Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Performance Pay and Wage Inequality
Evidence from Germany and Great Britain∗
Pieter De Vlieger
University of Michigan
May 28, 2016
Abstract
This paper tests of model of performance pay and wage inequality, proposed by Lemieux,
MacLeod and Parent, in Germany and Great Britain. In this model, changes in the return to skill
motivates employers to adopt performance pay schemes, which, in turn, translates into higher
wage inequality. I test the empirical predictions using the German Sozio-Oekonomisch Panel
(1984-2008) and British Panel Household Survey (1993-2008). I address non-standard issues of
misclassification to show I am less likely to detect the empirical predictions of the model. The
results are consistent with performance pay as a channel to deal with underlying changes in
returns to skill in Germany, but I find no evidence that supports this role for performance pay
in Germany. Using reweighting and novel RIF decomposition techniques, I assess the effect of
performance pay on the wage distribution. In Great Britain, I find performance pay has led
to a widening of the wage distribution in the upper half, which is consistent with the model of
performance pay. Despite a large increase in the dispersion of earnings in Germany, performance
pay does not seem to be a main contributor. Taken together, these results indicate that different
labor markets may deal with changes in the returns to skill in different ways, despite similar
trends in these returns.
∗[email protected]. I want to especially thank Stephen Machin, under whose supervision I started this project. Ialso want to thank John Bound, Charlie Brown, Gabor Kezdi, Jeff Smith, and Mel Stephens for useful commentsdiscussions and seminar participants at the University of Michigan. All remaining errors are my own.
1 Introduction
Performance pay forms has become increasingly important in the labor market, representing an in-
creasing fraction of the workforce and share of the total wage bill over time (Bryson et al. [2008]).
This has sparked a great deal of research interest in the impact of performance pay on a number
of labor market outcomes (see, for instance, Brown and Heywood [2002] for a cross-country anal-
ysis and Bryson et al. [2012] for more recent evidence on the US and Europe). One particular
outcome that has received a lot of attention is wage inequality. Piketty and Saez [2003] find that
performance pay workers are usually concentrated in the upper tail of the wage distribution where
increases in wage inequality have been particularly dramatic. Similarly, Bell and Van Reenen [2010]
provide evidence that performance pay and bonuses are associated with extreme wage inequality
in Great Britain, especially in the financial sector. However, also when taking a broader look at
the workforce, it is a well known result that performance pay jobs are associated with higher mean
wages compared to non-performance pay jobs (Pencavel [1978]; Brown [1990]; Booth and Frank
[1999]; Bryson et al. [2012]) and that performance pay leads to higher within-firm wage inequality
(Lazear [2000]).
Lemieux, MacLeod and Parent [2009], henceforth referred to as LMP, posit a model where
employers adopt performance pay schemes as an endogenous response to changes in the returns to
skills. On the one hand, technological advances have decreased monitoring costs, making it easier
for employers to implement performance pay schemes (Bloom and Van Reenen [2011]; Henneberger
et al. [2007]). On the other hand, Skill Biased Technological Change (SBTC) has increased the
productivity wedge between skilled and unskilled workers in recent decades (Autor et al. [2006]).
As introducing performance pay typically leads to sorting of more productive workers into these
types of jobs (see, for instance, Lazear [2000] for firm-level evidence), benefits may now exceed
technology investments. They test this model using PSID survey data from 1976 through 1998,
and find evidence that supports performance pay as an endogenous response to SBTC. Furthermore,
they show the increase in performance pay accounts for about a quarter of the increase in wage
inequality over this period.
It is not clear, however, whether performance pay is used as a means to channel changes in
the returns to skill across different countries – in other words, whether this model of performance
1
pay holds more generally in other countries. There are several reasons why it might not. First,
several institutional factors of the labor market, such as unions or the minimum wage, have been
shown to affect wage inequality (DiNardo et al. [1996]). These differ across countries and could
limit the extent to which performance pay can channel changes in the returns to skill. Second,
output is often hard to measure in many of the jobs that use performance pay schemes, casting
some doubt on monitoring as the main driver. Second, Finally, the SBTC hypothesis has been met
with some skepticism (Card and DiNardo [2002]), and it’s an open question whether this model of
performance pay and wage inequality can fit more recent trends where especially highly educated
workers have seen wage increases.
In this paper, I use micro survey data from Germany and Great Britain to test this model
of performance pay and wage inequality. As Machin and Van Reenen [1998] point out, rising
wage inequality is not strictly a US phenomenon, and the use of performance pay can be quite
heterogeneous across countries (Brown and Heywood [2002]). I focus on Germany and Great
Britain for two key reasons. First, these countries have had similar experiences to the US in
terms of wage inequality, but provide a different institutional setting and timing of wage inequality
trends. (Dustmann et al. [2009]). Second, the SOEP (Sozio-Oekonomisch Panel) in Germany
and the BHPS (British Household Panel Survey) in Great Britain provide data sources that are
very similar to the PSID, both in terms of scope and set-up. I focus on the period 1985-2008 for
Germany and 1991-2010 for Great Britain.
Additionally, I adjust my performance pay measures for misclassification error building on
the work of Bound et al. [2001] and Card [1996]. I show that I am less likely to detect the empirical
implications of the model, but also that several specifications may suffer from omitted variable
bias. I use outside information on the use of different types of performance pay jobs and try to
quantify the impact of misclassification on the final estimates. Finally, I investigate the impact of
performance pay jobs on the distribution of wages using a reweighting approach as proposed by
DiNardo et al. [1996] and use novel Recentered Influence Function (RIF) techniques as proposed
by Firpo et al. [2009] to decompose changes at various percentiles of the distribution.
I find two key results. First, this model of performance pay and wage inequality fits the
British data quite well. Wages in performance pay jobs are more closely linked to observable skills,
such as higher levels of education, and this relationship has become more pronounced over time.
2
In Germany, however, this is not the case. While wages in performance pay jobs are also typically
linked to productive characteristics, such as high education levels, this has not changed over time.
Second, the decomposition results further confirm this schism. Performance pay jobs in Great
Britain account for about 10% of the increase of the variance in the wage distribution from the
early 90s to the late 00s, and percentile decompositions confirm that these changes are concentrated
in the upper half of the wage distribution. While performance pay jobs in Germany account for
about 7% of the increase of the variance of the wage distribution between the mid 80s and late
00s, much of this dispersion is concentrated in the lower half of the distribution, a finding at odds
with the SBTC hypothesis. Overall, these results suggest that despite similar experiences in terms
of returns to skills and wage inequality, there are cross-country differences in how these returns to
skills are channeled through performance pay.
This paper proceeds as follows. Section 2 shortly describes the LMP model, and the key
empirical implications and econometric specification it gives rise to. Section 3 describes the data,
the performance pay measure, and the misclassification adjustment. Section 4 presents the empir-
ical tests fo the model. Section 5 presents RIF regressions and decomposition results. Section 6
concludes.
2 Performance pay: model and empirical implications.
This section provides a stylized overview of the model presented by LMP. In a first step, the model
considers wage setting at the job match level and distills four empirical implications. In a second
step, the model considers effects across the labor market and beyond the job-match alone.
2.1 Wages, selection rules and implications at the job level.
At the job match level, firms are free to choose between fixed and performance pay wage setting.
On the employer side, setting up performance pay contracts entails a cost-benefit analysis between
monitoring costs and returns to monitoring. Exerting effort also entails a cost-benefit analysis for
the worker. This situation can be represented by a utility function for worker i (equation 1) and
an output function for employer j (equation 2).
3
Uij = wij − exp(eij − αi) (1)
yij = kj + β · γj · eij (2)
As eij represents the effort level by the worker and αi represents the ability of that worker, the
utility function captures the incentive to exert effort according to one’s ability. The worker-specific
ability αi is modeled as a single-index type of ability and assumed to follow a normal distribution,
αi ∼ N(αi, σ2i ). The expected ability and variance are assumed to be common knowledge once
characteristics (such as schooling and experience) are observed.
The output function (equation 2) assumes a minimum level of output kj , regardless of
the worker providing effort. The parameters γj and β represent job-specific and market-specific
marginal returns respectively.1 This results in two key equations, where mj functions as an inter-
cept2 and Mj the monitoring cost: a fixed wage equation 3 and a performance pay wage equation
4.
wFWij = mj + β · γj ·(αi − σ2i
)(3)
wPPij = (mj −Mj) + β · γj · αi (4)
This leads to a selection rule, equation 5, that describes when, in expectation, wages under
performance pay contracts exceed wages in fixed wage contracts. Overall, performance pay jobs are
associated with high market-wide and job-specific returns to effort and a high (conditional) variance
of ability. Changes in the incidence of performance pay can happen either through a decrease of
monitoring costs (the right hand side) or an increase in benefits (the left hand side).
β · γj · σ2i ≥Mj (5)
In order to compare the returns to expected skills (such as education and experience), it is
1Some jobs are more sensitive to effort than others, which translates to heterogeneity in γj . One example, forinstance are cleaning services (higher return to effort) and parking guard (lower return to effort). Broad, market-wide changes to returns (such as SBTC) can change the return to effort in all jobs at the same time. For instance,machines can speed up the work professional cleaning services provide, while computers might make the work ofparking guards more efficients (e.g. when cars check out of the parking lot). It is these two different channels thatthese two parameters attempt to capture.
2More specifically, mj = kj + β · γj · log(β · γj).
