Does Sorting Play Any Role in the Incentive

∗

Pay-Output Uncertainty Relationship?

Bok Hoong Young Hoon 26 July 2010

Abstract A key prediction of agency theory is that incentive payments and uncertainty in output should be negatively related. Early empirical work on this issue using mainly CEO, agricultural, and franchising data has been unable to provide support for such a prediction, and in many instances nds a positive relationship. Because this work does not account for heterogeneity in agents' risk preferences, one source of this positive relationship might be the sorting of more risk-loving agents into contracts which pay for performance. Using data from the Panel Study of Income Dynamics and Compustat North America I examine whether this is the case for a representative sample of the male US labour force. I show that workers in jobs paying either bonuses, or commissions and/or piece rates are exposed to higher levels of earnings risk than workers in jobs that do not pay for performance. Furthermore, I provide evidence that more risk-loving workers are sorting into these types of jobs. Despite this, I nd that controlling for workers' risk preferences only accounts for a small share of the positive empirical relationship found between performance pay and uncertainty in output.

1

Introduction

The use of incentive-based contracts by corporations to elicit the desired level of eort from workers is a common feature of the US labour market. For example, the ∗

I would like to thank Daniel Parent, Jennifer Hunt, Hitesh Doshi, and Jason Dean for their

many helpful comments.

1

compensation packages of CEO's which, in many instances are linked to company 1

prots , the cash renting of land by farmers, and the payment of xed franchising fees by franchisees. In addition to its use in these upper managerial and self employment settings, such incentive payments also play an important role in the compensation of non-managerial, rank-and-le workers in the form of bonuses and piece rate pay. Work done by Lazear (2000), Paarsch and Shearer (2000) and Shearer (2004) studying the eect of performance-based payments on productivity nds that, after netting out increases in productivity due to ability sorting, such payments do indeed lead to increased levels of productivity. Despite this nding however, in only about 40% of all US jobs do workers receive performance pay (Lemieux, MacLeod, and Parent, 2009). If performance pay does bring about increased levels of productivity, why is there not more widespread usage? There are a number of reasons why rms may choose not to use performance pay. When jobs entail multiple tasks for example, tying compensation to only some of these tasks (perhaps due to dierences in the relative observability of these tasks) will result in workers directing an excessive amount of eort towards these tasks at the cost of equally productive, but less veriable, ones (Holmstrom and Milgrom, 1991). Additionally, instead of providing incentives by paying for performance, rms may choose to maintain a strict hierarchical structure in which workers' promotions are based on their performance. One generally well-received theoretical reason is based on the level of uncertainty in the production process: more uncertainty in production restricts employers' ability to base pay on performance. This will be the focus of this paper. Specically, I examine performance pay and non-performance pay jobs in the Panel Study of Income Dynamics to determine whether such a factor constrains the use of performance pay. Since the 1980s, researchers interested in evaluating agency theory have placed much attention on the relationship between incentives and output uncertainty. In general however, as noted by Prendergast (2002), this body of work (which mostly looks at this relationship as it applies to executives, the agricultural industry, and

1 Furthermore, CEO's are normally oered stock options, the value of which is closely tied to company performance. Core and Guay (1999) provide evidence that rms continuously adjust CEO stock options so that these are consistent with optimal equity incentive levels.

2

franchising) has been unable to provide support for the negative relationship predicted by principal-agent models and, in many instances, has found evidence of a positive one (Allen and Lueck, 1992, 1999; Garen, 1994; Stroh et al., 1996). A potential problem with much of this work is the failure to account for the (endogenous) sorting of workers into jobs which pay for performance. In particular, it is the sorting of more risk-loving workers into performance pay jobs that poses a problem for the predicted negative relationship between uncertainty and the use of performance pay.

2

If workers who are more risk-loving are also more likely to be found in perfor-

mance pay jobs then, unless this is taken into account, there is no reason to expect uncertainty and incentive pay to be negatively related. Presenting a principal-agent model in which risk preference is a function of uncertainty, Serfes (2005) shows that, if risk averse agents are matched with low risk principals (i.e. there is risk sorting), then a positive relationship between uncertainty and incentive pay is possible. Indeed Dohmen and Falk (2006) and Grund and Sliwka (2010), both provide evidence of such sorting in Germany.

Accounting for this endogeneity is not a straightfor3,4

ward task as risk preferences are generally not observable by the econometrician.

Nevertheless, more recent work using xed eects (Aggarwal and Samwick, 1999; Wulf, 2007; Hilt, 2008) and instrumental variable techniques (Ackerberg and Botticini, 2002) to address this problem does uncover the predicted negative relationship between uncertainty and incentive pay, although only in very narrow contexts (Aggarwal and Samwick (1999) - executives; Wulf (2007) - division managers; Hilt (2008) - the American whaling industry; and Ackerberg and Botticini (2002) - agricultural contracts in early Renaissance Tuscany). I contribute to this literature by looking at this relationship as it applies to a

2 Note that sorting by itself does not pose a problem but rather sorting along unobservable (to the econometrician) dimensions. See Lazear (1986, 2000) and Parent (1999) for work on the sorting of more productive workers into performance pay jobs and the problem this presents when measuring the incentive eects of performance pay.

3 Researchers have attempted to use proxies for risk aversion (for example, wealth). However,

Ackerberg and Botticini (2002) show that unless these proxies properly represent agents' risk preferences, the endogeneity bias will persist.

4 See Holt and Laury (2002) and Bellemare and Shearer (2006, 2010) for experimental work

directed towards revealing respondents' risk preferences.

3

representative sample of the male U.S. workforce. I am not aware of any other research that investigates this relationship for such a diverse group of workers for whom performance based payments have become increasingly important. This is also the rst work addressing this issue that uses actual measures of workers' risk preferences to account for sorting.

Because of this I am able to provide direct evidence that

more risk-loving workers are more likely to be in jobs with higher earnings risk and, more importantly for this study, in jobs that pay for performance. I consider separately two forms of performance-based payments commonly received by U.S. workers: 5

bonuses, and commissions and/or piece rates.

Using information on performance

pay and worker risk preferences contained in the Panel Study of Income Dynamics I nd evidence for the sorting of more risk-loving workers into jobs that pay either bonuses or, commissions and/or piece rates.

I nd also that jobs that use these

payments are associated with higher levels of income risk, both across workers in a given occupation and for a given worker over time. Furthermore, I present results indicating that commission and/or piece rate jobs are associated with higher levels of income risk and risk sorting than are bonus paying jobs. Having shown that there is sorting of more risk-loving workers into incentive pay jobs, I investigate whether this contributes to the positive relationship between incentive pay and output uncertainty. Importing information on industry specic levels of uncertainty faced by rms obtained from Compustat North America into the Panel Study of Income Dynamics I present results that indicate a positive relationship between performance pay and uncertainty, even when controlling for workers' risk preferences. While it does appear that accounting for risk sorting does somewhat temper the positive relationship, it is not sucient to produce one that is negative.

