Compensation Structure and Employer Learning∗ Bok Hoong Young Hoon 26 July 2010
Abstract In this paper I investigate whether employer learning about worker productivity depends on the choice of compensation structure. From an econometric standpoint, if employers are learning about worker productivity then, over time, measures that are correlated with productivity that are observed by the econometrician but not by the employer should play an increasingly important role in wage determination. This is the identi cation strategy used by Altonji and Pierret (2001). Applying this strategy to job spells in the Panel Study of Income Dynamics I nd that employers of performance pay jobs learn twice as fast as those of non-performance pay jobs. I nd that this result is driven entirely by learning in bonus pay jobs and that no learning occurs in commission/piece rate jobs. This latter result is consistent with the notion that explicit pay-for-performance compensation schemes are, by design, revealing workers' productivities.
1
Introduction
Researchers have, for some time, recognized the important role that imperfect information plays in the labour market. One strand of literature, employer learning, describes a process through which this informational de ciency is reduced. In models of employer learning it is assumed that employers set wages equal to workers' expected productivity. At the time of hiring, employers form a belief of a worker's ∗
I would like to thank Daniel Parent and Jennifer Hunt for their helpful comments and discussions
with many aspects of this paper. I am also grateful for the suggestions oered by participants of McGill's weekly PhD student seminar.
1
productivity based on a limited, but readily available set of information.
As the
employment relationship progresses, employers are able to update their beliefs by observing additional signals of the worker's productivity. A number of productivity signals have been considered in the employer learning literature. Waldman (1984) considers the job assignment of a worker in the current rm to act as a signal of the worker's ability to outside rms, while Greenwald (1986) and Gibbons and Katz (1991) both consider the problem of adverse selection, in which a worker's presence 1
in the secondhand (job-changing) market signals low ability.
More recent work in
this area considers output measures as signals of productivity (Farber and Gibbons, 2
1996; Altonji and Pierret, 2001; Pinkston, 2003, 2006).
Empirically, predictions of learning models tend to hold more strongly for certain groups of workers.
For example, Gibbons and Katz (1991) show that their model
holds for white-collar workers, while Bauer and Haisken-DeNew (2001) show that the implications of an Altonji and Pierret (2001) type learning model holds in German data only for blue collar workers at the lower end of the wage distribution. Similarly, Strobl (2004) nds that learning in Ghana's manufacturing industry occurs only for workers in jobs where there is no on-the-job training and who were hired through formal processes (i.e. workers who had no previous association with the rm or any of its employees). In this paper I apply the employer learning model of Altonji and Pierret (2001) to investigate whether employer learning is related to compensation structure.
In
particular I examine whether there is a connection between paying for performance and employer learning about unobserved worker ability. This is reasonable to expect since, as suggested by the performance pay literature, such compensation schemes are meant to provide workers with the incentive to produce at a level that re ects (and thus reveals) their true ability. Using data from an auto glass company, Lazear
1 Gibbons and Katz (1991) however make the distinction between a plant closure and a layo, noting that a lemons eect should apply only to those in the unemployed pool due to a layo.
2 Since output measures are generally not reported in commonly used datasets, these papers use
wages as signals of productivity. Under the assumption that wages equal expected output in each period, wages incorporate all available information, and thus can be used as signals equivalent to output.
2
(2000) nds that half of the increase in productivity that is observed upon switching to piece rate compensation is due to the incentive eects of this pay structure. Examining job spells in the Panel Study of Income Dynamics, Lemieux, MacLeod, and Parent (2009) show that show that compensation in performance pay jobs match more closely individual worker productivity than compensation in non-performance pay jobs. Finally, using the National Longitudinal Survey of Youth (NLSY79), Parent (1999) shows that most of the wage variation of piece rate workers is due to unobserved worker productivity. In assessing the relationship between performance pay and employer learning I address two questions. Firstly, do employers learn more quickly about workers who are paid for performance than about those who are not?
Secondly, does learning
about performance pay workers depend on the speci c way in which performance is rewarded? Using job spells constructed in the Panel Study of Income Dynamics (PSID) I nd that, while employer learning occurs in both performance pay jobs and non-performance pay jobs, learning in performance pay jobs takes place at a considerably faster rate. Further examination shows that the learning in performance pay jobs is driven completely by employers learning about workers in jobs which pay bonuses, and that there is no evidence of learning taking place in jobs where workers receive explicit forms of performance pay such as piece rates or commissions. The paper proceeds as follows. Section 2 outlines the employer learning model of Altonji and Pierret (2001). Section 3 describes the PSID sample used for the analysis. Section 4 presents the results on employer learning by compensation structure. Section 5 concludes.
2
Econometric Strategy
In this section I outline a simpli ed version of Altonji and Pierret's (2001) employer learning model. termined by
Assume that the productivity of worker
yit = exp{rsi + Λzi + ηi };
where
si
and
ηi
i
in period
t, yit ,
is de-
represent schooling level
and unobserved ability respectively, and are regularly addressed measures in labour
3
economics. In contrast,
zi
which denotes worker characteristics correlated with pro-
ductivity that are observed by the econometrician but not the employer, appears less commonly. Examples of such a measure are aptitude test scores such as the Armed Forces Qualifying Test (AFQT), and parents' education levels. Even though
zi
and
ηi
conditional expectations:
are not observed by the employer, employers know their
E (z | s)
and
E (η | s).
The true values of
zi
and
ηi
can
then each be expressed as the sum of their expected value given schooling and a random noise component,
v
and
e
respectively.
z = E (z | s) + v = γs + v
(1)
η = E (η | s) + e = αs + e
(2)
Using equations (1) and (2) one can see that worker productivity at the time of hiring is given by:
yt = exp {(r + Λγ + α) s + Λv + e} where
Λv + e
(3)
is the error in employers initial expectation of worker productivity.
This is the part of productivity that cannot be forecasted using
s.
Each period the
yt + εt . Since s is known, observing ωt = Λv + e + εt , a
employer observes a noisy signal of worker productivity observing the productivity signal is equivalent to
noisy signal of the error in the employer's initial expectation of productivity.