4
useful to rewrite equation 3 and 4. First, it is reasonable to assume the variance is increasing in
expected ability. For simplicity, LMP assume that σ2i = δ · αi and plug this in equation 3. Equation
4 is rewritten to be more comparable.
wFWij = mj + β · γj · (1− δ)︸ ︷︷ ︸Lower skill return
·αi (6)
wPPij = (mj −Mj)︸ ︷︷ ︸Lower Intercept
+β · γj · αi + β · γj · (αi − αi)︸ ︷︷ ︸Additional error component
(7)
This leads to three empirical implications of performance pay at the job level:
(I.1) The wage intercept is lower for performance pay jobs.
(I.2) Observable skills are rewarded more in performance pay jobs.
(I.3) The variance term of the worker ability is higher in performance pay jobs.
2.2 Implications across jobs and for error components.
The results so far hold the job match constant. Relaxing this condition, leads to testable implica-
tions for job characteristics and changes in market-wide returns to effort β. First, LMP argue that
job characteristics might be rewarded less in performance pay jobs. They find returns to tenure are
lower in performance pay jobs, and therefore interpret tenure as a job characteristic. Nevertheless,
they acknowledge tenure could possibly be a productive worker characteristic as well. Second, the
variance of the firm-specific component might also be rewarded less in performance pay jobs. Third,
increases in β leads to increasing returns to education being more pronounced in performance pay
jobs than in non-performance jobs. In other words, LMP expect to find returns to education to
increase faster in performance pay jobs than in non-performance pay jobs.
(I.4) The return to observable job characteristics may be lower in performance pay jobs.
(I.5) The variance of the firm-specific component may be smaller in performance pay jobs.
(I.6) If returns to education increase, they do so faster in performance pay jobs.
5
2.3 Econometric Specification.
The implications can be tested using split sample regressions and pooled interaction regressions.
The split sample regression can be written as
wgijt = agt + xit · bgt + zijt · cgt + εgijt (8)
where g indexes performance pay or non-performance pay jobs, wgijt stands for the wage for worker
i in jobmatch j at time t, and xit and zijt stand for the observable worker and the observable job
characteristics respectively. The model assumes the error term consists of different error compo-
nents.
εgijt = dgt · ζi + νgij + ugijt (9)
where ζi is the unobserved (demeaned) ability of the worker, dgt is the return to this unobserved
ability, νgij represents an error component for the jobmatch and ugijt represents the idiosyncratic
error.
Alternatively, when adding in a vector of control variables, it is possible to pool all observa-
tions and impose the coefficients across these control variables are equal. Using interactions, the
model can then be tested using an indicator PPijt taking on the value 1 when worker i is in a
performance pay job j at time t.
wijt = at + appt × PPijt + xit · bt + xit · bppt × PPijt + zijt · ct + zijt · cppt × PPijt + εijt (10)
Both regressions will be used to test the model’s implications on the coefficients. The variance
analysis on the error components will be performed using the error terms from the split sample
regressions reported in equation 9.
6
3 Data and empirical strategy.
3.1 The SOEP and BHPS.
The SOEP and the BHPS are similar in scope, spirit and subject to the PSID.3 Households that
constitute a representative sample of the population as a whole are surveyed on a yearly basis.
There is entry and exit, so these panels are typically unbalanced. The PSID, SOEP and BHPS
started in 1968, 1984 and 1991 respectively and with about 5,000 households each. I focus on the
period 1985-2008 for the SOEP and the period 1991-2008 for the BHPS, the years for which the
data was made available to me.
There are two key advantages in using the BHPS and the SOEP to test the model. First,
it provides representative samples for relatively long periods and, second, the data are broadly
comparable to the PSID. However, some of the issues that arise with PSID data also carry over.
First, outcomes are self-reported and may therefore suffer from measurement error. Second, the
survey is relatively small relative to the total population and participation to these surveys is
entirely voluntary. As Bell and Van Reenen [2010] point out, people at the top end of the wage
distribution are probably more likely to opt out and could therefore be underrepresented in these
samples. Similary, Dustmann et al. [2009] find some differences in terms of wage inequality pattern
when comparing the SOEP to administrative income data. Third, even though the design of
these household surveys is broadly similar, there are inevitable differences in questions and actual
practices across countries. Therefore, these datasets can only present broad comparisons between
these countries.
The sample selection follows LPM for both datasets. Males aged 20 to 65 are retained, and
public sector employees are excluded, since the public sector wage sector wage-setting process may
work differently. Additionally, there is evidence of differences in performance pay practices across
the private and public sector (see, e.g. Bryson et al. [2008] for some descriptive evidence on Great
Britain). Wages are available on a yearly (BHPS) or monthly basis (SOEP). It is natural to rescale
these numbers to hourly earnings to account for hours worked. This means we need positive wages
and working hours for the analysis. Following LMP, we restrict the hourly wages to be between 1.00
3There is a large body of research that uses these panel surveys for cross-country analysis (see, e.g., Lersch (2012),Jones et al. (2009), Brynin and Longhi (2009), among others). One particularly interesting study, Fraessdorf et al.(2008), investigates the effect of capital income inequality on income inequality.
7
and 100.00 pounds (euros) in Great Britain (Germany).4 The East-German sample of the SOEP is
excluded, as it is customary to focus on the West-German sample when analyzing the SOEP (see,
e.g., Dustmann and van Soest [1998]).
3.2 Defining performance pay.
I use the measurement framework of LMP to define performance pay jobs, which consists of two
steps. In a first step, survey responses are used to see whether workers are ever paid for performance.
In a second step, a job spell is categorized as a performance pay job if the worker ever gets paid
for performance over this spell.5 This framework likely underestimates the number of performance
pay jobs, as some workers on these contracts may never meet the requirements to get paid for
performance. These worker-job matches would incorrectly be categorized as non performance pay
jobs.
Both the SOEP and BHPS contain questions on variable wage schedules. I follow Booth
and Frank [1999] to identify performance payments in the BHPS, using the question whether the
respondent’s pay “ever includes incentive bonuses or profit related pay”. This question is available
in all waves. In the SOEP, I use a similar question that asks whether respondents’ pay ever includes
profit-related bonuses or “Gratifikation”. This question is available in all waves. These measures
may misclassify performance pay jobs, as both question refer to piece rates, profit sharing and
other types of perfomance pay.6 In Germany, other bonus measures (such as Christmas bonuses)
are available, but typically do not depend on performance and are therefore not included in the
analysis.
The definition also poses a key problem that is typical of duration data – the endpoint
4LMP restrict the sample to earnings between 1.50$ and 100.00$. This accounts for about XXX percent of thesample in Germany and 0.1 percent of the sample in Great Britain. Additionally, I restrict observations in GreatBritain to have annual hours worked between 1,000 and 4,500 hours and observations in Germany to have monthlyearnings above than 600 euros. These sample restrictions result in an additional loss of XXX percent of the sample inGermany and 3.5 percent of the sample in Great Britain. The excluded observations are mainly situated in the lowertail of the earnings distribution. The results are robust to the inclusion of the upper tail outliers – these upper tailoutliers are driven by low reported hours worked, not high earnings. To stay true to the setup of LMP, subsequentresults will be based on the restricted interval.
5Both the LMP and the tenure variables in the SOEP are based on the employer-employee relationship. TheBHPS, however, asks people how long they have held their current occupational status, treating promotions as an“occupational job change”. This means the tenure variable needs to be adjusted for promotion, as in Blundell et al.(2008). I do this using the job history files of the BHPS.
6The PSID sample in LMP also suffers from misclassification error. Yet, LMP argue they are more likely to classifya job as a performance pay job. Therefore, their misclassification problem is the reverse of the one discussed in thispaper.
8
problem. If a person moves out of a performance pay job after joining the panel survey or enters a
performance pay job just before leaving the panel survey. We may misclassify these job matches as
non-performance pay jobs if actual performance pay is received before or after the observed waves.
I adjust for this issue using the three step endpoint correction method used in LMP, and – in line
with their results – find that this adjustment does not adjust the results. Therefore, the reported
regressions results do not adjust for this endpoint problem.
3.3 Misclassification Error.
As mentioned above, the performance pay indicator covers not only piece rates and similar types
of performance pay, but also profit sharing and other types of pay where final renumeration does
not necessarily strictly depend on the worker’s effort level. I investigate what this misclassification
error means for the regression equations 8, 9 and 10 building on the methodology of Bound et al.
[2001] and Card [1996].7 For simplicity, I leave out job characteristics, although the results carry
over in a straightforward way.
Denote PP ∗it as the true performance pay status and PPit as the observed performance
pay status. I define π as the true fraction of performance pay workers and p as the observed
fraction of performance pay workers. In the scenario considered here, π < p. I further define
P(PPit = 1|PP ∗it = 1) ≡ q1 and P(PPit = 1|PP ∗it = 0) ≡ q0 with q0 < q1. Since misclassification is
only showing up for workers reporting being in a performance pay job, q1 = 1. The split sample
econometric specification now contains some error for both split samples. When considering the
return to the observable productive characteristics, I show the following relationships hold:
if PP = 1 : E[wPPit |Xit] = 1p{q1πb
PP ∗t + q0(1− π)bnPP
∗t }xit
if PP = 0 : E[wnPPit |Xit] = 11−p{(1− q1)πb
PP ∗t + (1− q0)(1− π)bnPP
∗t }xit
(11)
The coefficients are contaminated by misclassification, but the final extent of its effect depends
on the actual misclassification rates. Outside data is not available for Germany, but the WERS in
Great Britain provides some notion of possible misclassification.8 It covers years 1998 and 2004, and
7To be clear, I assume away any misclassification error that arises from survey response mistakes. Misclassificationonly arises from the questions pooling different types of performance pay. The technical details are presented in moredetail in appendix XXXX.