5 Note that I consider commissions and/or piece rates to constitute one type of performance pay. Although, this is primarily because considering them to be separate types would result in too few observations of each type of performance pay, it seems reasonable as both commissions and piece rates are based on objective measures of performance.

4

2

The Principal-Agent Model

In this section I briey review the setup and implications of the standard principalagent model. I illustrate how the predictions of the model imply the sorting of more risk-loving workers into performance pay jobs, and show how failing to properly account for this sorting may result in an empirical relationship between performance pay and output uncertainty that is inconsistent with theoretical expectations. The principal-agent model, as applied to the labour market, considers an employer (the principal) contracting a worker (the agent) to complete a given task producing output

y.

uncertainty in the production process so that output is given by

ε ∼ N (0, σε ). 2

Finally, the employer is risk neutral, and the worker is risk and eort

U = y −w, where w is the wage utility is given by U = −exp{−r[w − C(e)]} risk aversion, and C(e) is the cost function of

averse. Specically, the employer's utility is given by paid to the worker; and the worker's where

r

e, there is y = e + ε where

While output depends on the worker's eort level,

p

a

measures his or her absolute

eort and is increasing at an increasing rate.

It is these features that lead to the

standard moral hazard problem. The solution to this problem is a linear contract of the form

w = a + βy

where optimality requires

β=

1 . 1 + rσε C 00 (e)

(1)

2

According to this result performance pay should be negatively related to both risk aversion,

r,

and the level of uncertainty,

2 6

σε .

To see why this result implies that more risk-loving workers should be sorting into performance pay jobs consider a worker's utility in a performance pay job with an

Ua, = −exp{−r(a + 1+rσy2 C 00 − C)}, and in a non-performance ε pay job, Ua, = −exp{−r(w¯ − C)}. The worker will be indierent between the performance paying job and the non-performance paying job when Ua, = Ua, y y 1 which occurs at w ¯ = a + 1+rσ2 C 00 or r = σ2 C 00 [ w−a − 1]. To determine how a dierent ¯ optimal piece rate,

pp

npp

pp

ε

npp

ε

level of risk aversion would have inuenced the worker's choice of jobs, consider the

6 ∂β ∂r

σ 2 C 00 (e)

ε = − [1+rσ 2 C 00 (e)]2 < 0; ε

∂β ∂σε2

00

rC (e) = − [1+rσ 2 C 00 (e)]2 < 0 ε

5

relationship between utility and risk aversion for each type of job:

∂Ua, ∂r

pp

y ryσε C 00 y − C) − ]exp{−r(a + − C)} 00 00 1 + rσε C (1 + rσε C ) 1 + rσε C 00 2

= [(a +

2

2

∂U ∂r

a,npp

2

2

= (w¯ − C)exp{−r(w¯ − C)} > 0.

(2)

(3)

Equations (2) and (3) show rstly that, for the non-performance pay job, the relationship between utility and risk aversion is unambiguously positive. Secondly and more importantly, they show that, even though it is not possible to sign this relationship for the performance pay job, when there is indierence between the performance pay and non-performance pay jobs, the slope of the relationship will always be less for the performance pay job than for the non-performance pay job. This means that, at the point of indierence between the two types of jobs, if the worker was more risk averse he or she would choose to work in a non-performance pay job since it would make him or her strictly better o. For the same reason, if the worker was less risk averse he or she would choose to work in the performance paying job. I illustrate this in Figure 1 by showing the utility-risk aversion relationship for each type of job when the worker is indierent between the performance pay and the non-performance pay job. For values of risk aversion to the right of

r¯ =

1 [ ¯y σε2 C 00 w−a

− 1]

to be in a non-performance pay job, and for values to the left of

the worker prefers

r¯ he

or she prefers

to be in a performance pay job. Consider now the optimal piece rate given in equation (1).

Taking the total

derivative and re-arranging leads to the following expression for the relationship between performance pay and uncertainty:

rC 00 ∂σ ∂e ∂β σ C 00 rσ C 000 =− − · − · . 00 00 00 ∂r (1 + rσ C ) (1 + rσ C ) ∂r (1 + rσ C ) ∂r 2

2

2

2

2

2

2

2

2

(4)

Assuming that workers are not sorting into performance pay jobs according to their risk preferences, and assuming risk preferences do not aect the amount of eort workers exert, we get the standard negative relationship between performance pay and uncertainty that is indicated by principal-agent models:

6

∂β σ C 00 =− < 0. ∂r (1 + rσ C 00 ) 2

2

2

However, if there is the sorting of more risk-loving workers into jobs with greater uncertainty, then

∂r ∂σε2

<0

and

∂β σ C 00 rC 00 ∂σ =− . − · 00 00 ∂r (1 + rσ C ) (1 + rσ C ) ∂r 2

2

2

2

2

(5)

2

Equation (5) shows that, if there is sucient sorting (i.e. if

|

∂r ∂σε2

|

is large enough),

∂β then performance pay and uncertainty will be positively related ( 2 ∂σε

3

> 0).

Data

To test for the sorting of more risk-loving workers into performance pay jobs and to examine if this plays any role in the relationship between output uncertainty and performance pay I use data from the Panel Study of Income Dynamics (PSID) and Compustat North America. I test for the risk sorting of workers into performance pay and non-performance pay jobs using only the PSID. Determining the relationship between uncertainty and performance pay however, requires detailed information from rms and is usually not included in standard survey datasets.

As such, I supple-

ment information in the PSID sample with industry-level information aggregated from rms contained in Compustat North America. In this section I describe these samples and the construction of some of the more pertinent measures used in this study.

The Panel Study of Income Dynamics The Panel Study of Income Dynamics is a longitudinal study that, since 1968, has collected primarily demographic and economic information on a representative sample of the US population. In this study I consider a PSID sample that includes only

7

male, wage-earning, heads-of-household between 1993 and 2007 who are between the ages of 16 and 65, work more than 35 hours a week, and have wages between $1 7,8

and $100 (1979$).

The nal sample consists of 11,563 worker-year observations

for 2,486 workers. In considering the data, an important requirement is the identication of observations corresponding to performance pay jobs. Since each year the PSID asks workers whether or not they have received dierent forms of performance pay, one way to do this is to classify a given worker-year observation as performance pay if the worker received performance-based payments in that year.

9

A problem

with this approach is that, because workers in these jobs may not necessarily receive such payments every year in that job, worker-year observations which belong to true performance pay jobs but in which no performance pay is actually received, are misclassied as being non-performance pay.

