotionally to join. Perhaps the brittleness of morale is a • Trust and mutual affinity among members of the organization. A person may be judged to have internalized the objectives of their organization if they act to advance its interests without specific instructions and without any possibility of being monitored and rewarded. Identification is manifested by internalizing the organization’s objectives as well as by actions that demonstrate membership and by expressing feelings of belonging. Moods are states of mind that affect work habits and the pleasure or displeasure derived from work. Cooperation within an organization is fostered by a network of trusting relationships among employees. Cooperation may not be directed toward helping the organization, however, unless members accept its objectives. Though the social network is an important component of morale, it is not hurt by pay cutting, so I give it little attention.3 Identity and Internalization 3 Some industrial psychologists measure morale as the existence of effective groups (Blum, 1956, pp. 163-69). It is clear that human beings have the capacity to identify with organizations and to internalize codes of behavior and the interests of others. Experimentally, it is easy to induce people to identify with a group and to act in its interests (Tajfel, 1970, and Turner, 1987). Children show empathy for others at a young age and learn to internalize social and moral rules (Gleitman, 1995, pp. 550-58). It is impossible to know whether the capacities for empathy, morality, and group identity are accidental or if long ago they evolved in humanity because they increased chances of survival, nor do we need to know the answer to this question. What is important is that the capacities exist. A psychological theory of organizational identity should describe its function or purpose and the mental mechanism of which it is a part. Identity, in general, is a person’s image of who they are. One advantage of identity is that it simplifies mental processes by summarizing a person’s goals and providing a set of rules for how to behave. A great deal of what we do mentally is done F E D E R A L R E S E R V E B A N K O F S T. L O U I S 38 M AY / J U N E 19 9 9 self-protective reflex provoked by violation of fairness standards. Reading a psychology textbook, such as Gleitman (1995), makes it clear that many parts of the nervous system operate through offsetting pairs of activating and inhibiting signals. The teetering between group commitment and indignant rebellion may reflect just such a pairing that is built into the psyche. It remains to be explained why unjustified pay cuts impair identification with the employer. A superficial answer, given earlier, is that they are regarded as unfair. Fairness specifies rules of reciprocation and workers receive nothing in exchange for a pay cut that saves few jobs. It might be that in a different world, workers would view wages and salaries coolly as fluctuating market prices and would accept price declines as a normal part of business life, just as salespeople accept large income fluctuations. Most people do not think this way. I was told many times that workers do not view themselves as commodities and inevitably interpret pay cuts as statements about how satisfied the company is with them, because, in their experience, pay increases signal appreciation of workers’ contribution. with a lion hunting an antelope. In chasing an antelope, the lion expends energy, loses time, and risks injury. These costs must be weighed against the probability of catching the prey and the pleasure of eating it. Imagine that the lion unconsciously, or half consciously, weighs the costs against the benefits before deciding whether to chase his prey and before choosing the level of physical effort to expend on the pursuit. Once he makes his decision, the lion’s mind automatically adjusts his mood and level of nervous and physical arousal to handle the effort required. If the lion decides not to go after the antelope, he will not be aroused. He will feel lazy and may find running uncomfortable. In contrast, if he decides to try for a kill, he will be mobilized, excited, and will probably be exhilarated by the effort. Given this decision, he will consciously decide how much effort to put into the hunt and how to go about it. We may imagine that the lion’s unconscious mind chooses the mood and level of arousal so that his conscious mind chooses an effort level that optimizes a utility that is unconscious. This unconscious utility depends on the probability of success, energy expenditure, and risk of injury. The lion’s unconscious choice of mood may be constrained by his preexisting state of mind. If he just lost his wives to a rival, he may be discouraged and not feel like hunting, whereas if he is a hopeful young bachelor, he may feel vigorous. Another illuminating analogy may be that of a virtuoso pianist. For an appreciative and sensitive audience, she will probably play at her best and love doing so. If she hears snores and catcalls, she will no doubt feel her fingers stiffen, stumble, and hate playing. It is important, in my opinion, to recognize that mood automatically adjusts to fit the perceived net benefits of tasks. I believe it is general human experience that capacities to act and perceptions of pain or pleasure depend on circumstances. Danger stimulates us to fight or flee. Anger makes us ignore pain and danger. Though deprivation of life’s necessities causes us discomfort and unhappiness, we get used to prolonged hardship, probably so that we can cope with the unavoidable misery. Soldiers and prisoners living in frightful conditions Mood, Work Effort, and Its Disutility Managers and labor leaders did not usually speak of jobs as disagreeable, but assumed that employees liked to work. They said that one of the bad effects of layoff was loss of the pleasure of working and of social contacts on the job. If, in the standard model of work, however, we assume that effort brings positive rather than negative utility, then people should work hard, even if they have no financial incentive to do so. This implication conflicts with common sense. Though volunteer labor makes important contributions to society, it is hard to imagine that it would succeed in producing ordinary economic output. Other phenomena that are inconsistent with the usual model of work effort are the importance of mood to job performance and satisfaction from work. I was told that bad moods are distracting and that they increase fatigue, discomfort, and accidents. In order to make sense of these observations, it may be helpful to consider an analogy F E D E R A L R E S E R V E B A N K O F S T. L O U I S 39 M AY / J U N E 19 9 9 eventually cheer up and joke about their state, though, of course, they are not happy. It is a mistake to separate the disutility of labor from the utility of its reward, or to imagine that labor is normally perceived as disagreeable. The utilities of labor and its reward interact. level there. These conclusions seem consistent with reality. The model does not distinguish reward from punishment, despite the fact that this distinction is crucial in reality. There is no way of determining what level of utility marks the boundary between punishment and reward. I now modify the above model to obtain one that retains its plausible conclusions and yet does not represent labor as a burden. It gives a role to emotion in mobilizing and directing the powers of mind and body and includes a distinction between reward and punishment. I try to model mood, because it has an important role in the explanations of wage rigidity given by managers and labor leaders. The model is suggested by the analogies described in the previous section. Focus on an action (or program of actions), e, to be taken by a person over a fixed period of time. Though e may be thought of as effort, it is better to interpret it as productive activity. The action has an unconsciously felt mental and physical cost, measured as the number, C(e), and earns income w(e), which might be a wage paid by an employer.4 The unconsciously felt benefit to the worker of the wage is the number B(w(e)), and the net unconscious gain is A FORMAL MODEL In the usual incentive model, a worker expends effort, e, which is a non-negative number, and receives in exchange a wage, w(e), which is a non-decreasing function of e. The worker chooses e so as to maximize u(w(e)) – c(e), where both u and c are increasing functions and u is concave and c is convex. The first term is the utility of consumption purchased with the wage and the second term is the disutility of effort. In this model, the consumer prefers to expend as little effort as possible to earn a given income, so that if the wage does not increase with effort, the consumer expends none of it whatsoever. Because effort creates disutility, people acting according to the model would experience work as unpleasant, which is contrary to what most people say (Juster, 1985). If we try to escape this difficulty by assuming that c(e) is zero, then the worker offers the maximum effort possible if w(e) increases with e, an implication that contradicts common sense. This difficulty can be evaded by assuming that c(e) decreases with e until it reaches a certain level, beyond which it increases. If the functions u and c are differentiable, then the optimum level of effort satisfies the equation 4 5 The earnings, w(e), could be a vector including pay, praise, promotion, and other rewards. Here, and elsewhere, I choose the additively separable functional form for convenience of exposition, not out of conviction. B(w(e)) – C(e).5 Unconscious goals could include the basic psychological drives as well as fidelity to family, firm, or country. Assume that the function B is increasing and strictly concave. I propose that people unconsciously adjust their mood and general state of mobilization so that conscious choices maximize B(w(e)) – C(e). The conscious person does not choose e but makes a decision (or program of decisions), d. The actual action taken is e = E(d,m), where m is the person’s mood and state of mental and physical arousal. The decision d might correspond to the pace of work desired by the person, whereas E(d,m) is the realized pace of work; the person might actually work faster or slower than he or she intended. The person’s consciously experienced utility is U(w(E(d,m)), m) + V(E(d,m), m), where the first term is the utility of the earnings and du(w( e )) dc( e ) = , de de from which it is easy to see that increasing the level of the function w(e) by adding a positive constant decreases effort (or, more accurately, does not increase it), whereas optimal effort increases (or does not decrease) when the slope of w(e) at the optimum is increased without increasing the function’s F E D E R A L R E S E R V E B A N K O F S T. L O U I S 40 M AY / J U N E 19 9 9 Figure 1 the second is the utility from the action itself. The person chooses the decision, d = D(m, w), so as to solve the problem [ B ] max U( w( E( d , m )), m ) + V( E( d , m ), m ) , d ∈D Northeast Frontier where D is the set of possible decisions. The unconscious side of the person chooses the mood, m, so as to maximize the unconscious utility, that is, to solve the problem [ Optimum B – C = constant The Set of Possible Pairs, (–C(E(D(m, w), m))), B(w(E(D(m, w), m))) max B( w( E( D( m, w ), m ))) m ∈M − C( E( D( m, w ), m )) , ] where M is the set of possible moods. If the person has a preexisting state of mind or mood, then his or her unconscious self may not be able to choose m freely, but must chose from a subset, SM, of M. The subset SM may be thought of as representing the restrictions imposed by solution of a larger unconscious utility maximization problem that determines the context of the one under consideration. For instance, the person may be frightened by some danger, which may be escaped through the actions under consideration. When interpreted properly, the standard results mentioned earlier (regarding incentives) apply to the new model. Imagine a two-dimensional plot with –C(E(D(m, w), m)) on the abscissa and B(w(E(D(m, w), m))) on the ordinate, as in Figure 1. The unconscious self chooses m to maximize the sum of the two components, so that the northeast frontier of the plot is the relevant set of points. I compare two earnings functions, w and w′, and assume that –C If w9 is w plus a positive constant, then, because of the strict concavity of the function B, the northeast frontier of {(–C(e), B(w′(e))): e ∈ E} is no steeper than that of {(–C(e), B(w(e))): e ∈ E}. It follows that at the optimum the disutility of effort, C(e), is no higher with the earnings function w′ than with the function w. The second standard result regarding incentives is that making w steeper at the optimum does not decrease effort. In the new model, it is not possible to speak of the slope of w because the action variable, e, may not be a number. By an analogous definition, however, w′ is “at least as steep” as w at e if w′(e) = w(e), where e is the optimum for w, and if w′(e) ≥ w(e), whenever w(e) ≥ w(e), and w′(e) ≤ w(e), whenever w(e) ≤ w(e). Given this definition, it is obvious that if w′ is at least as steep as w at e, then –C(e′) ≤ – C(e ) and w′(e′) ≥ w(e), where e′ is the optimum with earnings function w′. That is, steepening the earnings function does not decrease the unconscious disutility of effort at the optimum. The utility V(E(d, m), m) may increase with effort if mood favors exertion, though V may decrease with effort when it is increased beyond a point appropriate for the mood. What I have in mind may be explained by returning to the usual model in which the effort variable is a number corresponding {E(D(m,w),m): m ∈ SM} = {E(D(m,w′),m):m ∈ SM} ≡ E, so that the set of possible actions achievable by manipulation of mood does not depend on the earnings function. Then, the twodimensional plots are of the sets {( −C(e), B(w( e ))): e ∈ E} and {( −C(e), B(w′(e))): e ∈ E}. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 41 M AY / J U N E 19 9 9 to the pace of work. In this spirit, assume for the moment that d and m are non-negative numbers, where larger values of m correspond to a better mood. Assume also that E(d, m) = d, and that w(e) = e, so that e and w can be suppressed. Finally, assume that U and V are twice differentiable functions satisfying the following conditions: rationality take account of limits on the ability of the conscious mind to reason and use information. No doubt, sentiment influences imperfect logic, so that a more realistic version of the above model should take account of the effect of mood on reasoning. The model accomplishes the objectives of giving mood a role in motivation and allows workers to enjoy positive utility from work, while preserving the obvious common sense results about the impact of financial incentives. The model fails to permit a distinction between reward and punishment, however, nor does it include morale or explain why pay reductions have such a severe impact on mood. ∂U( d,m ) ∂2U( d,m ) ∂2U( d,m ) > 0, < 0 , > 0, ∂d ∂m ∂d ∂d 2 ∂V( 0, m ) ∂2 V( d,m ) > 0, < 0, ∂d ∂d 2 ∂2 V( d,m ) and > 0, ∂m ∂d Reward and Punishment for all d and m. Let d = D(m) be the solution to the problem I now assume that the unconscious mind forms a notion of what is normal in terms of unconscious living standards. This idea of normality may be useful to an individual for two reasons: It tells the mind what to store as habits or mental subroutines, and it serves as a trigger level for alarm. The mind adapts habits to the way of life that it expects to be normal. Decline of living standards below a normal level signals the unconscious that something is wrong. This provokes anger, unhappiness, or distress. These moods, in turn, stimulate the conscious mind to find solutions for the problems that have arisen. It is not efficient for the conscious mind constantly to be stimulated and on the look-out for new solutions. Bad moods and the efforts they incite are exhausting. Therefore, bad moods should be called upon only when needed. The normal or expected path of welfare may grow, shrink, or fluctuate over time. For instance, salespeople expect their income to fluctuate sharply and probably react badly only to prolonged patterns of low income. A fall in welfare below the expected level may not trigger alarm if the conscious mind can persuade the unconscious that there is no reason to worry, the bad situation will soon be rectified, or there is nothing that can be done. The unconscious probably adapts gradually to lower welfare, as do the soldiers I mentioned earlier. max[U(d,m) + V(d, m)]. d $0 Under the given conditions, it is easy to see that D is a nondecreasing function of m and is increasing at values of m for which D(m) > 0. That is, improved mood increases effort. Notice also that at the optimum, ∂V(d, m) < 0, ∂d so that the worker finds increased effort unpleasant. From now on, I drop the assumption that d and m are numbers. Rationality It is natural to ask whether people behaving as in the above model are rational. Economists define people to be rational if they reason correctly and use all available information in order to maximize their utility. The model is consistent with rationality, if we allow utility maximization to occur at two levels, the conscious and the unconscious. The effect of mood on realized actions and on conscious objectives does not contradict rationality. In a loose sense, however, the model is inconsistent with rationality. Realistic models of conventional F E D E R A L R E S E R V E B A N K O F S T. L O U I S 42 M AY / J U N E 19 9 9 Rewards may be defined as payments that provide welfare in excess of the normal level, whereas punishments may be defined as payments that bring welfare below the normal level. Punishments have a greater impact than rewards because they provoke a powerful negative emotional reaction while rewards trigger no corresponding positive reaction. Rewards or punishments that are too frequent become normal and so lose their impact, a matter of concern to managers. A pay cut causes anxiety and discontent because the fall in the worker’s welfare below the normal level both triggers bad moods and requires the worker to adopt new habits that are appropriate to the new standard of living. Cuts that are perceived as justified are thought of as inevitable, so that they do not provoke a strongly negative mood. It is easy to incorporate a normal welfare level in an intertemporal version of the formal model. The external conditions of the person’s decision problem at one time, t, are defined by the earnings function, wt(e), and by the set of possible decisions, Dt. Let the function Dt(m, wt ) be the solution to the problem coercion precisely. Presumably, it implies a lack of freedom, but a person who is coerced into doing something, strictly speaking, also chooses to do it, for he or she could refuse to comply and suffer the consequences. Also, everyone works under some degree of compulsion. For instance, stealing from the company or punching the boss will usually lead to automatic dismissal. In a sense, people are forced not to do these things. Similarly, blatant insubordination can cause firing, so that, in this sense, workers are compelled to take orders. When managers spoke of coercion, they did not refer to cases such as these. A rough definition of what they had in mind might be to say that a worker is compelled to do something if not doing it results in punishment and if the worker would do something else if there were no threat of punishment and he or she had good morale. The key aspects of coercion that managers and labor leaders found demotivating were that they frightened people and diminished self-confidence. Though fear is understood by everyone to be a powerful and useful motivator, managers typically use threats only to discourage extreme behavior. They do not want workers to be preoccupied with fear, because it distracts and undermines self-confidence. The latter is important, because it frees the mind and body to act smoothly and efficiently. An apprehensive person consciously thinks through every step of what they do lest they make a mistake. Conscious thought overrides the mental subroutines that guide much of what people do. In relation to the formal model, lack of self-confidence limits the set of moods to a disadvantageous subset. Extreme forms of coercion may lead to the learned helplessness that I mentioned earlier. max [U ( wt ( E( d , m)), m) + V ( E( d , m), m)] . d ∈ Dt The unconscious welfare in period t is [ Wt = max B(wt ( E( Dt ( m, wt ), m ))) m ∈ SM − C( E( Dt ( m, wt ), m )) . ] The expected or normal welfare level may be assumed to be a constant, W , so that the person reacts with anger and discontent when Wt falls below W. Coercion and Freedom EXTENSION TO MORALE Managers and labor leaders stressed that workers are energized by the feeling that they control their lives. They are antagonized and made passive by compulsion and excessive control. These matters are beyond the scope of the model presented here and are not easy to think about carefully. For instance, it is difficult to define Recall that morale has two key aspects, mood and internalization of organizational objectives. Internalization may be expressed by including the firm’s objectives with those of the worker. This procedure is appropriate, since utility functions are inferred from behavior and workers who F E D E R A L R E S E R V E B A N K O F S T. L O U I S 43 M AY / J U N E 19 9 9 internalize their firm’s objectives act as if these were their own. Formally, let R(e) be the revenue the firm earns from a worker’s output, so that the firm’s profit is R(e) – w(e). Internalization may be expressed formally by adding multiples of the firm’s profit to the worker’s conscious and unconscious utility functions, so that these become w, is constant and that improvements in morale enlarge SM in such a way as to make available actions or effort levels, e, that increase R(e) for each level of C(e) and increase R(e) more, the greater is C(e). Imagine a twodimensional diagram, such as Figure 2, with –C(e) on the abscissa and B(w) + µ1[R(e) – w] on the ordinate. Then, improvement in morale causes the northeast frontier of the set of possible points (–C(e), B(w) + µ1[R(e)w]) to rise vertically in such a way that the vertical increase is greater, the larger is C(e) (i.e., the smaller is –C(e)). Because w is constant, it follows that the new optimum yields a higher value of R(e) and a lower value of –C(e). In other words, improved morale affects mood in such a way as to increase profits. Let us assume that the utility functions B and U are differentiable with respect to income, w, so that the unconscious and conscious marginal utilities of income, dB/dw and ∂U/ ∂w, respectively, are well-defined. It must be that B(w(e)) + µ1[R(e) – w(e)] – C(e) and U(w(e), m) + µ2[R(e) – w(e)] + V(e, m), respectively, where µ1 and µ2 are constants that are positive if morale is good. The impact of morale on mood may be expressed by varying the subset SM of possible moods available to the unconscious side of the person. Improvements in morale increase the size of the set SM, thereby giving the unconscious a larger selection of possible states of mind. Improvements in mood, resulting from improved morale, do not decrease and may indeed increase the maximized value of the unconscious objective function B(w(e)) + µ1[R(e) – w(e)] – C(e), because it is maximized over a larger set of moods. That is, if d = D(m, w) solves the problem [ [ (1) ] (2) ] then the value of max {B( w( E( D( m, w ), m))) + µ1[ R( E( D( m, w ), m)) − w( E ( D( m, w ), m )) ] − C( E ( D( m, w ), m))} 6 Akerlof and Kranton (1998) model the moral aspects of identity as internalized rules that restrict the utility function. µ2 < ∂U ( w( e ), m ) , ∂w for levels of e and m that are actually realized. If these inequalities did not hold, the worker would be indifferent to having his or her wage increased, or would prefer to have it reduced, consequences that are contrary to common sense. The inclusion of profit in the worker’s utility function does not portray the sort of good morale that inhibits theft. According to inequalities 1 and 2, workers could improve their welfare by stealing from the employer. In order to give a utilitarian interpretation to moral values, it is necessary either to include punishment, introduce a sense of guilt, or have people take into account the consequences of having other people break the moral codes they break themselves.6 Though the model cannot explain the impact of morale on morality, it does cap- − w( E( d , m )) + V( E( d , m ), m ) m ∈ SM dB(w(e)) dw and max U( w( E( d , m )), m ) + µ2 R( E( d , m )) d ∈D µ1 < increases as the size of the set SM increases. Without more assumptions, it is not possible to say whether the effect of improved morale on mood increases profits. A plausible set of assumptions is that the wage function, F E D E R A L R E S E R V E B A N K O F S T. L O U I S 44 M AY / J U N E 19 9 9 Figure 2 ture important consequences of good morale. Using the argument made in the previous section, it is easy to show that the inclusion of the terms µ1[R(e) – w(e)] and µ2[R(e) – w(e)] does not decrease profits. More precisely, profits do not decrease, provided µ1 is positive and provided the inclusion of these terms does not change the set of actions, e, which is achievable by varying mood, m. In order to see why, let e be the choice of e that maximizes B(w(e)) – C(e) and let e′ be the choice of e that maximizes B(w(e)) + µ1[R(e) – w(e)] – C(e). Then, B+µ1(R–w) Optimum with good moods from good morale Optimum with poorer moods from bad morale B(w(e′)) + µ1[R(e′) – w(e′)] – C(e′) ≥ B(w(e)) + µ1[R(e) – w(e)] – C(e) –C ≥ B(w(e′)) + µ1[R(e) – w(e)] – C(e′), incentives by assuming that w(e) = w0 + w1R(e), where w1 > 0. Let us assume that the firm varies w0 and w1 so that w0 + w1R(e) remains constant, where e is the worker’s choice of action. Now, hold w0, w0′, and w1′ fixed, where w1′ > w1. Let us also assume that B is linear, that is, B(w) = bw, where b is a positive number. By inequality 1, we must assume that b > µ1. I show that increasing w1 to w1′ increases profits. Let e and e′ be the worker’s optimal choices of e when the wage is w0+ w1R(e) and w0′+ w1′ R(e), respectively. By the optimality of e and e′ it follows that which implies that R(e′) – w(e′) ≥ R(e) – w(e), as is to be shown. Financial Incentives and Morale Next, I show that financial incentives and morale complement each other. An argument similar to the one that was just made shows that increasing µ1 does not decrease (and may increase) profits, for any wage function w, including ones offering financial incentives. In order to see how to make the argument, assume that µ1′ > µ1, notice that (3) b[w0 + w1R(e)]+ µ1[R(e) B(w(e)) + µ1′[R(e) – w(e)] – C(e) –w0 – w1R(e)] – C(e) = B(w(e)) + µ1[R(e) – w(e)] – C(e) ≥ b[w0 + w1R(e′)]+ µ1[R(e′) + (µ1′ – µ1)[R(e) – w(e)], –w0 –w1R(e′)] –C(e′) and assume that increasing µ1 to µ1′ does not change the set of actions achievable by varying mood. I show that if µ1 is positive, then increasing financial incentives increases profits, provided the function B is not too concave and provided the change in µ1 does not change the set of actions attainable by choice of mood. I introduce explicit and (4) b[w0′ +w1′R(e′)]+µ1[R(e′) –w0′ – w1′R(e′)]– C(e′) ≥ b[w0′ +w1′R(e)]+µ1[R(e) –w0′ – w1′R(e)]– C(e). F E D E R A L R E S E R V E B A N K O F S T. L O U I S 45 M AY / J U N E 19 9 9 Bn(wn(yn(en ))) +µ n1 [R(y(e1, ..... , eN )) These inequalities imply that – w1(y1(e1 )) – ..... – wN (yN (eN ))] – C(en ), (b – µ1)w1[R(e) – R(e′)] ≥ C(e) – C(e′) + µ1[R(e′) – R(e)] ≥ (b – µ1)w1′[R(e) – R(e′)]. and his or her conscious utility is Un(wn(yn(En(dn, mn))), mn) Because b – µ1 > 0 and w1′ > w1, the last inequalities imply that + µ n2[R(y(E1(d1, m1 ), ..... , EN (dN , mN ))) – w1(y1(E1(d1, m1 ))) – ..... R(e′) ≥ R(e). – wN(yN(EN(dN , mN )))]+V(En(dn, mn), mn). Since by assumption w0 + w1R(e) = w0′ + w1′R(e′), it follows that profits are not decreased by increasing incentives. Profits would be increased strictly if there were strict inequality in either of inequalities 3 and 4. In this case, profits still would increase if the function B(w) were a slightly concave approximation to the linear function bw. It is easy to make an example in which B is very concave and increased incentives decrease profits. completes the argument that increased incentives may increase profits, even when morale is good, just as improved morale increases profits even when workers receive financial incentives. In this sense, incentives and morale are complements. Interaction among the N workers suggests a coordination game, for they all derive utility from profits. In order to see the connection more clearly, let us assume that workers’ moods adjust so that the utility of labor, V(En(dn, mn), mn), is the same for all decisions dn actually adopted by the workers, so that this term may be ignored. In addition, suppress mood and the distinction between the conscious choice, dn, and the realized action, en, and focus on conscious utility, since cooperation is arranged deliberately. Suppose that the choice of action has two components, selection of a method of production and the selection of effort, which is thought of as the pace of work. Since effort is influenced by mood, which is governed unconsciously and almost automatically, it makes sense to ignore the effort part of actions and to think of the actions solely as production methods. Under these assumptions, the relevant utility functions are Cooperation The model can be used to demonstrate one reason good morale fosters cooperation among workers; it gives them a common objective. Let there be N workers and let the subscript n indicate variables and functions applying to the nth worker. The employer observes worker n’s output to be (5) Un(wn(yn(en ))) + µ n2[R(y(e1, ..... , eN)) – w1(y1(e1)) – ..... – wN(yN(eN ))], yn(en) = yn(En(dn, mn)) for n = 1, ..... , N. If the parameters µ n2 are positive and the functions wn are constant, as would be the case for truly fixed wages, then, in effect, the workers play a coordination game with payoff R(y(e1, ..... , eN)) for all players, and the obvious solution is to maximize this payoff jointly. Management normally gives workers at least some financial incentives linked to individual and pays him or her wn (yn(en)) = wn(yn(En(dn, mn))). The actual output of all N workers is y(e1, ..... , eN ) = y(E1(d1, m1 ), ..... , EN(dN, m N )). Worker n’s unconscious utility is F E D E R A L R E S E R V E B A N K O F S T. L O U I S 46 M AY / J U N E 19 9 9 performance, however, such as production targets, performance evaluations, promotion criteria, and piece rates. I was told that it is difficult to design incentives so that the workers’ financial interests are entirely consistent with those of the firm. An important function of good morale is to motivate workers to act in the firm’s interest, even when it conflicts with their own financial advantage. I show that the above model includes this function. More precisely, I argue that cooperation induced by internalization of the firm’s goals increases profits. Suppose that morale is neutral. That is, suppose that µ n 2 = 0, for all n. In addition, suppose that the wage functions, wn, include financial incentives. For each n, let en be that value of en that maximizes Un (wn (yn (en ))). With these choices of effort, the firm’s profit is > R( y(e 1 , ....., e N )) − w1( y1( e )) − ..... − wN ( yN ( e N )). 1 That is, internalization of the firm’s objectives increases profits. Information Sharing One of the reasons it is difficult to give workers incentives consistent with the firm’s objectives is that the conditions workers face change frequently, so that the actions that are correct, from the employer’s point of view, also change. If management knew conditions precisely, it could order workers to do exactly what was needed or it could include the conditions in the specification of incentives. Often, however, only the workers observe the relevant changes in circumstances. Managers said that one of the benefits of good morale is that it induces workers to share information with each other and with their superiors. This advantage can be introduced into the above model by having company revenues depend on random variables observed by the workers alone. For instance, assume that worker n observes the random variable θn and that company revenue depends on all the θn, so that utility function 5 becomes R(y(e1, ..... , eN)) – w1(y1(e1)) – ..... – wN(yN(eN)). In contrast, suppose now that morale is good, so that the µ n 2 are all positive. Though it is hard to say how the workers would behave, it would be to_their mutual _ advantage to choose actions (e 1, ..... , e N) that: • Were a Nash equilibrium for the game with payoffs as in function 5, Un(wn(yn(en))) + µ n2[R(y(e1, ..... , eN, θ1, ....., θN)) – w1(y1(e1)) – ..... – wN(yN(eN))]. • And gave each worker, n, a payoff exceeding If all workers reveal the values they observe of the θn , then workers can cooperate more effectively. Expected profits, and hence, the expected individual utilities that are earned from cooperation will not decrease (and may increase), provided the parameters µ n2 are all positive. Hence, workers have a positive incentive to share their observations with each other and with management. Un(wn(yn(en))) + µ 2[R(y(e1, ..... , eN)) – w1(y1(e1)) – ..... – wN(yN(eN))]. Suppose that such an equilibrium exists. Because of the form of utility function 5 and because µ n 2 is positive and e n maximizes Un(wn(yn(en))), for all n, it follows that Morale vs. Coercion R( e 1 ,....., e N ) − w1( e1 ) − ..... − wN ( e N ) Managers explained that the chief disadvantage of using threats to obtain cooperation is the loss of worker initiative. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 47 M AY / J U N E 19 9 9 max[pÎãI – w] Though force may succeed in making people work with great intensity, people working under such pressure may only make a show of cooperation and may not use their heads to help the firm. I was told that coercion works well for tasks that are easily monitored and when management knows what employees should do. Managers said that compulsion is inefficient when workers know best what they ought to do because of information they alone receive. I express these ideas formally using an example in which I suppress mood, the unconscious, and the distinction between decisions and realized effort, since these variables are irrelevant here. Suppose a worker may do one of two types of tasks, A and B, and that these are performed with intensities, IA and IB, respectively, where IA and IB are non-negative numbers. The action e = (i, I) is task i done with intensity I, where i = A or B. Let the disutility of doing either task, i, with intensity I be –V(i, I) = I2 and let the utility of wage w be simply U(w) = w. Suppose that one, and only one, of the tasks is profitable; management does not know which task is profitable, and the worker can learn which is profitable at a small cost in utility. Suppose further that management observes the intensity level; task A is profitable with probability p, where 1/2 < p < 1; and a task done with intensity I earns revenues R(i, I) = ÎãI, when it is profitable, and earns no revenue otherwise. Finally, suppose that to retain the worker, management has to offer a reservation utility level of at least 1/16. If the firm obtains cooperation through threats, morale is zero and the worker cannot be counted on to do the task that is profitable. In this case, optimal management strategy is to set the wage, w, equal to [1 + (2p)4/3]/16, to fix the task to be A, and to require work intensity, I, to be (2p) 2/3/4, for these values solve the profit maximization problem w,I s.t. w – I2 ≥ 1/16. The firm fires the worker if work intensity is less than (2p)2/3/4, in which case the worker earns his or her reservation utility level of 1/16. The firm’s expected profits are the positive number [3p(2p)2/3/8] – 1/16, and expected revenues are [p(2p)1/3]/2. Suppose management does not threaten, but depends on positive morale. Assume that in this case the morale parameter, µ2, equals 1/2. Then the worker’s total utility function is U(w) + V(I) + µ2[R(i, I) – w] = w – I2 + 0.5[R(i, I) – w], minus a small quantity if the worker verifies which task is profitable. Assuming the worker knows which task is profitable, he or she solves the problem max [– I2 + 0.5 ÎãI ], I so that work intensity is I = 1/4, which is less than the intensity in the previous case with compulsion and no morale. Because the worker chooses the profitable task, however, the firm’s expected revenues are ÎãI = 1/2, which exceeds the level with no morale. If the firm continues to pay wage w = [1 + (2p)4/3]/16, then total expected worker utility, U(w) + V(I) + µ2 [R(i, I) – w], is at least 1/16, and expected profits are higher with positive morale than with no morale, unless [p(2p)1/3]/2 > 1/2, that is, unless p > 0.51/4 ≅ 0.84. That is, coercion is more profitable than dependence on morale alone only if management knows with high probability which task is profitable. This result corresponds to the intuition I wish to express. max[pR(A, I) – w] Testing the Model w,I s.t. U(w) + V(I) ≥ 1/16. The proposed model of work motivation might be tested by psychological experiments. One implication of the model is that the utility or disutility of or F E D E R A L R E S E R V E B A N K O F S T. L O U I S 48 M AY / J U N E 19 9 9 work effort depends on expected reward. In testing this implication, it would not be correct to measure the disutility of effort by offering people a choice between effort and something else, such as having to pay a certain amount of money, for that choice would affect the context of the work, and hence, might affect mood. Something might be learned, however, by asking people how they feel about their efforts. By looking for consistency in people’s reactions to various work and reward situations, it might be possible to test for the existence of an unconscious utility function and even to estimate it. CONCLUSION The usual model describes a worker’s trade-off between financial reward and the disutility of labor and has no place for morale. Neither managers nor labor leaders, however, dwelled on the unpleasantness of work, but rather stressed its benefits. Managers spoke as if one of their primary tasks was to maintain good morale, and labor leaders also emphasized its importance. In view of these observations, it seems appropriate to replace the usual model with one more consistent with the observations of people running workplaces. I do not know whether my own suggestions are correct. Perhaps further empirical inquiry will give a firmer basis for theory. REFERENCES Akerlof, George A., and Rachel E. Kranton. “Economics and Identity,” Russell Sage Foundation Working Paper No.136, August 1998. Blum, Milton L. Industrial Psychology and Its Social Foundations, New York: Harper and Row, 1956. Gleitman, Henry. Psychology, New York: W. W. Norton, 1995. Juster, F. Thomas. “Preferences for Work and Leisure,” in Time, Goods, and Well-Being, F. Thomas Juster, and Frank P. Stafford, eds., Institute for Social Research, University of Michigan, 1985, pp. 333-51. Tajfel, Henri. “Experiments in Intergroup Discrimination,” Scientific American (November 1970), pp. 96-102. Turner, John C. Rediscovering the Social Group: Self-Categorization Theory, Oxford, UK: Basil Blackwell, 1987. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 49 M AY / J U N E 19 9 9 F E D E R A L R E S E R V E B A N K O F S T. L O U I S 50 M AY / J U N E 1 9 9 9 Lowell J. Taylor is a professor of economics at the H. John Heinz III School of Public Policy and Management at Carnegie Mellon University. Commentary Professor Bewley describes the central conclusion of his work as follows: From the interviews, I conclude that wage rigidity stems from a desire to encourage loyalty, a motive that superficially seems incompatible with layoffs. My findings support none of the existing economic theories of wage rigidity, except those emphasizing the impact of pay cuts on morale. Other theories fail in part because they are based on the unrealistic psychological assumptions that people’s abilities do not depend on their state of mind and that they are rational in the simplistic sense that they maximize a utility that depends only on their own consumption and working conditions, not on the welfare of others. Lowell J. Taylor I n preparing my comment on Truman Bewley’s work, I had the privilege of reading a draft of Professor Bewley’s forthcoming book, Why Wages Don’t Fall During a Recession. This book describes an extensive field study, in which Professor Bewley interviewed hundreds of business managers and labor leaders, seeking to understand the puzzling phenomenon of wage rigidity. The paper presented here is an outgrowth of this project. Before turning to the paper itself, I would like to make a few comments about the research enterprise more generally. I am quite enthusiastic about the manuscript I read. There are three reasons why you should read this book. First, you will learn a lot of economics. In the course of discussing ideas raised in his interviews, Professor Bewley cites and discusses a vast literature (his book cites more than 1,000 papers!). The book provides an engaging way of becoming acquainted with the issues currently under debate in the study of motivation, compensation, labor markets, and macroeconomics. Second, the research makes important headway in understanding the nature of rigidities in labor markets. Bewley makes a number of striking and controversial claims about labor markets, but always in a clear, well-reasoned, and convincing manner. Finally, the book will challenge you to think hard about how to “do economics.” Economic theorists generally implicitly follow the path laid out by Milton Friedman (1953): We hope that even though our assumptions are often self-consciously unrealistic, useful predictions about behavior can nonetheless be derived. Bewley’s book reminds us that when assumptions drift too far from reality, theory is a futile exercise with no hope of shedding light on behavior. The paper presented at this conference provides theoretical musings designed to encourage readers to think about a world in which morale and mood are part of human behavior. The idea here is to lay out a formal representation in which workers’ decisions depend not merely on balancing utility of earnings against disutility of effort, but depend as well on morale—an internalization of organizational objectives—and mood—the worker’s psychological state of mind. I do not know if the formalizations provided here or in the books are the most fruitful ways of incorporating morale and mood into models of worker-firm interactions.1 Nonetheless, the exercise is useful as a way of clarifying how explicit consideration of mood and morale can change the usual predictions of agency theory. Indeed, it is clear that people do pay attention to the well-being of others, and that the way in which they do this depends critically on their own “mood.” Bewley’s work persuades me that this general idea can be important in understanding many features of workplace behavior. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 51 1 The model and discussion presented here are somewhat different than those in the book (and those presented at the conference itself). Readers are again encouraged to look at the book as well. M AY / J U N E 1 9 9 9 A simple story of a “rude driver” might provide a useful way of thinking about these kinds of motivation. Suppose you are traveling by automobile to an important engagement for which you definitely do not want to be late. Construction has closed the left lane of the highway, and you are stuck waiting, not too patiently, in the right lane as traffic inches its way toward the construction site. You notice in your rearview mirror a lone car whizzing past the “merge right” signs, bypassing drivers—yourself included—who have dutifully pulled into the right lane. The driver proceeds to the very last spot at which the car can physically remain in the left lane and signals right, hoping to pull in front of you! What do you do? Do you let him in or not? The self-interest model normally employed by economists provides an easy answer. Letting the driver pull into your lane imposes a cost—it increases the probability that you will be late for your engagement—and gives you no benefit, so you do not let him in. Economists understand, though, that this simple way of viewing the problem is not exactly right. People are routinely altruistic; they pay attention to others’ well-being. There are many examples in which ideas are incorporated into formal economic models, including, for instance, Gary Becker’s (1991) work on the family and Matthew Rabin’s (1993) theory of fairness. In the current context, you would surely be willing to let the driver pull in front of you if you noticed a pregnant woman in the car who appeared to be in labor. The idea here is simple: People care about others and are often willing to sacrifice their own material well-being to benefit others. Moreover, in models like Rabin’s and in empirical work like Andreoni and Miller (1996), people incorporate others’ welfare into their own decision-making in generally predictable ways. For example, Andreoni and Miller conduct experiments which demonstrate that individuals’ willingness to sacrifice money varies with the extent to which other individuals benefit from the sacrifice. Returning to the scenario of the rude driver, though, it is clear that something besides even the altruism of Rabin is at work in most people’s decision-making. Your decision of whether to let the driver pull in front may well depend not only on your assessment of the potential benefits to the driver (e.g., does he have a pregnant passenger?), but also on your gut-level belief about the driver’s motives. If you believe the driver is indeed being rude or obnoxious, you may be anxious not to let him; you might even be willing to suffer some cost to avoid letting him in. On the other hand, if there is some indication that the driver was merely confused (perhaps he has an out-of-state license plate and a befuddled, contrite look on his face), you will be more inclined to let him in. The point is that your utility function evidently depends, at a minimum, on three factors: first, your own narrow, self-interested goal— to move through traffic as quickly as possible; second, your internalization of the other driver’s goals; and, third, your own mental state, which in this case follows from your assessment of the other driver’s motives. This last part of the decisionmaking process may be transparent at the time or may work at a largely subconscious level. It is self-evidently important in the rude driver scenario, and it may be important in other contexts as well. By section 4 of his paper, Bewley works up to a representation of worker preferences that perhaps parallels the observations we glean from the rude driver scenario. Bewley’s characterization of utility, both “conscious” and “unconscious,” depends, as usual, on compensation and effort. It incorporates also an internalization of the goal of the firm’s owners. This altruistic component is not generally included in theorists’ representations of preferences, but as just noted, it is sometimes incorporated; it appears, for instance, as a central component in George Akerlof’s (1982) gift exchange model of worker-firm interaction. Bewley pushes further, though, by positing a model that includes “morale,” which he suggests has two key aspects: an internalization of the firm’s goal (firm profit) and his mental state F E D E R A L R E S E R V E B A N K O F S T. L O U I S 52 M AY / J U N E 1 9 9 9 or mood. As in the rude driver story, an individual’s behavior is affected by his own personal objectives and by an inclination to at least consider the well-being of others (in this case the employer), but all of this is tempered by a more elusive construct— mood or mental state. Building from this basic representation of worker preferences, Bewley derives a number of implications that seem in accord with the observations he gathered in his field research. In so doing, he clearly demonstrates the viability and value of constructing a model that pays serious attention to motivations that (realistically) move beyond the usual assumption in principal– agent models. Bewley’s work covers lots of ground, making suggestions about how to incorporate mood and morale, and also exploring cooperation, information sharing, and the relative merits of managers’ use of morale and coercion. This is a valuable contribution to behavioral economics as applied to the workplace; it provides seed ideas that other economists will find useful in doing their own theoretical work. Economists may find alternative ways of constructing realistic models of work motivation, models that include a serious effort to incorporate features like morale and mood. In this enterprise, they will do well to follow Professor Bewley’s exemplary effort to see that their formal representations are firmly rooted in careful systematic observation of the real world. REFERENCES Akerlof, George A. “Labor Contracts as Partial Gift Exchange,” Quarterly Journal of Economics 96 (November 1982), pp. 543-69. Andreoni, James, and John H. Miller. “Giving According to GARP,” Working Paper, Carnegie Mellon University, (1996). Becker, Gary. A Treatise on the Family, Harvard University Press, (1991). Friedman, Milton. “The Methodology of Positive Economics,” Part 1 in Essays in Positive Economics, University of Chicago Press, (1953). Rabin , Matthew. “Incorporating Fairness into Game Theory and Economics,” American Economic Review 83 (December 1993), pp. 1281-302. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 53 M AY / J U N E 1 9 9 9 F E D E R A L R E S E R V E B A N K O F S T. L O U I S 54 M AY / J U N E 1 9 9 9 Wouter J. den Haan, Garey Ramey, and Joel Watson are associate professors of economics at the University of California, San Diego. Den Haan is also a member of the NBER. The authors thank Marty Eichenbaum and Bob Hall for helpful conversations. Ramey and Watson thank the NSF for financial support under grant SBR-965868. ContractTheoretic Approaches to Wages and Displacement influencing the form of worker compensation. Moreover, the responses of aggregate wages and employment to business-cycle shocks are sensitive to the structure of worker/firm contracting. Overall, our study establishes that the particular form of contracting imperfections can have major implications for economic outcomes. This highlights the importance of going beyond the reduced-form analysis of contracting that typifies much of the macroeconomics literature. Our key assumption throughout is that firms and workers maintain long-term contractual relationships, whereby a particular firm and worker transact repeatedly until their relationship is severed. Within a unified theoretical framework, we consider two types of contracting imperfections in labor relationships. First, relationships may be subject to limited verifiability, whereby external enforcement authorities are unable to compel payments conditioned on the full set of actions chosen by the contracting partners. For example, the authorities may be unable to ascertain whether severance of the relationship was due to the worker’s action or the firm’s action. Second, desirable contracts may be infeasible due to the worker’s limited liquidity, which prevents the worker from making payments to the firm.1 We demonstrate that privately inefficient breakup of relationships may occur in the presence of limited verifiability; that is, limited verifiability leads contracts to be fragile. The fundamental idea is that when a negative shock hits, the joint returns to cooperation between the firm and worker may be insufficient to offset the collective inducements to behave dishonestly, so there is no way to specify transfers between the firm and worker that can preserve the relationship. Increased verifiability leads to more robust relationships, more direct punishments for misbehavior, and a wider set of optimal compensation schemes. The worker’s relative bargaining position always Wouter J. den Haan, Garey Ramey, and Joel Watson M odels of moral hazard in labor relationships have proven to be useful in explaining a variety of important macroeconomic phenomena. The existence of involuntary unemployment has been linked to the need to provide incentives for workers to choose high effort (Shapiro and Stiglitz, 1984). Further, since wage levels are important for workers’ incentives, adjustment of wages in response to cyclical shocks may be subject to contractual constraints. This may help to explain the low observed variability of average wages relative to employment levels (Danthine and Donaldson, 1990, 1995; Strand, 1992; MacLeod, Malcomson and Gomme, 1994). More recently, contracting problems have been tied to inefficient severance of employment relationships, giving a mechanism whereby business cycle shocks may be magnified and made more persistent (Ramey and Watson, 1997a). This paper focuses on the contracttheoretic underpinnings of wage adjustment and worker displacement in moral-hazard models of the labor market. We show that contracting imperfections play a key role in determining the fragility of employment relationships in the face of shocks to productivity, as well as F E D E R A L R E S E R V E B A N K O F S T. L O U I S 55 1 Our analysis omits some other aspects of labor contracts that have been considered in the literature. First, we assume risk-neutral workers, so wage payments do not play any insurance role, in contrast to implicit contract models. Second, renegotiation of wage contracts is allowed, meaning that inefficient severance cannot occur as a consequence of failure to renegotiate. Implicit contract models are surveyed in Romer (1996, ch.10); see Boldrin and Horvath (1995) for a recent empirical implementation. Suppression of renegotation as a cause of displacement is considered in Hashimoto and Yu (1980) and Hall and Lazear (1984). M AY / J U N E 1 9 9 9 2 Our recent analysis of lending relationships and liquidity flows (den Haan, Ramey, and Watson, 1999a) incorporates a contracting framework that is a special case of the one considered here. More tangentially related is den Haan, Ramey, and Watson (1999b), which explores the theoretical foundation of stylized facts about compensation over a worker’s career and the experiences of displaced workers. 3 For example, one component of our theory is the view that discretionary transfers are subject to renegotiation. determines his total compensation, while the particular forms of payment may be influenced by the requirements of contracting. Performance bonding, for example, arises naturally in settings of unrestricted liquidity, but may be circumscribed when the worker is liquidity constrained. Moreover, since the form of compensation is identified by the timing and conditioning of payments, informal notions such as salary may be ambiguous. Our framework also allows for a more precise analysis of the role of wage premia in solving labor contracting problems. We say that a worker obtains an efficiency wage when, in contract negotiation, the firm and worker must directly weigh reducing the worker’s compensation against motivating effort. We demonstrate that, in the absence of liquidity constraints, effort incentives are driven by verifiability, bargaining power, and the state of the matching market, but not by the worker’s current period compensation. Thus, in a precise sense, efficiency wages play no role in helping to preserve relationships. When the worker is liquidity constrained, however, the incentive constraint may bind at the time of contract negotiation, as a consequence of the worker’s inability to make payments to the firm that would implement the unconstrained bargaining outcome. Thus, efficiency wages may emerge as added restrictions on wage determination when workers are liquidity constrained. To analyze how contracting imperfections affect market outcomes, we posit that relationships are formed on a matching market, as in Pissarides (1985) and Mortensen and Pissarides (1994). We consider an example in which a limited liquidity specification with efficiency wages but no fragility is contrasted with a limited verifiability specification that is subject to fragility. In response to a permanent shock to the distribution of productivity, the presence of limited liquidity does serve to dampen wage adjustment, relative to a completecontracting benchmark. However, the dampening is much more pronounced in the limited verifiability case, as the severance of low-productivity relationships greatly reduces the sensitivity of average wages to the shock. Moreover, the effect of the shock on total employment is greatly magnified as a consequence of fragility. Our example illustrates how models that emphasize displacement may be much more potent for explaining wage adjustment and propagation of shocks than models stressing wage effects within a given employment contract. The framework presented in this paper builds on the contracting model of Ramey and Watson (1997a), who first developed the theoretical foundation of fragility and the related magnification of shocks. The present paper is also related to our earlier work on endogenous destruction margins and propagation of shocks, as reported in den Haan, Ramey, and Watson (1997). Here we address a wider range of contractual imperfections (including liquidity constraints and a range of verifiability conditions), we incorporate wage determination, and we consider adjustment of average wages in market equilibrium.2 Our framework is also related to the work of MacLeod and Malcomson (1989, 1993, 1998). MacLeod and Malcomson focus on the timing and enforcement of compensation, using a model in which parties can make both externally enforced and discretionary transfers. Two contractual forms are emphasized: efficiency wages (which they define as the use of high wages with the threat of severance) and performance pay (defined by the use of discretionary bonuses). They demonstrate how the form of compensation depends on labor market conditions and equilibrium beliefs, interpreting the latter as a “social norm for a fair wage.” Our work, in contrast, addresses (a) a broader range of contracting settings, including different restrictions pertaining to verifiability and liquidity; (b) inefficient severance of relationships following shocks; and (c) issues of propagation in the macroeconomy. We also take a different approach to modeling contract determination, whereby negotiation (and renegotiation) between workers and firms is directly incorporated.3 F E D E R A L R E S E R V E B A N K O F S T. L O U I S 56 M AY / J U N E 1 9 9 9 Figure 1 On the issue of contractual form, we obtain results different from those of MacLeod and Malcomson. The main body of this paper is divided into four sections. The first introduces the basic model of an employment relationship. The second considers enforcement under various contracting environments, which differ in terms of what can be verified to a third party. The third section discusses efficiency wages and limited liquidity, while the fourth derives market outcomes from a matching setup. Timing of Actions in Employment Relationships In each period: THE MODEL Employment Relationships Employment relationships consist of one worker and one firm who interact in periods t = 1, 2, ... until their relationship is severed. The firm is represented by a manager. Both the worker and manager make a private effort choice (high or low) that contributes to production. In addition, the parties negotiate a contract specifying transfers as a function of verifiable information. If both agents exert high effort during production, then the cooperative output level z t is realized. We assume that zt varies randomly across periods, taking the value z G in the good production state and z B in the bad state, with z G > z B > 0. For simplicity, z t is assumed to be realized independently in each period, with r denoting the probability that z t = z B. The realization of z t , contracting, and effort choices within a period occur in three phases, as illustrated in Figure 1. In phase 1, the worker and manager observe the realized value of z t for that period, then negotiate a contract that governs current-period interaction. If they reach an agreement, the contract specifies which decisions the agents will make in subsequent phases, as well as a profile of contingent payments. Disagreement leads to severance of the relationship, with the worker and manager obtaining outside option values of bw + ww and b m + w m, respectively. The parameters b j reflect current-period benefits received when Phase 1: Contract Negotiation D Phase 2: Manager Effort Choice A Phase 3: Worker Effort Choice B C the relationship is severed in phase 1 (e.g., the worker may obtain unemployment benefits), while the parameters w j indicate the discounted values of future benefits, which may include returns from new relationships formed in the future. Severance as a result of disagreement will be referred to as outcome D. Further details of the contract negotiation are discussed below. The manager makes his effort choice in phase 2. Low effort leads to outcome A, where the manager obtains a current-period private benefit of x m, while the worker receives no benefit. Worker effort is selected in phase 3; there, low effort leads to outcome B, in which the worker receives a current-period private benefit of x w and the manager obtains zero. Under either low-effort outcome, output is zero and the relationship is severed at the end of the period. On the other hand, high effort by both agents induces the cooperative outcome C, in which case output is z t and the relationship continues into the next period. The manager is assumed to appropriate the output.4 We assume x j > b j for both j, meaning that agents gain more in the current period from staying in the relationship and putting out low effor t, than from leaving the F E D E R A L R E S E R V E B A N K O F S T. L O U I S 57 4 The model is easily extended to introduce imperfect monitoring of effort choices, as follows. In phase 2, outcome A is reached with a positive probability under either effort choice of the manager, with low effort implying a greater probability of reaching A. Similarly, in phase 3 the probability of outcome B is higher when the worker chooses low effort. Thus, the contract cannot be conditioned directly on the effort choices, but only on the observable outcomes. M AY / J U N E 1 9 9 9 where x = x w+x m. The first inequality in assumption 1 implies that the agents prefer the robust outcome under either production state, so that the robust outcome is efficient. The remaining two inequalities will determine the conditions under which the agents can find a contract that supports the robust and fragile solutions, as discussed below. relationship in phase 1. Observe, however, that when an agent chooses low effort, his partner forgoes the opportunity to obtain b j. We also assume that x j < b w + bm, meaning the agents will never agree in phase 1 to induce outcomes A or B. Interpretations for our assumptions are discussed below. We now compute the value of the relationship under various outcomes. First, the agents may choose high effort under both z G and z B in every period, in which case the relationship never breaks up. In this robust solution, the value of the relationship is given by Contracting At the start of each period, the worker and manager negotiate a short-term contract that specifies payments from the manager to the worker conditional on the productivity level z k, k = G, B, and on the outcome of productive interaction (A, B, or C). The set of feasible contracts is generally constrained by the limits of the external enforcement institution. Payments might also be subject to liquidity constraints. Let sCk denote the payment made to the worker in the event that outcome C is realized, under productivity level zk, k = G, B. Since the manager directly appropriates zk when outcome C is reached, his currentperiod payoff in this case is zk – sCk , while the worker obtains sCk . Transfers conditioned on outcomes A and B will be written sA and sB, respectively; these transfers will not need to depend on k. In addition, the agents may agree on up-front transfers s0k, made at the time of contracting in phase 1. We adopt the convention that negative values of sCk , sA, sB and s0k denote transfers from the worker to the manager. The worker and manager also formulate a joint plan for how they will behave in the future, which yields an expected continuation value g. For example, if the agents intend to implement a robust solution, then g=g R. We look for a specification of behavior, consisting of explicit contracts and individual actions over time, that satisfies four conditions. First, agents’ expectations about g accurately reflect the value of continuing the relationship. Second, in each period, agents make their effort choices in a utility-maximizing manner, given g and the values of transfers agreed to under the contract. Third, short-term (1 − ρ ) zG + ρ z B , 1− β where b is the agents’ common discount factor. We let g R denote the discounted continuation value of the relationship in this case: g = R [ β (1 − ρ )z G + ρ z B 1− β ]. Second, the agents may agree to select high effort when z G is realized, but to sever the relationship under z B. In this fragile solution, in each period the relationship breaks up with probability r. The discounted future value of the relationship in this case satisfies [ ( ] ) g F = β (1 − ρ ) z G + g F + ρ ( b + w) , where b = bw + bm and w= ww + wm. Solving for gF yields gF = β (1 − ρ )z G + ρβ (b + w) 1 − (1 − ρ )β . Finally, the agents may agree to sever the relationship under both z G and z B and the relationship breaks up in period 1, having value b + w. We impose a final assumption: (1) b + w < z B + g R < x + w < z G + g F , F E D E R A L R E S E R V E B A N K O F S T. L O U I S 58 M AY / J U N E 1 9 9 9 contract negotiation is resolved according to the Nash bargaining solution. Here, the agents recognize that they are implicitly bargaining over the total value of the relationship, which is the sum of currentperiod returns and the continuation value g (assuming the agents are able to maintain the relationship into the next period). The worker’s and manager’s bargaining weights are p w and p m, respectively, and the disagreement point is outcome D. The parameters p w and p m are nonnegative and satisfy p w + p m = 1. Fourth, the best equilibrium satisfying the first three conditions is selected by the firm and worker.5 In light of assumption 1, the firm and worker will chose the robust solution if it can be supported. Next in line is the fragile solution, followed by immediate severance. a larger private benefit than does unemployment (this is the assumption x j > b j, j = w, m). The most direct way to interpret this assumption is that employment relationships convey perks that are themselves attractive, apart from personal costs associated with high effort. Further, unemployment may involve private costs, such as psychic harm and search costs, that are not incurred within active relationships. Note that private benefits are zero for agents who exert high effort in a relationship, which serves to normalize utility.6 Low effort and severance. We have assumed that low effort by either the worker or manager leads the employment relationship to be severed. There are two basic motivations for this assumption. First, low effort may induce rapid decay in the productivity of the relationship, to the point where returns to continuation fall short of operating costs. For the manager, low effort might also be directly tied to liquidation; for example, the manager may sell off essential assets. Second, contractual enforcement mechanisms used by the partners to sustain cooperation may entail a costly and time-consuming dispute resolution process in the event that either agent chooses low effort; see Ramey and Watson (1997b) for a detailed discussion of such mechanisms. When dispute resolution costs are sufficiently high, the worker and manager will opt to sever their relationship. As another possibility for contractual agreement, the agents might seek to temporarily suspend their relationship when high effort is unsustainable, for example, through a layoff, in order to preserve match capital. Such suspensions will be infeasible, however, if returns from the relationship would experience rapid deterioration without active inputs of effort. For example, production equipment or organization may depreciate during the suspension, or market dominance may be permanently lost. Further, as will become clearer below, contracts that support temporary suspension will be infeasible if a third-party enforcement authority Interpretation Effort choices and unemployment benefits. Our model of employment relationships allows for effort choices by both workers and managerial personnel. These choices can be interpreted in a number of ways. The most familiar interpretation involves personal exertion, and here we augment the usual shirking model by specifying that, in addition to worker effort, managerial effort is also important for production. Further, low effort may entail activities that are directly harmful to production, such as theft. Managers may also abuse their power to direct workers’ activities, by unexpectedly assigning them to undesirable tasks. An agent obtains a current-period private benefit when he chooses low effort. Alternatively, the agents can agree to dissolve their relationship at the start of the period and obtain current-period benefits outside the relationship. A key assumption of our model is that these unemployment benefits become unavailable once agents have agreed to a contract and have proceeded to phase 2; that is, the agents must make a commitment to production activities that rule out outside benefits in the current period. We have also assumed that low effort conveys F E D E R A L R E S E R V E B A N K O F S T. L O U I S 59 5 The first three properties define a negotiation equilibrium, which is simply a specification of behavior consistent with private incentives and the Nash bargaining solution. Specifically, the Nash solution implies that, given g, the surplus of the relationship at the time of negotiation (which is z k + g – b – w ) is shared in proportions defined by the bargaining weights. The fourth property implies that, at the meta-level of negotiating over negotiation equilibria, the firm and worker select the equilibrium that maximizes their joint returns. For example, if they could sustain both the values g R and g F, then they are assumed to select the preferred plan yielding g R. In our framework, there will always be an equilibrium that is maximal in every period. 6 Our setup admits the standard shirking model, in which firms behave more passively. The standard model is obtained by m m setting b = x = 0, so that the manager obtains neither unemployment benefits nor benefits from low effort. In this case, the manager’s incentive to agree to the contract at phase 1 are identical to his incentive to choose high effort at phase 2, so in effect the manager does not make an effort choice. M AY / J U N E 1 9 9 9 7 MacLeod and Malcomson’s (1989, 1993, 1998) bonus payment is like sCk , although they assume it is discretionary; for example, the firm is not contractually committed to make the payment. In our framework, firms would never make discretionary transfers following production, and only what is enforceable matters. 8 In our setting, short- and long- term contracts differ only to the extent that agents can enforce a transfer conditional on outcome D occurring in the next period. To see this, consider two contracting environments: (a) short-term, as described in the text; and (b) long-term, with a recontracting option in each period. Fix the scope of what can be verified and enforced within a given period, and assume that the agents have symmetric information whenever they negotiate. Then (a) and (b) support exactly the same behavior over time, if in setting (b) the agents cannot condition transfers on outcome D occurring in the next period. Further, the latter restriction on setting (b) may be implied by limited liability, in that the legal institution might not enforce transfers conditional on severance unless there is cause for awarding damages. In most of the work presented here, options for long-term contracting do not affect our results. 9 This is without loss of generali- ty, given that in each period the agents maximize their joint value over feasible equilibria. EXTERNAL ENFORCEMENT AND VERIFIABILITY is unable to tell whether suspension resulted from a breach of contract by one of the parties. While severance following low effort is our benchmark case, the model can also cover situations in which temporary suspension is feasible. This is done by setting w j = g j – a j, j = w, m, where gw and g m give the discounted values to the worker and manager, respectively, of continuing the relationship into the next period, and a m and a w are the costs of maintaining the relationship while not producing. Full Verifiability The agents’ ability to find a contract that supports the robust solution will depend on whether they are able to enforce the needed contingent transfer payments. This, in turn, depends on what external enforcement authorities can observe about the currentperiod effort choices. We begin by considering the case of full verifiability, in which the external authority can perfectly observe which of the outcomes A, B, or C is realized. In this case, the robust solution is supported and, therefore, it is selected by the agents. This is confirmed by checking the four conditions of our contracting solution. Since the outcome must be C in every period under the robust solution, the worker’s total compensation is given by the stream of payments s G0 + s GC and s B0 + s CB for periods having the good state and bad state, respectively. Note that we are assuming the agents select the same contract in each period.9 Bargaining in each period determines the discounted value of this payment stream. This is characterized by Contracted transfer payments. The model allows for contracts specifying an up-front transfer to the worker, s0k, as well as a transfer that is made conditional on choices of high effort by both agents, sCk . The former can be interpreted as a “salary,” in that it is paid in return for the worker’s commitment to forgo his unemployment benefit and commit to production activities for some interval of time, while the latter represents a “performance payment,” received only after the successful completion of production.7 The transfers sA and sB are used to impose direct punishments for low effort and can be interpreted as damages stipulated by the contract for nonper- formance, or penalties imposed by an external legal or regulatory authority. (2) sk0 + sCk + g wR = π w zk + g R − b − w ( ) + b + w , k = G , B, w w where gwR indicates the discounted future value to the worker of continuing the relationship: Contract duration. We have assumed that agents can write only short-term contracts, specifying transfers that are enforceable within the current period. In this contracting environment, the transfers sA and sB can be thought of as severance payments (in addition to punishments), since the relationship is dissolved following low effort. Note that agents are free to sever their relationship following outcome C, but such a decision is made in phase 1 of the next period, after the current contract expires. Thus, by short-term contract we mean that the agents cannot stipulate to transfers conditional on whether they reach agreement in the negotiation phase of the next period.8 g wR = [ ( ) ( β (1 − ρ ) s0G + sCG + ρ s0B + sCB 1− β )] . To interpret equation 2, note that the left side is the worker’s value of continuing the relationship under the cooperative plan. The Nash solution dictates that s0k and sCk be set so that this value is equal to the worker’s outside option plus his share of the surplus of the relationship. The above two equalities capture the first and third conditions of equilibrium. To verify the second condition, note that outcome C F E D E R A L R E S E R V E B A N K O F S T. L O U I S 60 M AY / J U N E 1 9 9 9 is consistent with the agents’ private incentives at phases 2 and 3 if and only if solution can be supported. Adding the incentive conditions 3 and 4 gives k k R wR m m (3) z − sC + g − g ≥ x − s A + w (5) and (4) z k + g R ≥ x + w, which fails when k = B, given assumption 1. Thus, in the bad productivity state, either the manager or the worker will have an incentive to choose low effort, no matter what value of s is proposed. Limitations on verifiability, in the form of inability to condition severance transfers on the reason for severance, imply that the robust solution becomes infeasible. The key problem is that the joint surplus from cooperative behavior, given by z B + g R, falls short of the sum of the agents’ returns from low effort, which are x m + wm and xw + ww. Despite their inability to achieve the robust solution, the agents can find a contract that supports the fragile solution. We can specify s G0 + s GC to satisfy sCk + g wR ≥ x w + s B + ww . Inequalities 3 and 4 can be satisfied for each k by making sA sufficiently positive and sB sufficiently negative, that is, by imposing sufficiently large punishments for choosing low effort. Since the robust solution maximizes the joint value of the relationship in each period (from phase 1, where negotiation occurs), the fourth contracting condition also holds. Beyond the requirements on sA and sB, there is wide latitude for selecting salary and performance payments that satisfy equation 2, and there is essentially no distinction between the two kinds of payment. For example, contracts might involve performance payments only, or salaries only; in the latter case, the worker’s incentive to choose high effort is supported by the loss of future-period salary payments, rather than current- and future-period performance payments. (6) sG0 + sCG + g wF = π w zG + g F − b − w + b w + w w , ( ) where gwF gives the worker’s discounted future value of continuation in the fragile solution: g Limited Verifiability Now suppose the enforcement authority can enforce payments conditional on severance of the relationship due to low effort, but the authority cannot ascertain which agent’s low effort choice caused the separation. That is, the authority cannot distinguish between outcomes A and B. Remember that the agents cannot contract on severance following outcome C, since this would occur in the next period. However, the agents can still specify the transfer sCk contingent on C occurring. Further recall that, at the time of negotiation, there is no outstanding contract specifying transfers in the event of outcome D. Given the limitation on what can be observed, the contract can specify only a single severance transfer s, where sA = sB = s. Let us check whether the robust wF = [ ( ) ( β (1 − ρ ) s0G + sCG + ρ b w + w w 1 − β (1 − ρ ) )] . Thus, the first and third equilibrium conditions are satisfied. Since zG + gF > x + w, we can find a value sGC – s satisfying ( ) G G F wF m m (7) z − sC − s + g − g ≥ x + w and (8) (s G C ) − s + g wF ≥ x w + ww , and clearly each agent has an incentive to choose high effort in the good state. Thus, the relationship continues as long as the good state is realized, while in the bad state the relationship is severed. Finally, the fourth equilibrium condition follows from the fact that z G + g F > b + w; that is, the fragile solution is superior to immediate F E D E R A L R E S E R V E B A N K O F S T. L O U I S 61 M AY / J U N E 1 9 9 9 up-front transfer, s0G, and receives recompense sGC only in the event that high effort is realized. To the extent that sGC, is fixed by inequalities 7 and 8, higher values of p m correspond to larger bonding measures. severance, while the robust solution is infeasible. Importantly, severance is inefficient for the agents, since z B + g R > b + w implies that the agents would prefer the robust solution if it could be implemented. Observe further that there is a large range of payment profiles that can support the fragile solution; for example, if higher sGC is specified, then the severance transfers will be correspondingly increased to preserve inequalities 7 and 8, and sG0 will be reduced to maintain equation 6. As in the case of full verifiability, here the worker’s total compensation, driven by relative bargaining powers, does not determine the exact form of compensation. The analysis is similar for the case in which disagreement or low effort imply temporary layoff as opposed to severance. For example, suppose aw = a m = 0. In this instance, assumption 1 is replaced by b < zB < x < zG. Under the robust solution with temporary layoffs, we have w = gR ; since equation 6 continues to be necessary for satisfaction of the incentive constraints, it follows that the robust