8It is far from an ideal measure though, as the percentages apply to firms and establishments, not workers. Larger
9
indicates that about 35% of workplaces use profit-related pay, about 20% use employee ownership
schemes, and the use of performace-related pay in British workplaces went up from about 20% to
45% [Kersley et al, 2006].
To fix ideas, I assume about 50% of the reported performance pay jobs are actual performance
pay jobs, with the remainder consisting of profit sharing pay and similar payment schemes, so
q0 = 0.50. About 70% of my sample reports being in a performance pay job, so I set p = 0.70, with
π = 0.35. As mentioned before, q1 = 1.00. Under this scenario, the estimated coefficient on xit is
0.5× bPP ∗t + 0.5× bnPP ∗
t for the performance pay sample and bnPP∗
t for the non performance pay
sample. Therefore, misclassification only biases coefficients in the performance pay sample and are
less likely to confirm the testable implications.
The impact of misclassification error on the interacted regressions is somewhat more involved.
Suppose the misclassification error can be characterized as PP ∗it = γ0 +γ1PPit+xitγ2 +ηit. If γ2 is
zero, I show that γ1 = πp
[q1−p1−p
], or, under our assumptions, γ = 0.50. Nevertheless, it is reasonable
to assume γ2 is different from zero. Then the observable characteristics in equation 10 will show up
in this regression, but also in the performance pay indicator. This will lead to interactions of these
observables that are now part of the error term, leading to omitted variable bias. Additionally, the
formula for γ1 also changes to
γ1 =Cov(PP ∗it, PPit)− (q1 − q0)c′VXXcV ar(PP ∗it, PPit)− (q1 − q0)2c′VXXc
(12)
Whenever adjustments are made below, they will be made using the simplification that γ2 is
zero. A more completely discussion can be found in the appendix.
3.4 Descriptive statistics and evidence.
Table 1 presents descriptive statistics for both Great Britain and Germany, comparing sample
means between performance pay and non-performance pay jobs. Workers in performance pay jobs
typically earn more, work longer hours, are better educated and stay with their employer longer
than their counterparts in non-performance pay jobs. Unionization rates in Great Britain are higher
firms may not provide performance pay to all their workers, and even when they provide it to all workers, theseoutcomes do not reflect on workers directly. Nevertheless, they present a useful point of departure. See the appendixfor more detail.
10
for performance pay jobs, reflecting the well known positive association between unionization and
firm size.9 Unionization rate questions are not available for most waves in the SOEP (see, e.g.,
Goerke and Pannenberg [2008] that use NACE codes to identify unionization membership in the
SOEP).10 Differences in marital status and minority rates across performance pay workers and non-
performance pay workers are not very pronounced. Table 2 also highlights workers in performance
pay jobs tend to be higher educated.11
Figure 1 presents graphical evidence on trends and importance of performance pay in the
British (panel [a]) and German (panel [b]) workforce. On the one hand, performance pay and
performance pay jobs incidence in Great Britain increased in the early nineties, largely flattening
out in the subsequent years. This is largely consistent with the findings of Bryson et al. [2008]
and Bloom and Van Reenen [2011]. On the other hand, Germany has experienced a more steadfast
growth in performance pay and performance pay jobs, especially after 1995. Both figures also
highlight end-point problems do exist in the data, but affect performance pay job incidence in
a minimal way.13 Furthermore, panels [c] and [d] (figure 1) confirm that trends and changes in
performance pay incidence are robust to redefining our measure of performance pay jobs as having
received performance pay at least one out of five or two times while being observed on the job.
Figure 2 presents wage distributions and the change in the standard deviation of log wages
across performance pay and non-performance pay jobs. Panel [a] confirms that British performance
pay wages have a higher mean, median and variance than their non-performance pay counterparts.
Panel [b] shows that these differences are markedly stronger in Germany. This could be explained
in two ways. First, Booth and Frank (1999) point out that the wording of the BHPS question
we use to identify performance pay does not strictly distinguish between incentive pay and piece
rates or profit sharing. The extent to which this happens in Germany may be different. Second,
performance pay has permeated the labor market much more in Great Britain, while performance
pay in Germany is largely part of the rising upper tail inequality in the German wage distribution
9This is not the first paper to find higher unionization rates in Great Britain for performance pay jobs. Boothand Frank [1999] provide similar results based on BHPS data and use the same variable to identify performance pay.
10It is therefore worthwhile to note here that all regressions in the subsequent analysis will control for (1-digit)industry (10 in Great Britain, 17 in Germany) and (1-digit) occupation codes (10 in both Great Britain and Germany).
11For Great Britain, we keep the seven educational levels as defined by the BHPS.12 For Germany, we define fourdummy variables: University (university/ technical college, college abroad), Technical (technical college), Vocational(apprenticeship, vocational school, etc.), None.
13End-point problems in the period 2005-2008 pose no real problem in the BHPS.
11
(Dustmann et al. [2009]).
Panel [c] and [d] present the standard deviation of wages across performance pay jobs and
non performance pay jobs, and cast some doubt on the link between performance pay and wage
inequality in Germany. Increases in wage inequality seem to be concentrated in performance pay
jobs in Great Britain (panel [c]), but Germany exhibits strong increases in wage inequality in both
jobs. If anything, wage inequality increases faster in non performance pay jobs than in performance
pay jobs (panel [d]). presents graphical evidence of the link between wage inequality (as measured
by the standard deviation of log hourly earnings) and performance payment.
4 Regression framework.
4.1 Effect of performance pay on wages.
Table 3 presents a simple framework to analyze the effect of performance pay on wages. All five
regression specifications control for standard worker specific controls (education, a quadratic in po-
tential experience and dummies for ethnicity, marital status and for region) and job characteristics
(union status, a quadratic in job tenure and dummies for industry and occupation).14 The speci-
fications also control for year effects. The first four columns highlight that performance pay jobs
are associated with higher wages, even when controlling for receiving performance pay during the
current year (both under an OLS and a fixed effects framework). Column five introduces job match
fixed effects alongside worker fixed effects and estimates the effect of receiving performance pay
over the past year. The positive and statistically significant coefficients provide evidence that, even
after taking into account individual and job-specific characteristics, performance pay is associated
with higher wages and does not merely replace base pay in non-performance pay jobs. Overall, the
effects are consistent with positive sorting into performance pay as predicted by Lazear (1986) and
line up very closely with the results LMP find using the PSID.
14These regional fixed effects (or regional dummies) capture local determinants such as local unemployment levels.Industry dummies capture the union status of workers for Germany. Education is defined in terms of degrees andreported for three categories (high, medium, low). In Germany these are set to university (high), technical (medium)and vocational (low); in Great Britain, these are set to first or higher degree (high), HND, HNC or teaching (medium)and A-level (low). Lower levels of education in Great Britain are included in the regressions but not reported forpresentation issues. The estimated effects are consistent with the reported results unless otherwise noted.
12
4.2 Returns to worker and job characteristics.
Table 4 sets up an OLS regression framework and tests an identical specification on a split sample
(a sample of workers in performance pay jobs and a sample of workers in non-performance pay
jobs). The first two columns report results for Great Britain, the third and fourth column show
coefficients for Great Britain adjusted for misclassification. The final two columns report results
for Germany. The dependent variable, yit, is the log of hourly earnings and is indexed by worker (i)
and wave (t). The indicator PPit takes on two values; 1 if the worker i is in a performance pay job
during wave t and 0 if not. Education are sets of dummy variables for the relevant education levels
in Germany and Great Britain. The marginal effects of experience and tenure are evaluated at
one year. Overall, the results indicate that the returns to observable skills, such as experience and
education, tend to be higher under performance pay contracts than under non-performance pay
contracts. Whereas LMP found tenure to get rewarded less in performance pay jobs and therefore
interpreted it as a job-specific characteristic, there seem to be no real difference between the return
on tenure across both types of jobs in Germany or Great Britain. Adjusting for misclassification
error in Great Britain does not change the results substantially.
Table 5 pools all observations for each country in one sample, forcing the effect of the controls
(such as union or marital status) to be the same for both groups. The columns are as in table 4,
but the second column now adjusts for misclassification in the interacted indicator. A performance
pay job dummy and a set of interaction variables (education, experience and tenure interacted with
the performance pay job dummy) provide tests for the model implications.15
The first prediction, that performance pay jobs have lower intercepts than non performance
pay jobs, does not hold in any of the two countries. The intercept difference is minor and not
statistically insignificant in Great Britain. The high incidence of performance pay jobs in Great
Britain could possibly have leveled the intercepts. This intercept is much higher and statistically
significant at 1% for Germany, providing a much clearer rejection of the model. The second predic-
tion, a higher return to worker observables in performance pay jobs, is borne out in both countries
for high levels of education and experience. In contrast with the PSID results in LMP, tenure is
15These interaction variables entail six interaction variables for Great Britain and three interaction variables forGerman education levels. Two interaction variables are added for experience and tenure each, as the quadratic formsare also imposed on these interaction variables. As before, the effects of these quadratic interaction variables forexperience and tenure and evaluated at one year.
13
rewarded the same in both jobs. As the predictions of LMP were not as clearcut for this prediction,
I do not interpret this finding as support in favor or against the model. Adjusting for misclassi-
fication error does not change the point estimates in a meaningful way. The standard errors for
this column are bootstrapped, since the standard error computation for misclassification in the
interacted regression is not straightforward to compute (see appendix for details). Therefore, not
too much weight should be given to the standard errors.16 Overall, the predictions of the model
turn out to largely be confirmed and the misclassification correction shows there was a downward
bias in the original estimates.