An alternative way of classifying

observations as performance pay which avoids this problem is provided by Lemieux, MacLeod, and Parent (2009): construct individual job spells and classify a job spell as being performance pay if at any point during the job spell the worker has received performance-based payments. I follow this procedure. Although the PSID does not provide explicit indicators of specic employment relationships, it does provide rich enough employment information that makes it possible to apply various criteria by which to divide workers' labour market experience into individual job spells. Brown and Light (1992) evaluate these dierent criteria (which they refer to as partitions) and conclude that the most consistent way to partition workers' labour market experience into job spells is to indicate a new job whenever tenure is less than the time that has elapsed since the last interview. Applying this criterion to my sample I am able to identify 3,130 job spells. Of these

7 The PSID is available annually until 1997. After this it is available every two years. 8 I consider only heads of households because much of the relevant information (for example, information on performance pay and risk preferences) were collected only for these household members. I consider only males since the number of female heads of household were relatively small. I restrict the sample to include observations from 1993 to 2007 since, starting in 1994 (employment information pertaining to 1993), the PSID started asking more detailed questions about performance pay. This allowed for cleaner identication of performance pay jobs and, in particular, allowed for a cleaner classication of those that were bonus paid and those that were commission and/or piece rate paid than in previous years.

9 See the Data Appendix for details on how performance pay is identied in the PSID.

8

3,130 jobs spells, 912 are bonus paying, and 210 are commission and/or piece rate paying. Sample statistics are presented in Table 1. Rows (1) to (8) indicates that workers in jobs that pay for performance (especially bonus paying jobs) generally have more education, are more likely to be married, and have higher incomes (both in terms of wages and yearly income) than non-performance pay workers.

Risk Preference Measures While there are a number of appealing reasons to use the PSID to examine issues concerning performance-based payments, its use in this study is particularly appealing because it reports measures of workers' risk preferences; measures which are not commonly available in standard datasets. In this paper I use two measures of worker risk preferences.

Both measures are constructed using responses to the following

sequence of hypothetical questions:

[M1] Now I have another kind of question that guaranteed you income for life equal to your current, total income. And that job was [your/your family's] only source of income. Then you are given the opportunity to take a new, and equally good, job with a 50-50 chance that it will double your income and spending power. But there is a 50-50 chance that it will cut your income and spending power by a third. Would you take the new job? IF NO GO TO [M3] [M2] Now, suppose the chances were 50-50 that the new job would double [your/your family] income, and 50-50 that it would cut it in half. Would you still take the new job? IF YES GO TO [M5] [M3] Now, suppose the chances were 50-50 that the new job would double [your/your family] income, and 50-50 that it would cut it by 20 percent.

9

Then, would you take the new job? IF NO GO TO [M4] [M4] Now, suppose that the chances were 50-50 that the new job would double [your/your family] income, and 50-50 that it would cut it by 10 percent. Then, would you take the new job? [M5] Now, suppose that the chances were 50-50 that the new job would double [your/your family] income, and 50-50 that it would cut it by 75 percent. Would you still take the new job?

Responses to these questions, which dier only in the amount of risk to which a worker is exposed, can then be used in a straightforward way to construct a measure that is increasing in risk aversion. This measure, which is used by Brown and Taylor (2007) to look at the eect that risk preferences have on educational attainment and Brown et al.

(2007) to look at the relationship between risk preferences and self

employment, is as follows:

   0       1     2 RAi = 3        4     5

M 1 = yes & M 2 = yes & M 5 = yes M 1 = yes & M 2 = yes & M 5 = no M 1 = yes & M 2 = no

(6)

M 1 = no & M 3 = yes M 1 = no & M 3 = no & M 4 = yes M 1 = no & M 3 = no & M 4 = no.

As equation (6) shows, the most risk averse respondents are classied as those not willing to risk losing even 10 percent of their income, while the least risk averse respondents are classied as those willing to risk losing at least 75% of their income. Row (9) of Table 1 indicates that workers in performance pay jobs, and especially in commission and/or piece rate paying jobs, exhibit lower risk aversion than the average worker.

10

One problem with using

RA

i

to represent risk preferences is that, although it

does a good job in ordering respondents according to their risk preferences, it may not provide an accurate indication of the increase in risk aversion between respondents in dierent risk aversion groupings.

Because of this I also use an index of

risk tolerance constructed by Barsky et al. (1997) which reports the expected risk tolerance conditional on being in each of these groups. Under the assumption that utility is given by

U (c) =

1− 1

c q (where 1− 1q

q

is risk tolerance and

can identify the risk tolerance cut-points that divide to questions [M1]-[M5].

10

q

c

is consumption), one

into regions corresponding

The likelihood function is given by the product of each

individual's probability of being in a particular group of equation (6).

The risk

tolerance associated with each of these groups can be obtained by maximizing the likelihood function and calculating the expected means conditional on being in that group. As the average risk tolerance values presented in row (10) of Table 1 show, workers in performance pay jobs are able to tolerate higher levels of risk than the average worker. Again, this is especially true for those in commission and/or piece rate paying jobs.

Income Risk Measures In order to show that workers' risk preferences are playing a part in the sorting of more risk-loving workers into jobs that pay for performance, it is necessary to show that workers in these jobs do experience higher levels of risk.

To examine this I

consider only income risk (as opposed to other forms of risk such as physical risk) and construct two such measures using the PSID. The rst is a measure of the crosssectional variation in wages that is unexplained by observable worker characteristics within a given occupation. Specically, it is the variance of occupation-specic residuals constructed from a regression of log wage on education, experience, experience squared, union status, occupation, risk tolerance, year dummies, and the interaction between year and education. Such a measure is used by Bonin et al. (2007) to show

10 For example, the risk tolerance that separates respondents responding yes to [M1] from those that respond no is determined by solving

1 2 U (2c, q)

11

+ 12 U ( 23 c, q) ≥ U (c, q)

(Barsky et al., 1997).

the sorting of more risk-loving German workers into occupations with higher earnings risk. The second is a measure of intertemporal wage variation within each job of the sample. More specically, for each of the 3,130 job spells, it is the variance of residuals constructed from a regression of log wage on a job tenure and job tenure squared.

Robst, Deitz, and McGoldrick (1999) use both these measures to relate

income uncertainty to the probability of home ownership, nding that income uncertainty does reduce the probability that individuals own their home. As rows (11) and (12) of Table 1 show, workers in performance pay jobs do appear to experience higher levels of both cross-sectional and intertemporal risk than those in non-performance pay jobs. Comparing these rows shows also that workers face much less intertemporal risk than cross-sectional risk. Since the intertemporal risk measure nets out job-match specic eects this suggests that much of the dispersion in earnings is a result of job-match heterogeneity.