solution cannot be implemented as a consequence of zB < x. Further, it is easy to verify that the fragile solution, which involves layoffs in the bad state, can be implemented, and the assumption b < zB implies that the layoffs are inefficient. Noncontractible Worker Effort. The actions of some agents may be unobservable to the enforcement authority, even as transfers can be conditioned on the behavior of other agents. Consider the case in which the worker’s effort is noncontractible in this sense. Thus, the authority cannot disinguish between outcomes B and C, although A is still separately observable.10 In contrast to the case of limited verifiability, it is possible to implement the robust solution in this environment. First, the manager’s incentive constraint (inequality 3) can be satisfied by choosing sufficiently large sA. Since sB = s kC , however, the worker’s constraint (inequality 4) now becomes (9) Observe that current-period choices of s0k and sCk cannot affect whether inequality 9 is satisfied. It follows that the robust outcome is sustainable if and only if inequality 9 holds at values of the transfer payments that solve equation 2, which will tend to occur when p w is large or when xw is small. Thus, through their effect on the worker’s expected future compensation, bargaining weights have an impact on incentives, although they have no implications for the form of compensation (salary versus performance pay). Other Cases 10 It is implicit in this assumption that the authority cannot tell whether severance is the result of worker low effort in the current period or failure to reach agreement in the following period; for example, the current-period contract does not extend to cover separations that occur as a result of the worker’s low effort. g wR ≥ x w + ww . Limited Liability. Agents may be protected from liability for payments in the event that the relationship is severed. This serves as a further restriction on the case of limited verifiability, where sA = sB = 0 is now imposed. It is easy to see that there is a solution to expressions 6-8 satisfying this restriction: sGC is pinned down by inequalities 7 and 8, and sG0 is then chosen to satisfy equation 6. Interestingly, a contract of this form may involve bond-posting by the worker. For example, a high positive value of sGC may be specified in order to sustain the worker’s incentives to choose high effort, combined with a negative value of sG0 that implements the bargaining solution. Here the worker makes an Nonverifiability. Finally, consider the case in which the enforcement authority cannot distinguish between any of the outcomes A, B, and C. Thus, there is a single transfer payment sk that is enforced under all three outcomes. Note first that the robust solution cannot be forced in this case, as adding inequalities 3 and 4 for k = B implies violation of the assumption zB + gR < x + w. Next, the fragile solution can be enforced if the following conditions hold for the value of gwF determined by equation 6: F E D E R A L R E S E R V E B A N K O F S T. L O U I S 62 M AY / J U N E 1 9 9 9 (10) z G + g F − g wF ≥ x m + wm , (11) g wF ≥ x w + ww . mental to the idea of an efficiency wage is that motivating the worker to choose high effort places a binding constraint on wage setting, so that wages cannot be cut without inducing low effort. In other words, when the firm and worker negotiate over wages in a period, they confront a trade-off between the worker’s compensation and effort incentives. In this section we show, however, that such a trade-off never arises in the contracting setting considered thus far. Thus, there is no scope for efficiency-wage effects in contracting models of this form.11 Consider the incentive constraints for the manager and worker, which we can write generally as As in the previous case, the agents’ relative bargaining weights influence whether cooperation can be sustained. Note that these conditions are unaltered if it is instead assumed that the enforcement authority cannot enforce any transfers at all, since all needed transfers can be made using the up-front payment s0G. Summary Observability of actions within the relationship by external authorities plays a key role in determining how successful agents can be in solving their contracting problems. Full verifiability implies that the complete range of necessary transfer payments can be enforced, allowing the most efficient solution to be implemented. In contrast, nonverifiability rules out efficiency, and even the fragile solution becomes unenforceable for a range of parameter values. Between these two extremes, various possibilities arise. When verifiability is limited in the sense that severance payments cannot be conditioned on the reason for severance, only the fragile solution is implementable; when worker effort is noncontractible, the bargaining outcome determines the solution, and there will be no production in any state when the worker’s bargaining power is sufficiently small. Finally, except in the case of limited liability, the split of the worker’s compensation between salary and performance payments plays no role in implementing the various solutions. z k − sCk + g j − g wj ≥ x m − s A + wm and sCk + g wj ≥ x w + s B + ww . Observe that, in addition to the parameters zk, xm, xw, wm, and ww that are fixed from the perspective of the manager and worker, these constraints depend on three sets of parameters. First, there is the joint continuation value g j, which is maximized when the agents select the best equilibrium (either robust, fragile, or immediate severance). Since higher values of g j relax the incentive constraints, there is no trade-off between compensation and incentives at the level of equilibrium selection. Second, the constraints involve the manager and worker’s shares of the continuation value, described by g j – gwj and gwj. Given g j, these values are tied down by negotiation in future periods, which in turn is fully determined by bargaining weights and the fixed disagreement point D. In other words, from the agents’ perspective at the negotiation phase in any given period, they have no control over continuation values in a way that forces them to address a trade-off between compensation and incentives. The third set of parameters comprise the contracted transfers sA, sB, and sCk . These are directly controlled by the worker and firm in the current period. Note, however, that the up-front transfer s0k does EFFICIENCY WAGES Efficiency Wages and Contract Negotiation The literature on moral hazard in labor relationships has placed great emphasis on solving worker incentive problems through the payment of efficiency wages. Funda- F E D E R A L R E S E R V E B A N K O F S T. L O U I S 63 11 The term “efficiency wages” is also used in connection with the idea that incentive problems lead to involuntary unemployment. Regardless of incentive problems, however, employed workers fare better than unemployed workers whenever employment relationships entail quasi-rents (as when matching is costly/frictional) and workers have some bargaining power. Further, as argued by Carmichael (1985), involuntary unemployment is not a necessary consequence of incentive problems. M AY / J U N E 1 9 9 9 not appear in the incentive constraints. As a free parameter, s0k can be set to affect any division of the relationship’s value between the firm and worker, with no implications for the provision of incentives in the current period.12 As a result, during contract negotiation, there is absolutely no trade-off between compensating the worker and inducing high effort, and so there is no payment of efficiency wages.13 This is not to say that incentive constraints are unimportant. Our point is that consideration of incentives in employment relationships should center on the satisfaction of incentive constraints given the contracting and matching environment, which may or may not generate phenomena such as efficiency wages. Importantly, the contracting environment is described by bargaining powers, whether negotiation is recurrent, and the extent of verifiability. sCB + g wR < x w + ww ≤ sCG + g wR . Here the agents must agree to a higher value of s CB when the bad state is realized, in order to induce the worker to choose high effort. Correspondingly, sCG will be chosen at a lower value in order to maintain equation 2 in the good state. It may be that sCG must be lowered so much that inequality 12 becomes binding even in the good state. In any event, we have ( and it follows that the worker receives an efficiency wage in the bad state. Observe that the worker obtains a value strictly in excess of his outside option even when p w = 0; in this case, compensation is equal in both states, and efficiency wages are paid in both states. We conclude that efficiency wages may emerge when worker liquidity constraints rule out the use of direct penalties or worker bonding to enforce high effort. It should be noted that the manager must give up some of his bargaining surplus when efficiency wages are needed, which may lead to disagreement and inefficient severance despite the existence of full verifiability. Whenever inequality 12 is binding, the manager obtains a payoff of zk + gR – xw – ww, which can lie below his outside option value bm + wm even when agreement is reachable in the absence of liquidity constraints. A similar analysis may be carried out for the other contracting environments, where prospects for obtaining productive solutions are also reduced by the addition of a worker liquidity constraint.14 Worker Liquidity Constraints 12 Note that only in the case of nonverifiability is the value s0k constrained to equal one of the other contracted transfers. In this case, however, the externally enforced transfers disappear from the incentive constraints altogether. 13 This conclusion remains valid when effort choices are imperfectly monitored; see note 4. Imperfect monitoring affects the values of g j and g wj that can be achieved and alters the form of the incentive constraints, but it remains the case that sOk can be freely varied to effect any desired division of surplus. 14See Dickens, et al. (1989) for an informal discussion of legal and social constraints to bonding that can motivate payment of efficiency wages. These authors also consider the tradeoff between costly monitoring and performance payments as mechanisms for eliciting effort. Efficiency-wage effects emerge if the worker is unable to make payments to the manager, because of insufficient worker liquidity. A worker liquidity constraint can be introduced into the model by requiring that all transfer payments be nonnegative. Consider the implications of this constraint in the case of full verifiability. Since sB >_ 0, supporting the robust solution requires that inequality 4 be replaced by (12) ) sCB + g wR > π w z B + g R − b − w + b w + ww , sCk + g wR ≥ x w + ww , where gwR is determined by equation 2. Inequality 12 is made as slack as possible by setting the salaries s0k equal to zero and compensating the worker completely through performance payments. If inequality 12 still does not hold, then sCk must be raised above the value determined by equation 2 in order to induce high effort, so that inequality 12 becomes binding in sCk . In this case, a trade-off between compensation and incentives is clearly present, and we can say that an efficiency wage is paid in state k. As one possibility, suppose that equation 2 with s0k = 0 implies the following: Relation to Other Models In this section, we consider how the efficiency wage issue is treated in a few of the standard models of dynamic labor contracting found in the literature. The model of Shapiro and Stiglitz (1984) can be viewed as producing a trade-off between worker compensation and incentives by constraining the kinds of contracts that firms may offer to workers. In essence, firms are required to offer a single, stationary wage. Over F E D E R A L R E S E R V E B A N K O F S T. L O U I S 64 M AY / J U N E 1 9 9 9 multiple periods of time, firms are committed to the same transfer in each period, conditional on no discovery of shirking. In fact, firms would prefer to offer a low wage in the current period, with only the promise of higher wages later.15 MacLeod and Malcomson’s (1989, 1993, 1998) theory is designed to rectify such inconsistencies by explicitly modeling the contracting process. They provide a more rigorous foundation for the kinds of market phenomena of interest to the early efficiency-wage literature, such as involuntary unemployment. By tying prevailing labor contracts to a social norm, however, their model does not incorporate trade-offs between compensation and incentives at the level of individual employment relationships. Rather, compensation and incentives are together traded off against social sanctions.16 In Ramey and Watson (1997a), firm/worker pairs determine long-term contracts through direct bilateral negotiation, and they are not influenced by social norms. Like the model presented here, however, there is no tension between compensation and incentives, so efficiency wages are not at issue. Through the use of an up-front transfer, a firm and worker can manage any split of the relationship’s value, while implementing the best outcome that verifiability will allow. The present model takes the contracting framework a step further by incorporating the negotiation phase in each period of interaction, which implies that ongoing surplus division is moderated by bargaining weights.17 managers can elect to post vacancies at a cost of c > 0. For simplicity, we assume that unmatched workers bear no search costs. The flow of new matches in a period is given by a standard matching function, m(U,V), where U indicates the mass of unmatched workers and V gives the mass of managers who post vacancies.18 The matching process is assumed to take place in phase C at the same time as production occurs in active relationships. Thus, workers whose relationships are severed in phases D, A, or B can enter the current-period matching pool. Further, to ensure that the pool of unemployed workers does not become empty, we assume that, with probability r x, relationships are severed for exogenous reasons. Exogenous separations occur in phase 1, and workers who experience these separations can also enter the current-period matching pool. We consider two types of steady-state equilibria of the model, distinguished by whether contracting solutions within relationships are robust or fragile. For robust and fragile equilibria, respectively, the discounted future values of relationships are determined by MARKET OUTCOMES where wR and wF give the values of outside options in robust and fragile equilibria. The value of the worker’s outside option in either case satisfies {( )[(1 − ρ) z + ρ (b + w )} gR = β 1 − ρ x x and G + ρ zB + β g R ] R ) {( ) ( + [1 − (1 − ρ ) (1 − ρ )](b + w )}, g F = β 1 − ρ x (1 − ρ ) z G + β g F x We now describe how employment relationships are formed in steady-state matching equilibria. Assume that the labor market contains a unit mass of workers, each of whom either begins a period matched with a manager in an employment relationship or begins the period in a pool of unmatched workers seeking to locate a manager. In addition, there is a large number of potential managers. At the beginning of each period, unmatched w wj = m(U , V ) U F g wj m(U , V ) wj + 1 − βw , j = R, F , U where g wR and g wF are determined by equations 2 and 6, respectively. Because of F E D E R A L R E S E R V E B A N K O F S T. L O U I S 65 15 Resolving this issue requires a more complete model of contract determination, as Carmichael’s (1985) critique of the Shapiro-Stiglitz (1984) model indicates. 16 In MacLeod and Malcomson (1998), if a firm offers any contract not in accord with the norm, it is branded as a deviant, and workers at this firm shirk forever after. In MacLeod and Malcomson (1989), the social coordination role is modeled more abstractly in terms of prevailing equilibrium beliefs. Incidentally, since a matched firm and worker do not have direct control over their joint plan of behavior in the theory of MacLeod and Malcomson, total inaction can be supported as an equilibrium. 17 Ramey and Watson (1997a) also incorporate what one might call an efficiency investment: Since the firm makes a noncontractible investment that affects incentives and value, it faces a direct trade-off between compensating the worker (by raising or lowering the value of the relationship) and satisfying incentive constraints. 18 Added assumptions are ordinari- ly imposed to guarantee existence of equilibrium and to facilitate theoretical analysis of steady-states and dynamics. We do not lay out these assumptions here, since we restrict our attention to a single numerical example. M AY / J U N E 1 9 9 9 particular parameterization of the model. Equilibrium employment and average wages under the four cases are traced out as r rises from zero, at the upper righthand corner of all four curves, to 0.04. For comparison, the values at r = 0.02 are indicated by dots.19 Consider first the case of full verifiability, in which workers and managers are able to write robust contracts. The right-most curve in Figure 2 depicts employment and average wages for this case. A productivity shock taking the form of an increase in r shifts the outcome down the curve, so that employment and wages both fall. Since relationships are robust, the rise in r has no effect on the breakup probability. Employment is lower only because managers are less willing to post vacancies, given that average productivity is lower. The reduction in wages also reflects lower average productivity, as well as a reduced value of the worker’s matching probability. Next, the case of full verifiability with worker liquidity constraints, as discussed in the previous section, is considered for two values of the worker matching probability. For p w = 1/2, equilibria are robust, and the worker liquidity constraint binds in the bad productivity state; thus, efficiency wages are paid only in the bad state. In this case, wages adjust a little less relative to employment, when compared to the full verifiability case, but the effect is slight. Setting p w = 0 yields robust equilibria with efficiency wages paid in both states, and relative wage adjustment declines a bit more.20 In these cases, worker liquidity constraints restrict the decline in wages as r rises, and the dampening effect on wage adjustment is more pronounced as the liquidity constraint binds in a larger number of states. Finally, the case of limited verifiability, described in the external enforcement and verifiability section, is shown as the left-most curve. Equilibria are fragile in this case; in particular, inequality 1 holds at the value w = wF. As r rises, employment reductions become much sharper due to the increase in the probability of severance. Average productivity is also reduced relative to the Figure 2 Average Wages and Employment in Steady-State Equilibria Wages 0.865 Efficiency Wages (Worker’s Bargaining Weight = 0) 0.860 0.855 Limited Verifiability 0.850 Efficiency Wages (Worker’s Bargaining Weight = 0.5) 0.845 Full Verifiability 0.840 0.9125 0.9175 0.9225 0.9275 Employment 0.9325 0.9375 NOTE: This figure plots the combinations of employment and average wages that are obtained in the specified contracting environment for different values of r (the probability of the “bad” state occuring). The dots correspond to values of r = 0.02. free entry of managers into the vacancy pool, the value of managers’ outside option is zero, and we have the following freeentry condition: m (U , V ) ( ) g j − g wj = c , j = R, F. V Finally, the number of employed workers and the size of the unemployment pool are given by ( ) N = 1 − ρ x (1 − ρ ) N + m (U , V ) and [ ( ) ] U = 1 − 1 − ρ x (1 − ρ ) N. 19Parameter values are z G = 1, z B = 0.5, x w = 1.25, x m = 1.45, b w = b m = 0.2, b = 0.96, m(U, V) = 0.25U 0.5V 0.5, c = 0.157, and r x= 0.07 . We also have p w = 0.5 except for one case where we set p w = 0. 20We have renormalized the p w = 0 economy to equate employment and wages at r = 0 under the various cases. Observe that the latter two equations conform to the assumption that workers whose relationships are severed enter the unemployment pool in the current period. We model cyclical shocks as changes in r holding other parameters fixed. To keep the analysis simple, we look at comparative statics of steady states as a function of r. Thus, we approximate the business cycle by studying the economy’s long-run response to a highly persistent productivity shock. This is a good approximation when the cycle has a low frequency and/or response is rapid. Figure 2 shows results for four contracting environments under a F E D E R A L R E S E R V E B A N K O F S T. L O U I S 66 M AY / J U N E 1 9 9 9 earlier cases, since output in bad states is zero, rather than z B. The latter effect also tends to reduce employment, by depressing vacancies, and it causes wage reductions to be much greater. Thus, the higher breakup probabilities lead shocks to be significantly magnified when contracts are restricted by limited verifiability. Observe further that relative wage adjustment is substantially less when compared to the other cases, as the higher probability of breakups serves to shift the cross-sectional distribution of wages toward relatively high productivity relationships. Overall, this example demonstrates that efficiency-wage effects can dampen wage adjustments, as past authors have suggested, but that the scope for efficiency wages as a mechanism for propagating shocks is limited. Fragility effects deriving from limited verifiability, on the other hand, can produce large magnification of shocks; further, changes in the composition of jobs generates more dampening in the adjustment of wages.21 employment and the relatively dampened character of wage adjustments. Broad cyclical swings in job destruction rates have been documented by Davis, Haltiwanger, and Schuh (1996), who also highlight the large number of macroeconomic questions that may be linked to job creation and destruction. From the quantitative standpoint, den Haan, Ramey, and Watson (1997) show how fluctuations in job destruction rates can serve as an important mechanism for propagation of business cycle shocks. The heavy focus of much past work on wage-setting within a given set of employment contracts may be misplaced. Third, interactions between credit market imperfections and imperfections in labor contracting can yield interesting new implications. In this paper, we have linked the occurrence of efficiency wages with worker liquidity constraints. More broadly, the ability to solve contracting problems is closely tied to credit-market trading, and these ties may prove to be of central importance in accounting for macroeconomic phenomena. CONCLUSION On the basis of the preceding results, we offer three broad conclusions. First, the particular form of imperfections that are present in the contracting environment can have major implications for economic outcomes. In moving from limited liquidity to limited verifiability, for example, the implications for important variables such as employment, wages and job displacement probabilities can be radically altered. “Reduced-form” analysis of contracting imperfections that have been prevalent in much past macroeconomic literature may hide too much of the key underlying structure. Contractual outcomes depend on the way firms and workers meet and negotiate, and this demands a new theoretical perspective and reformulation of conventional notions, such as the idea of efficiency wages. Second, economic effects deriving from severance of employment relationships warrant very close attention as explanations for observed phenomena, including the occurrence of large cyclical fluctuations in REFERENCES Boldrin, Michael, and Michael Horvath. “Labor Contracts and Business Cycles,” Journal of Political Economy (October 1995), pp. 972-1004. Carmichael, Lorne. “Can Unemployment be Involuntary? Comment,” American Economic Review (December 1985), pp. 1213-14. Danthine, Jean-Pierre, and John B. Donaldson. “Efficiency Wages and the Business Cycle Puzzle,” European Economic Review (November 1990), pp. 1275-1301. ________ , and ________. “Non-Walrasian Economies,” Frontiers of Business Cycle Research, Thomas Cooley, ed., Princeton University Press (1995), pp. 217-42. Davis, Steven J., John C. Haltiwanger, and Scott Schuh. Job Creation and Destruction, MIT Press, 1996. Den Haan, Wouter J., Garey Ramey, and Joel Watson “Job Destruction and Propagation of Shocks,” Discussion Paper 97-23, UCSD, (October 1997). ________ , ________ , and ________ . “Liquidity Flows and the Fragility of Business Enterprises,” Discussion Paper 99, UCSD, (Februrary 1999a). F E D E R A L R E S E R V E B A N K O F S T. L O U I S 67 21As mentioned earlier, inefficient severance in the bad state may occur under full verifiability and worker liquidity constraints when the manager must give up too much surplus in order to elicit worker effort. In this case, fragile contracts would arise in equilibrium, leading to aggregate wage adjustment and magnification effects similar to those of the limited verifiability case. M AY / J U N E 1 9 9 9 ________ , ________ , and ________ . “Job Destruction and the Experiences of Displaced Workers,” manuscript, (May 1999b). Dickens, William T., Lawrence F. Katz, Kevin Lang, and Lawrence H. Summers, “Employee Crime and the Monitoring Puzzle,” Journal of Labor Economics (July 1989), pp. 331-46. Hall, Robert E., and Edward P. Lazear, “The Excess Sensitivity of Layoffs and Quits to Demand,” Journal of Labor Economics (1984), pp. 233-57. Hashimoto, Masanori, and Ben T. Yu. “Specific Capital, Employment Contracts, and Wage Rigidity,” Bell Journal of Economics (1980), pp. 536-49. MacLeod, W. Bentley, and James Malcomson. “Implicit Contracts, Incentive Compatibility, and Involuntary Unemployment,” Econometrica (March 1989), pp. 447-80. ________ , and ________. “Wage Premiums and Profit Maximization in Efficiency Wage Models,’’ European Economic Review (August 1993), pp. 1223-49. ________ , and ________. “Motivation and Markets,” American Economic Review (June 1998), pp. 388-411. ________ , ________ , and Paul Gomme. “Labor Turnover and the Natural Rate of Unemployment: Efficiency Wage versus Frictional Unemployment,” Journal of Labor Economics (April 1994), pp. 276-315. Mortensen, Dale, and Christopher A. Pissarides. “Job Creation and Job Destruction in the Theory of Unemployment,” Review of Economic Studies (July1994), pp. 397-415. Pissarides, Christopher A., “Short-Run Equilibrium Dynamics of Unemployment, Vacancies, and Real Wages,” American Economic Review (September 1985), pp. 676-90. Ramey, Garey, and Joel Watson. “Bilateral Trade and Opportunism in a Matching Market,” Discussion Paper 96-08, UCSD, March 1996. ________ , and ________. “Contractual Fragility, Job Destruction, and Business Cycles,” Quarterly Journal of Economics (August 1997a), pp. 873-91. ________ , and ________. “Contractual Intermediaries,” Discussion Paper 97-26, UCSD, October 1997b. Romer, David. Advanced Macroeconomics, McGraw-Hill, (1996). Shapiro, Carl, and Joseph E. Stiglitz. “Equilibrium Unemployment as a Worker Discipline Device,’’ American Economic Review (June 1984), pp. 433-44. Strand, Jon. “Business Cycles with Worker Moral Hazard,” European Economic Review (August 1992), pp. 1291-1303. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 68 M AY / J U N E 1 9 9 9 Christopher Foote is an assistant professor of economics at Harvard University. Commentary The authors contend that long-term employment relationships can survive even in bad aggregate states if the agents can successfully commit themselves not to shirk, but this requires low effort to be detectable by a court so that it can bring about the required intermediate exchanges. The setup reminded me of an unemployed worker in the Shapiro-Stiglitz model; such a worker would like to be hired at the prevailing wage, but his offer is not accepted because he cannot make a credible promise not to shirk. Here, in the case of perfect verifiability, the offer is accepted even when the production value is low because shirking is detected with probability one and can be proven in court. This brings about the required intermediate transfer from the worker to the firm. (Another difference between this model and the Shapiro-Stiglitz framework is that in this model the workers can post bonds for jobs via negative values of the initial payment s k0 , a point to which I will return below). When verification of effort is not possible in the DRW model, the intermediate-payment strategy for the robust employment contract fails. The constant benefit of shirking is too large and tempting for either the worker or the firm, relative to the low outcome from production. The model has several interesting microeconomic implications. But even though the possibility of initial, intermediate, and final payments makes the model very versatile, I am not sure that it is general enough to speak definitively on the issue of salary versus performance pay in an optimal compensation policy. The applicability of the model for this issue hinges on whether a negative payment to a worker in the initial period (s k0 < 0) can be thought of as a salary, or rather, as some sort of negative bond reflecting the ex ante terms of trade in the labor market. Because such bonds can eliminate involuntary unemployment when effort elicitation is a problem, such bonds have featured prominently in Christopher Foote T his stimulating paper by den Haan, Ramey, and Watson (DRW) seeks to contribute to two literatures. First, it helps answer the microeconomic question of when firms will use performance pay rather than salaries to motivate workers. Second, it examines the microeconomic issue of why employment relationships appear so fragile with respect to aggregate shocks. The overarching goal is to show that these two literatures are related, or, as the authors put it, “the particular form of imperfections that are present in the contracting environment can have major implications for economic outcomes.” The central result of the paper is that employment contracts often are robust with respect to aggregate shocks only if the effort choices of both the worker and the firm are verifiable and contractable. The paper is exceptionally clear and insightful; even though the model is simple, the authors can analyze a surprisingly large number of special cases. As a result, it is a great paper for a volume that seeks an integrated study of the macroeconomics of the labor market. I will comment first on the structure of the model and then discuss its micro and macro implications. The nicely parsimonious model is based around three possible exchanges between the worker and the firm. The two parties begin the period by negotiating the contract and making an initial exchange (s k0 ), which depends on the aggregate state k. In the negotiation process, they promise to make intermediate exchanges ( s A and s B ) if one of the parties shirks. Finally, if both parties provide the required levels of high effort, production completes satisfactorily and the firm receives the output, making a final exchange (s kC ) to the worker. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 69 M AY / J U N E 1 9 9 9 work of this kind. Yet the authors claim that in their model, a positive value s k0 is like a salary, while the final payment, sCk , resembles performance-based pay. Therefore, they suggest that their model can speak to the performance-pay versus salary issue. Yet the DRW model does not allow the worker both to post a bond (which would require s k0 < 0) and receive a salary (which would require s k0 > 0). As a result, I am not sure that the model can be considered a general case of the previous models in the effort-elicitation literature. An alternative interpretation of the salaryvs.-performance-pay distinction in the DRW model would be that some average value of the end-of-production payment (which could be denoted s C ) would be more like a salary, while the time-varying payments (sCk ) are similar performancebased pay. The point is that a salary is generally a constant payment received by the worker regardless of how much output is produced, yet if production does not take place, no salary is received. A larger microeconomic goal of the paper is to investigate the effects of nonverifiability of effort in employment relationships. To make the model tractable, the authors make some simplifying assumptions expressed in inequality 1. One implication of the assumptions is that the promise of continuing the job is generally not enough to preserve a current match in the bad state today, unless there is some verifiability of effort. Basically, firms and workers will not stick it out through the bad times in order to enjoy the good times again later on. As the paper shows, however, the continuation value of the job in the different states (the g’s) are themselves functions of the common discount factor (b), the probability of the bad state (r), and the relative levels of the productive outcomes (zG and zB). Therefore, underlying the assumptions in inequality 1 are implicit assumptions about preferences and the stochastic properties of aggregate shocks. Indeed, firms and workers may want to endure the bad times even without verifiability—if the discount rate is small enough or bad aggregate shocks are rare enough— but particular values of the underlying parameters, which imply that verifiability is required for a robust contract, are not obvious. A third comment on the micro-structure of the model involves the relationship of this model to work involving specific investments in the employment relationship. Papers by Caballero and Hammour stress the “fundamental transformation” that employment relationships undergo when searching workers and firms find one another, or when either side of the employment relationship invests in capital that is specific to the match. Both of these phenomena transform the employment relationship into a bilateral monopoly where the ex ante terms of trade may not carry over to the ex post Nash bargain over the surplus in the match. In reading this paper, I was curious to know whether there was a simple mapping between its shirking interpretation and the specific capital basis of other work. It may be that the shirking model is a particularly strong form of the specific capital model. For example, consider the type of specific capital that is created simply when a searching worker finds a specific job. Match capital is created because the firm no longer has to pay the costs of posting the vacancy and the worker can start earning wages rather than spend time looking for a job. The outside alternatives of the worker, the firm, and the exogenously determined bargaining weights (the p ’s) determine how the surplus is divided in both models; though in the DRW setup, the worker or the firm receives an additional reward (the benefit to shirking) if the match breaks up. In the specific capital model, a party who leaves the match does not receive this type of benefit. It would be interesting to know if there is a simple way to link both the shirking and specific-capital interpretations of the employment relationship. One way that ex ante terms of trade can be reflected ex post in the employment match is when the workers post a bond. One of the cases discussed by the authors is that of “limited liquidity,” which argues F E D E R A L R E S E R V E B A N K O F S T. L O U I S 70 M AY / J U N E 1 9 9 9 that low liquidity may prevent workers from paying a bond. The inability to pay a bond opens up the possibility that efficiency wages may be paid in order to elicit effort. Several economists, however, have pointed out that the utility cost of a bond payment may vary inversely to the worker’s liquid assets. Workers with low liquidity may find it hard to post a bond, but it is precisely these workers who will be quite averse to shirking and losing their bond if they get caught shirking. Of course, the power of very small bonds to motivate liquidity-constrained workers depends on marginal utility going to negative infinity as assets go to zero, so the outside benefits, b, may prevent this from occurring in the DRW setup. I now turn to the macroeconomic implications of the model. One of the key questions in cyclical macroeconomics is why so many employment relationships break up during recessions. Pioneering work by Davis, Haltiwanger and Schuh showed that (at least in manufacturing) the drop in employment that occurs when recessions begin comes not from a decline in the creation of new jobs, but rather, a large spike in the destruction of existing jobs. Several authors have suggested that the spike in job destruction at the onset of recessions may be linked to the economy’s amplification mechanism, by which moderate innovations to productivity or aggregate demand may bring about large movements in employment and output. Not surprisingly, a number of theories to explain the burst in job destruction have been advanced. One branch of the literature stresses “cleansing” effects of recessions. Large numbers of jobs are destroyed in recessions because a large number of jobs are typically close to the margin of unprofitability at any point in time. Convex job-creation costs for the aggregate economy mean that it is more efficient for these older jobs to be destroyed than for the rate of new job creation to drop. A second branch of the job destruction literature stresses the “pit stop” role of recessions. Just as the real business cycle literature contends that recessions are a good time to enjoy leisure, that pit-stop view suggests that recessions are good times to reorganize production. This paper can be placed in a branch of the economic literature that contends that the increase in job destruction is a primary result of some imperfection or friction in the labor market; here, the friction is nonverifiability. Other papers of this kind suggest that job destruction is high during recessions because wages cannot fall. Two potential causes of wage rigidity are the suppression of wage renegotiation (since bargaining is costly and may encourage opportunistic behavior) and efficiency wages (the need to motivate workers provides a floor through which wages cannot fall). The DRW model suggests that nonverifiability, rather than the suppression of negotiations or efficiency wages, can better explain the cyclical dynamics of the labor market. The authors implicitly argue against the suppression of the renegotiation model by having the worker and firm bargain at the start of every production period. The suppressed renegotiation models imply that the firm and the worker would like to renegotiate and stay together, but doing so would result in redoing the employment contract and thereby violate some social norm. The DRW paper suggests, in contrast, that firms and workers are not averse to renegotating, but they prefer to separate in bad times because they cannot make a credible promise not to shirk without external verifiability. The paper argues against the efficiencywage-model-with-verifiability more explicitly with the experiments displayed in DRW’s Figure 2. Recall that efficiency wages arise in the DRW framework if workers are unable to post a bond and the value of the final payment they can receive at the end of production is not large enough to prevent them from shirking. In this case, the firm must raise the worker’s total compensation (here, just the final payment) above the level implied by the worker’s bargaining weight, p w. In this way, the worker is encouraged not to shirk and an efficiencywage trade-off arises. This is due to the F E D E R A L R E S E R V E B A N K O F S T. L O U I S 71 M AY / J U N E 1 9 9 9 inverse relationship between the total compensation of the worker and his incentive to shirk. The firm is prevented from shirking because of the verifiability assumption. Should it fail to provide high effort, the court will assess a payment to the worker of sA. The efficiency-wage cases analyzed in Figure 2 are, therefore, ones of robust contracts, because the efficiency wage and the monitoring of the firm by the court combine to keep the parties honest. Figure 2 shows that the three verifiable contracts result in little amplification of shocks, which are modeled as an increase in the breakup parameter r. Even the two contracts that imply efficiency wages are robust, in part because firm behavior can be verified. On the other hand, the nonverifiable (“severance payment only”) contract results in substantial amplification of the increase in r. Employment relationships break up because no mechanism exists to ensure high effort. My main concern with this result is that I am not sure it portrays efficiency wages in the most familiar way. The efficiency-wage model in this paper is essentially backloaded compensation, which is paid only when the worker does not shirk, an event that is detected with probability one. (Of course, there may be an incentive for the firm to shirk when compensation is backloaded, but the courts are assumed to regulate the firm’s behavior since verification is assumed in these cases.) Another interpretation of efficiency wages, however, might affirm that they arise when worker misbehavior is detected only imperfectly and, therefore, the workers may consider the shirking decision differently. The differences in the two interpretations for macro behavior of the two views of efficiency wages are not obvious immediately. Another question I have is how general equilibrium effects work in the macro simulations of DRW’s Figure 2. The general equilibrium is important because the previous work by Caballero and Hammour has shown that high unemployment during a recession can result from wage rigidity engendered by specific-capital investments. The high unemployment is an equilibrium response of the economy to wage rigidity, as it disciplines the wage demands of insiders. I am curious to know whether a similar effect is operating here. All in all, I found this paper well worth the time and effort it took to study the subject carefully. And I hope the graduate students enrolled in my next “Macroeconomics of the Labor Market” course do so as well. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 72 M AY / J U N E 1 9 9 9 Gilles Saint-Paul is a professor at Universitat Pompeu Fabra and a fellow of the CEPR. He is grateful to the Bank of Spain and the Centre de Recerca en Economia Internacional, Universitat Pompeu Fabra, as well as Chris Waller and participants at the Federal Reserve Bank of Saint Louis annual conference, and seminar participants at the Bank of Italy, Rome and CEMFI, Madrid for helpful comments and suggestions. Assessing the Political Viability of Labor Market Reform: The Case of Employment Protection wage. In 1994, the Swedish government lost the elections because it had lowered the unemployment benefit replacement ratio from 90 percent to 80 percent. After reunification, the German government gave in to western unions’ pressure and allowed eastern wages to converge rapidly to western levels, despite large productivity differentials and the need to restructure the eastern economy, which led to substantially higher unemployment rates in the East than in the West. In my view, an understanding of the political determinants of labor market institutions is a crucial prerequisite for being able to implement structural reforms that are acceptable to those social groups that potentially may block these reforms. While we believe that the set of institutions that prevail in many European countries form a coherent whole, given the complexity of the issue it is often more convenient to analyze these institutions separately from each other. In this paper we focus on employment protection legislation (also called “firing costs”). We want to know who gains and who loses from such regulation, and what will be the equilibrium level of employment protection. We abstract from other rigidities—we do not ignore them, but take them as given, ignoring that they, too, are the outcome of the political process. Why firing costs rather than other institutions? This is partly a matter of taste and I have discussed other institutions elsewhere.1 But there are several reasons why employment protection is more relevant than other rigidities when one deals with the political economy of reform. First, it is regularly pointed out by employers as one of the most severe constraints on their incentives to create jobs. Second, it is somewhat more renegotiable than minimum wages or unemployment benefits. Some reductions in firing costs have been observed in various countries in the eighties and nineties. We have not seen similar reductions in unemployment Gilles Saint-Paul T here is somewhat of a consensus among economists that labor market rigidities are responsible for high unemployment in Europe, and in particular for its most alarming aspects such as its long duration and high incidence on youth. Unemployment benefits lower the incentive for job search and increase wage pressure by insiders. Minimum wages price the least skilled out of the market. Firing costs deter hiring, thus reducing labor demand, and hamper the economy’s ability to deal with uncertainty and structural change. This is why experts frequently recommend making the labor market more flexible, as is exemplified by the conclusions of the recent OECD Jobs Study (1995). But, in practice, few of the remedies economists advocate pass the test of political viability. In 1994, an attempt by the French government to lower the minimum wage for young workers was followed by violent demonstrations, and the government eventually withdrew its reform proposal. In 1995, in order to be elected, a French presidential candidate put on its platform an increase in the minimum F E D E R A L R E S E R V E B A N K O F S T. L O U I S 73 1 For example, Saint-Paul (1996a, b). M AY / J U N E 1 9 9 9 2 See Saint-Paul (1996b). 3 See, for example, Bentolila and Bertola (1990). 4 See Cohen, Lefranc, and Saint-Paul (1997), Blanchard and Portugal (1998). compensation or minimum wage laws. Unemployment benefits are seen as part of the “welfare state” and attempts to reduce them often are interpreted as a first blow to the whole welfare state, while the minimum wage is often an untouchable symbol.2 Third, while firing costs’ impact on employment is actually unclear,3 they clearly increase unemployment duration. If anything, the key difference between Europe and the United States is not so much the former’s higher unemployment rate—which partly reflects composition effects and a greater incentive to register as unemployed—as Europe’s much larger unemployment duration.4 Behind the political support for employment protection lies the existence of rents in favor of the employed, which arise due to imperfections in the labor market. We understand firing costs as a device to protect the rents of incumbent employees. The greater these rents, the greater their incentive to support protective measures. We define the “rent” as the welfare differential between an employed and an unemployed worker. In a perfectly competitive labor market, this differential should be equal to zero, because any worker looking for a job would find one instantaneously at the going equilibrium wage. Thus, there would be no welfare difference between the employed and the unemployed. In practice, the employed have rents, that is, they are strictly better off than the unemployed. The size of these rents depends on their bargaining power (their ability to prevent the unemployed from underbidding them, which itself is affected by labor market institutions), and also how closely their work effort can be monitored by employers. The rent is a measure of how far wage setting is from competitive behavior; the higher the rent, the less competitive wage formation and the higher the natural rate of unemployment. Most of the essence of labor market reform is about eliminating the rent. This is certainly true of any reform of the minimum wage and the bargaining process, or of any change that makes it easier for outsiders to compete with insiders: hiring rules, work rules, and many aspects of employment protection. Here, however, we take the workers’ bargaining power as given, and consider what happens when people vote on a firing cost that does not directly affect their bargaining power. Rents have important consequences for the political preferences of incumbent employees. This is because the rent tells us how much they lose if they lose their jobs, or how much they are willing to pay for keeping them. The greater the rent, the greater the aversion of insiders to unemployment and the greater the political support for employment protection legislation. Employment protection legislation is complex; it associates to each cause of firing a set of constraints imposed on the employer. These constraints include severance payments, administrative supervision, obligation to provide the displaced workers with job counseling and to give them priority over hiring by the same conglomerate, unions’ right of scrutiny and appeal, etc. To some extent, these constraints increase the employee’s bargaining power by making it more difficult for the employer to resist wage demands by refusing to employ the worker any longer. The direct effect of firing costs, however, is to make it more costly for the firm to adjust its labor force when facing a fall in demand. Because we want to isolate the pure employment protection effect of firing costs, we shall assume that it does not affect the workers’ bargaining power. Unlike my previous work on the same topic (Saint-Paul, 1993, 1997), this paper pays a lot of attention to the role of firing costs in the growth process when obsolescence—or “creative destruction”—is an important aspect of growth. In our vintage capital model, each match gradually becomes obsolete (because its productivity fails to catch up with the latest technology) until it is destroyed, at which time the worker becomes unemployed. We assume people vote between two levels of the firing cost (a “flexible” and a “rigid” one). In our model, firing costs increase the life span of any match by inducing firms to postpone the date of economic obsolescence. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 74 M AY / J U N E 1 9 9 9 In voting in favor of employment protection, incumbent employees trade off lower living standards (because employment protection maintains workers in less productive activities) against longer job duration. The support for employment protection will then depend on the value of the latter relative to the cost of the former. We highlight two key determinants of this trade-off: first, the workers’ bargaining power; second, the economy’s growth rate—more precisely its rate of creative destruction. Let us explain briefly the mechanisms that underlie the effect of these two parameters. The rent. The value of longer job duration to incumbent workers is proportional to the rent, or equivalently, their bargaining power; long job duration would not be valued if the employed were not earning rents above the unemployed. The cost of job loss would then be zero, and so would the support for employment protection. This result tells us that there exists a “complementarity” between firing costs and other labor market rigidities to the extent that the latter increases workers’ bargaining power. One important consequence is the existence of complementarities across policy reforms. A comprehensive labor market reform attacks those rigidities— one that increases workers’ bargaining power at the same time that it reduces firing costs—is more likely to be successful than one that only tackles the latter aspect. Creative destruction. Firing costs reduce the economy’s average productivity by maintaining a fraction of the workforce in vintages that are older than the most up-to-date technology. In equilibrium, this ends up reducing wages and living standards. Now, this effect will be stronger, the greater the productivity gap between old vintages and new vintages, that is, the greater the growth rate. A higher growth rate consequently reduces the political support for employment protection legislation, because it increases its cost in terms of lower wages. We show that the political support for firing costs typically comes from a fraction of the employed workers: those who work in matches that are not too old, nor too young. In the first case, workers are going to lose their jobs quickly; they are better off unemployed in a flexible society than employed in a rigid society. In the second case (which may be degenerate and reduced to an empty set), workers consider that the end of their job is pretty remote and are not willing to pay the cost of employment regulation. We supplement our analytical reasoning with some evidence suggesting that increases in firing costs tend to occur at times when workers’ weight in bargaining is high, and, conversely, reductions in firing costs take place when bargaining power is low. This is in accordance with our model. EMPLOYMENT PROTECTION IN A RENOVATING ECONOMY Let us consider a world with different vintages of capital.5 At any point in time t there is a state-of-the-art technology that allows production of at units of output with one unit of labor, where at is assumed to grow at a constant exogenous rate g, so that at = a0 e gt. There is free entry of firms (considered as hiring a single worker) in the state-of-the-art technology; but once a firm has entered, it cannot upgrade. It is stuck with the level prevailing at the time of entry. If exit were costless, firms would enter the market for a very small amount of time and then disappear, because competition by new entrants would constantly drive wages up to the state-of-the art technology level, thus making any old plant unprofitable. We assume, however, that exit is costly so that in order to close at time T the firm pays a firing cost in terms of output, equal to FaT .6 We assume that this firing cost is wasted. Firing costs imply that unprofitable plants, instead of closing, will continue until losses become so large that it is actually preferable to pay the firing cost and close the position. By the same token, for new jobs to be created, it must be the case that they run positive profits in the beginning of their lifetime, F E D E R A L R E S E R V E B A N K O F S T. L O U I S 75 5 This is in a fashion somewhat similar to Caballero and Hammour (1994). Another option would be to rule out growth and assume that once in the market, firms suffer shocks to their productivity level, and that they can freely enter the market at the maximum possible level, as in Mortensen and Pissarides (1994) or Saint-Paul (1997). 6 Dependence on the productivity trend implies that the firing cost grows at the same rate as the rest of the economy, and, therefore, does not become negligible relative to the surplus of a match. M AY / J U N E 1 9 9 9 to compensate for the future losses associated with the firing cost. Therefore, wages cannot be as high as the state-of-the art technology level, contrary to what would happen absent exit costs. Firms and workers can freely borrow and lend at the real interest rate r. We assume r > g, which guarantees that the present discounted value of income streams will be well defined. Workers are homogenous and negotiate wages in an imperfectly competitive fashion, thus being able to raise their welfare strictly above their outside opportunity—that is, the welfare of an unemployed worker. Consequently, in equilibrium, there is a positive stock of involuntarily unemployed workers who wait to find a job created by a new entrant. At any time t the net value of a firm that entered the market at date s is equal to the present discounted value of its profits minus the (discounted) firing cost it has to pay upon termination: (1) We now turn to wage determination. Our key assumption is that workers can appropriate a share of the surplus generated by the match gross of the firing cost. Formally, this is equivalent to (5) where ϕ is the share of the gross surplus that the worker is able to appropriate, bat is the income flow of an unemployed worker (for example, the unemployment benefits he is paid), θ t is the probability per unit of time that an unemployed worker finds a job, while (6) − r u− t J ( s, t ) = ∫tT ( s )( a s − w( s, u ))e ( )du where w(s, u) is the bargained wage between firms and workers at date u, which will be determined below. The firm sets its exit time optimally by maximizing expression 1 with respect to T(s). The first order condition is: 7 If the firing cost is large enough, it may be optimal for the firm to never fire the worker. In that case the condition is: (3) w(s, t ) − as < (r − g) Fat ∀t ≥ s 8 For a formal derivation of that wage formation schedule, see Saint-Paul (1998). w( s, T ( s)) − as = (r − g) FaT ( s ) . The left-hand side is the loss per period made by the firm while the right-hand side is the annuity value of the firing cost. Note that faster growth reduces the annuity value of the firing cost: As they are indexed on the economy’s growth trend, postponing dismissal increases the value of the firing cost. This effect reduces the opportunity cost of firing today.7 Finally, the free-entry condition implies that the net value of the firm is zero at the time it enters the market: (4) S(t ) = ∫ tT ( t ) at e − r ( u − t ) du is simply the present discounted value of the firm’s gross output. The meaning of equation 5 is as follows. The last term is the fraction of the match’s output appropriated by the worker. The first two terms are the worker’s “alternative wage,” or outside option, that is, the wage that would make him just indifferent between being unemployed and working for that firm. The first term represents the unemployed’s flow of income, while the second term represents the contribution to his welfare of the future rents he will appropriate from his next jobs. It is larger, the greater the probability of finding a job and the larger the share of the surplus appropriated by the worker.8 The probability of finding a job, θ t , is the key endogenous variable that determines the adjustment of the labor market. Its equilibrium value is determined by the free-entry condition, equation 4, that requires that the net value of a newborn firm be equal to zero. If the labor market were too tight, relative to that equilibrium value of θ , wages would be too high and new firms’ net value would be negative. Consequently, firms would not enter the market, which would reduce the jobfinding probability and push wages downwards to the point where the free-entry condition is met again. −Fa T ( s )e − r( T ( s )− t ) , (2) w( s, t ) = bat + θtϕ S(t ) + ϕ as , J (t, t ) = 0. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 76 M AY / J U N E 1 9 9 9 EQUILIBRIUM Figure 1 Equilibrium Determination We are now in a position to compute the equilibrium of our economy. To do so, we limit ourselves to a “steady state,” that is, a balanced growth path where wages and output grow at rate g, while unemployment, labor market tightness θ, and the duration of a job are constant. Let us call that constant duration x. To do so, we proceed as follows. First, note that the firm’s present discounted output, in steady state, is simply equal to u W P W P x (7) S(t ) = at 1− e r − rx . The term (1– e –rx) /r is simply the present discounted equivalent of a constant, unit flow of income over a time interval of length x. Substituting equation 7 into the wage equation 5 we get that (8) The left-hand side is the present discounted value of profits, gross of the firing cost,9 and the right-hand side is equal to the firing cost. The equation tells us that under free entry the cumulated profits must exactly cover the firing cost. As long as F < (1 − ϕ ) / r, equation 10 defines a unique equilibrium value of x. If F is greater than (1 − ϕ ) / r , then it is optimal for firms never to close, a case we rule out by assumption. The equilibrium is determined in Figure 1 by the intersection of two schedules, a downward sloping schedule WW defined by equation 9, and a vertical one PP defined by equation 10. The equilibrium has the following properties: • An increase in ϕ , the workers’ bargaining power, shifts PP to the right and WW downward (Figure 2a). Consequently, the duration of matches increases and labor market tightness declines. An increase in ϕ directly increases labor costs, which reduces incentives for job creation but makes it affordable to close later. At the same time, profits fall, so the job must last longer in equilibrium in order for cumulated profits to cover the firing cost. • An increase in F, the firing cost, shifts PP to the right and WW 1 − e − rx r − g(t − s ) ), +ϕe w (s, t ) = at (b + θϕ where, as previously, the first two terms represent the alternative wage and the last term the rent earned on one’s current job. Using that condition, we can rewrite the optimal closing condition, equation 2, as (9) 1 − e − rx r = ( r − g ) F. (ϕ − 1)e − gx + b + θϕ This is a first equation that gives us a relationship between x and θ . This is a decreasing equilibrium relationship that tells us that a tighter labor market pushes wages up, thus forcing firms to close at an earlier time. Next, using the free-entry condition 4 along with the wage equation 5 and with equation 7, then making use of equation 9, we get a second equation that determines x: − rx − gx 1 ge − re (10) (1 − ϕ ) + r (r − g) r = F. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 77 9 They are discounted at the closing date of the firm and expressed in productivity units at that time. M AY / J U N E 1 9 9 9 upward, (Figure 2b). A higher firing cost makes it optimal to postpone the closing time, and jobs must last longer for cumulated profits to cover the firing cost. Despite the upward shift in WW, θ unambiguously falls: Labor market tightness is reduced as increased firing costs discourage hirings. The upward shift in WW simply means that at any given job duration, one would require higher wages and therefore tighter labor markets for closing to be optimal. • An increase in g, the growth rate, unambiguously shifts PP to the left while WW shifts down (Figure 2c). Faster growth increases the pace of obsolescence via more rapid wage growth within existing matches, and also because the growth of firing costs is faster, as they are indexed on the economy’s average productivity level. The incentives to fire are therefore increased: Matches are shorter (PP shifts to the left), while the degree of labor market tightness that makes it optimal to fire falls (WW shifts down). The net effect on θ is ambiguous as WW is downward sloping. Next, it is possible to characterize the equilibrium unemployment rate and the steady-state distribution of employment across vintages. In steady state, the density of employment in firms aged z is constant and equal to 1/x. If l is total employment, then the number of jobs destroyed per unit of time is l/x. In steady state, this must be equal to the outflow from unemployment, which is equal to θ (1 – l). This allows us to compute the unemployment rate as a function of θ and x: Figure 2a Impact of an increase in worker's weight in bargaining u W P W P x Figure 2b Impact of an increase in firing costs u P W P W x Figure 2c Impact of an increase in the growth rate u W P (11) W u= 1/ x 1 = . θ + 1 / x 1 + xθ P x It should be noted that unemployment is not necessarily higher when firing costs are lower. A lower firing cost increases θ , F E D E R A L R E S E R V E B A N K O F S T. L O U I S 78 M AY / J U N E 1 9 9 9 which tends to reduce u, but reduces x, which tends to increase u. Job creation is higher but so is job destruction, so unemployment may either rise or decline. This is well known from the analysis of firing costs.10 One can further compute aggregate output in steady state. It is simply equal to the product of employment and average productivity. Given that employment is uniformly distributed over all vintages, the latter is simply equal to assume that there are only two alternatives, reflecting the fact that there is some indivisibility in the design of legislation and that political agendas are often formulated in a binary fashion. We will typically consider that the status quo is the “rigid society,” so that initially workers are distributed over plants aged between 0 and x R . We also assume, for simplicity, that when we do comparative statics the underlying values of the firing costs FR and FF are altered so as to maintain the two options invariant in terms of plant duration. at ∫0x e − gu du 1 − e − gx = at . x gx The Shape of Preferences for Employment Protection The first step is to compute the utility of the employed and unemployed voters as a function of the collectively decided firing cost. The utility of the unemployed is given simply by the present discounted value of the alternative wage, that is, of the first two terms in 8. Its value is At any point in time t output is therefore equal to yt = at θ 1 − e − gx . 1 + xθ g Note that in the extreme case, where the firing cost goes to zero, so does x, while θ goes to infinity. The labor market converges to a situation where both the job creation rate and the job destruction rate are infinite. As technology changes continuously and is embodied into new vintages, it is optimal to close firms an instant after they have been created. As a result, people move constantly between employment and unemployment: It is as if unemployment were equally shared among the workforce. Furthermore, because of free entry, the wage is always equal to at, the state-of-the art productivity, and is higher than it would be for any positive level of the firing cost. With that discussion, we conclude the characterization of equilibrium. We now proceed and discuss voting on firing costs. Vu ( t ; x ) = a 0e gt r−g 1 − e −rx × b + θϕ r . As future alternative wages grow at rate− g , this stream of income is discounted at rate r − g . Substituting in the equilibrium conditions, equations 9 and 10, we see that this is equivalent to (12) Vu ( t ; x ) = a 0e gt (1 − ϕ ) r−g 1 − e −rx . × 1 − g r VOTING ON FIRING COSTS This formula allows us to express the utility of the employed as a function solely of the equilibrium value of plant’s lifetime, x. As for the utility of employed workers, it depends on which plant they are working at. The older the plant, the lower the time left for reaping their rent, and the In the sequel, we will assume that society votes once and for all between two alternatives: a “rigid” society associated with a firing cost FR and a plant life x R , and a “flexible” society associated with FF < FR and x F < x R . Therefore, we F E D E R A L R E S E R V E B A N K O F S T. L O U I S 79 10Bentolila and Bertola (1990). M AY / J U N E 1 9 9 9 lower their utility. More precisely, an employed’s utility is the sum of an unemployed’s utility and the present discounted value of the employed’s share of the gross surplus between now and the closing time. The corresponding value is highest probability of finding a job and the highest wage, it is the one preferred by the unemployed. Turning now to the employed, equation 13 implies that their utility is always decreasing with z and increasing with x if and only if 1 − e − rx 1 − ϕ (13) Ve (z , t; x ) = a 0e gt 1 − g r r − g −r x −z 1−e ( ) ; z ≤ x. +ϕ e − gz r (14) e( r − g) z > g(1 − ϕ ) . ϕ (r − g) For any given x, workers in older firms have a lower utility than workers in younger firms, as they expect their rent from employment to be exhausted earlier. The marginal gain from increasing firing costs is larger, the older the vintage where the worker is working. This is because the remaining duration of their job increases more, in proportional terms, than those of workers at younger plants. Consequently, a marginal increase in firing costs would be supported by those workers whose vintage is greater than a critical z*, defined by The last term in the brackets represents the present discounted value of the rent to be earned until the current job elapses. It is larger, the larger the voted value of x, and smaller, the larger the current age of the job z. It is important to note that equation 13 is only valid if x ≥ z . Once people have voted on x, all firms with age z > x, if any, instantaneously disappear and fire their workers. Therefore, the utility of any worker in a plant older than x is by definition equal to the utility of an unemployed: g(1 − ϕ ) ln ϕ (r − g) z* = . r−g Ve ( z, t; x ) = Vu (t, x ), z ≤ x . This property tells us that in some sense workers at older plants like firing costs better, but it should be remembered that we actually rule out voting on a marginal increase in firing costs as we only consider two alternatives. Note that if In equations 12 and 13, F is treated as a function of x as defined by equation 10; that is, voting on F or voting on x are equivalent given the relationship between the two that must hold in equilibrium. In the sequel, we find it easier to consider that workers actually vote between two values of x. How do, now, the preferences of the people for firing costs depend on their labor market status? Beginning with the unemployed, equation 12 clearly implies that their utility is strictly decreasing with x. The unemployed prefer the lowest possible value of x (or, alternatively, F). In the F = 0 equilibrium, people move constantly between employment and unemployment so that it is as if the total amount of work were shared perfectly among people. The incumbent employee’s advantage for tomorrow’s jobs is eliminated; as this equilibrium yields the (15) ϕ > g / r, then the numerator is negative (or equivalently, the right-hand side of 14 is lower than 1), which implies that all employed workers benefit from a marginal increase in firing costs. In the sequel, we shall assume that condition 15 holds, and discuss the case ϕ > g / r later. Figure 3 illustrates how preferences depend on x for various types of workers in the case where ϕ > g / r . The down- F E D E R A L R E S E R V E B A N K O F S T. L O U I S 80 M AY / J U N E 1 9 9 9 ward sloping curve Cu represents the preferences of an unemployed worker. C0 represents the preferences of a worker at a newly created plant. Cz represents the utility of a worker at a plant of age z for x > z. It is important to note that for x < z, his utility is given by Cu . Cz ′ represents the utility of a worker with z ′ > z . As z increases, Cz shifts down and its slope shifts up. While people employed at old plants enjoy firing costs more at the margin than people working at new plants, they also are increasingly unhappy as the age of their plant increases. Figure 3 Preferences according to plant age and employment status C0 Cz C z’ Cu Voting Between Two Values of the Firing Cost z We now turn to the question of who will favor flexibility and who will oppose it when people choose between x R and x F. Figure 3 is a useful starting point. As the unemployed’s utility is monotonically decreasing in x, they clearly support the lowest value of the firing cost. What about the employed? They typically split into two groups, as illustrated in Figure 4. There exists a critical plant age z̃ such that workers in plants older than z̃ favor flexibility while workers below z̃ favor rigidity. In the rigid society, workers at plants of age z̃ get exactly the same utility as an unemployed worker of the flexible society. Workers of the first group are in a match that is about to expire. They have consumed most of their rent and expect to be soon unemployed and to suffer from the low job creation rate and the low productivity of the economy. They would be better off either with an increase in firing cost beyond x R, but such an increase is not on the political agenda, or with a decrease in firing costs. Thus, if the status quo is the rigid society, they end up voting for the flexible one. The reason why this “lost generation” prefers flexibility is that they will soon be constrained to a “new start” anyway, and the flexible society is the one that gives them the best chances. If the status quo is the flexible society, this group would not exist since z̃ is always greater than x R. z’ x Figure 4 Interest groups among the employed, >g/r ϕ C0 C ~z Group 1 Cu Group 2 xF ~ z xR Workers such that x F < x < z˜ prefer to maintain the rigid society: They will lose their jobs if the economy were to shift to flexibility, and their jobs will last long enough to make rigidity worthwhile for them. Workers such that x < x F will not lose their jobs if the economy becomes flexible. They prefer the rigid society because it increases the length of time over which they reap their rent, while the prospects of job loss is too remote for them to worry about the low rate of job creation. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 81 x M AY / J U N E 1 9 9 9 Effects of Growth and Workers’ Bargaining Power Their net loss from rigidity is decreasing with ϕ . As ϕ increases, other labor market rigidities are more important relative to firing costs in reducing the unemployed’s job prospects so that their welfare loss between the high and low firing cost societies is actually reduced. Finally, those workers who would lose their jobs if the economy were to shift from rigidity to flexibility have a gain from flexibility equal to How do the parameters of the model affect the outcome? The two parameters in which we are most interested are ϕ , the workers’ bargaining power, and g, the growth rate. To analyze the effect of these parameters, it is useful to distinguish between three groups: those who work at a vintage young enough (z < xF) so that they would be employed in both the rigid and the flexible world; the unemployed; and those who are employed in the rigid world but would lose their job if society decided to become flexible (z > xF). Let us start with an increase in the workers’ bargaining power ϕ. In general, for a given F, a change in ϕ affects x. Now, for simplicity, we assume that the two alternatives are specified in terms of x rather than F. That is, we assume that the two job lengths x R and x F do not change. As equation 13 shows, an increase in ϕ reduces the first term in brackets—the value of being unemployed—but increases the second term—the rent. As rents are higher, incumbent employees are more in favor of employment protection. Consider workers that would be employed in both worlds, that is, such that z < x F . Their net gain from being in the rigid world instead of the flexible one is Ge′ ( z, t; x ) = Ve ( z, t; x R ) − Vu ( z, t; x F ) e − rx F − e − rx R 1 − ϕ g = a 0e gt − r−g r r − r( x R − z ) 1−e + ϕ e − gx r ( ) = Vu (z , t; x R ) − Vu (z , t; x F ) ( + Ve (z , t; x R ) − Vu (z , t; x R ). This is the sum of the (negative) welfare gain of the unemployed, which, as we have seen, increases with ϕ and the employee’s rents in the rigid economy, which also clearly increase with ϕ . Thus, these people gain more, or lose less, from rigidity when ϕ increases. Therefore, in an economy with a high value of ϕ —powerful employees—a given individual, whether working in a plant of any age z or unemployed, will always be more in favor of rigidity than in a world were ϕ is low. Is it obvious, then, that the political support for the rigid society is greater? The answer is no. For it is also true that unemployment is higher when ϕ is large, which tends to push up the number of people who oppose rigidity, even though these people lose less from rigidity than if ϕ were small. What is clear, however, is that within the employed, the support for rigidity increases, meaning that the critical plant age increases. ( z̃ must satisfy Ge′ ( z˜ , t; x ) = 0 , and that function is decreasing with z and shifts up when ϕ increases.) If labor market institutions were mostly determined by the employed, say because they are Ge ( z, t; x ) = Ve ( z, t; x R ) − Ve ( z, t; x F ) e − rx F − e − rx R = r 1−ϕ g gt × ϕ e ( r − g )z − a 0e . r −g r This is clearly increasing with ϕ . As for the unemployed, their welfare difference between the two economies is Gu (t; x ) = Vu ( z, t; x R ) − Vu ( z, t; x F ) e − rx F − e − rx R 1 − ϕ g gt = − r − g r a0 e . r F E D E R A L R E S E R V E B A N K O F S T. L O U I S 82 M AY / J U N E 1 9 9 9 better organized collectively or because the unemployed have a low rate of participation in elections, then the political support for rigidity would unambiguously increase when ϕ rises. Another way to put it is to say that controlling for the unemployment rate, the high firing cost is more likely to be chosen when ϕ is higher. What happens, now, when the growth rate is larger? If we compare the high growth economy and its low growth counterpart at two points in time when they have the same technological level, which amounts to holding a0egt constant in our comparisons, the above formulae imply that an increase in g reduces Ge, Gu, and Ge′ . Consequently, faster growth unambiguously reduces the political support for employment protection if we hold the stock of unemployment constant. The growth rate acts in two ways. First, it increases the obsolescence rate and, therefore, the deadweight cost of maintaining relatively unproductive matches idle. This in itself reduces the support for employment protection. Second, faster growth tends to reduce the effective discount rate applied to the future: Incumbent workers put more weight on the lower job finding rate they will experience once their current match is dissolved, because future jobs pay more. This also tends to reduce the support for employment protection. Figure 5 Three groups in the ϕ <g/r case C 0 C z* C~ z Group 1 Group 2 z* x F Group 3 ~z x R plants older than z̃ are about to lose their jobs and oppose rigidity for the reasons already explained. Those who work in plants between z* and z̃ gain more from rigidity than what they lose. Thus, among the employed, labor market reform (in the sense of a deregulation) would be supported by an “extreme coalition” of people working in either the most dynamic plants or plants that soon will become obsolete. Another property of this case is that the maximum welfare point is actually attained at x = 0. Since workers at young plants are always happier, given x, than workers at old plants, and since those at plants just created (z = 0) have a utility which is decreasing with x, there would be unanimity in favor of a zero firing cost, if this is a feasible outcome. Conversely, if the status quo is x = 0 then all employed workers work in plants of age z = 0. A necessary condition for zero firing costs to remain a political equilibrium is therefore that workers at plants with z = 0 would be worse off if firing costs were higher, which given equation 14, is precisely equivalent to ϕ < g / r . Therefore, for a “flexible” society to be stable (in the sense that people will not want to change its institutions), it must be the case that the worker’s share The ϕ < g/r Case What happens now if ϕ < g / r ? Equation 14 implies that workers with sufficiently small z will have a utility strictly decreasing with x. As illustrated on Figure 5, the flexible society is preferred by a group of workers who work in the most recent plants. These workers lose more, in terms of lower wages, than they gain in terms of a postponed dismissal. There are now three interest groups among the employed. Those who work at plants younger than z* have a utility that is decreasing with the firing cost, so that they will always prefer the flexible economy. Those who work at F E D E R A L R E S E R V E B A N K O F S T. L O U I S 83 x M AY / J U N E 1 9 9 9 of the surplus does not exceed the ratio between the growth rate and the interest rate. This simple formula (which we are tempted to label the “golden rule of flexibility”) is a useful shortcut for thinking about the determinants of the political support for flexibility. approach would be to take the share of labor in national income. This share, however, typically reflects other phenomena as well, such as variations in factor inputs as a reaction to changes in factor prices. Bentolila and Saint-Paul (1998) show that such movements are associated with a relationship between the labor share and the capital/output ratio. To proxy for workers’ bargaining power, we just take the residual of a first-difference regression of the labor share on the capital/output ratio.11 Our results are conditional on the validity of that proxy, which clearly may be questioned. The capital/output ratio filters out many sources of movements of the labor share unrelated to bargaining power, but other sources remain.12 One should also keep in mind that our measure is positively related to the economic cycle, so that correlation between that measure and the timing of reforms may also capture other mechanisms. The results we present, therefore, should be interpreted with caution. Figures 6 through 10 represent the evolution of our measure of the workers’ bargaining power for the five largest European countries. These figures are not comparable across countries and the initial value cannot be interpreted. Only the evolution within each country is meaningful. The evolution of our measure is somewhat related to the reforms that actually took place.13 For example, in Spain, our measure dropped sharply between 1978 and 1984, suggesting the opening of a “window of opportunity” for reducing firing costs in 1984. It is precisely that year that a major reform was introduced with the liberalization of the use of temporary contracts. Prior to that reform temporary contracts mostly were restricted to work of temporary nature, as in many other European countries, and temporary contracts only accounted for 10 percent of the workforce. In 1984, however, the government made it possible to use those contracts over a wide range of circumstances. This amounted to a substantial reduction of firing costs as employers could simply hire a worker on a temporary contract ASSESSING THE VIABILITY OF LABOR MARKET REFORMS: DO GOVERNMENTS REDUCE FIRING COSTS WHEN BARGAINING POWER IS LOW? 11See Bentolila and Saint-Paul (1998) for more details. 