Let
Ωt = {ωt , ωt−1 , ωt−2 , . . .} represent the history of productivity signals so that Λv+e = E (Λv + e | Ωt ) + µt . Wages are equal to expected productivity implies:
where
wt = E (yt | s, Ωt ) = (r + Λγ + α) s + log (E (expµt )) + E (Λv + e | Ωt )
(4)
wt
Of
is the log wage in period
t and all other terms are as previously de ned.
critical importance is the last component of equation (4). This component indicates that employers are continuously making use of workers' productivity signals, encompassed in
Ωt ,
to update their beliefs on workers' unobserved ability,
Λv + e.
Taken
as a whole, equation (4) says that this updating process, which represents employer
4
learning, plays a role in the way workers' wages evolve. With data on
zi ,
the econometrician can estimate the conditional expectation:
E (wt | s, z, t) = bst s + bzt z.
(5)
A comparison of the actual wage generating process, equation (4), and the econometrician's speci cation, equation (5), suggests that OLS estimates of
bst
bzt
and
will be biased by the arrival of productivity signals which are unobserved by the 3
t, Bt , is given by the coe cients from −1 0 a regression of E (Λv + e | Ωt ) on s and z : Bt = (K K) K 0 E (Λv + e | Ωt ); where K is the row vector with elements s and z . Furthermore, by expanding B : econometrician.
"
Bst Bzt
#
The vector of biases at time
#−1 "
# cov (s, E (Λv + e | Ωt )) = cov (z, E (Λv + e | Ωt )) #" # " cov (s, E (Λv + e | Ωt )) σz2 −σsz 1 = σs2 σz2 − σsz σsz −σsz σs2 cov (z, E (Λv + e | Ωt )) "
σs2 σsz σsz σz2
s and E (Λv + e | Ωt ) cov (z, E (Λv + e)) = cov (v, E (Λv + e)),
one can see that, since
1 (−σsz ) cov (v, E (Λv + e | Ωt )) − σsz σsz 1 2 σ = cov (v, E (Λv + e | Ωt )) , s σs2 σz2 − σsz σsz
Bst = Bzt
are independent (equation (4)) and
(6)
σs2 σz2
indicating that all of the bias on
s
arises from its correlation with
correlated with the productivity signal.
(7)
z,
since
z
is
4
All of the studies (of which I am aware) that use this framework to identify employer learning, use measures of
z
(worker characteristics that are not observed
3 Altonji and Pierret (2001) also include a variable representing time-invariant information observed by the employer but not the econometrician.
The eect of this is an initial bias on the
coe cients.
4 Farber and Gibbons (1996) use the part of
z
that is uncorrelated with schooling (and all other
covariates) resulting in the absence of a time-varying bias on
5
s.
by the employer) that are positively related to
s
(education level). This study is no
exception. Under this condition, equations (6) and (7) indicate that the coe cient on schooling in equation (5) will be negatively biased (Bst
< 0)
and the coe cients on
worker characteristics unobserved by the employer will be positively biased (Bzt
> 0).
A key prediction of Altonji and Pierret's identi cation strategy is the time path followed by these biases. Equations (6) and (7) can be rewritten as
Bst Bzt
1 cov (v, E (Λv + e | Ωt )) = (−σsz ) cov (v, Λv + e) · σs2 σz2 − σsz σsz cov (v, Λv + e) cov (v, E (Λv + e | Ωt )) 1 2 = σ cov (v, Λv + e) · , σs2 σz2 − σsz σsz s cov (v, Λv + e)
(8)
(9)
where the rightmost term (in brackets) of each equation represents the amount the employer knows about unobserved worker productivity at time t. With the arrival of additional signals of worker productivity, this term increases until
Λv + e,
E(Λv + e | Ωt ) =
at which point the bias is at its maximum and learning ceases.
5
Notice that
cov(v,E(Λv+e|Ωt )) is responsible for the bias changing over time so that employer learning cov(v,Λv+e) ∂bst predicts 6 0 and ∂b∂tzt > 0. The basic intuition is that, as more is learnt about ∂t the worker's productivity, worker's wage, leading to
Λv + e becomes increasingly important in determining the biases on bst and bzt which gradually increase in absolute
value. Employer learning models predict that employers will learn faster if they observe cleaner, more informative signals of worker productivity.
6,7
It seems logical
to expect such signals to be observed more in performance pay jobs than in nonperformance pay jobs. This is because performance pay jobs, by de nition, rely to a great extent on measures of performance, and thus are more likely to implement monitoring and measurement devices, and use more indepth evaluation methods. In
50
cov(v,E(Λv+e|Ωt )) 6 1 and is nondecreasing in t. cov(v,Λv+e) 6 In Appendix 1 I show how, under Bayesian updating of employer beliefs, more informative (less
6
noisy) signals of productivity imply faster learning.
7 Pinkston (2003) uses a similar argument when comparing screening discrimination between
women and men: the noisier initial productivity signal of women compared to those of men implies that employers put less emphasis on this information when setting initial wages.
6
a case study examining performance based systems implemented by the pharmaceutical company Merck and Co., Kevin J. Murphy writes,
The key to any successful incentive program is the system that de nes performance and identi es high and low performers......The success of companywide incentive programs therefore depends on managers' eectiveness in monitoring and appraising their subordinates' performance
8
Framing this argument in terms of the econometric model outlined above indicates that, through the arrival of cleaner productivity signals, measures of productivity that are unobserved by the econometrician are incorporated into wages at a faster rate in performance pay jobs than in non-performance pay jobs.
As a result, the
biases given in equations (8) and (9) should increase faster in these jobs than in jobs in which compensation is not based on performance. Note however the following caveat: if performance is based on objective measures, then employers can set wages equal to realized productivity (rather than to expected productivity), and no employer learning takes place. In keeping with contract theory, I describe performance to be objectively measured if it is veri able. Such a measure then acts as a noiseless signal which employers can use to set an explicit contract of the form
w = a + by .
Examples of this type of performance pay job are commission
and piece rate jobs. Consider such a contract in which
a=0
and
b=1
so that log
wages are given by:
w = (r + Λγ + α) s + Λv + e re ecting the full realization of the worker's productivity. Biases on
(10)
bst
and
bzt
when
estimating equation (5) are then given by
1 (−σsz ) cov (v, Λv + e) − σsz σsz 1 2 = cov (v, Λv + e) . σ s σs2 σz2 − σsz σsz
Bs = Bz
σs2 σz2
(11)
(12)
8 Kevin J. Murphy, Performance Measurement and Appraisal: Motivating Managers to Identify and Reward Performance. in Performance Measurement, Evaluation, and Incentives, ed. William J. Bruns Jr., 37 (Boston, Massachusetts: Harvard Business School Press, 1992).