Compustat North America Compustat North America is a database containing fundamental and market information for more than 24,000 US and Canadian publicly held companies.

Data is

available on both an annual and a quarterly basis and includes information reported in income statements, balance sheets, and statement of cash ows. I use this information to construct two measures of industry-specic, rm revenue dispersion. These measures of rm revenue dispersion are constructed in a similar manner to the measures of income risk experienced by workers discussed in the previous subsection. The rst measures the level of cross-sectional income variation between rms in a given industry, that is unexplained by rm size. It is the variance of industry-specic residuals constructed from a regression of quarterly revenue on rm size, industry, and year and quarter dummies.

11

The second measure is the average

intertemporal variation in revenue experienced by rms within a given industry. The intertemporal variation in revenue for each rm is given by the variance of residuals

11 Firm size is represented by the value of the rm's assets.

12

constructed from rm-specic regressions of quarterly revenue on rm size, and year and quarter dummies. I use the entire sample of US rms to construct the crosssectional measure, and the sample of US rms for which there were observations on at least two quarters to construct the intertemporal measure. I take these measures to be proxies for the uncertainty in output that rms face. The basic idea is that because rm size explains so much of the variation in revenue between rms in a given industry (R

2

= 0.69),

any leftover variation can be attributed, for the most

part, to factors which are beyond the rm's control.

Table 2 shows the levels of

cross-sectional and intertemporal uncertainty by industry.

Notice that while both

measures agree in terms of the relative uncertainty faced by dierent industries, there is a much higher level of unexplained variation in income across rms than there is for a given rm over time.

Merging these measures into the PSID at the

industry level, rows (13) and (14) of Table 1 show that there is more uncertainty in performance pay jobs compared to non-performance pay jobs; this is especially the case for commission and/or piece rate paying jobs. One possible criticism of this measure is that it may not reect variation in output uncertainty across geographic regions. For example, a measure of uncertainty at the industry-state level may be more informative than one only at the industry level. Unfortunately Compustat North America, which reports information on rms with multiple locations, provides data only for the rm as a whole rather than for each of the individual rm locations, and only provides geographic information pertaining to the rm's head oce. Regardless though, I do not believe this is an important issue since is unlikely that a rm's compensation policy for a given production process shows signicant variation between regions.

12

12 Although I am unable to nd literature supporting this, I do not think that two workers doing the same job for the same company would be remunerated in dierent ways just because they are in dierent geographic regions. For example, I would nd it hard to believe that a Sandwich Artist employed at a Subway Sandwiches restaurant in London, Ontario (yes, this was my job title when I worked there!) would be paid for performance while one at a restaurant in Montréal, Québec would not.

13

4

Results

According to the principal-agent model outlined in Section 2, if there are no dierences in the risk preferences of workers in performance-based and non-performancebased jobs (that is, if there is no risk sorting), then there should be a negative relationship between measures of uncertainty in production and the use of performancebased payments. In Table 3 I present the correlation matrix between measures of uncertainty, the use of various forms performance-based payments, measures of earnings risk faced by workers, and measures of workers' risk preferences. Columns (1) and (2) show that, in contrast to the predictions of the model, there is a non-negative relationship between the use of performance pay and uncertainty. This holds more for commission and/or piece rate jobs than for bonus paying jobs and holds when considering either cross-sectional or intertemporal uncertainty. Note however that the positive correlations between risk tolerance and the use of bonuses and commissions and/or piece rate payments, and the negative correlations between risk aversion and the use of these performance-based payments shown in the last two rows of columns (3) and (4) suggest that the sorting of risk-loving workers into performance pay jobs may be taking place.

Positive correlations between being in a job that pays for

performance and both cross-sectional and intertemporal risk faced by workers (rows (5) and (6) of columns (3) and (4)) suggests further that it is the performance paid workers' higher tolerance for risk that is driving their sorting into these types of jobs. It is thus plausible, as shown in Section 2, that sorting might be leading to the nding of a positive relationship between uncertainty and the use of performance-based payments. I examine whether this is the case more formally in the following subsections by looking rst at the issue of sorting, then at the uncertainty-performance pay relationship, and the role that sorting plays.

14

Sorting To determine whether more risk-loving workers are sorting into performance pay jobs I use the following binary probit model:

P rob(P P J = 1) = f (r, X), where

PPJ

(7)

indicates being in a particular type of performance pay job,

of worker risk preferences, and

X

r is a measure

is other worker characteristics that have been shown

to be correlated with working in a performance pay job. This includes education, experience, and union status. I use the risk tolerance measure and the risk aversion measure described in Section 3 in separate specications as these measures exhibit a high degree of correlation (-0.9533). I examine risk sorting into bonus paying jobs and commission and/or piece rate paying jobs separately. This is consistent with the view taken in the performance pay literature that jobs paying explicit forms of performance pay (commissions and piece rate pay) are dierent from those paying implicit forms of performance pay (bonuses).

In general, although both bonus paying jobs and

commission and/or piece rate paying jobs reward workers for good performance, the subjective nature of bonuses makes bonus paying jobs more similar to jobs that do not pay for performance than are commission and/or piece rate paying jobs. I consider rst the sorting of risk-loving workers into bonus paying jobs.

The

positive coecient on risk tolerance in column (1) and the negative coecient on risk aversion in column (2) of Table 4a are indeed consistent with such risk sorting. In particular, a worker at the 75

th

percentile of the risk tolerance distribution has

more than a 5% higher probability of being in a bonus paying job than a worker th

at the 25

percentile (35.7% vs.

33.9%), and a worker at the 90

th

percentile has

almost an 11% higher probability of being in a bonus paying jobs than one at the 25

th

percentile (37.6% vs. 33.9%).

13

th

Similarly, a worker at the 25

percentile of the

risk aversion distribution has 6.4% higher probability of being in a bonus paying job than one at the 75

th

percentile (35.9% vs. 33.8%).

13 These and all further reported predicted probabilities are calculated holding the other covariates at their mean value.