12Again, we refer the reader to Bentolila and Saint-Paul (1998) for a detailed analysis. 13See Saint-Paul (1996b) for a discussion of these reforms and their determinants. In that paper, I actually ignore the role of workers’ rent, focusing on the employed’s exposure to unemployment and on the number of unemployed workers as factors important for the viability of reform. The above analysis suggests that there are three important determinants of reform, namely the workers’ bargaining power, the growth rate, and the interest rate. It is, therefore, tempting to take these predictions to the data and see how they square with reality. Now the question arises of how literally one can interpret our results. For example, the importance of the growth rate, in our analysis, captures the role of the rate of renovation of old plants in the long run. By contrast, macroeconomic data on growth mostly capture cyclical fluctuations and changes in the underlying trend of productivity that may not be associated with changes in creative destruction. The real interest rate may not play a big role if incumbent employees have only an imperfect access to capital markets. Furthermore, what is relevant for people’s voting behavior is not the current level of the growth rate and the interest rate, but the whole path that they are expected to follow in the future. For that reason, we prefer to focus on perhaps the most robust prediction of the model, that is, the positive relationship between the workers’ bargaining power and the support for employment protection. Our strategy is to construct a time series for that bargaining power for a selection of European countries and see if it bears a relationship with the timing of reforms. How can one construct a proxy for workers’ share in bargaining? One simple F E D E R A L R E S E R V E B A N K O F S T. L O U I S 84 M AY / J U N E 1 9 9 9 Figure 6 and fail to renew that contract when it expired if they wanted to get rid of the worker. The graph for Spain tells us that this reform came into effect at a time when the rent of the employed had substantially declined from the peak it had reached in the mid-seventies, so that the resistance of the insiders to such a reduction in firing costs was considerably lower than if one had attempted to implement it in, say, 1980. In the United Kingdom, the fall in workers’ bargaining power apparently occurred earlier than in Spain, so that the window of opportunity began in the late seventies/early eighties. Again, this squares with our theory, because this interval coincides with the rise to power of a conservative government, who subsequently engaged in comprehensive labor market reform, including a reduction of firing costs. Note that despite these reforms, workers’ bargaining power seems to go up again thereafter; this captures the high wage inflation of the second half of the eighties, but there was no reversal of the reforms. In France, the decline of workers’ bargaining power occurs somewhat later than in Spain and the U.K.; but again, the opening of the window of opportunity, 1986, coincides with the rise to power of a conservative government and a reduction in firing costs—namely, the suppression of the compulsory administrative approval for layoffs, which was established in 1974 (at a time of rising bargaining power but before it reached its peak). Our proxy, on the other hand, fails to account for an increase in firing costs that was implemented in 1989 when the Left returned to power. Reforms that reduce firing costs have been much milder in Germany than in Spain, perhaps reflecting a society that needs greater consensus to move ahead and is, therefore, more likely to stay where it is. Nevertheless, the timing of the reform matches our analysis well. As in Spain, temporary contracts were liberalized in 1984 (although this was much more timid than in Spain), after a sharp drop of our estimated workers’ Worker's Bargaining Power, Spain 0.08 0.06 0.04 0.02 0.00 -0.02 72 74 76 78 80 82 84 86 88 90 92 94 92 94 Figure 7 Worker's Bargaining Power, U.K. 0.04 0.02 0.00 -0.02 -0.04 -0.06 -0.08 -0.10 72 74 76 78 80 82 84 86 88 bargaining power. Of all the countries we deal with, Italy is the one most characterized by “stop-and-go” policies. Reductions in firing costs alternate frequently with increases in firing costs. For that reason, one should not expect our proxy to work too well. But, in fact, it does a reasonable job at explaining the twists of policy. Firing costs were reduced in 1977, 1984, F E D E R A L R E S E R V E B A N K O F S T. L O U I S 85 90 M AY / J U N E 1 9 9 9 1986, and 1987, following drops in our measure of the bargaining power. They were increased in 1989 and 1990, at times when the employed’s rent appears to be high. Finally, there was a further reduction in firing costs in 1991, a move that our proxy clearly fails to predict. Obviously, this evidence is only indicative and leaves a lot of room for qualifications, alternative interpretations, and further research. However, it is suggestive that one can actually identify some regularities in the timing of labor market reforms. Figure 8 Worker's Bargaining Power, France 0.08 0.06 0.04 0.02 0.00 -0.02 -0.04 -0.06 72 74 76 78 80 82 84 86 88 90 92 CONCLUSION 94 In this paper, we have studied the circumstances under which there will be sufficient support for a high level of employment protection. We have argued that two key determinants of such support are the employed’s share in bargaining and the rate of growth of the economy. The political viability of a reduction in firing costs is highest for low levels of the employee’s share and or high growth rates. The prediction of our model bears some resemblance to the real world’s experience, although one cannot hope for sharp empirical tests when dealing with political-economy models. Figure 9 Worker's Bargaining Power, Germany 0.06 0.04 0.02 0.00 -0.02 -0.04 72 74 76 78 80 82 84 86 88 90 92 94 REFERENCES Bentolila, Samuel and Giuseppe Bertola. “Firing Costs and Labour Demand: How Bad is Eurosclerosis?” Review of Economic Studies (July 1990), pp. 381-402. Figure 10 __________, and Gilles Saint-Paul. “Explaining Movements in the Labor Share,” Centre for Economic Policy Research, Discussion Paper No. 1958, September 1998. Worker's Bargaining Power, Italy 0.08 Blanchard, Olivier, and Pedro Portugal “What Hides Behind an Unemployment Rate: Comparing of Portuguese and U.S. Unemployment,” mimeo, MIT, June 1998. 0.04 Caballero, Ricardo and Mohamad Hammour. “The Cleansing Effect of Recessions,” American Economic Review (December 1994), pp. 1350-68. 0.00 Cohen, Daniel, Arnaud Lefranc, and Gilles Saint-Paul. “French Unemployment: A Transatlantic Perspective,” Economic Policy: A European Forum (October 1997), pp. 265-85. -0.04 -0.08 72 74 76 78 80 82 84 86 88 90 92 Mortensen, Dale and Christopher Pissarides. “Job Creation and Job Destruction in the Theory of Unemployment,” Review of Economic Studies (July 1994), pp. 397-415. 94 F E D E R A L R E S E R V E B A N K O F S T. L O U I S 86 M AY / J U N E 1 9 9 9 OECD. The OECD Jobs Study, Paris: OECD, January 1995. Saint-Paul, Gilles. “On the Political Economy of Labor Market Flexibility,” NBER Macroeconomics Annual 1993, Cambridge: MIT Press, 1993, pp. 151-87. __________. “Labour Market Institutions and the Cohesion of the Middle Class,” International Tax and Public Finance (July 1996a), pp. 385-95. __________. “Exploring the Political Economy of Labour Market Institutions,” Economic Policy: A European Forum (October 1996b), pp. 263-300. __________. “The Rise and Persistence of Rigidities,” American Economic Review (May 1997), pp. 290-94. __________. “The Political Economy of Firing Costs,” mimeo, Universitat Pompeu Fabra, Barcelona, October 1998. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 87 MARCH/APRIL 1999 F E D E R A L R E S E R V E B A N K O F S T. L O U I S 88 M AY / J U N E 1 9 9 9 Christopher J. Waller holds the Carol Martin Gatton Chair of Macroeconomics and Monetary Economics at the University of Kentucky. Commentary since this is cheaper than firing the workers and incurring the firing cost. Thus, not surprisingly, Saint-Paul finds that firing costs lengthen job matches and reduce the probability that the unemployed workers find jobs. The firing cost has three key effects on workers’ lifetime utility: Christopher J. Waller W ith the advent of the European Monetary Union (EMU), members will lose one policy tool (the exchange rate) for dealing with asymmetric shocks to their economies. Consequently, some other macroeconomic variable or market must become more flexible to eliminate excess demand or supply of output and employment. It is generally argued that labor markets must become more flexible in Europe to compensate for the loss of independent monetary policies as a stabilization tool. It is generally believed that firing costs in Europe are much higher than in the United States, hence, to make labor markets more flexible, firing costs need to be lowered. Firing costs are affected by labor legislation, and therefore, politics enter the picture when discussing whether or not to lower firing costs. While lowering firing costs will benefit unemployed workers, firing costs generate rents for employed workers and it is unlikely that workers would give up those rents willingly. Therefore, to study the political viability of policies aimed at lowering firing costs, we need an economic model that reflects the benefits and costs of firing costs in a dynamic model of employment. This is the task undertaken by Gilles Saint-Paul in his paper. Saint-Paul uses a pseudo-search/ matching model with exogenous productivity growth to study how firing costs affect the length of job matches. A worker’s productivity is constant during a match, although average productivity is rising in the economy. Upon being fired, a worker instantly acquires the average level of productivity in the economy. Firing costs generate “rents” for workers that must be paid out of firm profits: After some critical date, firms subsidize workers whose wage exceeds their productivity, • It lengthens the period they collect the rent (improves utility). • It lowers the average productivity, and thus, lifetime income (reduces utility). • It worsens the probability of finding a job when unemployed (reduces utility). Having ascertained the benefits and costs to workers from the existence of firing costs, Saint-Paul then asks, “Who would favor an increase in the magnitude of firing costs?” He considers a majority-rule voting equilibrium to see how individual workers would vote between two possible firing cost policies to see if a majority of workers would support changing the firing cost from its current level. The author then examines who would prefer moving towards a higher firing cost from the low firing cost regime and vice versa. It is clear that unemployed workers always favor lowering firing costs, since this would shorten job matches and increase their probability of finding a job. What about employed workers? They typically divide into three groups based on the relative age of their job. Because the rent is acquired towards the end of a match, the present discounted value of the rent in the future is of less value to workers in “young” matches. In general, the last two effects dominate for these workers so they oppose a policy that increases firing costs. Workers in old matches are collecting the largest rent so they benefit the most from the first effect. Because they are near F E D E R A L R E S E R V E B A N K O F S T. L O U I S 89 M AY / J U N E 1 9 9 9 the end of their current match, however, they worry the most about finding a job since they soon will be unemployed. This latter effect is dominant for them. Since the workers in relatively young and relatively old matches support lowering firing costs, who is left to support the policy? Well, the only workers left are those in middle-age matches. The middleage workers are the ones just entering the rent-collecting phase so reducing firing costs significantly reduces the rents they are about to collect. Furthermore, their match is still sufficiently young to make the future cost of lower job-finding probability of little importance. Consequently, this is the group that would actually support increasing firing costs. Saint-Paul shows that if high firing costs are the status quo, most of the cases correspond to a majority of workers favoring a move to lower firing costs. On the other hand, if a low-firing cost economy is the status quo, then the outcome is not so clear. Under some parameterizations of the model, workers would be in favor of moving to a more rigid economy, while under other parameterizations they will be opposed to it. Finally, they may be divided over whether or not to move to a more rigid economy. As a friend of mine is fond of saying, the value of discussing a paper is that you have to dig into the assumptions of the model and that is where all of the bodies are buried. Although all models have bodies buried in them, with Saint-Paul’s paper I often felt I was on the trail of a serial killer. So where are the bodies buried in this model? My first question is—What are “firms” in this model? Who owns them and does it matter? In perfect competition with no fixed costs or capital, the wage bill is equal to output, so workers extract the entire surplus from production. This is a common view of firms in all constant-returns-to-scale models of production—input payments exhaust the output. Hence, firms are merely a veil. With firing costs, however, the firms receive some surplus early in the match and then subsidize workers later in the match. What do firms do with the surplus and how do they pay for the subsidy that occurs later in the match? If they save it to finance the subsidy later, then there should be an intertemporal financing constraint on the firms, much like a firm faces with pension fund obligations. An intertemporal financing condition is missing in the model. If firms save the surplus early in the match and use it to pay the rent to workers later in the match, then workers (in effect) are paying for the firing costs via smaller shares of the output earlier in the match—something Saint-Paul was trying to avoid. If firms do not save their share of the surplus, what do they do with it? Do the owners of the firm consume it? Is it transferred to workers via lump-sum equiproportionate transfers? If matches were perfectly deterministic in length, then an intertemporal compensation scheme would require firms have zero surpluses on net. In this model, however, some matches end early for random reasons. These firms clearly end up with net surpluses from the match. What do they do with them? The model is completely silent on this entire issue. If firm owners receive some of the surplus from trade, then they too will have a stake, and presumably a vote, in any referendum on firing cost. But alas, in this model, only workers vote—owners of firms do not. Why do workers prefer this type of compensation scheme to some other? The author assumes that financial markets are perfect. This implies that workers can borrow and lend to achieve their desired consumption path over time regardless of the timing of income receipts. In this case, workers simply want the highest lifetime present discounted value of income, which occurs in a perfectly competitive labor market. Thus, why would workers want to distort markets by voting for firing costs? Another problem I have is with the voting analysis in the model. Since voting is over a single issue (firing costs), then the median voter model should work well. Simply determine the magnitude of the firing costs that maximizes the utility of F E D E R A L R E S E R V E B A N K O F S T. L O U I S 90 M AY / J U N E 1 9 9 9 each worker by match age and take the median value of those maximums. The only remaining question is whether the median voter’s preferred firing cost is zero. Unfortunately, in this model, preferences over firing costs are not single-peaked. The reason for this is that a vote by the majority to lower firing costs would cause workers in old matches to be fired immediately. Being fired lowers utility due to job loss but also causes an immediate increase in productivity that raises the probability of finding a job and, therefore, increases lifetime utility. Hence, there is a certain age of a match in which these two forces offset each other and workers are indifferent to lowering firing costs and leaving them unchanged. Consequently, preferences over firing-cost policies are not single-peaked, which creates multiplicity of voting equilibria and possibly Condorcet voting cycles.1 This greatly complicates the analysis of the voting equilibrium and requires numerical analysis to study the problem. There is an easy way around this voting complexity—grandfathering. The source of the voting complexity arises from the threat that older workers will be fired if firing costs are lowered. To eliminate this, simply grandfather all current matches against the change in firing costs and only apply the lower firing costs to new matches. Grandfathering current voters from the undesirable consequences of changed policies is an age-old method of pushing through socially desirable policies. As an example, a university I was once associated with wanted to lower its faculty contributions to TIAA-CREF to reduce generous labor benefits. Not surprisingly, this proposal was a nonstarter with the faculty until the administration grandfathered the current faculty from the benefit cuts and imposed the cuts on new hires from a certain date onward. This policy was supported by the faculty and implemented. The moral of the story is that grandfathering dramatically alters people’s voting behavior and can greatly simplify the voting outcome. Another problem I have with the model is that firms supposedly have free entry and do not face search frictions in finding workers, whereas workers do face some search frictions in finding firms. So, I am puzzled why anyone is unemployed in equilibrium since you need both sides to face frictions. Unemployment implies workers can’t find firms and vice versa. But if firms face no search frictions, then as soon as a match ends, a competitive firm should swoop in and instantly hire the worker. In fact, the current firm should simply rehire the worker instantly, since productivity increases occur instantly and costlessly upon separation.2 If firms do face search frictions, then it would seem that the free-entry condition does not produce a zero-valued unmatched firm and the solutions to the model are thus incorrect. In summary, I conclude that the author is working on an interesting and important labor problem, particularly as it applies to Europe. He also has adopted an interesting model for studying the issue and has obtained some interesting and plausible results for thinking about how voters will line up either in favor of or in opposition to changing labor laws. Unfortunately, for my tastes, there are some additional bodies buried in the model’s intellectual basement that need to be exhumed before I believe the author has accurately captured the essence of the problem. 1 Condorcet cycles arise when preferences over outcomes are intransitive, i.e., A is preferred to B, B is preferred to C, but C is preferred to A. 2 F E D E R A L R E S E R V E B A N K O F S T. L O U I S 91 A similar problem occurs in monetary search models with barter trade; pairs of traders who have a double-coincidence of wants meet and trade but then separate. It seems irrational to separate once traders pair up in successful matches but this is the typical assumption of money search models and, after reading this paper, apparently also is typical of labor search models. M AY / J U N E 1 9 9 9 F E D E R A L R E S E R V E B A N K O F S T. L O U I S 92 M AY / J U N E 1 9 9 9 Erica L. Groshen heads the domestic research function at the Federal Reserve Bank of New York and Mark E. Schweitzer is an economist in the research department of the Federal Reserve Bank of Cleveland. The authors appreciate the helpful comments of John Haltiwanger, Kenneth Troske, Joseph Ritter and conference participants at the St. Louis Fed and the Society for Labor Economists. They also thank Amanda Moses and Jennifer Ransom for excellent research assistance. Firms’ Wage Adjustments: A Break from the Past? (a period noted for both low inflation and unemployment rates) differed from historical patterns. Another interesting question is whether some subset of jobs tends to react first to inflationary or deflationary stimuli. For our investigation of these questions, we examine a long (39-year) time series of wages for a panel of mobile occupations for a set of employers in three Midwestern cities. We study wage changes during years with rising, falling, and steady inflation to identify regularities that could broaden our understanding of the inflationary process at the micro level. Inflation (as measured by changes in the Consumer Price Index) and nominal wage growth (as measured in the means of the data set we study, as well as in national series) are largely co-timed. In this paper, we treat wage changes as caused by inflation. This approach does not reflect a stand on whether inflation is primarily a price-pull or cost-push phenomenon. Rather, this perspective reflects the experience of inflation from the individual worker or firm’s point of view. That is, our approach is consistent with how human resource managers (the agents who propose and justify pay increases in most large U.S. firms) describe their salary-adjustment policies. Personnel managers typically report that they use local cost-of-living increases and the wages paid by other employers to guide their wage adjustments. Though potentially compatible with many economic theories of wage adjustment (including firms’ price-taking in labor markets), these policies suggest that wage changes react to inflation instead of driving it. At a macroeconomic level, the managers’ policies should tend to tie pay increases to inflation and productivity growth on a lagging or contemporaneous basis. The paper proceeds as follows. First we describe the wage-setting process in large firms and discuss the reasons why wage change distributions may not be neutral with respect to inflation. Then we Erica L. Groshen and Mark E. Schweitzer D espite advances in understanding the policies that cause inflation, economists know little about inflation’s manifestations and transmission in the marketplace. For example, how does inflation affect wages in an economy composed of heterogeneous agents making individual optimizing decisions? We know that there is a wide dispersion of wage changes in any year (Groshen and Schweitzer 1999). In this paper we ask whether inflation and its changes alter the distribution of wage shocks—rather than being neutral for the distribution as conventional theories of wage adjustment would suggest. Distributional effects on wage changes have been the subject of conjecture by academic, policy, and business economists, but rarely the subject of systematic inquiry. Altered distributions in the presence of inflation would indicate that simple wage models—i.e., ones based on representative or aggregate agents—are inadequate to describe the complexity of wage determination. Initially, characterizing the nature of this complexity allows us to identify the variety of labor-market responses to shocks. From there, we can develop and evaluate richer models of the wage-setting process. Insights into the distribution of wage changes should also be helpful for monitoring the economy. For example, one question of particular current interest is whether the wage-setting process during the 1990s F E D E R A L R E S E R V E B A N K O F S T. L O U I S 93 M AY / J U N E 1 9 9 9 describe the data. The fourth section describes our main results on the distributional effects of inflation. To test for robustness, we also consider the impact of unemployment and changes in returns to education on wage-change distributions. The fifth section investigates two policy-relevant questions: whether some jobs tend to be the first to respond to changes in inflation, and whether wage changes in the 1990s have deviated from historical patterns. The sixth section summarizes and concludes. and product prices.2 Management aims to maintain the company’s profitability by not over- or underpaying employees to prevent both excessively high labor costs and unwanted turnover. Many employers pursue this goal by maintaining some ongoing desired parity with other employers. During the second stage, each corporate division allocates its share of the salary budget among its workers to match market wages and reward performance. Employers often need to reconfigure wage differences among occupations in their divisions to respond to external influences. In a competitive labor market, an occupation’s wages reflect the amount and kind of training necessary, working conditions, and whether such workers are in short supply compared to the firms’ need for them. These circumstances can change as technology, products, demographics, or input prices shift. INFLATION IN THE LABOR MARKET—THE AGENTS’ PERSPECTIVES In this section, we describe the wagesetting practices of large U.S. employers, such as those observed in the CSS. Large employers are of particular interest for this study because they provide a majority of jobs (over half and not shrinking) in the U.S. labor market. In addition, their behavior is more likely to deviate from the competitive price-taking model than are small firms’ actions. Why Inflation Affects the Distribution of Wage Shocks The process described above can be incorporated into a formal wage-setting model that allows for period-by-period heterogeneity in wages and their changes.3 Crucially, though, as long as individuals optimize over leisure and consumption, a general, observed increase in the price level will shift the wage-change distribution equivalently for all firms. This uniform response to inflation is characteristic of any wage determination model with representative or aggregate agents. Hence, we must move beyond simple representative or aggregate agents to find factors that make the distribution of wage changes sensitive (non-neutral) with respect to inflation. We posit three main sources. First, if the firms’ inflation outlooks differ, their wage changes will differ (if contracting is nominal and fixed for a period of time). Any employer’s mistakes in projecting product price growth shows up uniformly in the wages of all its workers. Second, nominal wages may be rigid. That is, workers may experience a discrete rise in the disutility of their effort after Wage-Setting Practices in Large Firms 1 Compensation includes wages, benefits, and working conditions. For simplicity, we focus on wages in this analysis. Wages are the largest and most flexible part of compensation and are most subject to the effects of inflation. 2 In a unionized company, wage determination also involves negotiation with union leaders and a long (usually three-year) time horizon. 3 One example would be the Sparks (1986) model, which is itself a generalization of efficiency wage models of Shapiro and Stiglitz (1984). Inflation affects the labor market by influencing workers’ expectations and firms’ wage-setting practices and compensation schemes. In economies with competitive labor, capital, and product markets, comparable workers at equivalent jobs should be compensated similarly.1 If an employer sets wages too low, employee morale and productivity may suffer, and turnover may rise—all resulting in lower profits. If an employer pays too much, however, it will also experience lower profits or have to lay off workers because it will be unable to price products competitively and still be profitable. Thus, inflation is a key factor in workers’ and firms’ wage setting. The annual wage-setting process in a large firm typically has two stages. In the first stage, an employer’s senior management sets the average wage change for its work force—to reflect inflation forecasts, labor market surveys, and projections of sales F E D E R A L R E S E R V E B A N K O F S T. L O U I S 94 M AY / J U N E 1 9 9 9 nominal wage cuts. This story is consistent with prevalence of nominally priced contracts in the U.S. economy. If firms do avoid nominal wage cuts, the workers most affected are those whose occupation gets a negative shock, no matter what type of firm they are in. So, in an economy with downward rigidity, the variance of occupational wage changes rises with the level of inflation— until the rigidity no longer binds. Finally, business-cycle phenomenon may alter the supply of workers in other ways that are correlated with inflation— yielding further non-neutralities in the distribution of wage changes. fall—resulting in smaller nominal wage increases than typical. Thus, lower wage increases may occur more often or be associated with different conditions than in the past. Alternatively, others have argued that wage setting has been altered by the persistence of very low inflation (below 3 percent). In a low-inflation environment, competition could pressure participants to accept more flexible practices—particularly practices that permit nominal pay cuts. Examples of such innovations already exist and would proliferate, such as bonus and incentive pay, and contingent contracts. Widespread use of such pay schemes would overcome the constraints of downward nominal wage rigidity, allowing lower overall wage changes. In addition, the lowest wage changes for particular occupations within firms might be less restricted—that is, lower than expected, based on previous patterns. Have Things Changed in the 1990s? Two schools of thought argue that wage setting during the 1990s has been different than in previous years. One set of analysts suggests that workers have become more insecure since the 1980s, because of employer downsizing and the elimination of lifetime jobs in the U.S. The other points to changes due to the persistence of the low-inflation environment. According to a recent series of articles in the New York Times, the leading explanation of why inflation has been so limited these last three years—despite low unemployment rates—is that wage demands have been held down by an unusually high degree of “worker uncertainty.”4 Substantial research effort has gone into identifying and disputing the sources of this presumed insecurity in the face of a buoyant labor market. The most commonly mentioned reasons include the threat of middle-management layoffs, competition with foreign workers, and less unionization. These factors could reduce wage inflation by making workers think twice before requesting higher wages, even if their firms’ balance sheets have improved. If this is the case, then some employers that in the past would have maintained or elevated their market wage position, no longer feel the need to do so. In an efficiency wage model, alternative employers are exogenously less attractive to workers, so the efficiency wage firms’ offers should THE COMMUNITY SALARY SURVEY This study uses annual private salary data from a survey that the Federal Reserve Bank of Cleveland has conducted in Cleveland, Cincinnati, and Pittsburgh since 1927 to assist its annual salary budget process. The analysis data set reports wages for detailed occupations, by employer from 1957 through 1996. The data set has three major selling points for this study. First, the wages recorded here are less prone to random reporting error than household data because they are derived from administrative records. Second, the data are longer-lived than any source previously investigated. Third, because employer data records wages in the way most meaningful to firms, it is preferable to household or aggregate data for studying impacts on the firms’ wage setting. This perspective appropriately reflects the strategies used by firms to adjust wage bills (e.g., promotions, reassignments or reorganization), but not the potentially confounding means used by individual workers to adjust their earnings (e.g., taking second jobs or changing hours). F E D E R A L R E S E R V E B A N K O F S T. L O U I S 95 4 Peter Passell, “A Pulse that Lingers,” The New York Times, July 22, 1997, p. A1. M AY / J U N E 1 9 9 9 Table 1 well-developed markets.9 Many occupations are divided into grade levels, reflecting responsibility and experience. In the analysis, to avoid unnecessary restrictions, we consider each occupational grade in each city to be a separate occupation. Thus, the total number of occupations in Table 2 exceeds the number surveyed during any given year. For example, 83 occupational grades were surveyed in 1996, yielding 240 occupations across the three cities. On average, each employer reports wages for about 27 occupations. Although the CSS is conducted annually, the month surveyed has changed several times. Throughout the paper, results for any year refer to the time between the preceding survey and the one conducted in that year—usually a 12-month span, but occasionally not. When we examine data means for periods longer or shorter than a year, we annualize the changes so they can be compared directly across years. All data merged have been adjusted to the extent possible to reflect time spans consistent with those in the CSS. We have repeated most of the exercises reported in this paper on the subset of years that covered exactly a year and find no qualitative difference in results. We also incorporate standard measures of inflation and national output-per-hour in our analysis (see Table 3). As a measure of general inflation experienced in the country, we use percentage changes in the monthly averages of the Consumer Price Index (CPI) for all Urban Workers. Our labor productivity measure is the Nonfarm Business Sector Output per Hour Worked (pre-chain-weights). In order to investigate the distribution of wage adjustments under different inflationary environments, we use two schemes to differentiate among years. First, we label all years as years of increasing, stable or decreasing inflation, using a 60.5% cutoff for the CPI. For example, years when the inflation rate rose by more than 0.5 percentage points are considered years of increasing inflation. Second, we identify multi-year episodes of inflationary changes as periods where the economy experienced two or more consecutive years of increasing, stable o r Description of the Annual Wage Adjustment Data Set Drawn from the CSS, 1957-1996 Total Number of Job-Cell Wage Adjustments Observed 73,094 Number of Years of Changes 39 Average Number of Observations Per Year 1,874 Mean Log Wage Adjustment 0.048 Standard Deviation of Log Wage Adjustment 0.086 NOTE: All numbers reported are for the first-differenced data set. SOURCE: Authors‘ calculations from the Federal Reserve Bank of Cleveland Community Salary Survey. 5 6 Job-cell-year observations where the calculated change in log wages exceeds 0.50 in absolute value are deleted from the sample on the assumption that most of these arise from reporting or recording errors. Over 1,000 observations are imputed from cases where jobcells are observed two years apart. The imputed one-year changes are simply half of the two-year differences. Many of the results reported here were also run without the imputed observations. Their inclusion does not affect the results. Comparison of the coefficients estimated separately for means and medians for some years where both were available (1974 and 1981-1990) suggests that they are highly correlated (correlation coefficients of .97 to .99). Coefficients estimated with medians show more variation than those estimated on means and are more highly correlated over time, however this is consistent with medians being a more robust measurement of central tendency. Table 1 describes the dimensions of the CSS wage-change data set. From wage levels, we compute 73,094 annual wage changes for occupation-employer (job) cells observed in adjacent years.5 Each observation gives the change in the log of the mean or median salary for all individuals employed in an occupation-employer cell. Since medians should be more robust to outliers yet only means were recorded before 1974, our results use means through 1974 and medians for the years thereafter.6 Cash bonuses are included as part of the salary, although fringe benefits are not. Participants in each city are chosen to be representative of large employers in the area. Until 1995, the number of companies participating trended up from 66 to over 80 per year (see Table 2). On average, they stay in the sample for almost 13 years each. Since each participant judges which establishments to include in the survey, depending on its internal organization, we use “employer,” a purposely vague term, to mean the employing firm, establishment, division, or collection of local establishments for which the participating entity chooses to report wages.7 The industries included vary widely, although the emphasis is on obtaining employers with many employees in the occupations surveyed.8 The occupations surveyed (43 to 100 each year) are exclusively nonproduction jobs that are found in most industries, with relatively high inter-firm mobility, and F E D E R A L R E S E R V E B A N K O F S T. L O U I S 96 M AY / J U N E 1 9 9 9 Table 2 Description of CSS Data by Year End Number of: Year Job cells 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 Total 1,336 1,557 1,714 1,669 1,701 1,881 1,910 2,032 2,123 1,965 1,967 2,128 1,972 853 854 1,262 1,477 1,335 1,379 1,391 789 1,674 2,418 2,689 2,196 2,185 2,013 2,274 2,272 2,396 2,437 2,401 2,407 2,505 2,536 2,398 2,355 2,128 1,841 1,345 75,765 Occupations* 94 94 103 103 103 109 112 113 124 125 125 124 114 49 49 66 90 96 101 104 60 197 267 295 186 193 190 213 212 220 226 222 225 222 223 223 223 223 241 240 6,187 Mean Log Wage Adjustment in: Employers 73 83 88 86 88 93 90 96 95 89 89 94 97 36 36 38 57 73 73 72 72 68 75 79 83 82 75 80 79 82 80 82 81 84 89 84 89 84 69 51 3,002 Cleveland Cincinnati Pittsburgh 0.051 0.049 0.040 0.036 0.039 0.024 0.019 0.026 0.021 0.040 0.037 0.046 0.066 0.068 0.061 0.061 0.056 0.126 0.074 0.065 0.030 0.052 0.064 0.095 0.086 0.072 0.050 0.047 0.040 0.042 0.031 0.036 0.045 0.052 0.038 0.039 0.032 0.027 0.027 0.040 0.049 0.046 0.054 0.048 0.032 0.035 0.022 0.026 0.022 0.026 0.045 0.042 0.044 0.050 ** ** ** 0.095 0.084 0.063 0.057 0.021 0.063 0.071 0.074 0.089 0.092 0.055 0.058 0.044 0.044 0.037 0.037 0.041 0.046 0.045 0.042 0.026 0.029 0.031 0.032 0.048 0.045 0.050 0.070 0.034 0.036 0.024 0.024 0.023 0.010 0.038 0.035 0.042 0.049 ** ** ** ** 0.139 0.090 0.078 0.052 0.066 0.069 0.087 0.059 0.078 0.073 0.063 0.042 0.037 0.038 0.023 0.036 0.024 0.035 0.043 0.040 0.025 0.019 0.030 0.048 * Occupations are counted separately for each city. ** In 1970-72, the CSS is missing Cincinnati; in 1970-73, the CSS is missing Pittsburgh. SOURCE: Authors’ calculations from the Federal Reserve Bank of Cleveland Community Salary Survey, 1956-1996. decreasing rates of inflation. Table 4, which appears on page 100, shows how the years under investigation (1957-1996) are categorized by these criteria. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 97 7 Some include workers in all branches in the metropolitan area; others report wages for only the office surveyed. Since a participant's choice of the entities to include presumably reflects those for which wage policies are actually administered jointly, the ambiguity here is not particularly troublesome. 8 The employers surveyed include government agencies, banks, manufacturers, wholesalers, retailers, utilities, universities, hospitals, and insurance firms. 