7
The absence of time components in equations (11) and (12) makes it clear that no employer learning occurs in these jobs.
3
The Panel Study of Income Dynamics
To investigate the predictions of the previous section I use information provided in the Panel Study of Income Dynamics (PSID) between 1976 and 2005.
9
This is a
point of departure from existing empirical investigations on employer learning in the United States which generally use data from the National Longitudinal Survey of Youth (NLSY79), a panel sample of workers who were between the ages of 14 and 22 in 1979. There are two reasons for the widespread use of the NLSY79; rstly the belief that most learning takes place during employment relationships early in a worker's career, and secondly, the reporting of measures of ability that are deemed unobserved by the employer (for example the AFQT, parents' education, and wages of siblings). Unfortunately however, the NLSY79 does not contain information su cient for the identi cation of performance pay jobs that is required for the present study.
10
The
PSID, a study following a representative sample of the US population since 1968, overcomes this problem through a series of questions asking respondents about their dierent forms of income. I discuss this in more detail later in this section. I consider an initial sample consisting only of male heads of household between
9 The PSID is available yearly from 1968-1997, after which it is available biennially. The reason I use data starting only in 1976 is that certain variables are only available from this time.
In
particular, variables that allow the identi cation of workers who received performance pay, and variables indicating employer and/or position tenure, which is used in the construction of job spells.
10 The NLSY79 asked whether workers received various forms of performance pay in 1988, 1989,
1990, and every two years from 1996 onward; I consider this too short a sampling period over which to investigate employer learning. From 1979-1987 households were asked about the receipt of tips, bonuses, or commissions, but it is not possible to identify the household member who received it. As an alternative I tried using the 1997 cohort of the NLSY (NLSY97) which samples individuals who were aged 12-16 at the beginning of 1997. However, although the NLSY97 collects detailed information on various forms of performance pay, because this panel was recently started, respondents are still quite young (21-25 years old in 2006). After restricting the sample to include only wage earners who had been out of school for at least 3 years (to ensure a minimum degree of labour force attachment) this resulted in overly small sample sizes.
8
1976 and 2005 who were either a member of a family in the rst wave of the panel, or moved into one of these families at a later point in time.
I include only heads
of households because much of the relevant information was only asked of the head of household. I consider only males in order to minimize on career interruptions in the data.
11
A problem with using the PSID to investigate employer learning early
on in a worker's career is that, because it is representative of the US labour force, it includes workers with a signi cant amount of labour market experience. To get around this problem while at the same time avoiding restrictions that produce overly small samples, I focus only on workers' rst jobs (in what follows I use the terms job and employment relationship interchangeably). This points to another shortcoming of the PSID - the absence of direct indicators of job spells and job changes. This issue is addressed by Brown and Light (1992) who evaluate dierent criteria by which to divide workers' labour market experience into individual job spells (Brown and Light refer to this as partitioning the data into individual job spells). In this paper I apply their preferred partition, which indicates a job change whenever tenure is 12
less than the time since the last interview, to identify each worker's rst job spell.
I further restrict my sample to include only full-time, private sector wage earners with hourly wages between $1 and $100 (1979$), and job spells for which there are at least 2 observations. My nal sample consists of 15,871 worker-year observations for 3045 workers.
13
Sample statistics are presented in column (1) of Table 1a.
As noted earlier, a crucial component of my data is information on performance pay. I de ne performance pay to include any piecework pay, bonus, or commission, and classify as performance pay any worker-year observation in which such payments were received.
Following work done by Lemieux, MacLeod, and Parent (2009), I
de ne a performance pay job as one in which performance pay is received at any point during the employment relationship. In doing this I am able to classify 819 of my 3045 jobs as performance pay jobs. Sample statistics for performance pay and non-
11 A secondary reason for sampling only males is that the number of female heads of household is relatively small and may not be representative of the female population.
12 This is Partition T in Brown and Light (1992). 13 See the Data Appendix for a detailed discussion of the sample construction; including details
of the procedure used to construct job spells.
9
performance pay jobs are presented in columns (2) and (3) of Table 1a. Consistent with the performance pay literature, performance pay jobs are associated with higher wages, greater degree of wage dispersion, longer job tenure, higher proportion of salaried workers, lower rate of unionization, and more education. To test the employer learning prediction that no employer learning should occur in performance pay jobs using objective measures of performance pay I further classify performance pay jobs as based on either an explicit contract (commission and/or piece rate) or an implicit contract (bonus job).
Using a series of questions in the
PSID it is possible to determine whether workers received compensation in the form 14
of piecework pay or commissions.
I classify jobs as being explicit contract jobs if
workers report receiving commissions or being paid by the piece; regardless of whether or not they also received a salary or an hourly wage. Following from the de nition of performance pay jobs, this implies that all implicit contract jobs are those in which workers received bonuses
and
no commissions or piece rate pay.
Summary
statistics for explicit contract jobs and implicit contract jobs are presented in Table 1b. Compared to workers in explicit contract jobs, workers in implicit contract jobs receive a higher hourly wage, have longer job tenure, and have parents with more education. In order to implement the econometric speci cation of Section 2, I require information on worker characteristics that are related to productivity but not observed by the employer.
Previous studies on employer learning using the NLSY79 make
use of mainly the AFQT and, to a lesser extent, father and siblings' education, and the presence of a library card in the household at age 14. While the PSID does not contain information on the AFQT, it collects information on respondents' mother's and father's levels of education; I use these as my measures of worker characteristics that are related to productivity but that are not observed by the employer. Tables 1a and 1b indicate that parents' education levels are higher in (1) performance pay jobs relative to non-performance pay jobs, and (2) in implicit contract jobs relative to explicit contract jobs.