15

Although this provides evidence that risk-loving workers are more likely to be found in jobs that pay bonuses than those that do not, this can only be attributed to risk preferences if bonus paying jobs expose workers to higher levels of earnings risk. To examine this I include in columns (3) and (4) the two measures of earnings risk that were constructed using workers' earnings reported in the PSID. Positive coecients on both the cross-sectional and the intertemporal measures of workers' earnings risk indicates that bonus paying jobs are associated with higher levels of earnings risk than jobs that do not pay bonuses. Jobs that expose workers to earnings th

risk at the 75

percentile of the cross-sectional risk distribution are 53% more likely

to be bonus paying than those exposing worker to earnings risk at the 25

th

percentile

of the distribution (29.3% vs. 44.8%). For corresponding points of the intertemporal risk distribution, this value is 2.5% (33.9% vs. 34.8%). This is important because it suggests that risk-loving workers are sorting into bonus paying jobs because of their capacity to tolerate higher levels of earnings risk. I apply the same examination to the sorting of risk-loving workers into commission and/or piece rate paying jobs. Results presented in Table 4b indicate that not only are more risk-loving workers sorting into these types of jobs, but that this sorting is stronger for commission and/or piece rate paying jobs than for bonus paying jobs. th

For this type of performance pay, workers at the 90

and 75

th

percentiles of the

risk tolerance distribution have a 37.2% and 16.6% higher probability of being in a commission and/or piece rate paying job than workers at the 25 vs.

9.7% and 11.3% vs.

9.7% respectively).

th

percentile (13.3%

Comparing workers at these points

of the risk aversion distribution indicates a 16 % lower probability of being in a commission and/or piece rate paying job.

14

Furthermore, positive coecients on

the two measures of earnings risk presented in columns (3) and (4) provide strong evidence that commission and/or piece rate paying jobs exhibit high levels of earnings th

variation, both across workers and over time. Comparing risk at the 25 th

that at the 75

percentile to

percentile of the cross-sectional and intertemporal risk distributions

indicates an increase in probability of 463% (4.4% vs 24.7%) and 5.1% (8.5% vs

14 The predicted probability of being in a commission and/or piece rate paid job is the same at th and 75th percentiles of the risk aversion distribution. the 90

16

9.0%) respectively. Results presented in this subsection do appear to be in line with the sorting of more risk-loving workers into performance pay jobs. In fact, it is interesting to note that the extent of risk sorting into the very risky, commission and/or piece rate paying jobs is greater than the sorting into the somewhat less risky, bonus paying jobs.

Other results in Tables 4a and 4b are consistent with previous work done

on performance pay in the US: education and experience are positively related to being paid for performance (especially bonuses), whilst being in a union is negatively related to being paid for performance.

The Uncertainty-Performance Pay Relationship In this subsection I examine the relationship between a rm's use of performancebased payments and the level of output uncertainty it faces. I continue to use the binary probit model from the previous subsection, but include now the industry-specic measures of uncertainty in production faced by employers that were constructed using data from Compustat North America. I rst examine the relationship between performance pay and uncertainty without controlling for workers' risk preferences. This is to conrm that the positive relationship found in earlier studies that do not control for sorting, also holds for workers in the PSID. I present the results for bonus paying jobs in columns (1) and (2) of Table 5a, and those for commission and/or piece rate paying jobs in columns (1) and (2) of Table 5b.

15

Positive coecients on

both measures of output uncertainty for both bonus paying and commission and/or piece rate paying jobs conrm the ndings of earlier researchers that jobs with these kinds of payment schemes are exposed to higher levels of output uncertainty. For the cross-sectional measure of uncertainty, rms at the 90

th

th

and 75

percentiles of the

distribution are 15.8% and 3.5% more likely to pay bonuses (38.8% vs. 33.8% and 35.1% vs. 33.8%), and 207% and 32.9% more likely to pay commissions and piece

15 Because of the high correlation between the cross-sectional and intertemporal measures of economic uncertainty, I use separate specications for each measure (0.9941).

17

rates (21.2% vs. 6.9% and 9.2% vs. 6.9%) than rms at the 25 th

intertemporal measure of uncertainty, rms at the 90

th

and 75

percentile. For the

th

percentiles of the

distribution are 15.0% and 3.8% more likely to pay bonuses (38.8% vs. 33.8% and 35.1% vs. 33.8%), and 179% and 33.9% more likely to pay commissions and piece th

rates (20.3% vs. 7.3% and 9.8% vs. 7.3%) than those at the 25

percentile.

Results obtained in columns (1) and (2) of Tables 5a and 5b, while consistent with the ndings of earlier studies, are at odds with the negative relationship implied by the principal-agent model and, with the ndings of more recent studies which control for sorting. In columns (3) and (4) of these tables I include controls for workers' risk preferences. This is in order to investigate whether risk sorting may be leading to the positive relationship observed between performance pay jobs and uncertainty in the PSID. As the coecients on the measures of uncertainty presented in these columns show, even though controlling for workers' risk preferences does reduce the positive relationship between performance pay and output uncertainty that was found in columns (1) and (2), it is not sucient to produce the negative relationship that is predicted by the principal-agent model. Controlling for risk preferences reduces the coecient on uncertainty by a little over 3% for bonus paid workers but only around 0.3% for commission and/or piece rate paid workers. Furthermore, while the reduction in the coecient on uncertainty (both measures) is statistically signicant for bonus paying jobs, it is not for commission and/or piece rate paying jobs. Overall then, although I showed earlier that risk-loving workers are more likely to work in performance pay jobs, I am unable to nd evidence that this sorting behaviour plays a substantial role in the positive empirical relationship between performance pay and output uncertainty. It is possible that this might be due to worker heterogeneity in my Compustat measures of output uncertainty. As Lazear (1986) and Lemieux, MacLeod, and Parent (2009) note, rms that face a heterogeneous labour force may benet from the sorting role of a performance-based compensation system. Recent work by Bellemare and Shearer (2010) indicate that the ability of such systems to attract less risk-averse workers could lead to as much as a 25% increase in rm prots. If my measures of output uncertainty are positively related to worker heterogeneity, then this may be contributing to the positive relationship,

18

and the inability to control for this variation in heterogeneity may explain why this relationship persists even after accounting for worker risk sorting.

5

Conclusion

Agency theory predicts that incentive pay and uncertainty in output should be negatively related; a prediction which has been an important focal point of the empirical evaluation of this theory.

Much of the early work, which did not account for the

sorting of more risk-loving workers into incentive-based contracts, has had a dicult time showing this relationship. More recent work however, which does account for this sorting behaviour, has been able to provide evidence for this negative relationship, although only in very narrow contexts. In this paper I examine the issue of risk sorting into jobs that pay for performance, and the role that such sorting plays in the relationship between performance pay and uncertainty in output for a representative sample of the male U.S. workforce. Using information from the PSID I present convincing evidence that 1) performance pay compensation systems expose workers to higher levels of earnings risk and 2) workers in performance pay jobs are more risk-loving than those found in non-performance pay jobs. I show that even within the class of jobs that pay for performance, those types of performance pay jobs that expose workers to higher levels of earnings risk are associated with stronger sorting behaviour. I am however, unable to provide evidence of a negative relationship between performance pay and uncertainty in output, even after controlling for worker risk sorting. Note that this does not necessarily imply that such a relationship does not exist or that risk-sorting does not play a more substantial role in the observed positive relationship. Rather, as alluded to at the end of Section 4, because I use somewhat aggregated measures of uncertainty that I import from the Compustat into the PSID, it may point to the need for better measures of uncertainty in output that are purged of any determinants of performance pay. Finding that workers are sorting into jobs based on their risk preferences is important as it indicates the ecient allocation (at least along a risk-dimension) of

19

workers in the labour market. In particular, it suggests that rms that implement performance pay compensation schemes may be beneting from cost savings brought about by this compensation's risk-sorting function.