9 They include office (e.g., secretaries and clerks), maintenance (e.g., mechanics and painters), technical (e.g., computer operators and analysts), supervisory (e.g., payroll and guard supervisors), and professional (e.g., accountants, attorneys, and economists) occupations. Job descriptions for each are at least two paragraphs long. M AY / J U N E 1 9 9 9 Table 3 Means and Standard Deviations of CSS Wage Adjustment Components and Other Economic Indicators Mean Standard Deviation ∆ Occupation-Employer Log Wage Current U.S. CPI-Ua 0.048 0.046 0.084 0.034 ∆ Output/Hourb 0.016 0.062 0.016 0.014 College to High School (H.S.) Wage Premium 0.000 0.545 0.009 0.156 High School to Less than High School Premium 0.337 0.134 Percentage Change in College to H.S. Wage 2.18 7.38 Percentage Change in H.S. to Less than H.S. Wage 2.78 9.01 Variable Unemployment Ratec ∆ Unemployment Ratec a Change during salary survey year in the BLS Consumer Price Index for all Urban Workers (CPI-U) for the United States. b Change during salary survey year in the BLS Nonfarm Business Sector Output per Hour Worked. c U.S. civilian unemployment rate. SOURCES: Authors’ calculations from the Federal Reserve Bank of Cleveland Community Salary Survey, 1957-1996. U.S. Bureau of Labor Statistics (BLS). As a check for our results focusing on business cycle variables, we also control for the long-run rise in earnings inequality. Limited earnings inequality measures are available for the full period of this paper, 1957 to 1996. The best measures available are median earnings by education level. Even this series is missing a few years during the 1950s. We interpolate to fill in these gaps on the justification that these controls are offered to account for long run trends. inflation and productivity growth, versus 1.3 percentage points lower over the full sample. This suggests that the early 1990s had somewhat weaker than usual wage growth, given inflation and the measured gains in productivity. As for timing, at the annual frequency of CSS data, wages and prices can be described reasonably as changing contemporaneously. Compared to the contemporaneous correlation between inflation and mean wage growth of 0.82, the correlations are substantially lower for wage growth leading inflation by one year (0.59) or two years (0.35). The alternative— that wage growth follows inflation—is better supported. The correlation with wage growth lagging inflation by one year is 0.83. It falls to 0.69 with a two-year lag. It also is clear that during particular periods, wage growth exceeded inflation or CPI growth, with or without subsequent increases in the inflation rate. Overall, this source of detailed wage data supports a relationship between wage growth, inflation and productivity growth, at least at an aggregate level. Wage Adjustments and Inflation Figure 1 confirms that CSS wage changes are generally synchronized with inflation. The correlation between the mean CSS wage adjustment and inflation (CPI) is high (0.82). Overall, though, CSS wage growth has a higher mean (by 0.37) than the CPI, because it includes the benefits of productivity growth. Recent wage growth has averaged much closer to the inflation rate (wage growth led by only 0.08 percentage points in the 1990s). From 1990 to 1996 mean wage growth was 1.7 percentage points lower than the sum of F E D E R A L R E S E R V E B A N K O F S T. L O U I S 98 M AY / J U N E 1 9 9 9 Figure 1 Inflation and the Dispersion of Wage Changes Mean Log Wage Changes, Productivity, and Inflation Figure 2 relates the distribution of log wage changes in the CSS to the CPI during the period. The line with circles shows the percentage change in the CPI. The other lines show the 10th, 25th, median, 75th, and 90th percentile log wage changes for cells in the CSS. If inflation were neutral with respect to the distribution of wage changes, there would be no relationship between the level of inflation and the widening of the gap between the top and bottom lines on the figure. We would expect the lines to roughly parallel the level of inflation. Instead, the quantile lines show a marked tendency to widen as the level of inflation rises. For example, in 1996, the inflation rate was 3.0%. In the CSS that year, the median cell had a wage change of 3.4%, while the 10th and 90th percentiles had wage changes of –4.7% and 12.5%, respectively. Thus, factors that affect the size of percentile wage changes increase the value of a good shock or a bad one in a particular year. One aspect of interest for interpreting our findings is whether wage changes are correlated with wage levels. If the dispersion of wages remained constant over time, we would expect no correlation between wage levels and changes. Wages in the CSS, however, like those in other U.S. data sources, show a recent widening inequality (Groshen 1991). Thus, the overall correlation coefficient between log wage levels and changes in the CSS is 0.13. Annually, the correlations range from 0.33 in 1977 down to 0.06 in 1982. Thus, in all years, higher-wage workers tended to receive bigger proportional raises than did low-wage workers. Yet the correlation is fairly low, so our findings say more about what drives the size of good and bad wage shocks than about what happens to good versus bad jobs. 20% Inflation + Productivity Growth CPI-U Inflation CSS Mean wage Change 16% 12% 8% 4% 0% 57 61 65 69 73 77 81 Salary Survey Year 85 89 93 This figure shows annualized percentage change by salary survey year, which is not always equal to 12 months. Notably in 1974. Figure 2 Distribution of Log Wage Changes, from 1957 to 1996 Percentiles of Cell Wage Changes vs. Inflation Percentiles: 10, 25, 50, 75, & 90. Dots indicate inflation rate. Percent change 20 10 0 -10 60 65 70 75 80 Year 85 utions, we use quantile regressions of wage changes on various measures of inflation and other controls. Quantile regressions (developed by Koenker and Basset, 1978) estimate the correlates of wage changes in various parts of the distribution. Formally, the estimator minimizes a weighted sum of absolute deviations of the residuals: How Inflation Affects Wage Gains in the Tails To formally test for and explore the impact of inflation on wage change distrib- F E D E R A L R E S E R V E B A N K O F S T. L O U I S 99 90 95 M AY / J U N E 1 9 9 9 Table 4 Classi cation of Sample Years by In ation Direction and Episode Year Inflation (CPI) 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 0.036 0.004 0.015 0.015 0.011 0.011 0.014 0.012 0.028 0.026 0.039 0.053 0.061 0.044 0.035 0.048 0.108 0.079 0.055 0.064 0.085 0.118 0.153 0.106 0.072 0.025 0.047 0.036 0.016 0.038 0.039 0.050 0.048 0.044 0.032 0.028 0.028 0.028 0.030 Inflation Change (∆CPI) 0.000 – 0.033 0.012 – 0.001 – 0.004 0.000 0.003 – 0.002 0.016 – 0.002 0.014 0.013 0.008 – 0.017 – 0.010 0.013 0.059 – 0.029 – 0.024 0.009 0.021 0.034 0.035 – 0.047 – 0.034 – 0.047 0.023 – 0.011 – 0.020 0.022 0.001 0.011 – 0.002 – 0.005 – 0.012 – 0.003 – 0.001 – 0.000 0.002 Direction of Inflation* Stable Increase Decrease Episodes of Inflation** Stable Increase Decrease • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • * An increase in inflation is defined as an increase in ∆CPI equal to or larger than 0.5%. Likewise, a decrease in inflation is defined as a decrease in ∆CPI equal to or less than 0.5%. ** An episode of inflation stability is defined as a period of two or more consecutive years when inflation was stable. Similarly, an episode of increasing (decreasing) inflation is defined as two or more consecutive years of increasing (decreasing) inflation. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 100 M AY / J U N E 1 9 9 9 Table 5 Simple Quantile Regressions for Total Cell Mean Wage Changes in the CSS ∑ hi yi − ∑ β j x ij , i j 2q if yi − ∑ β j x ij > 0 j i where hi = , ( ) − − ≤ 2 1 0 q if y β x ∑ j ij i j (Standard Errors in Parentheses) Independent Variable yi and xij are the ith observation of the dependent and independent variables. βj is a vector of regression parameters. The estimates are for quantile of interest, q. The predictions of the estimator are the expected change in wages at the qth quantile conditional on the values of the independent variables xij. Thus, we can distinguish between conditions which raise (or lower) the upperend wage changes, and those that primarily affect lower-end wage changes. If the estimated model were parameterized identically over the distribution of wage changes, then an OLS regression would yield very similar coefficients. Indeed, this is the reason that the median regression often is recommended as a robust (less susceptible to outliers) alternative to OLS regression. Koenker and Basset (1982) show that differences in parameter estimates at alternative quantiles convert into a very general test for heteroscedasticity. The test offers advantages over more common tests because it is robust to nonGaussian errors. We prefer it because the quantile estimators help elucidate the nature of the heterogeneity. The test statistic (interested readers are referred to Koenker and Bassett, 1982, for the formula), focuses on whether coefficient differences are significant given the quantile estimator measure of distribution of residuals. We report three sets of results, with increasing complexity. The first set shows the simplest estimates—for the effect of CPI inflation alone. Under the null hypothesis of inflation’s neutrality on the distribution of wage changes, we expect a coefficient of one on the level of inflation for every quantile. In the next set of regressions, we also include inflation’s square, to allow for nonlinearity. Under the null, the coefficient on this should be 90th 75th Quantile 50th 25th 10th Inflation CPI 0.949 (0.020) 0.707 (0.008) 0.555 (0.005) 0.432 (0.009) 0.067 (0.023) Constant 0.084 (0.001) 0.049 (0.000) 0.025 (0.000) -0.001 (0.001) -0.034 (0.001) Pseudo R2 0.060 0.065 0.046 0.015 0.000 KoenkerBasset χ2 T = 713.0 degrees of freedom = 2 Number of observations = 73,094 zero for all quantiles. Two additional variables capture any incremental influence of the level of inflation when inflation is falling (by more than 0.5 percentage points) or rising. Under the null, these coefficients should also be zero. In addition, we include the unemployment rate, the change in the unemployment rate, output per hour and its square in the regressions to control for the business cycle and real wage gains. Table 5 shows the simplest results. The first row shows how the level of inflation affects wage gains by quantile in the distribution of wage changes. As expected, and as we saw in Figure 2, wage changes in the 90th percentile rise almost one-forone with inflation. That is, the coefficient on CPI is 0.949. Wage gains in the lower tails amount to only a fraction of the inflation rate, however. The corresponding coefficient for the 10th percentile is 0.067; showing surprisingly low sensitivity to changes in prices. Thus, the disparity between wage changes in the upper and lower tails rises with inflation. Does this mean that the model predictions imply a growing disconnect between wages levels and prices? No, for two reasons. 1) The estimates for the intercept term are positive and statistically significant (except at the 25th and 10th percentiles), allowing most F E D E R A L R E S E R V E B A N K O F S T. L O U I S 101 Prob. < 0.005 M AY / J U N E 1 9 9 9 wage changes to keep up with the average level of inflation. This combination results in wage change predictions that are less variable than inflation, but similar in their mean levels—as implied by Figure 2. Estimated constants do decline from the 90th to 10th percentile, preserving a distinct pattern of divergent outcomes. 2) The regression results are for wage changes. If the set of affected jobs vary substantially from period-to-period, then being behind in one period may be made up in another. This issue will be explored in Section 5 of this paper. While the apparent explanatory power of the regressions is fairly low—particularly for the lower quantiles—we detect some very robust statistical relationships. In evaluating the results, it is crucial to realize that the psuedo-R2 we report is not directly comparable to the traditional R2. This measure, just decreased, the wage distribution will be narrower than it would have been otherwise. Raises of most workers are essentially insensitive to inflation drops in the first year after inflation declines. That is, the sum of the two coefficients on CPI and its negative change is close to zero for the 25th, median and higher quantiles. The workers in the 10th percentile, however, actually gain higher raises than they would have under last year’s inflation rate, all else being equal. Thus, inflation decreases tend to narrow the distribution of wage changes. Inflation increases are associated with additional wage gains in all quantiles. These bonuses are smallest for the median (0.377), but higher for workers at both extremes. Since the bonus coefficient for the 90th percentile (0.507) is smaller than the gain for the 10th percentile (0.792), inflation increases moderately narrow the distribution of wage changes, all else being equal. That is, while higher inflation rates widen the distribution, either increases or decreases modestly narrow the distribution in the year they are sustained. By contrast to the higher sensitivity of upper quantile wage gains to inflation levels, unemployment exerts most of its influence on the lower quantiles of wage growth. High unemployment depresses wage gains sharply in the bottom quantiles, with little effect on upper quantile raises. The coefficient of 0.701 on unemployment for the 10th percentile predicts that wage gains in the bottom decile will be 0.7 percentage points lower if unemployment is one percentage point lower, all else being equal. The opposite-signed coefficients on change in unemployment suggest that the effect of unemployment on wage growth is subject to a lag. Finally, the results for our proxy for productivity growth show a nonlinear relationship with wage changes at all quantiles. The coefficients for output-per-hour are positive with little variation among quantiles. This suggests that when productivity growth is slow, workers receive 30 to 50 percent of productivity gains in their paychecks. The coefficients on the quadratic term, however, suggest that this effect is sum of weighted deviations about the estimated quantile , pseudo R = 1 − sum of weighted deviations about the raw quantile 2 only approaches 1 when each observation is predicted as a conditional quantile. Thus, the estimator can yield accurate predictions of the quantile with a low psuedo-R2, as long as the weighted deviations are symmetric around the prediction. Table 6 adds considerable flexibility to the ways in which inflation can affect wage changes, as well as controls for unemployment and productivity. The bottom row shows that the addition of these terms does improve the fit of the equations, but by less than half in all cases. Thus, the level of inflation alone is a key element in predicting the size wage changes among quantiles. Crucially, the first row of the table shows that the basic decline in sensitivity to inflation as wage shocks get worse is maintained in the more complex model. Accelerating and decelerating inflation, per se, also have modest effects on the distribution of wage changes. For any given inflation level, if inflation has F E D E R A L R E S E R V E B A N K O F S T. L O U I S 102 M AY / J U N E 1 9 9 9 Table 6 Quantile Regressions for Total Cell Mean Wage Changes in the CSS With Controls for Productivity and Unemployment (Standard Errors in Parentheses) Independent Variable 90th 75th Quantile 50th 25th 10th Inflation CPI 0.962 (0.069) 0.766 (0.030) 0.634 (0.017) 0.547 (0.011) 0.216 (0.086) Inflation Squared 100*(CPI)2 -0.011 (0.004) -0.010 (0.002) -0.008 (0.001) -0.015 (0.001) -0.016 (0.005) Decreasing Inflation (∆CPI≤-0.05)*∆CPI -0.896 (0.078) -0.913 (0.033) -0.684 (0.019) -0.558 (0.013) -0.662 (0.097) Increasing Inflation (∆CPI≥0.05)*∆CPI 0.507 (0.102) 0.628 (0.044) 0.377 (0.026) 0.448 (0.018) 0.792 (0.131) Unemployment Rate 0.182 (0.057) -0.105 (0.025) -0.049 (0.015) -0.250 (0.010) -0.701 (0.070) Change in Unemployment Rate -0.051 (0.098) 0.075 (0.042) 0.142 (0.025) 0.380 (0.016) 0.736 (0.121) Productivity Growth ∆Output/Hour 0.371 (0.123) 0.479 (0.052) 0.478 (0.030) 0.381 (0.020) 0.410 (0.154) Prod. Growth Sqd. 100*(∆Output/Hour)2 -0.087 (0.030) -0.096 (0.013) -0.108 (0.007) -0.090 (0.005) -0.051 (0.037) Constant 0.062 (0.004) 0.041 (0.002) 0.019 (0.001) 0.007 (0.001) -0.007 (0.005) Pseudo R2 0.071 0.081 0.060 0.022 0.007 Koenker-Basset χ2 T = 953.8 degrees of freedom = 2 Prob. < 0.005 Number of observations = 73,094 attenuated when productivity growth is fastest. Nevertheless, workers in the lowest quantile (with its coefficient of –5.060) may benefit more from higher productivity than do the upper quantiles, narrowing wage adjustment distributions when productivity growth is faster. Are these differences statistically significant? Testing for heteroscedasticity in wage changes according to the level of inflation yields a strong rejection of the null hypothesis. Despite the inclusion of controls for the direction of inflation changes and other business cycle factors, the Koenker-Basset test for heteroscedasticity yields values well beyond conventional levels of statistical significance. Summarizing broadly, the highest wage changes in a year increase with inflation. Wage changes at the lower tails, however, are more influenced by the unemployment rate. Given statistical significance of these differences, we now turn to the question of whether the effects are economically relevant. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 103 M AY / J U N E 1 9 9 9 Figure 3 and constant unemployment and productivity growth for the sample period. We also overlay the actual values for the percentile (the line without circles). For the median and 90th quantile, the fit is very close—information on inflation alone is sufficient to produce a reasonably close fit. The fit is markedly worse for 10th percentile wages, however. Until the mid-1970s, wage growth at the bottom is underpredicted. Then the model overpredicts wage changes until the late 1980s. This figure illustrates the points that median and upper tail wage changes are highly responsive to the inflationary environment—much more so than are wage changes at the lower tails. Most strikingly, however, this figure shows that the response of the various quantiles to inflation captures most of the path of the dispersion of wage shock over time. Thus, inflation can be seen as the main driving factor in the variation of wage shocks over time. Figure 4 illustrates the point further by showing how the full set of quantiles in Table 6 would respond to a hypothetical inflation path. Suppose that over a forty-year span, inflation started at zero, then rose by one percentage point per year until it reached fifteen percent at year sixteen. After being stable at fifteen percent for four more years, then it fell by one percent per year, until it reached zero at year 36 and was stable until year 40. Figure 4 shows the five predicted paths of quantile wage changes for this scenario. The contrast among the paths is quite stark. The higher the quantile, the more responsive wages are to inflation. Indeed, wages in the 10th percentile show very little response at all. We now repeat these exercises to illustrate the impact of unemployment. The exercise shown in Figure 5 is analogous to that in Figure 3, but with inflation held constant and the unemployment rate allowed to follow its historical path from 1957 to 1996. Again, the line with the circles shows the model predictions under these circumstances, while the unmarked line represents actual values. Overall, the relationship with unemployment is a less accurate predictor of quantile wage changes Model Predictions When Only Inflation Rate Varies Percentiles: 10, 50, & 90. Circles indicate model. Percent change 20 10 0 -10 60 65 70 75 80 85 90 95 Year Figure 4 Model Predictions for Rising and Falling Inflation Rate Percentiles: 10, 25, 50, 75, & 90. Percent change 20 10 0 -10 0 10 20 Year 30 40 Isolating Factors’ Effects on the Distribution of Wage Changes Since the model estimated in Table 6 is complex, we construct some illustrative scenarios to gauge the total impact of inflation and unemployment on wage changes. Figure 3 compares the impact of inflation on wage gains in the 10th, 50th, and 90th percentiles. For each percentile, we plot predicted values of wage changes (shown as circles), given realized inflation rates F E D E R A L R E S E R V E B A N K O F S T. L O U I S 104 M AY / J U N E 1 9 9 9 Figure 5 than is inflation. In contrast however, variations in unemployment predicting wage changes do much better for the 10th decile than they do for the median or 90th percentile. Figure 6 constructs a hypothetical scenario to illustrate the differing responsiveness of wage change deciles to unemployment paths. In this exercise, we begin with an unemployment rate of four percent, raise it by 0.5 percentage points per year until it reaches ten percent. Then we hold it steady for five years, followed by a 0.5 percentage point per-year drop until it reaches four percent and stays constant for ten years. Again, the contrast in responsiveness among the quantiles is stark. But unemployment (in contrast to inflation) has its most potent impact on the lowest quantiles of wage changes. The median shows very little response, and the 90th percentile even has a counterintuitive pattern—albeit a muted one. These figures highlight the differing responses of the quantiles to inflation and unemployment shock. They illustrate the generalization that wage gains of those in the higher quantiles rise steadily with inflation, while wage gains of those in the lower tails (that is, those suffering the largest negative shocks) are determined mostly by the unemployment rate. They also show that during the period from 1957 to 1996, inflation was the main determinant of the dispersion of wage shocks. The finding that the impact of these factors on wage changes varies substantially by quantile suggests that even our relatively detailed model of how wages react to inflation and other business-cycle variables doesn’t capture all of the important issues. Indeed, a complete econometric model would need to predict widely varying levels of matching nominal wage growth to inflation and employer responsiveness to general slackness in the labor market. Nonetheless, this statistical representation of wage change provides a useful description of typical patterns. Model Predictions When Only Unemployment Rate Varies Percentiles: 10, 50, & 90. Circles indicate model. Percent change 20 10 0 -10 60 65 70 75 80 85 90 95 Year Figure 6 Model Predictions for Rising and Falling Unemployment Rate Percentiles: 10, 25, 50, 75, & 90. Percent change 10 0 -10 0 10 20 Year 30 sation. Many researchers have documented a substantial increase in earnings inequality in the United States during the period studied. This rise in inequality also occurs in the CSS (Groshen 1991). While this increasing inequality must be reflected in wage changes, the exact nature of the relationship is unclear. Perhaps rising inequality raises the variance of wage changes because the distribution of desired wages is more dispersed, allowing for larger possible changes. Or, wage adjustments might be larger during periods when some shock to the labor market is increasing inequality. Rising Earnings Inequality? The path of inflation is not the only systematic trend that might affect compen- F E D E R A L R E S E R V E B A N K O F S T. L O U I S 105 40 M AY / J U N E 1 9 9 9 10 Schweitzer (1997) shows that educational differentials are the most substantial measured factor in the rise in earnings inequality. In addition, it is possible that inequality rose in ways that did not affect the distribution of wage changes. For example, the correlation of individual wage changes over time might rise, leaving the size distribution of wage changes unaffected. Given our focus on inflation, the rise in earnings inequality argues for conducting probes with suitable control variables. To this end we reestimate our quantile regressions with controls for the ratios of median earnings of workers of different education levels. This measure of inequality is available back further than other inequality series. In addition, these ratios are highly correlated with the variance of log wages over the period when microdata is available (starting in 1972).10 The two included wage ratios are college graduates versus high school graduates and high school graduates versus high school dropouts. The CSS includes occupations that employ workers at each of these three levels, although it is slanted toward more skilled occupations. Since we are uncertain about how rising earnings inequality alters the distribution of wage changes, we introduce controls for both the level and the percentage change in the education wage differentials. Adding these earnings inequality variables to the previous estimates is intended to show what relationships are robust to the inclusion of these variables. Table 7 shows the results. First, we note that differences in the estimated wage changes by quantile remain. Indeed, the heteroscedasticity test based on the difference between the inflation coefficient at the 25th, 50th, and 75th percentiles continues to be significant, because the difference in the coefficient estimates at the 75th and 50th percentiles are still large. Thus, control for inequality adds support to the conclusion that the wage change distribution reacts nonuniformly to labor market shocks. Nevertheless, wage inequality does appear to influence the distribution of wage changes. Coefficient estimates on the inequality measures are significantly different from zero in almost all quantiles. Inclusion of the level of wage inequality and its trend improve the fit of the quantile regressions (the psuedo-R2s rise) in Table 6. The fit of the upper half of the distribution is improved more substantially by the inclusion of inequality controls than is the fit in the lower half. In addition, although most signs on the coefficients estimated in Table 6 are preserved, some point estimates change markedly. Two general patterns stand out. First, including inequality controls does not substantially alter the role of inflation on wage changes. While the coefficients on the level of inflation for the lower quantiles are now larger, they remain smaller than those of the high quantiles. Furthermore, the size of the negative coefficients on their quadratic terms also are substantially larger. Similarly, the impact of sharp changes in the inflation rate on wage changes is changed little for decreases and slightly muted for increases. Replicated Figures 2, 3 and 4 using the inflation coefficients from Table 7 are parallel those shown above, although muted differences in the response to inflation between upper and lower quantiles are evident in the analog to Figure 3. Second, both the productivity and unemployment variables appear to be more heavily related to the inclusion of inequality in their impacts on the distribution than does inflation. Coefficient changes were larger and their patterns were more strongly altered. Overall, inequality controls do not remedy the inability of a single equation model (of the type estimated here) to describe the factors that determine wage adjustments consistently across the distribution of wage adjustments. These controls do point out a relationship between unemployment and productivity variables and the rise of inequality in the United States. This interesting, but possibly spurious, relationship suggests an area for further study. TWO POLICY-RELEVANT QUESTIONS Are there Bellwether Jobs? One possible explanation for the finding that wage changes are highly vari- F E D E R A L R E S E R V E B A N K O F S T. L O U I S 106 M AY / J U N E 1 9 9 9 Table 7 Quantile Regressions for Total Cell Mean Wage Changes in the CSS, Including Inequality Variables (Standard Errors in Parentheses) Independent Variable 90th 75th Quantile 50th 25th 10th Inflation CPI 0.819 (0.078) 0.830 (0.036) 0.737 (0.019) 0.767 (0.051) 0.553 (0.086) Inflation Squared 100*(CPI)2 0.006 (0.005) -0.010 (0.002) -0.011 (0.001) -0.022 (0.003) -0.031 (0.005) Decreasing Inflation (∆CPI≥-0.05)*∆CPI -1.297 (0.100) -0.986 (0.044) -0.743 (0.023) -0.680 (0.064) -0.522 (0.110) Increasing Inflation (∆CPI≥0.05)*∆CPI 0.405 (0.103) 0.393 (0.048) 0.217 (0.026) 0.317 (0.061) 0.591 (0.118) Unemployment Rate 0.335 (0.089) 0.217 (0.041) -0.202 (0.022) -0.029 (0.051) -0.226 (0.098) Change in Unemployment Rate -0.700 (0.151) -0.285 (0.069) 0.151 (0.037) -0.027 (0.108) 0.506 (0.167) Productivity Growth ∆Output/Hour -0.116 (0.137) 0.080 (0.062) 0.225 (0.033) 0.170 (0.080) 0.173 (0.152) Prod. Growth Sqd. 100*(∆Output/Hour)2 0.026 (0.035) -0.037 (0.015) -0.066 (0.008) -0.042 (0.021) -0.027 (0.037) Col. to H.S. Ratio of median wage 0.053 (0.017) 0.008 (0.008) 0.016 (0.004) 0.036 (0.011) 0.022 (0.019) ∆ Col. to H.S. ∆ Ratio of median wage -0.070 (0.020) -0.038 (0.005) -0.026 (0.003) -0.007 (0.003) 0.021 (0.012) H.S. to Dropout Ratio of median wage -0.063 (0.023) -0.052 (0.010) -0.052 (0.006) -0.071 (0.016) -0.093 (0.026) ∆ H.S. to Dropout ∆ Ratio of median wage 0.042 (0.008) 0.018 (0.004) -0.010 (0.002) -0.007 (0.004) 0.010 (0.009) Constant 0.046 (0.008) 0.036 (0.004) 0.011 (0.002) -0.010 (0.005) -0.023 (0.009) Pseudo R2 0.075 0.086 0.063 0.023 0.008 Koenker-Basset χ2 T = 281.7 degrees of freedom = 2 Prob. < 0.005 Number of observations = 71,537 able is that the wage adjustments of certain occupations, employers, or occupationemployer cells are continually more responsive to inflation than are others. The CSS measures wages in nonproduction occupations with the thickest, best- F E D E R A L R E S E R V E B A N K O F S T. L O U I S 107 M AY / J U N E 1 9 9 9 Table 8 Spearman Rank Order Correlations of Wage Changes Across Years, by Type of In ationary Episode A. EPISODES OF STABLE INFLATION Years Within-Episode, One-Year Correlations Between-Episode First-Year Correlations 1st, 2nd 2nd, 3rd 3rd, 4th 4th, 5th 1961 1961-65 -0.125 (0.000) -0.191 (0.000) -0.086 (0.001) -0.118 (0.000) 1991-92 0.071 (0.002) - - - 1991 -0.055 (0.497) 1993-96 -0.057 (0.017) -0.019 (0.460) 0.085 (0.004) - 1993 -0.040 (0.644) 1991 -0.044 (0.053) B. EPISODES OF INCREASING INFLATION Years Within-Episode, One-Year Correlations Between-Episode First-Year Correlations 1st, 2nd 2nd, 3rd 3rd, 4th 1968 1968-70 -0.100 (0.000) -0.325 (0.000) - 1974-75 0.129 (0.000) - - 1974 -0.005 (0.890) 1977-80 -0.008 (0.839) -0.123 (0.000) -0.158 (0.000) 1977 -0.163 (0.002) 1974 -0.027 (0.502) C. EPISODES OF DECREASING INFLATION Years Within-Episode, One-Year Correlations 1st, 2nd 2nd, 3rd 3rd, 4th 1971-72 -0.012 (0.745) - - 1975-76 0.141 (0.000) - - 1981-84 -0.089 (0.000) 1985-86 -0.025 (0.311) -0.060 (0.017) - Between-Episode First-Year Correlations 1971 0.012 (0.619) - defined, inter-industry markets. Thus, it should capture mobile workers—those likely to be most sensitive to market conditions. 1975 1975 0.043 (0.403) 1981 -0.021 (0.784) -0.013 (0.800) 1985 -0.061 (0.457) -0.103 (0.065) 1981 -0.052 (0.058) In addition, the large employers in the CSS are arguably more able to track relevant market changes than smaller employers. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 108 M AY / J U N E 1 9 9 9 For monitoring and policy purposes, tracking bellwether jobs could provide useful signals of inflationary pressures. To investigate whether such bellwether jobs are likely to exist, we look for evidence of serial correlation in wage changes within and between types of inflationary episodes. Table 8 presents the results. The top panel focuses on the three periods of stable inflation during our sample time frame. The stability during these times provides a basis for comparison for the periods of rising and falling inflation. The first four columns present correlation coefficients between consecutive years during these three episodes. Were the majority of divergences in wage changes during these periods reflective of long-term divergent trends in occupation or employer differentials, these correlations would be positive—an above-average change during one year is likely to be followed by a similar one during the next year. On the other hand, if they reflected errors and corrections, or normal compositional changes in the workforce (promotions, hires, etc.) the correlations would be negative: An unusually big average increase in one year is likely to be followed by a below-average adjustment next year. During the stable periods, most (five out of eight) of the one-year correlations are statistically significant and negative, suggesting the importance of error, corrections and compositional shifts in the wage changes we observe. Across episodes, the correlations are essentially zero, suggesting that no particular type of job tends to benefit (or lose out) more than others during periods of stable inflation. The middle panel repeats the exercise for periods of increasing inflation during the sample years. Again, most of the correlations are statistically significant and negative—providing no evidence in support of bellwether jobs. Indeed, it looks as though deviations from the median during rising inflation are even more likely to be compensated for later on than if they occur during periods of stability. And across episodes, jobs that were early, fast movers in one period of Figure 7 Model Predictions vs Actual Quantities: 1990-1996 Percentiles: 10, 25, 50, 75, & 90. Circles indicate model. 15 Percent change 10 5 0 -5 90 92 94 Year inflation are, if anything, less likely to lead the way during subsequent episodes. The bottom panel looks at periods of declining inflation. When inflation is declining, the evidence of mean reversion seen in the upper two panels is attenuated. Most of the correlation coefficients are small and poorly identified, suggesting an even more random process. And again, across episodes, there is no evidence to suggest the existence of bellwether jobs. Thus, the evidence thus far argues strongly against the existence of bellwether jobs whose wage changes could signal inflationary changes. If bellwether jobs exist, they are a very small proportion of jobs in occupations or firms typical of the CSS. That is, they may be in smaller firms, or in production occupations, for example. In the CSS, being out on a tail is often preceded or followed by an opposite-tail wage change during the previous or following year. Which jobs land in one of the tails appears to be idiosyncratic, however, rather than a permanent feature of the job. Are the 90s Different? Our last empirical exercise examines whether the wage changes during the 1990s deviated from historical patterns, as some analysts suggest. We compare the actual F E D E R A L R E S E R V E B A N K O F S T. L O U I S 109 96 M AY / J U N E 1 9 9 9 path of wage-change quantiles during the 1990s to predictions based on the historical model estimated in Table 6. We want to see if the lower quantiles had much less wage growth during the 1990s than expected, given the underlying rates of inflation, productivity and unemployment. Figure 7 shows the results of the exercise. Each quantile is represented with two lines: its actual wage change (the unmarked line) and the model prediction (the line with circles). For most of the period, the model fits quite well. Only for 1994, 1995 and 1996 does the model miss much. During those years, the actual wage change was lower than the model predicted for the 10th percentile wage change by one to two percentage points. For the other parts of the distribution, the model performs quite well. Thus, the evidence of a sea change in wage-setting behaviors finds little support in the CSS so far. (highest) wage shocks in any year rises almost one-for-one with the level of inflation. • The lowest wage changes in any year do not rise much with inflation. 2. Other factors (including unemployment, inequality, and productivity growth) also affect the dispersion of wage changes. In particular: • Bad wage shocks are mitigated when unemployment is low. In addition, from a monetary policy or monitoring perspective, we add two intriguing findings: 1. Wage changes are slightly negatively autocorrelated over time. • Negative autocorrelations refute the notion of bellwether jobs (i.e., occupations or firms that regularly lead the way when prices rise) and suggests that inflation causes errors and corrections. CONCLUSION We have examined the Federal Reserve Bank of Cleveland Community Salary Survey from 1957 to 1996 for the impact of inflation on the size of good or bad wage shocks. Most importantly, our exploratory exercise uncovers strong evidence that the pattern of wage changes is not neutral with respect to inflation and other economic conditions. This finding suggests that the influence of errors and corrections, nominal rigidities, or business-cycle influences on wage-setting varies substantially within the labor market. These regularities provide a new window for comparing the behavior of wages with model predictions in our competitive economy. In particular, we find that representative or aggregate agent models abstract from important determinants of wage changes. We summarize our main findings as follows: 1. The dominant factor in predicting the distribution of wage changes is the inflationary environment. In particular, wage change dispersion is higher if inflation is higher because: • Small autocorrelations refute the existence of a permanent competitive fringe of firms or occupations and suggests that many jobs sustain occasional wage shocks. 2. There are no apparent changes in the early 1990s. The pattern of wage growth was predictable for the low levels of inflation and unemployment during the period. Under standard models of wage determination, many of these findings are puzzling. As such, they open the door to new areas for exploration. The next steps are to examine other wage data to confirm the patterns visible here, to refine our understanding of the patterns, and to test the predictions of particular variants of wage-setting models against observed patterns. • The magnitude of the best F E D E R A L R E S E R V E B A N K O F S T. L O U I S 110 M AY / J U N E 1 9 9 9 REFERENCES Groshen, Erica L. “Rising Inequality in a Salary Survey: Another Piece of the Puzzle,” Federal Reserve Bank of Cleveland Working Paper 9121, December 1991. _______, and Mark E. Schweitzer. “Identifying Inflation’s Grease and Sand Effects in the Labor Market,“ in The Costs and Benefits of Price Stability, Martin Feldstein, ed., University of Chicago Press, 1999, pp. 273-308. Koenker, Roger, and Gilbert Basset, Jr. “Regression Quantiles,” Econometrica (January 1978), pp. 33-50. _______, and _______. “Robust Tests for Heteroscedasticity Based on Regression Quantiles,” Econometrica (January 1982), pp. 43-61. Schweitzer, Mark E. “Workforce Composition and Earnings Inequality,” Federal Reserve Bank of Cleveland Economic Review (2nd Quarter 1997), pp. 13-24. Shapiro, Carl, and Joseph E. Stiglitz. “Equilibrium Unemployment as a Worker Discipline Device,” American Economic Review (June 1984), pp. 433-44. Sparks, Roger. “A Model of Involuntary Unemployment and Wage Rigidity: Worker Incentives and the Threat of Dismissal,” Journal of Labor Economics (October 1986), pp. 560-81. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 111 M AY / J U N E 1 9 9 9 F E D E R A L R E S E R V E B A N K O F S T. L O U I S 112 M AY / J U N E 1 9 9 9 John C. Haltiwanger is a professor of economics at the University of Maryland. Commentary tigates the relationship between the distribution of wage adjustments at the micro level and inflation. The paper is primarily an empirical exercise. The question at hand is whether changes in the rate of inflation have a neutral effect on the distribution of wage adjustments at the micro level. A simple view is that inflation should affect all participants similarly (i.e., all relevant parties simply care about real wages) and thus inflation should have little or no impact on the distribution of wage adjustments. The striking finding that emerges is that inflation is dramatically non-neutral in terms of its impact on the distribution of wage adjustments. Moreover, the pattern of nonneutrality is quite interesting. Wage changes at the upper tail of the wage change distribution respond to a much greater extent than wage changes in the lower tail of the wage distribution. The authors also investigate two related interesting questions about the nature of this non-neutrality. First, they ask the question as to whether there are bellweather jobs—in the sense that perhaps the non-neutrality is such that wages respond to inflation for some types of jobs more quickly than for others. They find little or no evidence for bellweather jobs. Second, they ask the question about whether this non-neutrality has changed over time, with a particular emphasis on the 1990s. The motivation for the focus on the latter is the popular perception that wage responses to inflation have been mitigated during the 1990s due to increased job insecurity and that this would, in turn, impact the nature of the non-neutrality. They find little or no evidence of changes in the non-neutrality. While I find the basic facts and related empirical exercises quite interesting, this work is somewhat difficult to interpret, given the lack of much of an overall conceptual framework to help us understand the possible sources of connections between changes in inflation and changes in the John C. Haltiwanger T his fine paper fits into a growing literature in macroeconomics that emphasizes the idea that it is difficult, if not impossible, to understand aggregate fluctuations without understanding the underlying behavior of heterogeneous microeconomic agents. It is self-evident that individual households, workers and businesses have heterogeneous characteristics and are subject to idiosyncratic events that yield dramatically different outcomes at the microeconomic level. This heterogeneity in microeconomic outcomes typically dwarfs aggregate fluctuations so that for most households and businesses, the macro economy is a relatively unimportant factor in determining their fortunes. In spite of this overwhelming micro heterogeneity, macroeconomists have traditionally abstracted from this heterogeneity because the common view is that the micro heterogeneity washes out in the aggregate. Thus, macroeconomists have traditionally developed models describing the behavior of the typical firm or the typical worker and worried relatively little about the differences in outcomes across economic agents. The growing availability of micro panel data on households and businesses (and in some cases linked employer-employee micro data) has made it increasingly clear that this traditional approach misses important aspects of aggregate fluctuations. That is, there is often a strong connection (albeit with questions about the direction of causality) between the aggregate fluctuations and the nature and extent of heterogeneity of outcomes across agents. Technically, the issue is often whether there is a connection between the fluctuations in the first and higher moments of the distributions of outcomes. In the current paper, this is precisely the question, as the paper inves- F E D E R A L R E S E R V E B A N K O F S T. L O U I S 113 M AY / J U N E 1 9 9 9 distribution of wage adjustments. This is not really a criticism of the current paper, but rather illustrates the need for a conceptual framework to help interpret these interesting findings. Put differently, we need to consider the sources of heterogeneity in wage adjustments at the most basic level and how this heterogeneity is likely to interact with changes in the rate of inflation. Many factors may be at work in the underlying distribution of wage adjustments. Changes in relative labor supply and labor demand for workers of different characteristics (both those that are easily observable to the researcher and those that are not) are obviously important in this context. The institutional structure (e.g., unionization) and differences in the manner that wages are determined by sector or firm also are likely to be important. In considering these alternative possible factors, in light of the findings in this paper, it is useful to consider what we know about changes in the structure of labor markets during the sample period for this analysis. One of the primary recent empirical findings from applied labor economics research is the observation that there have been systematic increases in the dispersion of wages across workers during the last few decades. While the sources of this rising wage inequality are still somewhat in dispute, there is a growing consensus that this rising wage inequality is due to a rising relative demand for skilled workers. The sources of the latter might be changing technology (broadly defined) or changing world markets but, nevertheless, the return to being skilled has risen during this period of time. These fundamental changes in the dispersion of wages are closely linked to the changes in the distribution of wage adjustments. Moreover, the rising wage inequality was especially dramatic during the 1970s and 1980s—a period in which the rate of inflation is high and there are large associated changes in the distribution of wage adjustments. Thus, one question that arises is whether any aspects of their findings are spurious. Perhaps what is driving the results are the underlying factors that cause rising wage inequality and that the timing of these factors corresponds to a period with many dramatic changes in the U.S. economy. Another related and relevant hypothesis is that it is no coincidence that the observed long-run structural adjustments in the labor market were bunched during this period of volatile business-cycle fluctuations. That is, either the business-cycle fluctuations caused a change in the timing of the structural adjustment, or the businesscycle fluctuations were partly due to the intense period of structural adjustment. Moreover, since this period of turbulence in labor markets is also associated with high and volatile rates of inflation, this may underlie the connection between inflation and the distribution of wage adjustments. All of this discussion is speculative, however. The main point is that it will be difficult to sort out the factors that generate this paper’s interesting results without a conceptual structure (and associated empirical analysis) to help us understand the factors driving the distribution of wage adjustments and the potential link to inflation. More generally, the question is whether the results are driven mostly by the turbulent events of the 1970s and 1980s— a period in which there were substantial fluctuations in macro variables like inflation and unemployment, and a period of substantial structural change in the economy and labor market. I have some other relatively minor concerns about specific aspects of the analysis. While the CSS appears to be a very rich and unique dataset, there are concerns about the representativeness of the sample. It is intended to be representative of large employers in the area. Since the sample period here is so long, these concerns may be especially important. That is, not only is one concerned about how representative the sample is at a given moment, but also whether its representativeness has changed over time. A somewhat related concern is that their analysis is in terms of the unweighted wage adjustment distribution —an interesting alternative would be to consider the hours-weighted wage-adjust- F E D E R A L R E S E R V E B A N K O F S T. L O U I S 114 M AY / J U N E 1 9 9 9 ment distribution. If their results are driven by occupation/employers with small hours weights, then the results are of less interest. Finally, the authors make a relatively big deal about the finding of weak or negative autocorrelations in wage adjustments. They want to interpret this as evidence against bellweather occupation and jobs. This interpretation may be correct, but there may be a number of factors underlying the weak or negative autocorrelations in changes observed in the data. For example, it may be that wage adjustments are lumpy at the micro level (due perhaps to some rigidities or fixed adjustment costs) which can lead to weak and negative autocorrelation. This would yield a very different interpretation of the findings. To sort out these alternatives, we need more structure and further analysis. To sum up, this paper represents an installment on a very nice research agenda with a rich and unique dataset. This particular installment offers some interesting new “facts.” While there may be some concerns about the robustness of these facts to measurement concerns and about whether the results are idiosyncratic to the turbulence of the 1970s and 1980s, they are interesting and deserve further consideration. We need more structure to interpret and understand these new facts, but that awaits another installment from these authors or in studies stimulated by these new facts. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 115 M AY / J U N E 1 9 9 9 F E D E R A L R E S E R V E B A N K O F S T. L O U I S 116 M AY / J U N E 1 9 9 9 Kenneth J. McLaughlin is an associate professor in the Department of Economics at Hunter College and the Graduate School of the City University of New York. The author wishes to thank David Lebow, Gil Maduro, Joseph Ritter, Richard Startz, and participants at the Federal Reserve Bank of St. Louis’ Twenty-Third Annual Economic Policy Conference, and participants in seminars at Hunter College and the Federal Reserve Banks of Atlanta and New York. Are Nominal Wage Changes Skewed Away From Wage Cuts? Correlations of these properties with inflation also help to identify skewness away from nominal wage cuts. Although these papers have much in common, the specific techniques, data sets, and even conclusions vary. With a series of simple calculations on a single data set, I intend to integrate the main results from this new and exciting area of research to shed light on an important question for macroeconomic policy and economic theory: Are nominal wage changes skewed away from wage cuts? In particular, does downward nominal rigidity censor some wouldbe wage cuts, transforming some wage changes that would be negative into zerowage changes? To answer this question, I document key properties of the distribution of wage changes in panel data. I show that tests based on the familiar skewness coefficient are particularly weak in the presence of fat-tailed distributions, such as the distribution of wage changes, so I introduce symmetrically differenced histograms, a convenient way to detect asymmetries visually. I also apply mean-median differences and sign tests of symmetry to the wage change data. Evidence of skewness of wage changes is decisive; however, establishing that wage changes are skewed away from wage cuts requires more than evidence of skewness of wage changes. Is downward nominal rigidity the source of the skewness of wage changes, or is the distribution of wage changes more generally skewed? To sort this out, I calculate measures of thinning, a reduction in the frequency of wage change observations below zero (Lebow, Stockton, and Wascher 1995; Card and Hyslop 1997; and Kahn 1997); the calculations do point to the thinning of wage cuts. A complete explanation also must account for two other features of wage change distributions. First, by focusing on wage changes close to the median, I show that wage changes are skewed right over a range that has nothing to do with Kenneth J. McLaughlin R eal-wage cuts are much more common than nominal wage cuts. Why? By definition, real cuts must be more common if inflation is positive. Yet there might be more to it. Perhaps workers suffer from money illusion. Maybe managers cannot cut pay in nominal terms, but they can cut real wages. As a result, a low-inflation economy might be a high unemployment economy. And moderate inflation might “grease the wheels” of the labor market. These issues are being addressed in a burgeoning body of literature on wage changes in panel data (McLaughlin 1994; Lebow, Stockton, and Wascher 1995; Craig 1995; Akerlof, Dickens, and Perry 1996; Card and Hyslop 1997; Kahn 1997; Altonji and Devereux 1997; Crawford and Harrison 1997; and Christofides and Stengos 1998). To detect the existence of downward rigidity of nominal wages, this literature identifies properties of the distribution of wage changes: • The frequency of wage cuts • A spike or mass-point of observations with no change in pay • Skewness • Thinning of the distribution below zero • Holes in the distribution around zero F E D E R A L R E S E R V E B A N K O F S T. L O U I S 117 M AY / J U N E 1 9 9 9 downward nominal rigidity at zero. This violates the mirror-image assumption of one thinning estimator; consequently, Lebow, Stockton, and Wascher; and Card and Hyslop overestimate the thinning of nominal wage cuts. Second, I show that skewness of wage changes is basically unrelated to inflation; that is, higher inflation does not reduce the impact of any nominal wage rigidity. I begin by introducing the issues and methods in the context of the literature. intrayear data from the Survey of Income and Program Participation (SIPP); Card and Hyslop (1997) using the PSID and matched samples from the Current Population Survey (CPS); and Kahn (1997) using the PSID confirms my findings of surprisingly frequent reported wage cuts even for stayers. Regarding the complementary question of downward rigidity, my analysis was limited to computing (a) the size of the spike at no change in pay and (b) skewness coefficients. I found that an additional 7 percent of the stayers report no change in pay from year to year, and that the distribution of wage changes was skewed to the right, away from wage cuts (McLaughlin 1994).1 Lebow, Stockton, and Wascher; Card and Hyslop; and Kahn also document the spike, although Card and Hyslop (1997, note 13) find a substantially larger spike by analyzing hourly workers. Some of the spike at zero could be due to variation in interview dates in the PSID. Lebow, Stockton, and Wascher estimate that 1 percentage point of an 8-percentage-point spike is attributable to interviews occurring within a year. They also estimate that an additional 3 percentage points are due to rounding of wage reports. Card and Hyslop (1997, p. 83), on their CPS sample of hourly workers, also estimate that about half of the spike is attributable to rounding errors. This literature takes a variety of approaches to estimating skewness of wage changes. Using the skewness coefficient, I found that the overall distribution of wage changes is skewed to the right; however, I also reported that the distributions of wage changes of nonunion workers, nonminimum-wage workers, and salaried workers are dead-on symmetric (McLaughlin 1994, Table 2). Similarly, Lebow, Stockton, and Wascher report small positive skewness coefficients for all stayers, and sizable positive skewness coefficients for hourly workers. (See also Crawford and Harrison (1997) and Christofides and Stengos (1998) for skewness estimates of wage changes in Canadian union contract data.) The subsequent literature focuses on the complementary question by estimating ONE LITERATURE, TWO QUESTIONS 1 Card and Hyslop (1997, pp. 7576) conjecture that I failed to detect nominally induced asymmetries because I pooled annual distributions of real-wage changes. But pooling the distributions had no bearing on my conclusions. I reported the spike at zero and positive skewness; I also found that wage changes vary closely with inflation, skewness does not vary with inflation, and wage changes of nonunion and nonminimum-wage workers are symmetric (McLaughlin 1994). These findings led me to conclude that there was little evidence of downward nominal wage rigidity. One can go a long way toward reconciling disparate conclusions in the literature by drawing a single distinction. That is, one must distinguish between the level or frequency of nominal wage cuts and the sensitivity of nominal cuts to downward rigidities. There is a distinct possibility that nominal wage cuts are common and wage changes are skewed away from wage cuts. The first property of nominal wage changes answers the question, “How common are wage cuts in nominal terms?” The second property answers a complementary but distinct question: Is there evidence of downward rigidity reducing the frequency of nominal wage cuts? One could conclude that nominal wage cuts are common and that they would be more common if nominal wages were not downwardly rigid. Wage Cuts Are Common, But Are They Common Enough? In McLaughlin (1994), I documented that real-wage cuts are frequent and nominal wage cuts are not rare. Using survey-week data from the Panel Study of Income Dynamics (PSID), I found that 43 percent of workers who do not change employers (i.e., stayers) suffer real cuts in straight-time pay (hourly wage or salary) on the main job. For about 17 percent of the sample, the wage cuts are nominal. The subsequent literature focuses on nominal wage changes. Research by Lebow, Stockton, and Wascher (1995) using the PSID; Craig (1995) using F E D E R A L R E S E R V E B A N K O F S T. L O U I S 118 M AY / J U N E 1 9 9 9 whether nominal wage cuts are too rare; that is, whether the left side of the distribution is too thin below zero (Lebow, Stockton, and Wascher 1995; Card and Hyslop 1997; Kahn 1997). Lebow, Stockton, and Wascher’s measure of skewness subtracts the proportion of observations below zero from the proportion above twice the median. Since zero and twice the median are the same distance from the median, this thinness measure would be zero for symmetric distributions and positive for right-skewed distributions. Their skewness measure is 6.8, so the left tail below zero is about 7 percentage points thinner than its mirror image on the right side of the distribution. Card and Hyslop also assume that the right tail would be the mirror image of the left in the absence of nominal rigidity. For each year, they provide kernel estimates (basically smoothing) of the actual and counterfactual histograms and find a range of thinning from 6 to 14 percentage points, depending on the year. These are in line with Lebow, Stockton, and Wascher’s estimate of 10-percentage-point thinning for hourly workers. By using year-to-year variation in the position of the wage change distribution, Kahn (1997) estimates the extent of nominal rigidities without imposing the mirrorimage assumption. One checks whether bars of the wage-change histogram tend to be shorter in years when those bars lie below zero. For instance, is the third bar below the median shorter in those years when it lies below zero? Kahn’s regression estimates, which weight the effects across bars, imply that 9 percent of hourly workers’ would-be wage cuts are censored at zero. The only evidence of downward rigidity for salaried workers is early in her sample period (before 1982), but this evidence appears to be dominated by more frequent than expected nominal cuts after 1982. With the two questions distinguished, the papers in this literature share much in common. Nominal wage cuts are not rare, but there is evidence of a spike at zero, positive or right skew of distribution, and thinning of the distribution below zero. The evidence is much weaker, however, for salaried workers and nonunion workers. Another common feature is that about three-quarters of would-be wage cuts actually occur. Removing downward nominal rigidity would increase the frequency of reported wage cuts of stayers from the observed 17 percent to 22 percent (using Kahn’s estimates of thinning) or from 18 percent up to 24 percent (using Lebow, Stockton, and Wascher’s estimates of thinning). Even in the CPS sample of hourly wage workers, about three-quarters of the predicted wage cuts appear in the data (Card and Hyslop 1997). If nominal wages are downwardly rigid, then there should be less skewness and thinning in high inflation periods. The evidence on this important point is mixed. I found that the skewness coefficient is positively correlated with anticipated and unanticipated inflation, which is not consistent with inflation relaxing the impact of nominal wage rigidity (McLaughlin 1994, note 12). Lebow, Stockton, and Wascher’s (1995) thinness measure is not significantly correlated with inflation on their sample of all stayers, although the correlation is negative for hourly workers. On the CPS sample of hourly workers, Card and Hyslop (1997) find less thinning in high inflation periods. Hence any evidence of inflation reducing the impact of nominal wage rigidity is limited to hourly workers. But Are Wage Cuts Common? Perhaps survey reports of wages are riddled with error and the appearance of nominal wage cuts is illusory. The wage variables in these surveys refer to straighttime pay on the main job during the survey week or the most recent pay period; wages are not generated by dividing annual earnings last year by annual hours worked. Indeed, for the SIPP data on hourly workers, a respondent is asked what the hourly wage rate was on his last pay stub (Craig 1995). But the role of measurement error in inflating reported wage cuts deserves scrutiny. To identify the frequency of true wage cuts, I estimated the variance of the measurement error component and applied a mean-preserving compression to correct F E D E R A L R E S E R V E B A N K O F S T. L O U I S 119 M AY / J U N E 1 9 9 9 the distribution of wage changes (McLaughlin 1994). I took three very different tacks to estimate the error variance: industry, occupation, location, and year. Since very few of the reported wage cuts align with union wage concessions, Shea concludes that most nominal wage cuts in the PSID are attributable to measurement error. A problem with Shea’s method is that wages of union workers change without corresponding changes in union pay scales. Union wages are typically assigned to jobs, and workers regularly move from job to job in some union firms. This was the case in the large manufacturing firm used in the PSID’s Validation Study; indeed, workers in the validation study had great difficulty reporting hourly wages because they changed job assignments week-to-week and even day-to-day (Bound, Brown, Duncan, and Rodgers 1994).3 Altonji and Devereux (1997) provide maximum-likelihood estimates of an empirical model of wage rigidity and measurement error. They exploit cross-sectional variation in the position of the wage change distribution to estimate thinning. That is, Altonji and Devereux replace Kahn’s (1997) time-series variation with cross-sectional variation in the distribution’s position, and they add a distributional assumption in the process. In addition, Altonji and Devereux simultaneously estimate the variance of the measurement component. For their model to account for the spike at zero nominal wage growth, small wage cuts must be censored. Hence, the presence of small wage cuts in the data must (in their model) be attributed to measurement error. Altonji and Devereux conclude that about 80 percent of observed wage cuts are an artifact of measurement error; however, this would overstate the extent of measurement error if some small wage cuts were genuine. The extent of measurement error is important for assessing the frequency of nominal wage cuts but not for the complementary question posed in this paper: Is there evidence of downward rigidity reducing the frequency of nominal wage cuts? Adding a symmetric measurement-error component would not bias any of the skewness tests or thinning measures. • First, I drew reliability measures of wage change variables from validation studies (Bound and Krueger 1991; Bound, Brown, Duncan, and Rodgers 1994). • Second, I used the frequency of reported nominal wage cuts of minimum-wage workers, which were assumed to be due to measurement error. • Third, I associated measurement error with the stationary component of wage change residuals, and the random walk component of wages was classified as true variation. 2 Akerlof, Dickens, and Perry (1996, note 10) criticize my use of a mean-preserving compression to correct the distribution of wage changes. Their criticism builds from the assumption that the true distribution of wage changes is asymmetric. Wage changes are not severely skewed, however, and Akerlof, Dickens, and Perry do not assess how sensitive my corrections are to mild asymmetries. In addition, one can redo my calculations on symmetric subsamples without affecting any of my conclusions (McLaughlin 1999). 3 This point was driven home recently when a colleague of mine reported that his salary for the new academic year had dropped 25 percent. The union pay scale governing his employment has not changed in years, but he had completed his term as acting dean. The three methods pointed to a single conclusion. Although significant measurement error is present in wage changes, wage cuts remain fairly common in the corrected distributions.2 In particular, my most conservative measurement error correction reduced the frequency of nominal wage cuts from 17 percent to 12 percent. Suspicious that measurement error is the source of reported nominal wage cuts, Akerlof, Dickens, and Perry (1996) asked respondents in a Washington, D.C., telephone survey whether they had experienced wage cuts during the previous year. That is, rather than differencing wage responses across years, Akerlof, Dickens, and Perry asked a single qualitative question directly. About 3 percent of the stayers reported cuts in base pay. Akerlof, Dickens, and Perry conclude that the frequency of wage cuts in panel data is an artifact of measurement error. It is well known, however, that survey respondents under report embarrassing personal information, so this survey instrument probably undercounts wage cuts. In a comment on Card and Hyslop (1997), Shea (1997) assesses the reliability of wage reports of union workers in the PSID. Shea matches union workers in the PSID to union wage settlements by F E D E R A L R E S E R V E B A N K O F S T. L O U I S 120 M AY / J U N E 1 9 9 9 WAGE CHANGES IN THE PSID Table 1 To determine whether wage changes are skewed away from wage cuts, I use data from the PSID, which has followed thousands of households since 1968. The PSID includes annual observations covering survey week pay with the main employer. My measure of wages is the respondent’s report of his survey week pay on his main job. For hourly workers, I use the straighttime hourly wage rate. For salaried workers; I do not convert salaries into hourly wage rates; hours-induced wage variability might mask salary rigidity. This is particularly important in light of the errors in reported hours of work (Bound, Brown, Duncan, and Rodgers 1994). Wage changes are the annual differences of log wages times 100. My data set combines the PSID’s 1992 cross-year individual file with 22 annual family files. Because downward wage rigidities are not expected to be important for workers who change employers, the sample is limited to household heads (since 1971) and spouses (since 1979) who stayed with their employers since the previous year (i.e., stayers). To be included in the sample, a worker must also report his wage in adjacent years.4 In the resulting sample, 5,887 persons contribute 34,633 person-year observations on wage changes —an average of nearly six wage-change observations per person. Nominal Wage Changes and Inflation Panel Study of Income Dynamics, 1971-92 Inflation Process ARIMA (0,1,1) AR(3) ARIMA (0,1,1) Variable Intercept Inflation 3.309 (.602) .840 (.103) Anticipated Inflation Unanticipated Inflation R2 No. of Observations Unit of Observations .777 21 Annual Avg. 3.063 (.589) .880 0(.101) .584 0(.176) .810 21 Annual Avg. 2.847 (.645) .928 0(.113) .592 0(.183) .805 21 Annual Avg. a Least-squares regressions with standard errors in parentheses. Nominal wage changes are for stayers in the PSID; inflation is based on the GDP Deflator. All variables are computed as annual differences of logs. b Additional regressors include years of age and education, as well as indicators of sex, race, marriage, disability, occupation, industry, and union status, and change in union status; an intercept is also included. of stayers, I compute the average rate-ofchange of nominal wage over the previous year. Table 1 contains the results of regressing this annual wage change variable (in percentage terms) on the rate of inflation, based on the GDP Deflator. Consistent with my earlier findings (McLaughlin 1994, p. 403), nominal wages move closely with the price level. The estimated effect is .84, which is 1.6 standard deviations from one, so the hypothesis that nominal wages and prices move one for one is not rejected. The test is not decisive, so a suspicion of incomplete indexing of wages to prices might remain. However, suspicion of incomplete indexing or money illusion vanishes if the inflation rate is partitioned into anticipated and unanticipated components. To reach this conclusion, I use time-series methods to generate one-step-ahead forecasts of inflation (i.e., anticipated inflation) and forecast errors (i.e., unanticipated inflation). Like Pearce (1979), and Fama and Gibbons (1984), I find that the inflation rate is the sum of a random-walk component and a stationary component; in par- Do Wage Changes Reflect Money Illusion? The bigger issue is money illusion. Downward wage rigidities associated with infrequent wage cuts would constitute money illusion, but placing the issue in a wider context is essential. The broader question is: How do nominal wages move with the price level? Or equivalently: How does the rate-of-change of nominal wages vary with the inflation rate? Perhaps some nominal wage cuts are censored at zero, but the overall distribution of wage changes could be tightly linked to inflation. This is the case in longitudinal data from the PSID. For each annual sample F E D E R A L R E S E R V E B A N K O F S T. L O U I S 121 .918 0(.049) .717 0(.100) .018 34,633 Individual b 4 Also excluded are top-coded observations, which were common during the mid-1970s. See McLaughlin (1999) for details of the sample exclusions. M AY / J U N E 1 9 9 9 5 The result also is confirmed in aggregate data (e.g., Startz, this issue). The principal advantage of using panel data is in controlling for sample composition; in addition, using panel data at the individual level allows for additional regressors. 6 Ahmad and Li (1997) have proposed a formal test of symmetry that is closely related to symmetrically differenced histograms, and their test has been applied to Canadian wage-change data by Christofides and Stengos (1998). The test amounts to summing the squares of the values of the symmetrically differenced histogram. Ahmad and Li prove that the kernel-based test statistic is asymptotically normal. Subtracting the Left Side from the Right Side ticular, the first difference of the inflation rate follows a first-order moving-average process (i.e., MA(1)), so the inflation rate is an integrated first-order moving-average (i.e., ARIMA(0,1,1)) process. (Details are available on request.) Table 1 displays the results of regressing nominal wage changes on anticipated and unanticipated inflation rates. Nominal wage changes move one-for-one with anticipated inflation, and there is a strong (but weaker) positive relationship between nominal wage changes and unanticipated inflation. The results are robust to alternative specifications of the time-series process. For instance, in Table 1, the relationship between nominal wage changes and inflation components is essentially unchanged if a third-order autoregressive model is used to compute anticipated and unanticipated inflation rates. So incomplete indexing of nominal wages to prices appears to reflect that inflation is not fully anticipated. This result is confirmed on individuallevel data from the PSID.5 Here I regress an individuals’ nominal wage change on anticipated and unanticipated inflation rates, as well as on the labor economist’s standard set of regressors (including years of age and education and indicators of sex, race, occupation, industry, and union status). The estimated effects of the inflation components, which are reported in Table 1, confirm the pattern. From the relationship between nominal wage changes and components of inflation, there is no evidence of money illusion. In terms of monetary policy, this suggests no role for moderate inflation in greasing the wheels of the labor market. Suppose, however, that one detects evidence of nominal rigidity by focusing on the nominal wage change distribution around zero. Such evidence of money illusion might be treated as a problem to be solved by monetary policy, but downward rigidity around zero should be treated as a higher-order problem. The wider context—that the overall distribution of nominal wage changes moves one-for-one with anticipated inflation—must not be dropped. Visual inspection of a distribution can sometimes reliably detect departures from symmetry. But visually inspecting the histogram, matching bars on each side of the distribution, can be misleading if departures from symmetry are not severe. This is particularly important in the current context, where stayers’ wage change histograms in the PSID appear to be fairly symmetric (McLaughlin 1994). To aid the eye, I present symmetrically differenced histograms. I compute the histogram based on equal-sized intervals around the median, flip the left side of the distribution over onto the right, and difference the two.6 Here I subtract the left from the right, so a positive (negative) value indicates that the right (left) side of the distributions is thicker than the left (right) side over that particular interval. Since the histogram sums to 100 percent (with 50 percent on each side of the median), the symmetrically differenced histogram sums to zero. For a sample drawn from a symmetric distribution, the symmetrically differenced histogram would be a scatter of points along the zero line. Alternatively, suppose the left side of the nominal wage change distribution was thinned below zero, with these would-be wage cuts piling up at zero. Such a density function is illustrated in the upper-left panel of Figure 1, and its symmetrically differenced histogram in the lower-left panel. Since the spike at zero lies below the median, the symmetrically differenced histogram has a negative spike at twice the median. (A reference line is drawn at the median.) Values are zero up to twice the median, and positive values beyond this point reflect that the left tail is thinner than the right. In addition, perhaps some small wage increases or decreases are censored at zero. The right two panels of Figure 1 illustrate such a density function and its symmetrically differenced histogram. By censoring half of the small wage changes, two positive spikes surround the negative spike in the lower-right panel. So if nominal rigidities F E D E R A L R E S E R V E B A N K O F S T. L O U I S 122 M AY / J U N E 1 9 9 9 Figure 1 Censoring Wage Cuts and Small Wage Changes -45 -30 -15 0 15 30 45 -45 60 -30 Some Wage Cuts Censored 15 30 45 60 20 30 40 50 -4 0 0 10 20 30 40 50 -2 -4 -6 median 10 Fraction X 100 0 -2 -6 0 2 median Fraction X 100 2 0 -15 Some Wage Cuts and Small Wage Changes Censored -8 -8 Symmetrically Differenced Histogram censor some wage cuts and some small wage changes, the symmetrically differenced histogram should resemble the lower-right panel of Figure 1. For the sample of stayers in the PSID, the symmetrically differenced histogram in Figure 2 does resemble the lower-right panel in Figure 1. (The stayers’ histogram of wage changes is also depicted in Figure 2.) Small positive spikes surround a large negative spike at twice the median, which indicates some censoring of small wage changes. Values tend to be positive for wage changes beyond twice the median, which indicates that, below zero, the left side of the distribution is thinner than the right. But this symmetrically differenced histogram reveals more than partial censoring of small wage changes and wage cuts. Two negative values appear just to the right of the median reference line. Thus, wage changes just to the left of the median are more common than those just to the right of the median. This prop- Symmetrically Differenced Histogram erty is common for right-skewed distributions, such as the log-normal. But the presence of this property in the context of wage changes is important. There is more to the skewness of the wage change distribution than is implied by the censoring of some wage cuts and some small wage changes. Test Statistics Visual evidence from the symmetrically differenced histograms can be put to formal tests by computing the skewness coefficient, the mean-median difference, and the sign test statistic. On a sample of size N from distribution function F, I test the null hypothesis H0 that the distribution of wage changes x is symmetric; that is, F(x)=1–F(–x) for all values of the random variable x. The skewness coefficient, perhaps the most familiar measure of symmetry, is the ratio of the third central moment of x to F E D E R A L R E S E R V E B A N K O F S T. L O U I S 123 M AY / J U N E 1 9 9 9 Figure 2 tions severely inflate the variance of the test statistic; that is, fat tails produce high-variance skewness coefficients. This explains why skewness coefficients applied to wage change distributions, which have fat tails (McLaughlin 1994), tend to jump around from sample to sample (Crawford and Harrison 1997) and are sensitive to tail observations (Lebow, Stockton, and Wascher 1995). Properly computed standard errors reflect this, but long-tails render the test weak in detecting even strongly skewed alternative distributions. Indeed, the skewness coefficient has trouble detecting the asymmetry of the log-normal distribution (McLaughlin 1999). Table 2 provides a good illustration of this problem with the skewness coefficient. Despite the apparent skewness of the wage change distribution in the PSID, a test based on the skewness coefficient fails to reject the null hypothesis of symmetry. The estimated skewness coefficient is .36, but its standard error is .30; however, since the distribution of wage changes has fat tails, this test’s failure to detect skewness is not surprising. Most tables of critical values for the skewness coefficient are based on a normal distribution of x. If the distribution of x has fat tails, such critical values are biased toward zero, which generates a problem with false positives. For comparison, Table 2 also includes the standard error of the skewness coefficient under normality. This standard error is as small as .01. Symmetry is clearly rejected, but at this point, whether the rejection is valid or spurious remains unknown. A simple by-product of positive (negative) skewness is that the mean lies to the right (left) of the median, which motivates the mean-median difference (Hotelling and Solomons 1932) as a test. Under the null hypothesis of symmetry, the difference between the mean and the median is expected to be zero; furthermore, if the median m is treated as known, then by the Central Limit Theorem, the meanmedian difference is asymptotically normal with variance s2/N. The mean- Distribution of Nominal Wage Changes for Stayers 0.12 Fraction 0.09 0.06 0.03 0 -40 -30 -20 -10 0 10 20 Histogram 30 30 40 40 50 0 0 10 20 50 -2 -4 median Fraction X 100 2 -6 -8 Symmetrically Differenced Histogram the cubed standard deviation of x. Under the null hypothesis H0 , the asymptotic distribution of the skewness coefficient is normal with mean zero and variance: 9 µ6 − 6 µ 4 s 2 + N Ns 2 where s is the standard deviation of x, and µk denotes the kth central moment of x (Gupta 1967). If F were the normal distribution, then the variance of the skewness coefficient would simplify to 6/N. As a test of symmetry, the skewness coefficient has problems with false negatives. First, some skewed distributions have zero third moments (Mood, Graybill, and Boes 1974, p. 76). Second, because the skewness coefficient is sensitive to tail observations, the fourth and sixth moments associated with fat-tail distribu- F E D E R A L R E S E R V E B A N K O F S T. L O U I S 124 M AY / J U N E 1 9 9 9 Table 2 median difference reported in Table 2 is .81 percentage points, which clearly rejects symmetry in favor of a wage change distribution skewed away from wage cuts. Indeed, the mean is estimated to be 8.4 standard errors to the right of the median. What lies between the mean and the median also contributes to a test of symmetry. The sign test statistic (e.g., Gastwirth 1971) counts (and signs) the observations between the mean and median. Under H0, the number of observations below the mean is distributed binomial with parameters N and 1/2. By invoking the normal approximation to the binomial, the number of observations between the mean and the median is approximately normal with mean zero and variance N/4. In Table 2, results of the sign test applied to wage changes in the PSID also reject the null hypothesis of symmetry. In a sample of this size, it would not be surprising to find 100 or so observations between the mean and the median. More than 1,200 observations lie between the mean and median, however. Table 2 also contains a check of the sensitivity of the symmetry tests to the presence of the spike at zero. If all the zero-wage-change observations were would-be wage cuts or small wage changes censored at zero, these observations would belong in the calculations. Alternatively, these observations could reflect the rounding of wage responses or the timing of the survey (with some wages changing soon after the survey date). Neither factor constitutes an asymmetry of wage changes. By removing the spike at nominal zero, the skewness tests can isolate the contribution of “thinning” the distribution below zero. In the lower half of Table 2, I report the skewness test statistics on the sample that excludes observations with no change in nominal wages. All three skewness test statistics fall, providing weaker rejections of symmetry. But wage changes remain skewed right. Skewness Test Statisticsa Panel Study of Income Dynamics, 1971-92 Sample Including the Spike at Zero ±48 Point Band Around the Median ±5 Point Band Around the Median Excluding the Spike at Zero ±48 Point Band Around the Median ±5 Point Band Around the Median a Skewness Coefficient 0.335 (0.298) [0.013] 0.077 (0.024) [0.013] 0.058 (0.012) [0.020] 0.248 (0.291) [0.014] –0.023 (0.024) [0.014] 0.151 (0.012) [0.020] Mean – Median 0.811 (0.096) Sign Test 1,242.5 (93.1) Thinness Measure 7.89 0.687 (0.070) 1,068.5 (91.8) 7.85 0.065 (0.022) 200.5 (59.8) 0.586 (0.104) 837.5 (89.2) 4.11 0.470 (0.075) 725.5 (87.9) 4.04 0.206 (0.023) 302.0 (60.1) The sample contains 34,633 observations on the wage changes of firm stayers. Standard errors of the test statistics are reported in parentheses; displayed in brackets are the standard errors of the skewness coefficient under the assumption of normality. ness away from wage cuts. To address this question, Table 2 includes Lebow, Stockton, and Wascher’s (1995) measure of thinning, the proportion above twice the median minus the proportion below zero. The wage change distribution below zero is nearly 8 percentage points thinner than its mirror image on the right side of the distribution. If zero wage change observations are excluded, the thinness measure falls to about 4 percentage points. Intertemporal variation in the wage change distribution provides another way to identify thinning of the distribution below zero (Kahn 1997). This idea is simple and powerful. Take an interval a few percentage points below the median. When nominal wage growth is high (i.e., when inflation tends to be high), that interval lies above zero. In low nominal wage growth years, that interval might lie below zero. (In some years, the interval might span zero.) Kahn’s idea is to compare the histogram’s values for that interval when it lies to the right and left of zero. Thinning of the Distribution Below Zero While these tests reveal skewness, they do not address the question of skew- F E D E R A L R E S E R V E B A N K O F S T. L O U I S 125 M AY / J U N E 1 9 9 9 7 Composition of my sample from the PSID does vary: The universe of wage respondents widened beginning in 1976, and household spouses are included beginning in 1979. 