14 See the Data Appendix for details on the procedure used to identify commission and piece rate workers.
10
4
Estimation Results
In this section I implement the identi cation strategy of Section 2 using the log wage equation:
ln wit = β0 + β1 si + β2 t + β3 si · t + β4 zi + β5 zi · t + Xit γ + εit , where
si
is the number of years of schooling,
t
(13)
is potential experience, and
Xit
is a
vector of covariates that includes a quartic in potential experience, year dummies, 15
and year dummies interacted with education.
zi
refers to worker characteristics
that are related to productivity but not observed by the employer. As discussed at the end of section 3, I use individuals' mother's and father's levels of education to represent such characteristics. Employer learning implies that
β3 6 0
and
β5 > 0.
Because I have multiple observations per worker, I cluster all standard errors at the worker level. In Table 2 I present estimates for the full sample of workers. education as my
zi
zi
I use father's
variable in columns (1) and (2), and mother's education as my
variable in columns (3) and (4).
16
Results obtained using either of these measures
suggest that employer learning is taking place. Coe cients on father's education in column (1) and mother's education in column (3) are signi cantly positive, indicating that unmeasured ability has a positive overall eect on wages.
Note that under
employer learning it is not assumed that parents' education level plays any part in wage determination. Rather the positive relationship between these
zi
measures
and wages arises because of the bias induced by their (parents' education) positive correlation with unmeasured ability. By allowing the relationship between parents' education and wages to vary with experience, columns (2) and (4) show that ability actually plays no role initially, but becomes more important with time. This shows up
15 I attempted to construct a relatively more precise measure of experience by determining the actual experience obtained after 1976 and adding it to the potential experience as of 1976. However, roughly half the sample is not observed in 1976.
16 I do not include both mother's education and father's education simultaneously in estimating
equation (13). This is because, as shown by Altonji and Pierret (1997), when not necessarily the case that
β3 6 0
or
β5 > 0.
11
zi
is a vector, it is
as positive coe cients on interactions of parents' education levels with experience, and represent biases on father's and mother's education that gradually increase. This is interpreted as the market accumulating, and incorporating into wages, new information on the worker.
4.1
Employer Learning and Compensation Structure
The estimates presented in Table 2 for the full PSID sample are consistent with employers learning about unmeasured aspects of worker productivity. From the discussion in Section 2 however, it is likely that, based on dierences in the quality of productivity signals, learning might be occurring at diering rates across compensation structures. If employer learning about worker ability is occurring dierently across compensation structures, then this should be evident in the experience path followed by measures of cross-sectional wage dispersion.
In particular, under em-
ployer learning, the dispersion of residuals constructed from cross-sectional wage regressions should increase with experience. In Figure 1 I show experience-speci c measures of residual inequality for three dierent types of compensation structures: those used in commission/piece rate jobs, those used in bonus jobs, and those used in non-performance pay jobs.
Residual inequality is measured by the variance of
the residual distribution. Residuals are constructed from basic wage regressions controlling for education and experience, including a quartic in experience. Overall the trends illustrated in Figure 1 suggest (1) that no learning is taking place in commission/piece rate jobs, and (2), that while learning is taking place in both bonus jobs and non-performance pay job it appears to be faster in bonus jobs. In this subsection I use equation (13) to formally investigate whether employer learning does in fact dier across compensation structures.
12
4.1.1 Learning in Performance Pay vs. Non-Performance Pay Jobs I rst look at dierences in employer learning between performance pay and nonperformance pay jobs.
I argue that while learning is expected in both types of
jobs, it should take place at a faster rate in performance pay jobs since productivity signals in these jobs are more informative (for reasons described in Section 2). I report estimates of equation (13) for performance pay jobs in Table 3a, and for non-performance pay jobs in Table 3b. These estimates are generally consistent with employers in each subsample learning about workers' unmeasured productivity. For performance pay jobs, coe cients on interaction terms of both father's education with experience (column (2)), and mother's education with experience (column (4)), are positive and signi cant, indicating an increasingly important role played by worker ability in wage determination. Note also that the coe cients on interactions between schooling and experience in columns (2) and (4) are negative although not signi cant at conventional levels (p-values of .158 and .106 respectively). In the sample of non-performance pay jobs, the coe cient on the interaction term between father's education and experience is also positive and signi cant, although now that on the interaction term between mother's education and experience is only marginally signi cant (p-value = .111). I turn now to comparing the rate of learning in performance pay and nonperformance pay jobs.
zi
To do this I use the coe cients on experience interacted
variables as indicators of the rate of employer learning since they re ect the rate
at which new information is being incorporated into wages. looking at equation (8) which gives the bias on the
zi
This can be seen by
variable in period
t.
If new in-
formation is quickly being incorporated into wages, then the rate of change of be high, leading to a large coe cient on the interaction between
zi
Bzt will
and t. Consistent
with faster learning taking place in performance pay jobs than in non-performance pay jobs, the estimate on the interaction of mother's education with experience in the performance pay sample (column (4) of Table 3a) is more than three times that of the non-performance pay sample (column (4) of Table 3b). Similarly, column (2) of each table shows that the estimate on experience interacted with father's educa-
13
tion in the performance pay sample is almost twice the size of the estimate in the non-performance pay sample (although now the dierence is only signi cant at a 20% level (p-value = .193)). Note that, although the dierence between coe cients on the experience interacted
zi
variables in the performance pay and non-performance pay samples are
relatively large, .0123 (.0096) and .0181 (.0104) for father's education and mother's education respectively, they are potentially underestimates of the true dierence in employer learning between these samples.
The reason for this is, as noted by
Lemieux, MacLeod, and Parent (2009), a misclassi cation of job types which is biased towards misclassifying true performance pay jobs as non-performance pay jobs. To see this recall that a job is considered to be based on performance pay if such payments are received at any point during the employment relationship.
Hence,
once performance pay is received, the job is (correctly) classi ed as a performance pay job. On the other hand, the absence of performance pay does not necessarily imply a non-performance job: it is possible for workers in performance pay job to never receive such payments, especially if the employment relationship is only over a short period. As a result, it is likely that true performance pay jobs included in the non-performance pay sample may be mechanically raising the rate of employer learning in this sample above its true level. In particular, this is more likely to be a problem in jobs that either started before the sampling period (1976), or those that ended after the sampling period (2005), since for these jobs, not all years of the employment relationship are observed.