Finally, in light of the debate

on the role of risk and risk preferences in contract theory, it provides evidence that these are important factors which need to be taken into consideration in order to fully understand contract structures.

20

References [1] Daniel A. Ackerberg and Maristella Botticini. Endogenous matching and the empirical determinants of contract form. Journal of Political Economy, 110(3):564 591, 2002. [2] Rajesh K. Aggarwal and Andrew A. Samwick. The other side of the trade-o: The impact of risk on executive compensation. Journal of Political Economy, 107(1):65105, 1999. [3] Douglas Allen and Dean Lueck. Contract choice in modern agriculture: Cash rent versus cropshare. Journal of Law and Economics, 35(2):397426, 1992. [4] Douglas W. Allen and Dean Lueck.

Risk preferences and the economics of

contracts. American Economic Review, 85(2):447451, 1995. [5] Douglas W. Allen and Dean Lueck. The role of risk in contract choice. Journal of Law, Economics, and Organization, 15(3):704736, 1999.

[6] Robert B. Barsky, F. Thomas Juster, Miles S. Kimball, and Matthew D. Shapiro. Preference parameters and behavioral heterogeneity: An experimental approach in the health and retirement study. Quarterly Journal of Economics, 112(2):537 579, 1997. [7] Charles Bellemare and Bruce Shearer. Sorting, incentives and risk preferences: Evidence from a eld experiment, 2006. [8] Charles Bellemare and Bruce Shearer. Multi-dimensional heterogeneity and the economic importance of risk and matching: Evidence from contractual data and eld experiments, 2010. [9] Holger Bonin, Thomas Dohmen, Armin Falk, David Human, and Uwe Sunde. Cross-sectional earnings risk and occupational sorting: The role of risk attitudes. Labour Economics, 14(6):926937, 2007.

[10] James N. Brown and Audrey Light.

Interpreting panel data on job tenure.

Journal of Labor Economics, 10(3):219257, 1992.

[11] Sarah Brown, Michael Dietrich, Aurora Ortiz, and Karl Taylor. Self-employment and risk preference, 2007. [12] Sarah Brown and Karl Taylor. Education, risk preference and wages, 2007.

21

[13] John Core and Wayne Guay. The use of equity grants to manage optimal equity incentive levels. Journal of Accounting and Economics, 28(2):151184, 1999. [14] Thomas Dohmen and Armin Falk.

Performance pay and multi-dimensional

sorting: Productivity, preferences and gender.

Technical report, Institute for

the Study of Labor (IZA), 2006. [15] John E. Garen. Executive compensation and principal-agent theory. Journal of Political Economy, 102(6):11751199, 1994.

[16] Christian Grund and Dirk Sliwka. Evidence on performance pay and risk aversion. Economics Letters, 106(1):811, 2010. [17] Eric Hilt. The negative trade-o between risk and incentives: Evidence from the american whaling industry. Explorations in Economic History, 45(4):424444, 2008. [18] Bengt Holmstrom and Paul Milgrom. Multitask principal-agent analyses: Incentive contracts, asset ownership, and job design. Journal of Law, Economics, and Organization, 7:2452, 1991.

[19] Charles A. Holt and Susan K. Laury. Risk aversion and incentive eects. The American Economic Review, 92(5):16441655, 2002.

[20] Edward P. Lazear. Salaries and piece rates. Journal of Business, 59(3):405431, 1986. [21] Edward P. Lazear.

Performance pay and productivity.

American Economic

Review, 90(5):13461361, 2000.

[22] Thomas Lemieux, W. Bentley MacLeod, and Daniel Parent. Performance pay and wage inequality. Quarterly Journal of Economics, 124(1):149, 2009. [23] Ming-Ching Luoh and Frank Staord. Estimating risk tolerance from the 1996 psid, 2005. [24] Harry J. Paarsch and Bruce Shearer. Piece rates, xed wages, and incentive effects: Statistical evidence from payroll records. International Economic Review, 41(1):5992, 2000. [25] Daniel Parent. Methods of pay and earnings: A longitudinal analysis. Industrial and Labor Relations Review, 53(1):7186, 1999.

22

[26] Canice Prendergast. The tenuous trade-o between risk and incentives. Journal of Political Economy, 110(5):10711102, 2002.

[27] John Robst, Richard Deitz, and KimMarie McGoldrick.

Income variability,

uncertainty and housing tenure choice. Regional Science and Urban Economics, 29(2):219229, 1999. [28] Konstantinos Serfes.

Risk sharing vs. incentives: Contract design under two-

sided heterogeneity. Economics Letters, 88(3):343349, 2005. [29] Bruce Shearer. Piece rates, xed wages and incentives: Evidence from a eld experiment. The Review of Economic Studies, 71(2):513534, 2004. [30] Linda K. Stroh, Jeanne M. Brett, Joseph P. Baumann, and Anne H. Reilly. Agency theory and variable pay compensation strategies. Academy of Management Journal, 39(3):751767, 1996.

[31] Julie Wulf. Authority, risk, and performance incentives: Evidence from division manager positions inside rms. Journal of Industrial Economics, 55(1):169196, 2007.

23

Tables Table 1. Sample Statistics, PSID (1)

(2)

Full Sample

Bonus

(3) Commiss/Piece Rate

1. Age 2. Years of Education 3. Potential Experience 4. Married

13.7

(2.1)

(2.1)

(2.0)

19.2

18.6

20.2

(9.5)

(8.9)

(9.0) 0.81

0.78

0.82

0.88

(0.41)

(0.38)

(0.32)

0.17

0.10

0.08

(0.37)

(0.29)

(0.28)

8.96

10.70

10.32

(5.79)

(7.05)

(6.35)

47927.26

58573.56

56052.42

(45213.05)

(51883.92)

(46140.91)

3.098

2.993

2.925

(1.584)

(1.590)

(1.644)

0.289

0.300

0.311

(0.159)

(0.164)

(0.172)

0.196

0.208

0.232

(0.047)

(0.047)

(0.055)

0.016

0.020

0.035

(0.054)

(0.055)

(0.094)

11. Cross-Sectional Earnings Risk 12. Intertemporal Earnings Risk

2503011

2533392

3877910

(2892938)

(2926042)