8 Unfortunately, properties of symmetry test statistics can reverse as the band around the median narrows. For instance, with a log-normal distribution, as the band around the median becomes very narrow, the skewness test statistics change sign. The ±5 percentage point band appears to be wide enough to avoid this problem, but I prefer to rely on the clear evidence in Figure 2. If the value of the histogram on the interval is smaller when it lies below zero, there is evidence of wage changes being skewed away from wage cuts. Kahn estimates econometric specifications that essentially weight all the intervals that move above and below zero. But her idea can be implemented most directly by picking a few intervals that pass zero. My symmetrically differenced histograms use intervals two percentage points wide. The third and fourth intervals below the median (i.e., 4-6 and 6-8 percentage points below the median) lie above zero in high inflation years and below zero in low inflation years. For each interval, Table 3 reports values of the histogram in years when the interval was above and below zero. Wage change observations on interval 3 are 2.1 percent more common when that interval lies above zero. For interval 4, the difference of the histogram values is only .4 percent. As with any difference estimator, there is the question of a control group. If the sample in low wage-growth years has lower variance of wage changes, the tails of the distribution would be thinner even if would-be wage cuts were not censored at zero. Perhaps the composition of the sample differs when nominal wage changes are higher, or as Card and Hyslop (1997, p. 86) argue, perhaps the dispersion of wage changes is affected by inflation.7 This issue can be addressed with a difference-in-difference estimator. I use the change in the corresponding interval above the median as the control. Difference-in-difference estimates in Table 3 are a bit larger: 2.5 on interval 3 and 1.9 on interval 4. Overall, these histogram difference estimates do point to a thinning of tails below nominal zero, with a thinning of one-third to one-half of would-be cuts near zero. these observations would not be affected by any downward nominal rigidity at zero. That is, the distribution’s right side is predicted to be heavier than its left side over the range of wage cuts, but not over the entire range of wage changes. These implications can be checked directly. First, to check whether skewness is limited to tail observations, I exclude wagechange observations that contribute to the bottom and top histogram bars in Figure 2. This eliminates 1.19 percent from the left side and 1.44 percent from the right, so these tail observations contribute to the right skew. Skewness test statistics are computed on the remaining wage-change observations, which lie within 48 percentage points of the median. The results in Table 2 reveal that positive skewness survives trimming the tails. Extreme wage change observations do not account for all of the skewness, a result consistent with downward nominal rigidities. Second, based on the symmetrically differenced histogram in the lower panel of Figure 2, wage changes just to the left of the median are more common than those just to the right of the median. This property is typical of right-skewed distributions such as the log-normal. In Table 2, on the sample of wage changes within 5 percentage points of the median, the test statistics reject symmetry.8 Since this band does not include wage cuts, the spike at zero, or small wage increases, the source of skewness of wage changes is not limited to the censoring of would-be wage cuts and small wage changes. Skewness seems to be a more general property of wage changes. Skewness close to the median presents a problem for the thinning estimates of Lebow, Stockton, and Wascher (1995) and Card and Hyslop (1997); these estimates rely on the mirror-image assumption that the distribution of wage changes would be symmetric if not for downward nominal rigidity. If the overall distribution were skewed right, which is consistent with the evidence close to the median, then the mirror-image assumption would be violated, and their estimates of thinning would overstate the extent of downward nominal rigidity. Are Censored Wage Cuts the Only Source of Skewness? If thinning of the wage-change distribution below zero were the only source of skewness, then (a) extreme tail observations would not be the main source of skewness, and (b) wage changes close to the median would be symmetric, since F E D E R A L R E S E R V E B A N K O F S T. L O U I S 126 M AY / J U N E 1 9 9 9 Table 3 How severe is the bias? This depends on how skewed the overall distribution would have to be to generate the observed skewness near the median. One could generalize from the estimates of skewness near the median to generate an overall distribution that is skewed for reasons unrelated to downward nominal rigidities. In particular, the BoxCox transformation that renders the observations close to the median symmetric could be applied to the overall distribution to generate a counterfactual distribution.9 A corrected estimate of thinning differences the proportions of wage cuts in the actual and counterfactual distributions.10 Although skewness near the median might not appear strong in Table 2, observations near the median are strongly skewed, more skewed than can be explained by even a log-normal distribution. In particular, sending the Box-Cox parameter to zero is not sufficient to produce symmetry near the median. The implied counterfactual distribution would be so strongly skewed that “corrected” estimates of thinning would identify upward nominal rigidity. Although these unreported results do not qualify as serious corrections, they are instructive. Skewness near the median is too severe to ignore; and estimates of thinning by Lebow, Stockton, and Wascher and Card and Hyslop probably overstate the extent of downward nominal rigidity substantially. Histogram Difference Estimates of Downward Nominal Rigiditiesa Panel Study of Income Dynamics, 1971-92 Sample Left of Median Years When Interval Is Above Zero Years When Interval Is Below Zero Difference Right of Median Years When Intervalb Is Above Zero Years When Intervalb Is Below Zero Difference Difference-in-Difference Interval 3 Interval 4 -6.29 -4.23 -2.06 -3.36 -2.96 -0.40 -4.43 -5.82 –0.39 -2.45 -4.57 -6.04 –1.47 -1.87 a Table entries are histogram values and differences in histogram values. The length of each interval is two percentage points. Intervals 3 and 4 lie 4-6 and 6-8 percentage points away from year-specific medians. b Histogram values to the right of the median are computed separately for years when the indicated interval to the left of the median lies above and below zero. tion periods. To test this, I compute each test statistic (i.e., skewness coefficient, mean-median difference, sign test, and thinness measure) annually. (The four test statistics are highly correlated across years.) The results of correlating the annual test statistics with inflation are reported in Table 4. The correlations are small and statistically insignificant, and only the thinness measure’s correlation is negative. So wage changes in the PSID contain no evidence of less skewed distributions of wage changes in years when inflation was higher. This result is robust to partitioning inflation into anticipated and unanticipated components. For inflation to grease the wheels, it must shift the distribution of wage changes away from zero, and this is more likely for anticipated inflation. Table 4 contains correlations of the skewness statistics separately with anticipated and unanticipated inflation. Based on the three main skewness measures, neither anticipated nor unanticipated inflation reduces the skewness of nominal wage changes. Only the thinness measure correlates significantly negatively with anticipated inflation. The negative correlations of inflation and anticipated inflation with the thinness Does Inflation Reduce Skewness Away From Wage Cuts? The evidence that wage changes are skewed is definitive, but inflation’s role remains to be determined. Does inflation render less restrictive a constraint against nominal wage cuts? Or, if the nominal rigidity at zero is the source of skewness away from wage cuts, does inflation reduce the skewness of the wage change distribution? By identifying how the test statistics vary with inflation, this can be checked directly. If inflation relaxes the constraint, then skewness of the distribution of wage changes would be less severe in high-infla- F E D E R A L R E S E R V E B A N K O F S T. L O U I S 127 9 The Box-Cox transformation is (y –1)λ / λ, where y is 1+∆w. For 0 ≤ λ < 1, this transformation is strictly concave, which reduces skewness to the right. 10 This is a difference-in-difference estimator. Wage-change observations close to the median constitute the control group, which are unaffected by the treatment of downward nominal rigidity. These observations are used to estimate how much thicker the right side would be than the left in the absence of downward nominal rigidity. This adjustment factor is then differenced from Lebow, Stockton, and Wascher’s or Card and Hyslop’s difference estimates. M AY / J U N E 1 9 9 9 not sensitive to skewness unrelated to downward nominal rigidity. The evidence of strong skewness near the median leaves little role for downward nominal wage rigidity, and this is confirmed by little evidence of negative correlations between the skewness test statistics and inflation. Alternatively, if I had found strong negative correlations with inflation, then the detected skewness near the median would not be sufficiently strong to account for overall skewness and thinning. The complementarity of these tests—near the median and correlations with inflation— is particularly useful in explaining differences across groups. With only downward nominal rigidity, groups with strong skewness should have strong negative correlations with inflation. This link would be broken if there were evidence—in the form of skewness near the median—of an intervening factor, another source of skewness. Table 4 Correlations of Skewness Statistics With Inflation Panel Study of Income Dynamics, 1971-92 Sample and Test Statistic All Workers Skewness Coefficient Mean-Median Difference Sign Test Statistic Thinness MeasureUnion Workers Skewness Coefficient Mean-Median Difference Sign Test Statistic Thinness MeasureNonunion Workers Skewness Coefficient Mean-Median Difference Sign Test Statistic Thinness MeasureHourly Workers Skewness Coefficient Mean-Median Difference Sign Test Statistic Thinness MeasureSalaried Workers Skewness Coefficient Mean-Median Difference Sign Test Statistic Thinness Measurea Inflation Anticipated Inflation Unanticipated Inflation -0.30 -0.07 -0.08 –0.33 -0.23 –0.14 –0.08 –0.50 0.10 0.36 0.27 0.31 -0.03 –0.55 –0.53 –0.60 –0.14 –0.69 –0.66 –0.68 0.34 0.36 0.34 0.27 -0.29 -0.21 -0.16 –0.10 -0.24 -0.11 -0.07 –0.12 0.06 0.18 0.16 0.05 -0.02 –0.04 –0.10 –0.65 -0.05 –0.37 –0.38 –0.79 0.06 0.72 0.60 0.31 -0.44 -0.14 -0.24 -0.27 -0.37 -0.08 -0.17 -0.06 0.13 0.12 0.14 0.42 Do Unions and Method of Pay Matter? My results cover the sample of stayers. Yet the literature has drawn a sharp distinction between hourly and salaried workers (i.e., by method of pay), and perhaps a distinction should be drawn between union and nonunion workers. Do unions and method of pay matter? Table 5 contains the skewness test statistics, including difference-in-difference estimates based on histogram shifts, by union status and method of pay. Wage changes of both union and nonunion workers are skewed right, although the wage changes of nonunion workers seem to be more highly skewed. This contradicts my evidence, based on the skewness coefficient, that wage changes of nonunion workers are symmetric (McLaughlin 1994). The high variance of the skewness coefficient in the presence of a fat-tailed wagechange distribution resolves the paradox. A sharper distinction emerges between hourly and salaried workers. The evidence of right-skewed wage changes for hourly workers is overwhelming; wage changes of salaried workers are also skewed, but the Correlations are computed on 21 annual observations. The sign test statistic is normalized by its standard deviation. 11 Since Card and Hyslop (1997) use the same thinness measure, their evidence of less thinning in high inflation periods suffers from the same bias. measure are probably biased down. The median wage change varies with inflation, hence thinning is measured farther out in the distribution’s tails in high inflation years. Maintain the mirror-image assumption, and the correlation of thinness with inflation detects inflation’s role in relaxing any constraint of nominal wage cuts. Drop the mirror-image assumption, and the correlation is almost surely biased down. Suppose the distribution of wage changes is skewed right for reasons unrelated to downward rigidity. Then moving farther out in the tails surely reduces the measure of thinness even in the absence of downward nominal rigidity. Consequently, the correlation of Lebow, Stockton, and Wascher’s thinness measure with inflation is biased down.11 Correlations of the other symmetry test statistics with inflation are F E D E R A L R E S E R V E B A N K O F S T. L O U I S 128 M AY / J U N E 1 9 9 9 Table 5 Skewness Test Statistics by Union Status and Method of Paya Panel Study of Income Dynamics, 1971-92 Sample b Entire Histogram Union Workers Sample Size Skewness Coefficient Mean – Median Sign Test Thinness Measure 28,013 0.087 (0.232) 0.226 (0.362) 1.674 (1.341) 0.049 (0.288) 0.615 (0.151) 0.773 (0.126) 1.106 (0.108) 0.326 (0.175) 286.5 (44.7) 721.5 (75.4) 839.5 (59.7) 161.0 (60.7) 26.23 0.90 1.09 28.24 2.55 0.60 10.85 NA 1.79 24.74 1.21 0.03 0.028 (0.023) 0.062 (0.015) 0.047 (0.017) 0.022 (0.020) 0.011 (0.043) 0.057 (0.028) 0.055 (0.031) 0.083 (0.038) 16.5 (31.0) 140.0 (46.9) 120.0 (42.0) 265.0 (35.3) Nonunion Workers 22,749 Hourly Workers 14,271 Salaried Workers 14,758 5 Point Band Around the Median Union Workers Difference-in-Difference Interval 3 Interval 4 23,845 Nonunion Workers 28,796 Hourly Workers 27,056 Salaried Workers 24,996 a 34,633 observations are the wage changes of firm stayers. Standard errors of the test statistics are reported in parentheses. b Observations that change union status are excluded from the analysis by union, and those that change method of pay are excluded from the analysis by method of pay. departure from symmetry is much weaker. For instance, thinning is less than half as severe for salaried workers. Table 5 also contains “near the median” skewness tests by union status and method of pay. Again, skewness near the median does not reflect downward nominal rigidity. Near the median, wage changes of union workers are symmetric, while those of nonunion workers are clearly skewed. This contrasts with the results by method of pay. Both hourly and salaried workers’ wage changes are weakly skewed right near the median. So, except in the case of union workers, there is evidence that the estimates of Lebow, Stockton, and Wascher as well as Card and Hyslop overstate the role of nominal rigidity in thinning the wage change distribution below zero. If the source of skewness were downward rigidity, then the skewness coefficients would be negatively correlated with inflation, and this is confirmed in Table 4 for union and hourly workers. On the union and hourly samples, correlations of three of the four test statistics with anticipated inflation (in the second column) are negative and fairly strong. There is no evidence of inflation relaxing downward nominal rigidity for nonunion and salaried workers. Integrating these results, I find that the skewness of union workers’ wage changes is all attributable to nominal rigidity. There is no evidence of skewness near the median, and the skewness statistics are negatively correlated with anticipated inflation. The source of the strongly skewed wage changes of nonunion workers, however, is not nominal rigidity. The skewness statistics are not correlated with anticipated inflation, and strong skewness of nonunion workers’ wage changes near the median confirms the result. The wage changes of hourly workers are strongly skewed right, and some of this is unrelated to nominal rigidity (based on skewness near the median); since anticipated inflation reduces the skewness of hourly workers’ wage changes, some of the skewness is a consequence of nominal rigidity. There is F E D E R A L R E S E R V E B A N K O F S T. L O U I S 129 M AY / J U N E 1 9 9 9 no evidence of downward nominal rigidity for salaried workers. Indeed, salaried workers’ wage changes exhibit only mild skewness; that mild skewness is present near the median, and none of the skewness statistics is negatively correlated with inflation. The union/nonunion comparison highlights the complementarity between the “near the median” and “inflation correlation” tests. Although the wage changes of nonunion workers are more skewed than those of union workers, correlations with inflation offer no evidence of downward nominal rigidity for nonunion workers. This would be anomalous if not for the evidence of strong skewness of nonunion workers’ wage changes near the median. And this constitutes evidence of an intervening factor, another source of skewness. more likely to be truncated by turnover. By analyzing the sample of stayers, we introduce a bias toward right skew in the distribution of wage changes. And the size of the bias is unknown. Third, pooling of samples with different distributions can generate spurious skewness. For instance, pooling samples of workers from different industries, as I have done herein, mixes the industry wage-change distributions. Mixing does not preserve symmetry (McLaughlin 1999), so pooling industry samples might skew the distribution of wage changes even if each industry’s distribution were symmetric. Although this bias is potentially serious, I find that it is not the source of skewness in the overall distribution (McLaughlin 1999). IF NOT NOMINAL RIGIDITY, THEN WHAT? WHAT DOES THIS MEAN FOR MONETARY POLICY? A theme emerges: There is more to skewness of the wage-change distribution than downward nominal rigidity. If the source of skewness is not nominal rigidity, then what is it? Consider three possibilities. First, perhaps there is an aversion to wage cuts in real terms, and this thins the left side of the distribution of real wage changes. Since inflationary expectations vary across employment matches, a focal point at zero real-wage change is not implied. Downward real-wage rigidity simply implies that wage changes would be skewed to the right. Second, self-selection skews wage changes to the right (Weiss and Landau 1984). The economic intuition is simple. We observe the distribution of accepted wage offers. Some wage offers are not accepted, and these are more likely to come from the left side of the wage-change distribution. Offer a worker a 20 percent wage increase, and he would be likely to accept it; offer that worker a 20 percent cut in pay and he would be likely to quit. Indeed, rather than offer a worker a 20 percent cut, the employer would probably just lay him off. Hence, wage changes from the left side of the distribution are My purpose has been to identify common patterns and themes in the burgeoning literature on wage changes in panel data. Using data from the PSID, I confirm that wage changes are skewed to the right, there is a spike at no change in nominal pay, and below zero, the left side of the distribution is thinner than the right side. These patterns have been identified in the literature, but I cast new light on the subject. I use several test statistics to detect asymmetry, and I show that, with the exception of the skewness coefficient, skewness tests are not sensitive to the choice of a test statistic. I identify the source of the problem others have found with the skewness coefficient: The fat tails of the wage change distribution tend to produce high-variance skewness coefficients. In response to Card and Hyslop’s (1997) criticism of Kahn’s (1997) histogram difference estimator of thinning, I estimate a difference-in-difference version of Kahn’s estimator, which strengthens her results. Since as much as half of the spike at zero change in nominal wages might be attributable to rounding errors and the timing of survey interviews, I retest for skewness excluding these observations. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 130 M AY / J U N E 1 9 9 9 The evidence of skewness is weaker, but the null hypothesis of symmetric wage changes is still rejected. There is more to the skewness of wage changes than can be attributed to downward rigidity at zero. Here I break with the literature. First, I place the issue of downward rigidity in a wider context by documenting that nominal wage changes move one-for-one with anticipated inflation. Second, checking the mirror-image assumption of Lebow, Stockton, and Wascher (1995), and Card and Hyslop (1997), I find that wage changes near the median are skewed. This implies that the estimates of thinning of nominal wage cuts by Lebow, Stockton, and Wascher, as well as Card and Hyslop, overstate the extent of downward nominal rigidity. It also means that wage changes are skewed for reasons unrelated to nominal rigidity, perhaps as a result of self-selection associated with efficient turnover. Third, confirming my suspicion that skewness is not attributable to nominal rigidity, I find that the skewness of wage changes is unrelated to inflation, anticipated inflation, or unanticipated inflation. Fourth, although nonunion workers’ wage changes are skewed to the right, the source of the skewness is not downward rigidity. Yet the weakly skewed wage changes of union workers do reflect downward rigidity, because union workers’ wage changes near the median are not skewed, and the test statistics on the union sample are negatively correlated with anticipated inflation. Few areas of economic research have more direct implications for economic policy than this one. The thorn in the side of the policy recommendation of stable prices (i.e., zero inflation) has been the labor market. Workers are subject to money illusion in the form of downward rigidity of nominal wages, or so the story goes. I do detect some nominal rigidity for union and hourly workers, but the magnitudes are smaller than others have found. And one must always remember the wider context: Nominal wages of workers who do not change employers do move one-for-one with anticipated inflation. Consequently, the labor market is not much of a thorn in the side of zero-inflation monetary policy. REFERENCES Akerlof, George, William Dickens, and George Perry. “The Macroeconomics of Low Inflation,” Brookings Papers on Economic Activity (1996), pp. 1-59. Altonji, Joseph, and Paul Devereux. “The Extent and Consequences of Downward Nominal Wage Rigidity.” Manuscript, (October 1997). Ahmad, Ibrahim, and Qi Li. “Testing Symmetry of an Unknown Density Function by the Kernel Method,” Journal of Nonparametric Statistics 7 (1997), pp. 279-93. Bound, John, Charles Brown, Greg Duncan, and Willard Rodgers. “Evidence on the Validity of Cross-Sectional and Longitudinal Labor Market Data,” Journal of Labor Economics 12, (July 1994), pp. 345-68. ________, and Alan Krueger. “The Extent of Measurement Error in Longitudinal Earnings Data: Do Two Wrongs Make a Right?” Journal of Labor Economics 9, (January 1991), pp. 1-24. Card, David, and Dean Hyslop. “Does Inflation ‘Grease the Wheels of the Labor Market’?” In Reducing Inflation: Motivation and Strategy, Christina Romer and David Romer, eds., National Bureau of Economic Research Studies in Business Cycles, v. 30, University of Chicago Press, 1997, pp. 114-121. Christofides, Louis, and Thanasis Stengos. “The Symmetry of the WageChange Distribution: Non-Parametric Tests Using Contract and Survey Data.” Manuscript, (July 1998). Craig, Ben. “Are Wages Inflexible?” Economic Commentary, Federal Reserve Bank of Cleveland (April 1, 1995). Crawford, Allan, and Alan Harrison. “Testing for Downward Rigidity in Nominal Wage Rates,” Price Stability, Inflation Targets and Monetary Policy, Bank of Canada (1997), pp. 179-218. Fama, Eugene, and Michael Gibbons. “A Comparison of Inflation Forecasts,” Journal of Monetary Economics 13 (May 1984), pp. 327-48. Gastwirth, Joseph. “On the Sign Test for Symmetry.” Journal of the American Statistical Association 66 (December 1971), pp. 821-23. Gupta, Milan K. “An Asymptotically Nonparametric Test of Symmetry.” Annals of Mathematical Statistics 38 (1967), pp. 849-66. Hotelling, Harold, and Leonard Solomons. “The Limits of a Measure of Skewness,” Annals of Mathematical Statistics 3 (1932), pp. 141-42. Kahn, Shulamit. “Evidence of Nominal Wage Stickiness from Microdata,” American Economic Review 87 (December 1997), pp. 993-1008. Lebow, David, David Stockton, and William Wascher. “Inflation, Nominal Wage Rigidity, and the Efficiency of Labor Markets,” Board of Governors of the Federal Reserve System, Finance and Economics Discussion Series: 94-45, (October 1995). F E D E R A L R E S E R V E B A N K O F S T. L O U I S 131 M AY / J U N E 1 9 9 9 McLaughlin, Kenneth. “Rigid Wages?” Journal of Monetary Economics 34 (December 1994), pp. 383-414. __________. “Testing for Asymmetry in the Distribution of Wage Changes.” Manuscript, (1999). Mood, Alexander, Franklin Graybill, and Duane Boes. Introduction to the Theory of Statistics, 3rd ed., McGraw-Hill, New York, 1974. Pearce, Douglas. “Comparing Survey and Rational Measures of Expected Inflation: Forecast Performance and Interest Rate Effects,” Journal of Money, Credit, and Banking 11 (November 1979), pp. 447-56. Shea, John. “Comment,” In Reducing Inflation: Motivation and Strategy, Christina Romer and David Romer eds. National Bureau of Economic Research Studies in Business Cycles, v. 30, University of Chicago Press, 1997, pp. 114-121. Startz, Richard. “Discussion of ‘Are Nominal Wage Changes Skewed Away From Wage Cuts?’” This Review, this issue. Weiss, Andrew, and Henry Landau. “Mobility and Wages,” Economics Letters 1,5 (1984), pp. 97-102. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 132 M AY / J U N E 1 9 9 9 Richard Startz is a professor of economics at the University of Washington. Commentary inflation rate. The argument for a zero inflation rate is that zero is a magic number for political reasons and reasons of transparency. The argument for 2 percent is that nominal wages are downward rigid, and that 2 percent allows for more flexible real wages. So McLaughlin’s paper bears precisely on the central question of mediumterm monetary policy. Second, using micro data is exactly the right way to answer this sort of question. It also is a lot of work. You have to worry about measurement error. You have to worry about exact definitions of survey data. And so forth. McLaughlin’s paper is very well done and deserves a great deal of appreciation for both the quantity and the quality of the work. By way of final preface, where does wage rigidity fit into macro? Specifically, in a recession, why don’t wages drop to clear the market? Let me give the old-fashioned answer. There are at least four places where wage rigidity fits: First, from an old-fashioned Keynesian viewpoint, the labor market is driven by the demand side as in: Richard Startz W hen I was an undergraduate, I was told that maybe a little inflation was a good thing because nominal wages are downward rigid. What’s more, back then this accounted in part for the Phillips curve being curved—being relatively flat at low inflation rates and steeper at high inflation rates. At low inflation rates, you had to have a lot of unemployment to get rid of a little inflation because nominal wage cuts are so painful. Downward wage rigidity is regarded as one of those truths that are so self-evident that there is no need to look at the data. Professor McLaughlin has actually looked at the facts and said, “If it’s so self-evident, why doesn’t it show up in the data?” Or, at the very least, said, “It’s not nearly so simple.” Let me outline six points for consideration. • Where does wage rigidity fit into macro? • Do nominal wage changes move onefor-one with inflation? • Are wage changes skewed? • Does the skewness change with inflation? • Is the spike at zero big? Does its size change with inflation? • What about the spread of the distribution? Does it change with inflation? w L = LD . p If the real wage doesn’t drop, too much unemployment results. In this situation, if nominal wages are downward rigid, then real wages surely will be. Second, consider a model in which the real money supply matters: perhaps quantity theory, perhaps Keynesian with a Pigou effect, or perhaps ISLM. (1) Let me begin by saying that Professor McLaughlin’s topic is really important, that it is really important to the Federal Reserve, and that it is really important at exactly this point in history. If you take a policy window of the last two or three years and the next two or three years (assuming a similar economy) a critical question for monetary policy is whether we should aim for a steady zero inflation rate or for a steady 2 percent (2) Y = v M or Y = C Y , M + I + G p p or ISLM (3) P = µW Suppose firms set prices as a markup over wages—maybe a competitive markup, maybe something different. Here, nominal F E D E R A L R E S E R V E B A N K O F S T. L O U I S 133 M AY / J U N E 1 9 9 9 data. This is neither better nor worse than using the usual macro data except that it is limited to 21 annual data points. We all agree that, in the long run, the real wage is neutral with respect to inflation. But the short to medium run matters, and 21 data points are not enough to answer the question. I’m not saying that the answer is wrong, just that you cannot get a definitive answer this way. I decided to do a very small amount of sensitivity analysis. The numbers in Table 1, 0.84 and 0.88, are the two numbers I think the author wanted to emphasize. Regressing wage on price inflation results in a coefficient that is marginally statistically different from one. The coefficient also says that 7 percent inflation lowers the real wage about 1 percent. I guess that’s a small effect—but I’d like to know if a sustained 7 percent inflation would lower the real wage 1 percent every year. We then split the effect into anticipated and unanticipated and see that the anticipated effect is somewhat smaller. Eschewing the daunting task of trying to replicate the micro data, I chose the first likely looking variable from the DRI (Data Resources, Inc.) database. The left-most panel of Table 1 shows McLaughlin’s numbers. The middle panel replicates his regressions using my data to demonstrate that the data differences are not important. In the right-hand panel, I augmented the author’s specification with very simple dynamics in the form of a Koyck lag. With the augmented specification, 7 percent inflation lowers the real wage by 2.3 percent, which is a lot. Even using anticipated inflation, the coefficient is statistically below one and economically is really far from one. This does not mean the right panel is better than the left; it just means that 21 annual data points are not the right way to answer this question. Are wage changes skewed? Absolutely! Some people have taken this as evidence that the lower tail is truncated. Of course, while truncation may imply skewness, skewness need not imply truncation. The author did something clever; he looked for skewness far from zero and found it. So while there is some truncation, there are other factors producing skewness. So we Figure 1 Illustration of Phillips Curves with Nominal vs. Real Wage Rigidity w w High π e Nominal wage rigidity Low π e High π e Real wage rigidity u u* wage rigidity causes nominal price rigidity and prevents the consumer price index from dropping to pull the economy out of recession. The third place where wage rigidity shows up is in the slope of the Phillips curve at low inflation rates. (See Figure 1.) If it is hard to force down wages, then the sacrifice ratio is really bad at low or negative inflation rates. One way to think about real vs. nominal rigidity is to ask whether the sacrifice ratio worsens specifically near zero or just at inflation rates close to expected inflation. Fourth, and last, rigidity just messes up allocational efficiency. The paper is excellent. Nonetheless, nobody gives a discussant credit for saying nice things; so I want to discuss the one part with which I disagree and then make a few suggestions for other ways of working with the data. The only part of the paper with which I take issue is the conclusion that “nominal wage changes move one-for-one with anticipated inflation, and are even closely linked to unanticipated inflation.” This is a question about averages, that is, macro data. The author has averaged his micro wage change F E D E R A L R E S E R V E B A N K O F S T. L O U I S 134 M AY / J U N E 1 9 9 9 Table 1 Regression of Wage Inflation on Price Inflation Variable Inflation Author Inflation Process Macro Data Inflation Process Macro Data Inflation Process ARIMA(0,1,1) ARIMA(0,1,1) ARIMA(0,1,1) AR(3) AR(3) .840 (.103) AR(3) .865 (.083) .666 (.123) Anticipated Inflation .880 (.101) .928 (.113) .876 (.082) .930 (.098) .722 (.139) .719 (.186) Unanticipated Inflation .584 (.176) .592 (.805) .678 (.161) .710 (.152) .554 (.191) .625 (.172) Lagged-Wage Inflation .284 (.139) .183 (.174) .238 (.186) AR(1) -.060 (.260) .169 (.331) .026 (.333) Notes on macro data: price = GDP price deflator (DRI GDNFPC) wage = Hourly compensation nonfinancial corp. (DRI LCPB) should be careful about interpreting skewness per se as evidence of wage rigidity. Is this skewness correlated with inflation? The author’s Table 4 says no. If skewness results from wage rigidity and inflation reduces wage rigidity, then inflation should reduce skewness. Strikingly, there is no evidence for inflation reducing skewness. I wonder if there is a way to estimate the power of these tests, in the economic rather than statistical sense. One suggestion is to write a simple model of nominal wage setting calibrated twice— once for parameters where all would agree that nominal rigidity is important and one with the opposite assumption. Then one would generate simulated data from each, compute the statistics in Table 4, and determine the extent to which the two simulation results differ. Is there a big spike at zero? There is a spike, but what metric tells us whether the spike is large? In addition to the size of the spike, the author discussed the amount of censoring. It would be useful to have a table or some guided comparison of the zero spike and estimated censoring against inflation to see if it tells the same story. Let me turn to the question of the variance of wage changes rather than the skewness. Take an example from life. I wear two hats. I spend 80 percent of my time as an economist and another 80 percent as department chair. I assign wage changes in my department, so for the University of Washington I know what the process is. The University of Washington faculty is about a third the size of the PSID sample. Straight-time wages are downward rigid. By this I mean there is no place on the form I complete to lower wages. It literally cannot happen. And I hazard the same thing is true for half the attendees of the October meeting. When I give raises I have a dollar-budget constraint as well as a non-negativity constraint. So downward rigidity cannot have any effect on the average wage change. The histogram in Figure 2 gives a fictional, but I think accurate, picture of what happens at different inflation rates. At a 10 percent average increase, there are maybe two raises at 6 percent to one raise at 18 percent. At a 4 percent average increase, there might be two raises at 2 percent to one raise at 8 percent. At high inflation, there’s a 12 percent change in relative real F E D E R A L R E S E R V E B A N K O F S T. L O U I S 135 M AY / J U N E 1 9 9 9 scientific work that is enormously informative on a fundamental policy question; in this case, “How much inflation should we have?” Figure 2 Illustrative (but Fictitious) Histogram of Wage Increases in High vs. Low Inflation Regimes Number of raises High inflation regime Low inflation regime 0 2 4 6 10 12 8 Rate of wage increase 14 16 18 20 wages. At 4 percent inflation there is only a 6 percent change in relative real wages. The former correctly reflects productivity. Thus, inflation does give us more efficient wage setting, but lower inflation in our case squeezes both tails. I’m not sure the squeezing of both tails would show up in any of the measures used in this paper. Further research might include comparing data from institutions with known nominal wage rigidity with data from institutions known to have flexible wages. During the discussion at the October meeting, Bill Poole offered a telling objection to my illustrative histogram. Some people are denied tenure; effectively, their wage change is –100 percent! Much of the wage rigidity literature carefully looks only at “nonmovers.” If we think carefully about the role of wages in labor markets, we know that labor force adjustments occur on both the intensive (hours) and extensive (hire/ fire) margins. McLaughlin’s paper focuses on nonmovers, as it should; but perhaps there is more work to be done on the linkage between wage rigidity and the intensive versus extensive margin of labor force adjustment. In summary, this is a stimulating paper that leaves the reader begging for more. It also is a great example of narrowly focused F E D E R A L R E S E R V E B A N K O F S T. L O U I S 136