Consider for example, an employment
relationship that started in 1973 and ended in 1979, paying a bonus only in 1975; such a job would be classi ed as a non-performance pay job even though it is really a performance pay job. This type of misclassi cation error is likely to be a particular cause for concern in the present analysis which focuses on jobs observed primarily at the beginning of the sampling period. To account for this problem I use a correction procedure proposed by Lemieux, MacLeod, and Parent (2009). The procedure entails rebalancing the sample of performance pay and non-performance pay jobs so that the incidence of performance pay jobs early in the sample (for which some years of the employment relationship are not observed) are more re ective of the
14
incidence of performance pay jobs in the middle of the sample (for which all years of the employment relationship is observed). As results reported in Table 4 show, accounting for the misclassi cation of performance pay jobs results in larger (about 40% larger when using either
zi
variable), and more precisely measured dierences
in the rate of learning between performance pay and non-performance pay jobs.
4.1.2 Learning in Dierent Types of Performance Pay Jobs Having shown in the previous section that employers learn faster in jobs that reward workers for performance, I now examine whether this learning depends on the speci c way in which performance is rewarded.
Speci cally I investigate whether
there are dierences in employer learning between performance pay jobs that use explicit compensation contracts, such as commission and piece rate jobs, and those that use more implicit contracts, such as jobs which pay bonuses. I present results for commission/piece rate jobs in Table 5a and for bonus jobs in Table 5b. Looking rst at commission/piece rate jobs, columns (2) and (4) provide no indication that employer learning is taking place - coe cients on father's education and mother's education interacted with experience are not statistically signi cant. This is not to say that unobserved worker ability is playing no role in wage determination. In fact, the large coe cient on father's education in the absence of its interaction with experience, suggests that worker ability is indeed an important component of wages; this is very much in line with strong positive selection into these types of jobs (Lazear, 1986, 2002). In general, results provided in Table 5a are consistent with the intuition of Section 2 that, while unmeasured aspects of worker ability are important in determining wages, because wages re ect realized productivity rather than expected productivity, no employer learning takes place. In contrast to the results for commission/piece rate jobs, results presented in Table 5b for bonus jobs are much more consistent with employer learning.
Posi-
tive coe cients on father's education and mother's education in columns (1) and (3) respectively indicate an important role for worker ability in wages (although a smaller role in these jobs than in commission/piece rate jobs). Positive coe cients
15
on these measures interacted with experience in columns (2) and (4) suggest that employers are gradually learning about workers' ability and incorporating this into wages. Taken together, the results presented in Tables 5a and 5b indicate that all the learning in performance pay jobs is due to learning in bonus paying jobs only.
5
Conclusion
In this paper I investigate whether employers are learning faster about workers in jobs that reward performance than about workers in jobs that do not. According to models of employer learning, employers will learn at a faster rate if they are able to observe cleaner signals of worker productivity. It is reasonable to argue that, because of the nature of performance-based compensation, productivity signals of workers in jobs that use this form of compensation are likely less noisy than those that do not. Consistent with this, I nd that employers learn considerably faster about workers in performance pay jobs than about those in non-performance pay jobs. Furthermore I nd that this result is driven entirely by learning about workers in jobs that pay bonuses, and that no learning takes place in piece rate and commission based jobs.
16
References [1] Altonji, Joseph G., and Charles R. Pierret. "Employer Learning and Statistical Discrimination." Quarterly Journal of Economics 116 1, (2001): 313-50. [2] Altonji, Joseph G., and Charles R. Pierret. "Employer Learning and Statistical Discrimination." National Bureau of Economic Research Working Paper No. 6279, November 1997. [3] Altonji, Joseph G., and Lewis M. Segal. "Small-Sample Bias in GMM Estimation of Covariance Structures." Journal of Business and Economic Statistics 14 3, (1996): 353-66. [4] Baker, George, Robert Gibbons, and Kevin J. Murphy. "Subjective Performance Measures in Optimal Incentive Contracts." Quarterly Journal of Economics 109 4, (1994): 1125-56. [5] Bauer, Thomas K., and John P. Haisken-DeNew. "Employer Learning and the Returns to Schooling." Labour Economics 8 2, (2001): 161-80. [6] Brown, James N., and Audrey Light. "Interpreting Panel Data on Job Tenure." Journal of Labor Economics 10 3, (1992): 219-57. [7] Cadsby, C. B., F. Song, and F. Tapon. "Sorting and Incentive Eects of Pay for Performance:
An Experimental Investigation." Academy of Management
Journal 50, no. 2 (2007): 387-405. [8] Farber, Henry S., and Robert Gibbons. "Learning and Wage Dynamics." Quarterly Journal of Economics 111 4, (1996): 1007-47. [9] Gibbons, Robert, and Lawrence F. Katz. "Layos and Lemons." Journal of Labor Economics 9 4, (1991): 351-80. [10] Gibbons, Robert, Lawrence F. Katz, Thomas Lemieux, and Daniel Parent. "Comparative Advantage, Learning, and Sectoral Wage Determination." Journal of Labor Economics 23 4, (2005): 681-723. [11] Greenwald, Bruce C. "Adverse Selection in the Labour Market." Review of Economic Studies 53 3, (1986): 325-47. [12] Lange, Fabian. "The Speed of Employer Learning." Journal of Labor Economics 25 1, (2007): 1-35.
17
[13] Lazear, Edward P. "Performance Pay and Productivity." American Economic Review 90 5, (2000): 1346-61. [14] . "Salaries and Piece Rates." Journal of Business 59 3, (1986): 405-31. [15] Lemieux, Thomas, W. Bentley MacLeod, and Daniel Parent. "Performance Pay and Wage Inequality." Quarterly Journal of Economics 124 1, (2009): 1-49. [16] MacLeod, W. Bentley, and Daniel Parent. "Job Characteristics, Wages, and the Employment Contract." Federal Reserve Bank of St. Louis Review 81 3, (1999): 13-27. [17] MacLeod, W. Bentley, Daniel Parent, and Solomon W. Polachek. "Job Characteristics and the Form of Compensation." In Research in Labor Economics. Volume 18, 177-242. Associate Editor: John Robst. Stamford, Conn.: JAI Press, 1999. [18] Murphy, Kevin J., Performance Measurement and Appraisal: Motivating Managers to Identify and Reward Performance, Performance Measurement, Evaluation, and Incentives, edited by William J.Bruns, Jr. (Harvard Business School Press, Boston, 1992). [19] . "Incentives, Learning, and Compensation: A Theoretical and Empirical Investigation of Managerial Labor Contracts." RAND Journal of Economics 17 1, (1986): 59-76. [20] Parent, Daniel. "The Eect of Pay-for-Performance Contracts on Wages." Empirical Economics 36 2, (2009): 269-95. [21] . "Matching, Human Capital, and the Covariance Structure of Earnings." Labour Economics 9 3, (2002): 375-404. [22] . "Methods of Pay and Earnings: A Longitudinal Analysis." Industrial and Labor Relations Review 53 1, (1999): 71-86. [23] Pinkston, Joshua C. "A Model of Asymmetric Employer Learning with Testable Implications." Review of Economic Studies 76 1, (2009): 367-94. [24] . "Screening Discrimination and the Determinants of Wages." Labour Economics 10 6, (2003): 643-58. [25] . "A Test of Screening Discrimination with Employer Learning." Industrial and Labor Relations Review 59 2, (2006): 267-84.