(3519249)

357676.7

359957.4

540892.4

(417577.8)

(423101.2)

(513457.8)

3130

912

210

11563

3911

949

15. No. of Jobs 16. No. of Observations

(8.8)

14.0

(0.39)

10. Risk Tolerance

Uncertainty

(8.9)

13.5

0.83

9. Risk Aversion

14. Intertemporal Output

(9.3)

(0.38)

7. Hourly Wage (1979$)

Uncertainty

39.9

0.80

6. Union

13. Cross-Sectional Output

38.6

(0.40) 5. White

8. Yearly Income

38.6

Notes: Standard deviation in parentheses. 24

Table 2. Measures of Income Risk Constructed from Compustat North America Industry

Cross-

Intertemporal

Sectional 1. Agriculture, Forestry, Fisheries 2. Mining 3. Construction

91889

16158

173536

25776

4035

851

4. Manufacturing

1875926

316697

5. Transport, Communication, Publ

1413018

159601

6. Wholesale and Retail Trade

8221348

1168494

7. Finance, Insurance, and Real Estate

Utilities 1018391

16694

8. Business and Repair Services

40557

7248

9. Personal Services

10989

3045

10. Entertainment and Recreation

79856

8077

411628

24634

Services 11. Professional and Related Services

Notes: The cross-sectional measure is the industry-specic variance of residuals constructed from a regression of quarterly revenue on rm size, and year and quarter dummies. The intertemporal measure is the industry-specic average of residuals constructed from rm-specic regressions of quarterly revenue on rm size, and year and quarter dummies.

25

1.0000

11388 0.0039

11388 0.1315*

11388

1.0000

11388 0.9941*

11388 0.0076

11388 0.1425*

11388

Uncertainty

Intertemp

Uncertainty

Bonus

Job

CPR

Job

Uncertainty

Uncertainty

CS

Intertemp

CS

Table 3. Correlation Matrix, PSID

-0.0117

11388

-0.0148

11388

Risk

Aversion

11418

-0.0487*

11418

0.0487*

11418

0.0565*

11418

0.1826*

11418

0.0229*

11418

1.0000

Bonus Job

11418

-0.0319*

11418

0.0403*

11418

0.1049*

11418

0.2300*

11418

1.0000

CPR Job

11418

-0.1171*

11418

0.1207*

11418

0.0680*

11418

1.0000

CS Risk

11418

0.0022

11418

0.0007

11418

1.0000

Risk

Intertemp

* denotes a signicance level of 5%.

# observations in bold.

Intertemp Risk refers to the intertemporal measure of earnings risk experienced by workers.

CS Risk refers to the cross-sectional measure of earnings risk experienced by workers.

CPR Job refers to commission and/or piece rate jobs.

Intertemp Uncertainty refers to the intertemporal measure of uncertainty in output experienced by rms.

Risk

11418

-0.9533*

11418

1.0000

Tolerance

Note: CS Uncertainty refers to the cross-sectional measure of uncertainty in output experienced by rms.

0.0127

11388

0.0173

11388

11388

Risk

11388

0.0451*

0.0452*

Intertemp

Tolerance

11388

11388

Risk

Risk

0.1333*

0.1613*

CS

26

11418

1.0000

Aversion

Risk

Table 4a. Sorting by Bonus Paid Workers, PSID (1) Risk Tolerance

(2)

0.2360***

0.1589**

(0.0772)

(0.0853)

Risk Aversion

-0.0121**

(0.0059)

(0.0063)

Intertemporal Risk

Potential Experience

Potential Experience

2

Married

(4)

-0.0195***

Cross-Sectional Risk

Education

(3)

3.9217***

3.9358***

(0.6816)

(0.6819)

1.8886***

1.8885***

(0.4319)

(0.4345)

0.0999***

0.1001***

0.0607***

0.0608***

(0.0147)

(0.0148)

(0.0073)

(0.0074)

0.0203***

0.0202***

0.0183***

0.0182***

(0.0086)

(0.0087)

(0.0089)

(0.0089)

-0.0005***

-0.0005***

-0.0005***

-0.0005***

(0.0002)

(0.0002)

(0.0002)

(0.0002)

0.1048

0.1030

0.0924

0.0910

(0.0648)

(0.0655)

(0.0597)

(.0604)

-0.4477***

-0.4475***

-0.3518***

-0.3514***

(0.0804)

(0.0805)

(0.0517)

(0.0519)

0.1208**

0.1209**

0.0920

0.0922

(0.0657)

(0.0656)

(0.0646)

(0.0644)

-2.0511***

-1.9231***

-2.2431***

-2.1617***

(0.1636)

(0.1773)

(0.1307)

(0.1629)

Observations

10832

10832

10832

10832

2 Pseudo R

0.0459

0.0457

0.0609

0.0608

Union

White

Constant

Notes: Dependent variable is an indicator variable indicating observations belonging to jobs which pay bonuses. Specications include year dummies. Standard errors in parentheses are clustered at the occupation level. *** indicates p-value < 0.01 ** indicates p-value < 0.05 * indicates p-value < 0.10

27

Table 4b. Sorting by Commission/Piece Rate Paid Workers, PSID (1) Risk Tolerance

(2)

0.3118**

(0.1458)

(0.1527)

Risk Aversion

-0.0265

(0.0173)

(0.0188)

Intertemporal Risk

Potential Experience

Potential Experience

2

Married

Union

White

Constant

Observations

2 Pseudo R

9.7213***

9.7478***

(2.0549)

(2.0516)

2.5689***

2.5732***

(0.5651)

(0.5688)

0.0502***

0.0507***

0.0550***

0.0551***

(0.0094)

(0.0094)

(0.0108)

(0.0108)

0.0438***

0.0436***

0.0452***

0.0450***

(0.0061)

(0.0062)

(0.0108)

(0.0108)

-0.0008***

-0.0008***

-0.0009***

-0.0009***

(0.0002)

(0.0002)

(0.0002)

(0.0002)

-0.0076

-0.0144

-0.0238

-0.0275

(0.0730)

(0.0726)

(0.0758)

(0.0753

-0.4870**

-0.4907***

-0.2109

-0.2121

(0.2394)

(0.2390)

(0.2521)

(0.2504)

0.3963***

0.3984***

0.2754**

0.2760**

(0.1181)

(0.1184)

(0.1197)

(0.1194)

-2.7047***

-2.4699***

-3.1854***

-3.0105***

(0.5327)

(0.5019)

(0.5653)

(0.5798)

7885

7885

7885

7885

0.0458

0.0447

0.1491

0.1488

Notes: Dependent variable is an indicator variable indicating observations belonging to jobs which pay commissions and/or piece rates. Specications include year dummies. Standard errors in parentheses are clustered at the occupation level. *** indicates p-value < 0.01 ** indicates p-value < 0.05 * indicates p-value < 0.10