Table 6 provides coefficient estimates that tests implication I.6, whether returns to education
have increased more rapidly in performance pay jobs than in non performance pay jobs. The
education variables are interacted with time period dummies for 3-year bins. Only the final period
interactions are reported for both countries to keep the table parsimonious, following LMP. The
prediction that returns to education have increased at a faster pace in performance pay jobs hold in
Great Britain, but less so in Germany. In the BHPS, these returns are largely captured by highly
educated workers. This is in line with prediction I.6, as it is mostly these education levels that have
witnessed large returns over this period in Great Britain. Misclassification error is not adjusted for
in these regressions, but would be largely have the same effect as it does in table 5.
In conclusion, we tested four of the six implications. The evidence on observable job charac-
teristics (I.4) was inconclusive in both countries, in part because the prediction was not clearcut.
Performance pay jobs in both countries reward observable skills more, in line with prediction I.2.
The intercept and SBTC predictions (I.1 and I.6 respectively) were firmly rejected in Germany,
while prediction I.6 holds up well in Great Britain. There seems to be sorting of workers across
both types of jobs in Germany, but changes in the returns to education do not seem to have changed
this as predicted by the LMP model.
16The point estimates on the experience and tenure variables are somewhat different. This is largely because itwas not straightforward to implement the bootstrap procedure where these variables were evaluated at one year.Therefore, these are the first order effects of the polynomial only (i.e. the linear part of the polynomial). The valuesfor these linear components are not very different between the adjusted and unadjusted regression before multiplyingthe relevant coefficients and standard errors by the relevant factor.
14
4.3 Variance component analysis of unobservables.
In order to test the final two predictions of the model, I perform an error component analysis using
the predicted error terms from equation 8. Conceptually, the idea is to estimate and compare the
variance of the worker and the jobmatch components from these error terms. Splitting up these
error components as in equation 9, I can rewrite
εppijt = dppt · ζi + νppij + uppijt
εnppijt = dnppt · ζi + νnppij + unppijt
The first equation is the error term of the regression that uses the performance pay job
sample; the second is the error term of the regression that uses the non-performance pay job
sample. The subscript j is added to capture jobmatch fixed effects. The structure imposed in this
step assumes an unobserved ability component (ζi) of worker i, a firm-specific wage component
(vij) and an idiosyncratic error term uijt. We present two different specifications, a model with a
worker component and a model with a worker and a job component, that are estimated separately
across two samples. One sample includes all workers, while a second sample focuses on people who
switched jobs only. Focusing on this switchers subsample is interesting as the underlying variance of
the person-specific component ζi is, in essence, the same for both performance and non-performance
pay jobs.
Table 7 presents the results of these different specifications for both the BHPS and the
SOEP. Consistent with prediction I.3, the variance of the worker-specific component is larger in
performance pay jobs than in non performance pay jobs. This holds for both samples and for both
the German and British datasets. Similar to the job characteristic findings in the previous section,
we do not find supportive evidence for prediction I.5, namely that the variance component of the
job-specific error component is smaller for performance pay jobs. Similar to the finding for job
characteristics, this could be interpreted as an inconclusive result that neither rejects or supports
the model’s predictions. Taken together, the predictions of the model hold up relatively well for
Great Britain, but are largely rejected by the German data.
15
4.4 Robustness checks.
Table 8 presents robustness checks for Great Britain. Since the performance pay job measure may be
crude, I use the two different definitions that are presented in figure 1 – having received performance
pay at least one out of five or two times while being observed on the jobmatch (column 1 and 2).
As there are some worries about the wording of the BHPS question, a subsample analysis of the
last 10 waves is performed using another, somewhat cleaner, question that asks about performance
pay (column 3).17 Finally, I also consider adding public sector workers to the sample (column 4).
The results show are robust to these different specifications, with the exception of the intercept.
Whereas it was insignificant before, it is not positive and significant, contradicting prediction I.1.
Table 9 presents robustness checks for Germany. The first two columns report the same
robustness checks as the first two columns in table 8. The final two columns include public sector
workers and females into the sample. Overall, the results for Germany are robust to these different
specifications.
Finally, it is possible the interval set out for the dependent variable (between 1 and 100 pounds
or euros) may be poorly chosen. However, the results are robust to winsorizing or trimming the
dependent variable or including the outliers above 100 pounds or euros (results not presented).
5 Decomposition of the wage distribution.
The central question this section explores is how changes in performance pay relate to wage inequal-
ities. A novel decomposition technique based on reweighting and Recentered Influence Functions
(RIFs) proposed by Firpo et al. [2009] is used, since this technique allows for decompositions at
percentiles and standard deviations, whereas the classic decomposition proposed by Oaxaca-Blinder
hold at the mean only [1973].
5.1 Rif-regressions and reweighting.
A RIF-regression is a standard regression where the dependent variable is replaced by the Recen-
tered Influence Function (RIF) of a statistic of interest of that same dependent variable (e.g. in
this case, a percentile or the variance of the log hourly earnings). A RIF is defined as RIF (y; v) =
17“Does your pay include performance related pay?” This question is only available for the last 10 waves.
16
v(Fy) + IF (y; v), where y is the dependent variable of interest (for instance, earnings), v(Fy) is
the statistic of interest of y and IF (y; v) denotes the Influence Function, defined such that the ex-
pected value of the RIF aggregates back to the statistic of interest. Assuming that the conditional
expectation of this RIF is a linear function of the covariates and that can be estimated using OLS,
this approach allows to now use classic Oaxaca-Blinder methods to go beyond the mean.
E[RIF (y; v)|X] = Xω
Yet, this does not provide a complete strategy to assess the differences in compositions and in
returns. Consider the overall difference of distributional statistic v(Fy) across the performance and
the non-performance pay group;
4vO = v(FY1|PP=1)− v(FY0|PP=1)︸ ︷︷ ︸
4vS
+ v(FY0|PP=1)− v(FY0|PP=0)︸ ︷︷ ︸4v
X
where 4vS is the wage structure effect and 4v
X denotes the composition effect. In order to ob-
tain a complete strategy to estimate these two effects, the counterfactual distributional statistic
v(FY0|PP=1) needs to be obtained. This is the distributional statistics that would have prevailed
if performance pay workers had been paid as non-performance pay workers. A popular method to
obtain this counterfactual is the reweighting method proposed by DiNardo et al. [1996], and it the
method used in the analysis.18
5.1.1 Quantile RIF-regressions.
The recentered influence function for quantile regressions is represented by the sum of the distri-
butional statistic of interest (Qτ ) and the influence function:
RIF (y;Qτ ) = Qτ +τ − 1[y ≤ Qτ ]
fY (Qτ )
This RIF can be estimated for given samples using sample quantiles and kernel methods (to
estimate the local densities). Once this sample RIF is obtained, it can be equated to the linear
18This counterfactual is straightforward to compute and consistent under the ignorability assumption [Firpo, 2007],which states that the treatment effect (in this case performance pay and non-performance pay) and the error termof the linear specification are conditionally indepent.
17
function of covariates to the obtain the estimated coefficients of this linear function:
ˆω(g, τ) = [∑i∈G
XiXti ]−1[∑i∈G
ˆRIF (Ygi;Qg,τ )Xi], g = PP,NPP
The results of these RIF quantile regressions are presented in figure 4 (for Germany) and
figure 5 (for Great Britain) and show a relatively strong consistency across both countries. The
quantile regressions for education levels highlight that bonus performance pay jobs are typically
associated with higher returns to schooling for higher degrees (panel [a] and [b] for Great Britain
and panel [a] and [b] for Germany). Vocational degrees Germany typically get higher returns to
education in non-performance pay jobs across the entire wage distribution, unless workers are in the
20% of the wage distribution. In this case, the returns are identical. Finally, the quantile regressions
in Great Britain portray lower returns in performance pay jobs compared to non-performance pay
jobs for all degrees in the bottom deciles of the wage distribution (10-15%).
5.1.2 Reweighting.
The reweighting approach applied is that of DiNardo et al. [1996], where the reweighting factor
can be defined and rewritten as
ψ(X) =dFXPP
dFXNPP
=Pr(PP = 1|X)/Pr(PP = 1))
PR(PP = 0|X)/Pr(PP = 0)
These probabilities can be obtained by taking the sample shares of performance and non-performance
pay jobs for the unconditional probabilities and by running a logit or probit model for the condi-
tional probabilities.19 Plugging in this reweighting factor in the distribution of non-performance
pay jobs gives:
FY C,PPNPP
(y) =
∫∫FYNPP |XNPP
(y|PP,X)ψ(PP,X)dFXNPP(X) (13)
=
∫∫FYNPP |XNPP
(y|PP,X)dFXPP
(X)
dFXNPP(X)
dFXNPP(X) (14)
=
∫∫FYNPP |XNPP
(y|PP,X)dFXPP(PP |X)dFXNPP
(X) (15)
19As proposed by DiNardo et al. (2006) we use a model that uses the standard set of controls used in the regressionsof table 4 and adds a set of interaction variables between education levels and experience and ethnicity.
18
In order to assess the changes over time, we compare reweighted results for performance pay at the
beginning of the sample to reweighted results at the end of the sample. Observations are pooled
across different years to minimize reweighting errors. In practice, the first three (1991–1993) and
last six (2003-2008) years are pooled together for the BHPS and the first five (1985–1989) and last
five (2004–2008) years are pooled together for the SOEP.20 Figure 3 plots the actual wage densities
by compensation and the counterfactual wage distribution if workers in performance pay jobs werer
paid as in non performance pay jobs.