18
[26] Schonberg, Uta. "Testing for Asymmetric Employer Learning." Journal of Labor Economics 25 4, (2007): 651-91. [27] Strobl, Eric. "Do Employers Use Education as a Signal for Ability in Developing Countries?
Evidence from Ghana." Applied Economics Letters 11 4, (2004):
259-61. [28] Waldman, Michael. "Job Assignments, Signalling, and E ciency." RAND Journal of Economics 15 2, (1984): 255-67. [29] Zellner, Arnold. An Introduction to Bayesian Inference in Econometrics, Wiley Series in Probability and Mathematical Statistics. New York: J. Wiley, 1971.
19
Tables Table 1a: Sample Statistics
Hourly Wage (1979$) Average Tenure (years) Potential Experience
(1)
(2)
(3)
All Jobs
PP Jobs
Non PP Jobs
8.61
9.58
8.09
(4.62)
(5.50)
(3.97)
11.1
11.5
11.0
(8.5)
(8.2)
(8.7)
18.0
17.2
18.4
(11.8)
(11.0)
(12.2)
12.7
13.2
12.4
(2.4)
(2.3)
(2.4)
(years) Education Mother's Education Father's Education Age White
3.5
3.7
3.4
(1.6)
(1.6)
(1.6)
3.4
3.7
3.2
(1.9)
(1.9)
(1.8)
36.7
36.4
36.8
(11.2)
(10.6)
(11.5)
73.4
80.1
69.7
Married
83.9
84.8
83.4
Average Annual Hours
2310
2369
2279
(428)
(433)
(422)
Union
29.3
19.3
34.7
Salaried
37.2
48.0
31.9
Hourly Paid
55.8
40.6
63.1
# workers
3045
819
2226
15871
5547
10324
# observations Notes: PP denotes Performance Pay jobs.
Average Tenure refers to the average tenure over all worker-year observations. Mother's and father's education: 3 corresponds to to 9-11 grades (some high school); junior high; 4 corresponds to 12 grades (completed high school); high school . Shares of salaried and hourly paid workers do not sum to one because about 10% of the observations are missing data on whether the worker was salaried or hourly paid.
20
Table 1b. Summary Statistics - Performance Pay Bonus
Commission/ Piece Rate
Hourly Wage (1979$) Average Tenure (years) Potential Experience (years) Education Mother's Education Father's Education Age
9.41
9.31
(5.33)
(5.80)
11.5
10.5
(8.3)
(7.8)
17.1
17.7
(11.1)
(11.0)
13.1
13.1
(2.3)
(2.2)
3.8
3.6
(1.7)
(1.4)
3.7
3.4
(1.9)
(1.7)
36.3
36.7
(10.6)
(10.5)
White
79.4
84.3
Married
83.8
88.0
Average Annual Hours
2335
2497
(411)
(506)
Union
20.0
15.8
Salaried
51.8
25.2
Hourly Paid
46.8
15.1
# workers # observations
718
225
4708
1449
Notes: Average Tenure refers to the average tenure over all worker-year observations in the relevant sample. Mother's and father's education: 3 corresponds to to 9-11 grades (some high school); junior high; 4 corresponds to 12 grades (completed high school); high school . Shares of salaried and hourly paid workers do not sum to one because about 10% of the observations are missing data on whether the worker was salaried or hourly paid.
21
Table 2: Employer Learning, Full Sample Independent: ln wage (1) Education .1056*** (.0073)
(2) .1133*** (.0072)
(3) .1090*** (.0072)
(4) .1140*** (.0071)
-.0635** (.0263)
-.0933*** (.0266)
.2207*** (.0525)
-.0220 (.0787)
Education x experience/100
-.0525** (.0266)
-.0917*** (.0263)
Father's education/10
.2723*** (.0494)
-.0073 (.0694)
Father's education x experience/10
.0174*** (.0044)
Mother's education/10 Mother's education x experience/10 R # observations # clusters 2
.0149*** (.0047) .3835 15871 3045
.3873 15871 3045
.3801 15871 3045
Notes: Dependent variable is the natural log of wage. Speci cations include potential experience, a quartic in potential experience, dummies for married, white, and union, year dummies; and the interaction of year dummies with education. Standard errors in parentheses are clustered at the worker level. *** indicates p-value<0.01 ** indicates p-value<0.05 * indicates p-value<0.10
22
.3823 15871 3045
Table 3a: Employer Learning in Performance Pay Jobs Independent: ln wage (1) (2) Education .1061*** .1128*** (.0147) (.0141) Education x experience/100 Father's education/10
-.0216 (.0506)
-.0658 (.0465)
.2451*** (.0843)
-.1378 (.1245)
Father's education x experience/10
(3) .1084*** (.0148)
(4) .1153*** (.0143)
-.0309 (.0508)
-.0778 (.0485)
.1704* (.0965)
-.2582* (.1437)
.0253*** (.0084)
Mother's education/10 Mother's education x experience/10 R # observations # clusters 2
.0262*** (.0091) .3919 5547 819
.3984 5547 819
.3883 5547 819
Notes: Dependent variable is the natural log of wage. Speci cations include potential experience, a quartic in potential experience, dummies for married, white, and union, year dummies; and the interaction of year dummies with education. Standard errors in parentheses are clustered at the worker level. *** indicates p-value<0.01 ** indicates p-value<0.05 * indicates p-value<0.10
23
.3943 5547 819
Table 3b: Employer Learning in Non-Performance Pay Jobs Independent: ln wage (1) (2) (3) Education .1059*** .1128*** .1102*** (.0083) (.0084) (.0081) Education x experience/100
-.0689** (.0300)
-.1014*** (.0316)
Father's education/10
.2847*** (.0590)
.0670 (.0795)
Father's education x experience/10
-.0810*** (.0294)
-.0976*** (.0315)
.2269*** (.0591)
.0952 (.0899)
.0130*** (.0045)
Mother's education/10 Mother's education x experience/10 R # observations # clusters 2
(4) .1132*** (.0082)
.0081 (.0051) .3841 10324 2226