(4)

-0.0336**

Cross-Sectional Risk

Education

(3)

0.4444***

28

Table 5a. Uncertainty-Performance Pay Relationship for Bonus Paying Jobs, PSID (1)

6 Cross-Sectional Uncertainty/10

Intertemporal Uncertainty/10

(2)

0.0173**

0.0169**

(0.0081)

(0.0080)

6

Potential Experience

Potential Experience

2

Married

(4)

0.1179**

0.1153**

(0.0609)

(0.0608)

Risk Tolerance

Education

(3)

0.2229**

0.2235**

(0.1118)

(0.1120)

0.1044***

0.1047***

0.1027***

0.1031***

(0.0213)

(0.0211)

(0.0213)

(0.0211)

0.0205***

0.0205***

0.0206***

0.0205***

(0.0060)

(0.0060)

(0.0059)

(0.0060)

-0.0005***

-0.0005***

-0.0005***

-0.0005***

(0.0002)

(0.0002)

(0.0002)

(0.0002)

0.1026*

0.1027*

0.1082*

0.1084*

(0.0579)

(0.0581)

(0.0592)

(0.0593)

-0.4384***

-0.4393***

-0.4359***

-0.4368***

(0.0685)

(0.0683)

(0.0677)

(0.0674)

0.1230***

0.1231***

0.1203***

0.1204***

(0.0421)

(0.0421)

(0.0416)

(0.0416)

-2.0845***

-2.0881***

-2.1329***

-2.1365***

(0.2659)

(0.2610)

(0.2599)

(0.2548)

Observations

10802

10802

10802

10802

2 Pseudo R

0.0463

0.0463

0.469

0.468

Union

White

Constant

Notes: Dependent variable is an indicator variable indicating observations belonging to jobs which pay bonuses. Specications include year dummies. Standard errors in parentheses are clustered at the industry level. *** indicates p-value < 0.01 ** indicates p-value < 0.05 * indicates p-value < 0.10

29

Table 5b. Uncertainty-Performance Pay Relationship for Commission/ PR Paying Jobs, PSID (1)

6 Cross-Sectional Uncertainty/10

Intertemporal Uncertainty/10

(2)

0.0835***

0.0833***

(0.0162)

(0.0161)

6

Potential Experience

Potential Experience

2

Married

Union

White

Constant

Observations

2 Pseudo R

(4)

0.5449***

0.5435***

(0.1431)

(0.1428)

Risk Tolerance

Education

(3)

0.4230**

0.4244**

(0.2114)

(0.2116)

0.0637***

0.0654***

0.0612***

0.0629***

(0.0096)

(0.0096)

(0.0097)

(0.0097)

0.0492***

0.0484***

0.0490***

0.0482***

(0.0075)

(0.0075)

(0.0073)

(0.0073)

-0.0009***

-0.0009***

-0.0009***

-0.0009***

(0.0002)

(0.0002)

(0.0002)

(0.0002)

-0.0031

-0.0042

0.0179

0.0168

(0.0607)

(0.0615)

(0.0615)

(0.0622)

-0.4470***

-0.4542***

-0.4340***

-0.4412***

(0.1433)

(0.1421)

(0.1424)

(0.1410)

0.4179***

0.4167***

0.4082***

0.4071***

(0.1236)

(0.1226)

(0.1239)

(0.1229)

-3.0728***

-3.0665***

-3.1800***

-3.1739***

(0.2092)

(0.1914)

(0.2237)

(0.2042)

7859

7859

7859

7859

0.0774

0.0732

0.0795

0.0754

Notes: Dependent variable is an indicator variable indicating observations belonging to jobs which pay commissions and/or piece rates. Specications include year dummies. Standard errors in parentheses are clustered at the industry level. *** indicates p-value < 0.01 ** indicates p-value < 0.05 * indicates p-value < 0.10

30

Figures

Figure 1

31

Data Appendix In this data appendix I describe how job spells are constructed, and how the two types of performance pay are identied, in the Panel Study of Income Dynamics (PSID). All data was obtained from the PSID's website: http://psidonline.isr.umich.edu/. The period of interest covers the years 1993-2007 and comprises respondents that were either part of the cross-sectional sample drawn from the Survey Research Center (SRC) or the Survey of Economic Opportunity (SEO) conducted by the Bureau of the Census for the Oce of Economic Opportunity. I omit respondents entering as part of either the Latino (1990/1992) or Immigrant (1997/1999) samples.

Construction of Job Spells Identifying jobs spells in the PSID is dicult because, unlike the National Longitudinal Surveys (NLS), the PSID does not provide information linking yearly worker observations to specic 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 identied 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 16

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, 2004 and 2006. 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 1993. . . 1997, 1998, 2000, 2002, 2004, 2006.

However there is neither tenure, nor time since last interview information

available in 1998, 2000, 2002, 2004 or 2006. I get around this problem by applying

16 Job tenure is constructed using 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.

32

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 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 Indicators Although the PSID has been collecting information on performance pay since the early 1970s, much of this considered only a broad denition of performance pay, making it dicult to distinguish between dierent forms of performance pay.

Be-

ginning in survey year 1994, the PSID started asking respondents more detailed information about specic forms of performance pay. This is the information I use in this paper since it likely leads to a cleaner identication of the dierent types of 17

performance pay jobs. Specically, respondents are rst asked:

[ER3140] (In addition to this,) did you have any income from bonuses, overtime, tips, or commission? If the response is yes, the following questions determines which of these payments the respondent received: [ER3141] Which was that?  BONUSES [ER3142] Which was that?  OVERTIME [ER3143] Which was that?  TIPS [ER3144] Which was that?  COMMISSIONS The following questions were then asked to determine how much of each of these payments were received: [ER3146] How much was from bonuses? [ER3147] How much was from overtime? [ER3148] How much was from tips? [ER3149] How much was from commissions?

17 All variable names below correspond to the survey year 1994.

33

To conrm these indicators, and to possibly identify additional observations that belong to performance pay jobs it is also possible to classify observations using information on payment method: [ER2081] (On your main job,) are you (HEAD) salaried, paid by the hour, or what? 1. Salaried 2. Salary plus commission 3. Paid by the hour 4. Hourly plus tips 6. Hourly plus commission 7. Other Finally, those who reported receiving neither a pure salary nor a pure hourly rate (and in particular, those who responded other to the previous question) were given the option of classifying their earnings as one of the following [ER2086]: 1. Piecework; hourly plus piecework/production 2. Commission 3. Self-employed; farmer; prots 4. By the job/day/mile 18

5. Other.

18 Some years allow for additional categories.

34