Table 10 presents the results for Great Britain (upper panel) and Germany (lower panel).
The effect of performance pay on the dispersion of wages can be calculated by computing what
share the difference of column (6) and (3) represents of the difference of column (4) and (1). We find
that in Great Britain about 30% of the increased dispersion in the variance of wages can be acribed
to performance pay, while for Germany this number is about 20%. Yet, when considering the
percentile changes, there is a stark difference in experiences between Great Britain and Germany.
In Great Britain, the overall contribution in the 90-10 percentile log wage gap, or the difference
between column (6) and column (3), can be largely attributed to changes in the upper half of the
wage distribution. This is where the returns to education have increased substantially over this
time period and therefore consistent with the model proposed by LPM. In contrast, the overall
contribution in the 90-10 percentile log wage gap in Germany is largely explained by the lower half
of the wage distribution. Since returns to education largely increased in the upper half of the wage
distribution, this finding is somewhat at odds with the model proposed by LPM. Therefore, these
decomposition results are consistent with the model of performance pay and wage inequality doing
a reasonably good job for the British data, while doing a poorer job in Germany.
5.2 RIF decompositions.
The results above imply that increasing returns to education play a role in the increased incidence
of performance pay in Great Britain, while the results for Germany are somewhat moot. For Great
Britain, we would then expect that the increase in percentile or variance gaps is driven by changes in
returns rather than in composition. This can be explored by rewriting the aggregate decomposition
20Changing these periods alters the reweighting errors as the reweighting factor is consistent for large samples, butthe results are robust to using different time periods.
19
as
4vO = v(FY1|PP=1)− v(FY0|PP=1)︸ ︷︷ ︸
4vS
+ v(FY0|PP=1)− v(FY0|PP=0)︸ ︷︷ ︸4v
X
where 4vS is the wage structure effect and 4v
X is the composition effect. Given the property of
RIF that the expected value of the function equals the distributional statistic of interest and that
this expectation is linear in the covariates, it is possible to rewrite the wage structure effect and
the composition effect as (with the superscript c denoting counterfactual);
4vX = (XC
PP − XNPP ) · ωNPP,v + XCPP · (ωCPP,v − ωNPP,v) = 4v
X,p + 4v
X,e
4vS = XPP · (ωPP,v − ωCPP,v) + (XPP − XC
PP ) · ωCPP,v = 4v
S,p + 4v
S,e
It can be shown that the sum of these two equations equals the total difference, 4vO. Fur-
thermore, in both the equations the second term, 4v
X,e and 4v
S,e reflect error terms. Firpo et al.
[2011] define the first term as the “specification error” and the second term as the “total reweight-
ing error”. The reweighting error should go to zero in large samples as the reweighting factor is
estimated consistently; non-zero values reflect that the reweighting of the non-performance pay
sample does not result in a perfect counterfactual.21 The specification error results from assuming
the linear model. As in the previous subsection, observations are pooled over the same years to
minimize these errors.
The results are presented in table 11 for Great Britain and table 12 for Germany. For
both tables, the upper panel shows the results for the first period, while the lower panel shows
the results for the final period. The specification and reweighting errors are not presented. The
impact of performance pay can be assessed by comparing the upper panel to the lower panel. For
instance, the first two columns of table 11 highlight that the wage structure has become a more
important determinant of the 90-10 percentile log wage gap over time, as it largely explains the
wage gap in the final period. When considering columns (3) through (6), it is clear that this effect
is almost entirely explained by the upper half of the wage distribution. The gap in the lower part
of the distribution has narrowed, which is mostly explained by compositional effects, not wage
structure effects. The final two columns report the results for the variance and largely confirm
21Firpo [2010] shows that the reweighting method is efficient, as long as the reweighted functional is smooth.
20
this interpretation. This strengthens the hypothesis that performance pay has had an effect on
the dispersion of wages, largely driven by wage structure effects in the upper half of the wage
distribution. This is consistent with changes in the returns to skill and the model proposed by
LMP.
The results for Germany are less straightforward to interpret. The upper panel, presenting the
results for years 1985 through 1989, highlight that the wage structure effects are relatively similar
when comparing the actual distribution to the counterfactual distribution where everybody is paid
as workers in non performance pay jobs. The composition effects are also similar, with an exception
being the 90-50 percentile log wage gap. When considering the lower panel, though, the modest
unadjusted gaps mask substantial changes in the importance of composition and wage structure
effects. These changes, however, are not explained by performance pay. The gaps between the
changes (both compositional and wage structure) are minor when comparing the actual distribution
to the counterfactual distribution where all workers are not paid for performance. This highlights
that, while there have been important changes over time in Germany, the channel of performance
pay does not seem a major channel through which changes firms (or the labor market as a whole)
have dealt with changes in the returns to skill. One important implication is that, even though
countries such as the United States, Great Britain and Germany have witnessed similar trends in
wage inequality (see, for instance, Dustmann et al. [2009]), caution is advised when extrapolating
ways in which labor markets deal with changes in the return to skill. While performance pay seems
to have played a role in the United States and Great Britain to channel these changes, it has not
in Germany. Which channels have been important, remains an open question.
6 Conclusion.
In this paper I investigate the link between performance pay and wage inequality. To my knowledge,
there is currently no research evidence relating this link for Great Britain and Germany. I test the
model proposed by LMP on survey panel data in these countries, and investigate the effect of
misclassification on the empirical strategy proposed by LMP. I find that for certain regression
strategies, misclassification can lead to omitted variable bias, and that misclassification in general
decreases the likelihood of confirming the empirical predictions proposed by LMP. When adjusting
21
for these issues, I find that performance pay seems to play a similar role in Great Britain and the
United States, where it is used to channel changes in the returns to skills. While Germany has had
somewhat similar trends (see Dustmann et al. [2009]), performance pay does not play a similar
role in this labor market.
RIF decompositions, that allow me to decompose changes at various distributional statistics
of the wage distribution, show that performance pay in Great Britain can explain about 30% of
the increase in the dispersion of wages. As LMP point out, this is not a causal effect, but can be
interpreted as the additional increase in dispersion that was the result from workers and employers
switching to performance pay contracts. Consistent with the LMP model, most of this increase
is explained by changes in the top of the wage distribution , where the returns to education have
increased over this same period, and changes in the wage structure. For Germany, I find performance
pay jobs explain about 20% of the increase in the dispersion of wages. Nevertheless, this increase
is driven by changes in the bottom of the distribution, which does not lign up with the model of
LMP. Decomposing these results highlights that Germany has seen important changes in the wage
distribution, and that both compositional and wage structure effects are important. Nevertheless,
performance pay jobs explain very little of these changes.
Taken together, the results indicate the labor markets may deal with changes in the returns
to skill in different ways. The three countries considered in this study and LMP have witnessed
similar trends in wage inequality and changes in the returns to skills and education. Nevertheless,
the single model of performance pay and wage inequality does not seem to hold uniformly. In the
United States and Great Britain, performance pay seems to have been an important channel to
deal with these changes in the returns to skills and education. In Germany, on the other hand, it
has had very little impact, indicating firms and the labor market have channeled changes in the
returns to skill in different ways. How these changes have been channeled, and the importance of
different institutions, is not the subject of this paper, but possibly an interesting avenue for future
research.
22
References
1. Autor, D. H., Katz, L. F. and Kearny, M. S., 2006. The polarization of the U.S. labor market.
American Economic Review, 96 (2), pp. 189-194.
2. Bell, B. and Van Reenen, J., 2010. Bankers’ pay and extreme wage inequality in the UK.
CEP Special Paper No. 21.
3. Bloom, N. and Van Reenen, J., 2011. Human resource management and productivity. Hand-
book of Labor Economics, Elsevier.
4. Blundell, R., Brewer, M. and Francesconi, M., 2008. Job changes and hours changes: un-
derstanding the path of labor supply adjustment. Journal of Labour Economics, 26 (3), pp.
421-453.
5. Booth, A. L. and Frank, J., 1999. Earnings, productivity, and performance-related pay.
Journal of Labor Economics, 17, pp. 447-463.
6. Bound, J., Brown, C. and Mathiowetz, N. 2001. Measurement error in survey data. Handbook
of Econometrics, 5, pp. 3705-3843
7. Brown, C., 1990. Firm’s choice of method of pay. Industrial and Labour Relations Review,
43, pp 165S-182S.
8. Brown, M. and Heywood, J. S., eds. 2002. Paying for performance: an international com-
parison. Armonk, N.Y.: M. E. Sharpe.
9. Brynin, M. and Longhi, S., 2010. Occupational change in Britain and Germany. Labour
Economics, 17 (4), pp. 655-666.
10. Bryson, A., Pendleton, A. and Whitfield, K., 2008. ‘The changing use of contingent pay at the
modern British workplace’. NIESR Discussion Paper No. 319. London: National Institute of
Economic and Social Research.
11. Bryson, A., Freeman, R., Lucifora, C., Pellizzari, M. and Perotin, V., 2012. Paying for
performance: incentive pay schemes and employees’ financial participation. CEP Discussion
Paper No. 1112. London: Centre of Economic Performance.
12. Card, D. 1996 The effect of unions on the structure of wages: a longitudinal analysis. Econo-
metrica, Vol. 64(4), 957-979.
13. Card, D. and J. E. DiNardo. 2002. “Skill-Biased Technological Change and Rising Wage
Inequality: Some Problems and Puzzles.” Journal of Labor Economics, 20 (October): 733-
83.