.3865 10324 2226
.3805 10324 2226
Notes: Dependent variable is the natural log of wage. Speci cations include potential experience, a quartic in potential experience, dummies for married, white, and union, year dummies; and the interaction of year dummies with education. Standard errors in parentheses are clustered at the worker level. *** indicates p-value<0.01 ** indicates p-value<0.05 * indicates p-value<0.10
24
.3811 10324 2226
Table 4: Learning in Performance Pay and Non-Performance Pay Jobs. Independent: ln wage 1. Non-PP 2. Education 3. Non-PP x education 4. Education x experience/100 5. Non-PP x education x experience/100 6. Father's education/10
Father's
Mother's
Education
Education
-.0635
-.0980
(.2305)
(.2328)
.1128***
.1153***
(.0141)
(.0142)
.0001
-.0028
(.0171)
(.0172)
.0658
-.0777
(.0465)
(.0484)
.0013
.0256
(.0620)
(.0623)
-.1378 (.1243)
7. Non-PP x father's education/10
.2180 (.1461)
8. Father's education x experience/10
.0253*** (.0084)
9. Non-PP x father's education x
-.0173*
experience/10 (.0097) 10. Mother's education/10
-.2582* (.1434)
11. Non-PP x mother's education/10
.3982 (.1689)
12. Mother's education x experience/10
.02616*** (.0091)
13. Non-PP x mother's education x
-.0253**
experience/10 (.0106) 2
R
.3682
.3657
# observations
15871
15871
3045
3045
# clusters
Notes: Dependent variable is the natural log of wage. Other included regressors: potential experience; a quartic in potential experience; dummies for union, white, married, year; interaction of year dummies with education; full set of interactions with dummy indicating non-performance pay job. Stata desmat package used to generate interactions. Standard errors in parentheses are clustered at the worker level. *** indicates p-value<0.01 ** indicates p-value<0.05 * indicates p-value<0.10
25
Table 5a: Employer Learning in Commission/Piece Rate Jobs Independent: ln wage (1) (2) (3) Education .0893*** .0920*** .0954*** (.0249) (.0256) (.0263) Education x experience/100
.0384 (.1239)
.0212 (.1280)
Father's education/10
.4215** (.1835)
.2232 (.2899)
Father's education x experience/10
.0131 (.1247)
.0112 (.1308)
.2335 (.1955)
.2161 (.3441)
.0111 (.0151)
Mother's education/10 Mother's education x experience/10 R # observations # clusters 2
(4) .0956*** (.0266)
.0010 (.0170) .4007 1449 225
.4020 1449 225
.3912 1449 225
Notes: Dependent variable is the natural log of wage. Speci cations include potential experience, a quartic in potential experience, dummies for married, white, and union, year dummies; and the interaction of year dummies with education. Standard errors in parentheses are clustered at the worker level. *** indicates p-value<0.01 ** indicates p-value<0.05 * indicates p-value<0.10
26
.3912 1449 225
Table 5b: Employer Learning in Bonus Jobs Independent: ln wage (1) Education .1000*** (.0162)
(2) .1083*** (.0153)
(3) .1021*** (.0164)
(4) .1126*** (.0155)
-
-.0525 (.0481)
.0099 (.0540)
-.0762 (.0504)
.2507*** (.0902)
-.1734 (.1286)
.2107** (.1038)
-.3178** (.1496)
Education x experience/100 Father's education/10 Father's education x experience/10
.0283*** (.0090)
Mother's education/10 Mother's education x experience/10 R # observations # clusters 2
.0323*** (.0099) .4020 4708 718
.4104 4708 718
.3992 4708 718
Notes: Dependent variable is the natural log of wage. Speci cations include potential experience, a quartic in potential experience, dummies for married, white, and union, year dummies; and the interaction of year dummies with education. Standard errors in parentheses are clustered at the worker level. *** indicates p-value<0.01 ** indicates p-value<0.05 * indicates p-value<0.10 - indicates a value smaller than 10e-04
27
.4088 4708 718
28
Figure 1: Cross-Sectional Residual Dispersion
Appendix 1 Notice that we can write Bayes' Law as
posterior pdf ∝ prior pdf ×likelihood f unction.
Assume that the distribution of ability (from which signals ω of worker productiv2 ity are drawn) is normal with mean µ and variance σ . Assume also that employers
µipr (µ − µpr ) ,
have prior beliefs of worker ability that are normally distributed with mean
√ 1 exp 2πσpr
2 variance σpr . Using the implied prior pdf,
h
− 2σ12 pr
and
and p (µ) = 1 2 likelihood function, p (ω | µ, σ ) = √ exp − 2σ1 2 (ω − µ) , the posterior (updated) 2πσ 2 +σ 2 2 +µ σ 2 2 σ ωσ pr pr pr 2 pdf can be represented by p (µ | ω, σ ) ∝ exp − µ − σ2 +σ2 . Notice 2σ 2 σ 2 pr
pr
that representing the posterior belief of worker ability in this way makes it explicit 2 +µ σ 2 ωσpr pr that this updated belief is distributed normally with mean and variance 2 +σ 2 σpr 2 2 σpr σ 2 2 one can write the updated belief of 2 +σ 2 . Finally, dividing throughout by σpr σ σpr −1
worker ability as
ω (σ 2 )
−1
2 +µpr (σpr )
2 (σ 2 )−1 +(σpr )
−1
, a weighted average of the signal and the prior
belief; where the weights are the inverse of the prior variance and signal variance 2 respectively. Thus the greater the precision of any given signal (the smaller is σ ), the more is learnt from that signal, leading to faster overall learning.