23
14. DiNardo, J., Fortin, N. M. and Lemieux, T., 1996. Labor market institutions and the distri-
bution of wages, 1973-1992: a semiparametric approach. Econometrica, 64, pp. 1001-1044.
15. Dustmann, C. and van Soest, A., 1998. Public and private sector wages of male workers in
Germany. European Economic Review, Elsevier, 42 (8), pp. 1417-1441.
16. Dustmann, C., Ludsteck, J. and Schoenberg, U., 2009. Revisiting the German wage structure.
Quarterly Journal of Economics, 124 (2), pp. 843-881.
17. Firpo, S., 2007. Efficient semiparametric estimation of quantile treatment effects. Economet-
rica, 75, pp. 259-276.
18. Firpo, S., 2010. Identification and estimation of distributional impacts of interventions using
changes in inequality measures. EESP-FGV. mimeo.
19. Firpo, S., Fortin, N. M. and Lemieux, T., 2009. Unconditional quantile regressions. Econo-
metrica, 77 (3), pp. 953-973.
20. Firpo, S., Fortin, N. M. and Lemieux, T., 2011. Decomposition methods in economids. In
Ashenfelter, O. and Card, D. (Eds). Handbook of Labor Economics, Volume 4. Elsevier.
21. Fraessdorf, A., Grabka, M. and Schwarze, J., 2008. The impact of household capital income
on income inequality: a factor decomposition analysis for Great Britain, Germany and the
USA. IZA Discussion Papers 3492, Institute for the Study of Labor (IZA).
22. Frick, J. R. and Grabka, M. M., 2003. Imputed rent and income inequality: a decomposition
analysis for Great Britain, West Germany and the US. Review of Income and Wealth, 49(4),
pp. 513-537.
23. Gibbons, R., Katz, L. F., Lemieux, T. and Parent, D., 2005. Comparative advantage, learn-
ing, and sectoral wage determination. Journal of Labor Economics, 23, pp. 681-723.
24. Goerke, L. and Pannenberg, M., 2007. Trade union membership and works councils in West
Germany. IZA Discussion Paper No. 2635, Institute for the Study of Labor (IZA)
25. Henneberger, F., Sousa-Poza, A. and Ziegler, A., 2007. Performance Pay, Sorting, and Out-
sourcing. IZA Discussion Paper No. 3019, Institute for the Study of Labor (IZA).
26. Jones, A. M., Rice, N. and Roberts, J., 2009. Early retirement and inequality in Britain and
Germany: How important is health? SOEPpapers on Multidisciplinary Panel Data Reserach
188, DIW Berlin.
27. Lazear , E. P., 1986. Salaries and piece rates. Journal of Business, 59, pp. 405-431.
28. Lazear, E. P., 2000. Performance pay and productivity. American Economic Review, 90 (5),
pp. 1346-1361.
24
29. Lemieux, T., MacLeod, W. B. and Parent, D., 2009. Performance pay and wage inequality.
Quarterly Journal of Economics, 214(1), pp. 1-49.
30. Lersch, P. M., 2012. Long-distance moves and labour market outcomes of dual-earner couples
in the UK and Germany. SOEPpapers on Multidisciplinary Panel Data Research 469, DIW
Berlin.
31. Machin, S. and Van Reenen, J., 1998. Technology and changes in skill structure: evidence
from seven OECD countries. Quarterly Journal of Economics, 113, pp. 1215-1244.
32. Oaxaca, R., 1973. Male-Female wage differentials in urban labor markets. International
Economic Review, 25, pp. 693-709.
33. Pencavel, J., 1978. Work effort, on-the-job screening and alternative methods of remunera-
tion. Research in Labor Economics, 33, pp. 225-258.
34. Piketty, T. and Saez, E., 2003. Income inequality in the United States, 1913-1998. Quarterly
Journal of Economics, 118, pp. 1-39.
25
Tab
le1:
Su
mm
ary
stat
isti
cs
BHPS
SOEP
1991
–200
819
85–2
008
Non
-PP
job
sP
Pjo
bs
Non
-PP
job
sP
Pjo
bs
Hou
rly
earn
ings
8.23£
10.1
8£11
.61e
17.4
0e
(4.3
2)(6
.07)
(5.4
1)(8
.49)
Exp
erie
nce
(yea
rs)
21.9
920
.57
19.2
219
.01
(13.
04)
(11.
95)
(11.
36)
(10.
02)
Ten
ure
(yea
rs)
5.15
5.68
10.4
412
.61
(6.7
1)(6
.21)
(9.5
8)(9
.63)
Age
39.3
438
.45
40.4
741
.37
(11.
95)
(10.
83)
(10.
90)
(9.5
9)M
arri
ed0.
590.
600.
730.
76(0
.49)
(0.4
9)(0
.45)
(0.4
2)N
onw
hit
e0.
030.
020.
130.
03(0
.18)
(0.1
5)(0
.34)
(0.1
8)U
nio
niz
ed0.
210.
25—
—(0
.41)
(0.4
3)A
nnu
alh
ou
rsw
orke
d2,
341
2,36
12,
299
2,38
7(5
24)
(467
)(3
93)
(433
)
Work
ers
3,35
93,
714
7,77
32,
786
Job
matc
hes
5,15
06,
535
10,7
493,
066
Ob
serv
ati
on
s9,
726
18,5
9740
,287
20,3
10
Notes.
The
sam
ple
consi
sts
ofm
ale
sb
etw
een
20
and
65
yea
rsold
that
are
emplo
yed
inth
epri
vate
sect
or.
The
hourl
yw
ages
are
rest
rict
edb
etw
een
1and
100
pounds
or
euro
s.In
Gre
at
Bri
tain
,th
esa
mple
imp
ose
san
addit
ional
rest
rict
ion
on
the
annual
hours
work
edto
be
bet
wee
n1,0
00
and
4,5
00.
InG
erm
an,
the
sam
ple
imp
ose
san
addit
ional
rest
rict
ion
on
the
month
lyea
rnin
gs
tob
eab
ove
600e
.T
he
rep
ort
edfigure
sre
pre
sent
sam
ple
mea
ns.
Educa
tion
inth
ista
ble
isre
port
edin
appro
xim
ate
yea
req
uiv
ale
nts
.In
Gre
at
Bri
tain
,(p
ote
nti
al)
exp
erie
nce
isdefi
ned
as
age
min
us
educa
tion
min
us
6.
InG
erm
any,
exp
erie
nce
isav
ailable
as
vari
able
.
26
Table 2: Summary statistics: Education in Great Britain and Germany
BHPS SOEP
1991–2008 1985–2008
Non-PP jobs PP jobs Non-PP jobs PP jobsHigher degree 0.02 0.03 University 0.06 0.17
(0.14) (0.17) (0.24) (0.37)First degree 0.09 0.13 Technical 0.05 0.12
(0.29) (0.34) (0.22) (0.32)HND, HNC, training 0.08 0.09 Vocational 0.68 0.63
(0.27) (0.29) (0.47) (0.48)A level 0.22 0.26 None 0.21 0.08
(0.41) (0.44) (0.41) (0.28)O level 0.28 0.27
(0.45) (0.44)CSE 0.07 0.07
(0.26) (0.25)None 0.23 0.16
(0.42) (0.36)
Notes. The academic qualifications for Great Britain are declared as reported in the BHPS. The academicqualifcations for Germany are defined as follows. University is defined as university, technical college(TH) or college abroad; Technical is defined as technical college (Fachhochschule); Vocational is definedas apprenticeship, vocational school, health care school, civil service training, other training.
27
Table 3: Regression estimates of the effect of Performance Pay on earnings
Estimation method
OLS Fixed Effects(1) (2) (3) (4) (5)
BHPS
Performance Pay job 0.0895*** 0.0595*** 0.0257*** 0.0192*** —(0.0074) (0.0088) (0.0062) (0.0067)
Performance Pay received — 0.0449*** — 0.0105** 0.0087*
in current year (0.0069) (0.0044) (0.0046)Worker fixed effect No No Yes Yes YesJob-match fixed effect No No No No YesN 28,323 28,323 28,323 28,323 28,323Jobmatches 11,685 11,685 11,685 11,685 11,685
SOEP
Performance Pay job 0.1521*** 0.1214*** 0.0721*** 0.0639*** —(0.0064) (0.0067) (0.0082) (0.0083)
Performance Pay received — 0.0840*** — 0.0221*** 0.0213***
in current year (0.0067) (0.0036) (0.0035)Worker fixed effect No No Yes Yes YesJob-match fixed effect No No No No YesN 60,597 60,597 60,597 60,597 60,597Jobmatches 13,815 13,815 13,815 13,815 13,815
Notes. Standard errors (clustered at the jobmatch level) are reported in parentheses. Significancelevel are starred at 10% (*), 5% (**) and 1% (***). All specifications include a full set of industrydummies (10 for Great Britain and 17 for Germany), occupation dummies (9 for Great Britain and10 for Germany), year dummies (18 for Great Britain and 24 for Germany), regional dummies (19for Great Britain and 16 for Germany) a quadratic in experience and tenure, degree levels (as definedin the text), dummies for being married, for being nonwhite and for union status (in Great Britainonly). “Performance pay job” indicates whether performance pay had been received during the jobmatch, “performance pay received in current year” indicates whether performance pay had beenreceived that actual year.
28
Table 4: Skills-related wage differentials and Performance Pay jobs
BHPS BHPS adjusted SOEP1991–2008 1991–2008 1985–2008
PP jobs non-PP jobs PP jobs non-PP jobs PP jobs non-PP jobs(1) (2) (3) (4) (5) (6)
Constant 1.6647*** 1.6396*** 1.6899*** 1.6396*** 1.8237*** 1.4380***
(0.0830) (0.0590) (0.1762) (0.0590) (0.0810) (0.0704)Experience 0.0073*** 0.0056*** 0.0116*** 0.0056*** 0.0039*** 0.0027***
(0.0005) (0.0006) (0.0018) (0.0006) (0.0007) (0.0003)Tenure 0.0106*** 0.0103*** 0.0119*** 0.0103*** 0.0069*** 0.0078***
(0.0012) (0.0015) (0.0034) (0.0015) (0.0007) (0.0005)High Education 0.3995*** 0.3422*** 0.4569*** 0.3422*** 0.3221*** 0.2249***
(0.0205) (0.0250) (0.0480) (0.0250) (0.0234) (0.0197)Medium Education 0.3052*** 0.2226*** 0.3878*** 0.2226*** 0.2510*** 0.1534***
(0.0208) (0.0240) (0.0479) (0.0240) (0.0236) (0.0180)Low Education 0.1776*** 0.1608*** 0.1944*** 0.1608*** 0.0721*** 0.0571***
(0.0156) (0.0165) (0.0353) (0.0165) (0.0169) (0.0067)
N 18,597 9,716 18,597 9,716 20,310 40,287
Notes. Standard errors (clustered at the jobmatch level) are reported in parentheses. Significance level are starred at10% (*), 5% (**) and 1% (***). All specifications include a full set of industry dummies (10 for Great Britain and 17for Germany), occupation dummies (9 for Great Britain and 10 for Germany), year dummies (18 for Great Britainand 24 for Germany), regional dummies (19 for Great Britain and 16 for Germany) a quadratic in experience andtenure, degree levels (as defined in the text), dummies for being married, for being nonwhite and for union status (inGreat Britain only). The reported effects of experience and tenure are evaluated at 20 and 10 years respectively usingthe polynomial models imposed. All specifications are OLS (Ordinary Least Squares).
29
Table 5: Interaction models across the different survey panels
BHPS BHPS adjusted SOEP1991–2008 1991–2008 1985–2008
OLS OLS OLS(1) (2) (3)
PP job 0.0127 0.0084 0.0651***
(0.0260) (0.1290) (0.0214)Experience 0.0062*** 0.0219*** 0.0026***
(0.0005) (0.0043) (0.0003)Experience x PP job 0.0019** 0.0104 0.0042***
(0.0010) (0.0098) (0.0015)Tenure 0.0104*** 0.0189*** 0.0076***
(0.0014) (0.0063) (0.0004)Tenure x PP job 0.0012 0.0022 -0.0011
(0.0021) (0.0150) (0.0014)High Education 0.3065*** 0.3111*** 0.2236***
(0.0228) (0.0672) (0.0188)High x PP job 0.0801*** 0.1932 0.1085***
(0.0282) (0.1314) (0.0262)Medium Education 0.2025*** 0.2042*** 0.1498***
(0.0233) (0.0678) (0.0170)Medium x PP job 0.0813*** 0.2038 0.1058***
(0.0295) (0.1422) (0.0262)Low Education 0.1565*** 0.1644*** 0.0524***
(0.0163) (0.0577) (0.0067)Low x PP job 0.0250 0.0500 0.0229
(0.0219) (0.1064) (0.0173)
N 28,290 28,290 60,597
Notes. Standard errors (clustered at the jobmatch level) are reported in paren-theses. Significance level are starred at 10% (*), 5% (**) and 1% (***). Allspecifications include a full set of industry dummies (10 for Great Britain and 17for Germany), occupation dummies (9 for Great Britain and 10 for Germany),year dummies (18 for Great Britain and 24 for Germany), regional dummies (19for Great Britain and 16 for Germany) a quadratic in experience and tenure,degree levels (as defined in the text), dummies for being married, for being non-white and for union status (in Great Britain only). The specifications also includeinteractions between the performance-pay dummy and the educations levels (7 inGreat Britain and 3 in Germany), experience (quadratic) and tenure (quadratic).The reported effects of experience and tenure are evaluated at 20 and 10 yearsrespectively using the polynomial models imposed (both for the interacted andthe non-interacted variables). The acronyms OLS and FE stand for OrdinaryLeast Squares and fixed effects (worker fixed effects).
30
Table 6: Interaction models across the different survey panels including time effects
BHPS SOEP1991-2008 1985-2008
OLS OLS(1) (2)
PP job 0.0119 0.0746***
(0.0266) (0.0219)Experience 0.0060*** 0.0027***
(0.0005) (0.0003)Experience x PP job 0.0025** 0.0038**
(0.0010) (0.0015)Tenure 0.0104*** 0.0075***
(0.0014) (0.0004)Tenure x PP job -0.0001 -0.0011
(0.0021) (0.0014)High Education 0.4148*** 0.2418***
(0.0447) (0.0219)High x PP job 0.0159 0.1587**
(0.0552) (0.0728)High Education x 2006–2008 -0.1738*** 0.0520
(0.0597) (0.0648)High x PP job 0.1435** -0.0753x 2006–2008 (0.0689) (0.0776)Medium Education 0.2376*** 0.2280***
(0.0514) (0.0351)Medium x PP job 0.0120 0.0582
(0.0685) (0.0521)Medium Education x 2006–2008 -0.0391 -0.0405
(0.0683) (0.0449)Medium x PP job 0.0817 0.0598x 2006–2008 (0.0851) (0.0586)Low Education 0.1862*** 0.0647***
(0.0277) (0.0105)Low x PP job -0.0216 -0.0088
(0.0365) (0.0211)Low Education x 2006–2008 -0.0571 -0.0010
(0.0408) (0.0205)Low x Performance Pay job 0.0742* 0.0469**
x 2006–2008 (0.0450) (0.0182)
N 28,323 60,597
Notes. Standard errors (clustered at the jobmatch level) are reported in parentheses. Significance levelare starred at 10% (*), 5% (**) and 1% (***). All specifications include a full set of industry dummies (10for Great Britain and 17 for Germany), occupation dummies (9 for Great Britain and 10 for Germany),year dummies (18 for Great Britain and 24 for Germany), regional dummies (19 for Great Britain and 16for Germany) a quadratic in experience and tenure, degree levels (as defined in the text), dummies forbeing married, for being nonwhite and for union status (in Great Britain only). The specifications alsoinclude interactions between the performance-pay dummy and the educations levels (7 in Great Britainand 3 in Germany), experience (quadratic), tenure (quadratic) and a set of interaction variables betweena period dummy and the performance pay dummy and the different education levels. A period is definedas three years, so there are six periods for Great Britain and eight periods for Germany. The reportedeffects of experience and tenure are evaluated at 20 and 10 years respectively using the polynomial modelsimposed (both for the interacted and the non-interacted variables). The acronyms OLS and FE standfor Ordinary Least Squares and fixed effects (worker fixed effects).
Table 7: Variance component models by type of job
Performance Pay jobs Non-performance pay jobsVariance component (1) (2) (3) (4)
BHPS
Full sampleWorker 0.1001*** 0.0851*** 0.0785*** 0.0663***
(0.0030) (0.0031) (0.0030) (0.0033)Jobmatch — 0.0261*** — 0.0235***
(0.0012) (0.0021)Idiosyncratic 0.0541*** 0.0423*** 0.0648*** 0.0537***
error (0.0006) (0.0006) (0.0012) (0.0012)
Job-changers sampleWorker 0.0844*** 0.0707*** 0.0664*** 0.0540***
(0.0040) (0.0041) (0.0040) (0.0043)Jobmatch — 0.0240*** — 0.0266***
(0.0018) (0.0031)Idiosyncratic 0.0525*** 0.0419*** 0.0673*** 0.0540***
error (0.0009) (0.0008) (0.0018) (0.0017 )
SOEP
Full sampleWorker 0.0792*** 0.0492*** 0.0614*** 0.0350***
(0.0026) (0.0039) (0.0013) (0.0016)Jobmatch — 0.0327*** — 0.0343***
(0.0034) (0.0013)Idiosyncratic 0.0284*** 0.0267*** 0.0332*** 0.0269***
error (0.0003) (0.0003) (0.0003) (0.0002)
Job-changers sampleWorker 0.0686*** 0.0463*** 0.0690*** 0.0507***
(0.0044) (0.0062) (0.0049) (0.0054)Jobmatch — 0.0262*** — 0.0274***
(0.0050) (0.0033)Idiosyncratic 0.0276*** 0.0255*** 0.0351*** 0.0264***
error (0.0006) (0.0006) (0.0011) (0.0009)
Source PSID. Lemieux et al. (2009)
Notes for BHPS and SOEP. Standard errors (clustered at the jobmatch level) arereported in parentheses. Significance level are starred at 10% (*), 5% (**) and 1%(***).
32
Tab
le8:
Rob
ust
nes
sch
ecks
inB
HP
S
20%
PP
ormore
50%
PP
ormore
Alt.question
PP
Publicwork
ers
(1)
(2)
(3)
(4)
Per
form
ance
Pay
job
0.03
92**
0.04
71***
0.036
9***
0.0
616
***
(0.0
186)
(0.0
097)
(0.0
092
)(0
.023
4)E
xp
erie
nce
0.00
65***
0.00
64***
0.00
50***
0.00*
**
(0.0
005)
(0.0
005)
(0.0
005)
(0.0
00)
Exp
erie
nce
xP
erfo
rman
ceP
ayjo
b0.