Data Appendix In this data appendix I describe in detail how I constructed my nal sample of workers using data from the Panel Study of Income Dynamics (PSID). All data was obtained from the PSID's website: http://psidonline.isr.umich.edu/. covers the years 1976-2005.
The period of interest
My initial PSID sample comprises all respondents of
either the cross-sectional sample drawn by the Survey Research Center (SRC) or the Survey of Economic Opportunity (SEO) conducted by the Bureau of the Census for the O ce of Economic Opportunity. More speci cally, I omit respondents entering as part of either the Latino (1990/1992) or Immigrant (1997/1999) samples. This gives an initial sample size of 60,844 respondents. From my initial sample I drop all female respondents, and any respondent who was not a head of household in at least one of the years between 1976 and 2005 inclusive. This leaves 12,268 respondents which, in long-form produces a balanced panel of 318,968 worker-year observations.
17
I further restrict the sample to include
only private sector (not government workers) wage earners (not self employed) who
17 There is data only for 26 years; from 1998 onwards the PSID is only available every two years.
29
work at least 35 hours a week, are between the ages of 16 and 65 inclusive, and who earn between $1 and $100 (1979$) an hour. For reasons discussed in the paper, for each worker I keep only the rst identi ed job for which there are at least two observations. The nal sample consists of 15,871 worker-year observations covering 3045 jobs. I discuss some of the important steps of this sample construction below.
Construction of Job Spells Identifying jobs spells in the PSID is di cult because, unlike the National Longitudinal Surveys (NLS), the PSID does not provide information linking yearly worker observations to speci c employers. Nevertheless, it is possible to use information that is provided in the PSID (for example interview dates, tenure with employer, tenure in current position, reason for position change etc) to identify the job spells that make up each worker's employment history.
As Brown and Light (1992) note however,
because of measurement error, the way in which employment histories are divided into job spells (which they refer to as partitioning the sample into job spells) depends heavily on which information is used. Evaluating dierent partitioning methods by comparing them to properly identi ed job spells using the NLS, they conclude that the most consistent way to identify a job change is whenever job tenure is less than time since last interview. I use this criteria (which Brown and Light (1992) refer 18
to as Partition T) to identify jobs in the PSID.
For each worker this numbered
job spells consecutively starting from one. Special care is taken when numbering jobs appearing in 1998, 2000, 2002, and 2004.
Recall that after 1997, data in the PSID is only available every two years
and that most of the relevant employment information is collected retrospectively. This means that my sample years are really 1976. . . 1997, 1998, 2000, 2002, 2004. However there is neither tenure, nor time since last interview information available in 1998, 2000, 2002, or 2004. I get around this problem by applying the following logic: for each of these years (call these the middle years ), if the job numbers are the same in adjacent years, then this is the job number in the middle year; if this is not the case then, if tenure in the following year is greater than 12 months, then the
18 Job tenure is constructed using tenure with employer for most years between 1976 and 1993. However, in 1979 and 1980 (and 1978 for workers over 45) this variable is not available so tenure in position is used instead. Finally, after 1994 separate variables are available for tenure years (tenure in year), tenure months (tenure in months), and tenure weeks (tenure in weeks). time since last interview is constructed using the variable interview month in the following way: time since last interview = (interview month[yr] - interview month[yr-1]) + 12 if yr61997 time since last interview = (interview month[yr] - interview month[yr-1]) + 24 if yr>1997.
30
job number in the middle year is the same as in the following year; if tenure in the following year is less than 12 months, then the job number in the middle year is the same as the previous year.
Construction of Performance Pay and Commission/Piece Rate Indicators In all years of the sample, a variable indicating receipt of bonus, commission, or overtime is available.
19
Note however that in the paper overtime is not considered as
performance pay. Following Lemieux, MacLeod, and Parent (2009) I use additional information available in the PSID that asks whether or not overtime was received to classify an observation as pertaining to performance pay if the respondent received a bonus, commission, or overtime, but did not receive overtime. It is also possible to classify observations as performance pay via information on payment method. Respondents are asked whether they are paid a salary, by the hour, or neither. Those who reported receiving neither a pure salary nor a pure hourly rate were given the option of classifying their earnings as one of the following: 1. Piecework 2. Commission 3. Tips; tips and salary 4. Salary plus commission 20
5. Other
From this list I classify performance pay workers as those indicating receipt of commissions and piece rates. Furthermore, a job is considered a performance pay job if some form of performance payment was received at any point during the job spell. I follow the procedure in the previous paragraph to identify observations in my commissions/piece rate sample only.
Similarly, a job is considered a commis-
sion/piece rate job if any such payments were received at any point during the job spell.
19 From 1976-1991 this asked for an amount; from 1993 onwards this asked whether any of these were received or not. In 1992 separate questions were used to identify receipt of each of these forms of payment.
20 This variable [V4515 in 1976] is available in all years between 1976 and 2005 although in some
years there are additional categories.
31
Data-Cleaning Procedures A number of data-cleaning procedures were necessary to maintain consistency in variable coding over this 30-year period. I list and discuss brie y some of the more important procedures below.
•
Brown and Light (1992) note that returns to tenure tend to be sensitive to the internal consistency of tenure (although they note that the method applied to bring about this internal consistency does not seem to matter). I thus force the tenure of valid worker-year observations to be internally consistent with the year of the rst valid observation for a job.
•
Following from the discussion above of missing information in years 1998, 2000, 2002, and 2004, I impute this information from adjacent years (for which the information is collected). For time invariant variables I impute them directly; for time varying variables I make the necessary adjustments.
•
In 1981 the coding of industry and occupation categories switches from 2digit to 3-digit categories.
The codebook provides su cient information to
consistently aggregate these into 12 occupation and 12 industry categories over the entire 30-year period.
•
A note in the codebook indicated that for the years 1994-1996, 1999, and 2001, questions collecting information on mother's education, father's education, degree, race, highest degree, degree year, and degree month, were not re-asked if the head of the household had not changed. For the respondents in my sample, these should be time-invariant variables so that I simply brought forward the information from the previous year (or 2 years if post 1997) if there was no change in the head of household [variable labeled same indicating the head of household had not changed from the previous year; this was asked every year].
32