0013
*0.
0015
**
0.002
5*0.
0(0
.000
8)(0
.000
7)(0
.0007
)(0
.00)
Ten
ure
0.01
06***
0.01
06***
0.01
19***
0.0*
**
(0.0
014)
(0.0
014)
(0.0
016
)(0
.00)
Ten
ure
xP
erfo
rman
ceP
ayjo
b0.
0011
0.00
12-0
.0002
-0.0
*
(0.0
021)
(0.0
021)
(0.0
024
)(0
.00)
Hig
h0.
3149
***
0.31
56***
0.31
03***
0.334
4***
(0.0
220)
(0.0
214)
(0.0
246
)(0
.017
7)H
igh
xP
erfo
rman
ceP
ayjo
b0.
0660
**
0.06
37***
0.07
87***
0.0
533
**
(0.0
261)
(0.0
244)
(0.0
274
)(0
.023
1)M
ediu
m0.
2094
***
0.20
95***
0.19
69***
0.231
7***
(0.0
228)
(0.0
225)
(0.0
260
)(0
.017
9)M
ediu
mx
Per
form
ance
Pay
job
0.06
85**
0.06
73**
0.09
60***
0.0
513
**
(0.0
261)
(0.0
373)
(0.0
313
)(0
.024
7)L
ow0.
1636
***
0.16
29***
0.15
77***
0.186
4***
(0.0
158)
(0.0
152)
(0.0
177
)(0
.014
9)L
owx
Per
form
an
ceP
ayjo
b0.
0120
0.01
270.
026
1-0
.005
6(0
.020
1)(0
.018
2)(0
.0209
)(0
.020
2)
N28
,290
28,2
9021,
103
36,6
80
Notes.
Sta
ndard
erro
rs(c
lust
ered
at
the
jobm
atc
hle
vel
)are
rep
ort
edin
pare
nth
eses
.Sig
nifi
cance
level
are
starr
edat
10%
(*),
5%
(**)
and
1%
(***).
All
spec
ifica
tions
incl
ude
afu
llse
tof
indust
rydum
mie
s(1
0),
occ
upati
on
dum
mie
s(9
),yea
rdum
mie
s(1
8),
regio
nal
dum
mie
s(1
9)
aquadra
tic
inex
per
ience
and
tenure
,deg
ree
level
s(a
sdefi
ned
inth
ete
xt)
,dum
mie
sfo
rb
eing
marr
ied,
for
bei
ng
nonw
hit
eand
for
unio
nst
atu
s.C
olu
mn
1(2
)te
sts
the
spec
ifica
tion
of
table
5,
but
defi
nes
the
dep
enden
tva
riable
as
hav
ing
rece
ived
per
form
ance
pay
at
least
one
out
of
five
(tw
o)
tim
esw
hen
obse
rved
on
asp
ecifi
cjo
bm
atc
h.
Colu
mn
3use
sa
subsa
mple
of
the
last
10
wav
esand
adiff
eren
tques
tion
todefi
ne
per
form
ance
pay
(see
text
for
det
ails)
.C
olu
mn
4ex
pands
the
sam
ple
by
incl
udin
gpublic
sect
or
work
ers.
All
spec
ifica
tions
are
esti
mate
dusi
ng
OL
S(O
rdin
ary
Lea
stSqaure
s).
33
Tab
le9:
Rob
ust
nes
sch
ecks
inS
OE
P
20%
PP
ormore
50%
PP
ormore
Publicsecto
rFemale
work
ers
(1)
(2)
(3)
(4)
Per
form
an
ceP
ayjo
b0.
0849
***
0.10
14***
0.09
90***
0.044
2***
(0.0
103)
(0.0
105)
(0.0
202
)(0
.0173
)E
xp
erie
nce
0.00
26***
0.00
25***
0.00
31***
0.003
0***
(0.0
003)
(0.0
003)
(0.0
003
)(0
.0003
)E
xp
erie
nce
xP
erfo
rman
ceP
ayjo
b0.
0048
***
0.00
59***
0.002
7*0.0
049
***
(0.0
012)
(0.0
012)
(0.0
014)
(0.0
012)
Ten
ure
0.00
76***
0.00
76***
0.00
78***
0.007
6***
(0.0
004)
(0.0
004)
(0.0
004
)(0
.0004
)T
enu
rex
Per
form
ance
Pay
job
-0.0
004
-0.0
003
-0.0
028
**
-0.0
003
(0.0
014)
(0.0
014)
(0.0
014
)(0
.0012
)H
igh
0.22
61***
0.22
55***
0.25
41***
0.229
8***
(0.0
187)
(0.0
187)
(0.0
138
)(0
.0153
)H
igh
xP
erfo
rman
ceP
ayjo
b0.
0950
***
0.09
38***
0.07
49***
0.108
8***
(0.0
236)
(0.0
236)
(0.0
224
)(0
.0218
)M
ediu
m0.
1522
***
0.15
13***
0.14
58***
0.142
0***
(0.0
169)
(0.0
169)
(0.0
138
)(0
.0140
)M
ediu
mx
Per
form
an
ceP
ayjo
b0.
0929
***
0.09
99***
0.10
10***
0.101
2***
(0.0
237)
(0.0
232)
(0.0
238
)(0
.0225
)L
ow0.
0518
***
0.05
01***
0.05
04***
0.057
7***
(0.0
066)
(0.0
066)
(0.0
063
)(0
.0058
)L
owx
Per
form
ance
Pay
job
0.02
45*
0.03
27**
0.0
171
0.0
325
**
(0.0
136)
(0.0
132)
(0.0
166)
(0.0
142
)
N60
,597
60,5
9779
,059
84,4
01
Notes.
Sta
ndard
erro
rs(c
lust
ered
at
the
jobm
atc
hle
vel
)are
rep
ort
edin
pare
nth
eses
.Sig
nifi
cance
level
are
starr
edat
10%
(*),
5%
(**)
and
1%
(***).
All
spec
ifica
tions
incl
ude
afu
llse
tof
indust
rydum
mie
s(1
7),
occ
upati
on
dum
mie
s(1
0),
yea
rdum
mie
s(2
4),
regio
nal
dum
mie
s(1
6)
aquadra
tic
inex
per
ience
and
tenure
,deg
ree
level
s(a
sdefi
ned
inth
ete
xt)
,dum
mie
sfo
rb
eing
marr
ied
and
for
bei
ng
nonw
hit
e.C
olu
mn
1(2
)te
sts
the
spec
ifica
tion
of
table
5,
but
defi
nes
the
dep
enden
tva
riable
as
hav
ing
rece
ived
per
form
ance
pay
at
least
one
out
of
five
(tw
o)
tim
esw
hen
obse
rved
on
asp
ecifi
cjo
bm
atc
h.
Colu
mn
3and
4ex
pand
the
sam
ple
by
incl
udin
gpublic
sect
or
work
ers
and
fem
ale
(pri
vate
sect
or)
work
ers
resp
ecti
vel
y.A
llsp
ecifi
cati
ons
are
esti
mate
dusi
ng
OL
S(O
rdin
ary
Lea
stSqaure
s)
34
Tab
le10:
Eff
ect
ofp
erfo
rman
cep
ayjo
bs
onm
easu
res
ofw
age
ineq
ual
ity
BHPS
1991–1993
2006–2008
Act
ual
Cou
nte
rfac
tual
Eff
ect
ofA
ctu
alC
ounte
rfac
tual
Eff
ect
of%
Exp
lain
edd
isp
ersi
ond
isp
ersi
onP
Pjo
bs
dis
per
sion
dis
per
sion
PP
job
sby
PP
job
s(1
)(2
)(3
)(4
)(5
)(6
)(7
)
Vari
an
ce0.
2259
0.20
630.
0196
0.25
530.
2270
0.02
83(0
.008
2)
(0.0
100)
(0.0
129)
(0.0
100)
(0.0
126)
(0.0
161)
Per
centi
legap
s90-
101.1
875
1.15
520.
0323
1.23
331.
1520
0.08
13(0
.021
2)
(0.0
356)
(0.0
415)
(0.0
254)
(0.0
348)
(0.0
431)
50-
100.5
746
0.56
690.
0077
0.57
620.
5786
-0.0
024
(0.0
157)
(0.0
257)
(0.0
301)
(0.0
174)
(0.0
273)
(0.0
323)
90-
500.6
129
0.58
830.
0246
0.65
710.
5734
0.08
37(0
.018
1)
(0.0
254)
(0.0
312)
(0.0
186)
(0.0
242)
(0.0
305)
G-S
OEP
1985–1989
2004–2008
Act
ual
Cou
nte
rfac
tual
Eff
ect
ofA
ctu
alC
ounte
rfac
tual
Eff
ect
ofd
isp
ersi
ond
isp
ersi
onP
Pjo
bs
dis
per
sion
dis
per
sion
PP
job
s(1
)(2
)(3
)(4
)(5
)(6
)
Vari
an
ce0.
1225
0.09
090.
0316
0.22
480.
1810
0.04
38(0
.005
0)
(0.0
057)
(0.0
076)
(0.0
073)
(0.0
093)
(0.0
118)
Per
centi
legap
s90-
100.8
596
0.71
670.
1429
1.22
481.
0536
0.17
12(0
.020
6)
(0.0
268)
(0.0
338)
(0.0
232)
(0.0
323)
(0.0
398)
50-
100.3
742
0.34
220.
0320
0.60
580.
5504
0.05
54(0
.012
2)
(0.0
155)
(0.0
197)
(0.0
191)
(0.0
221)
(0.0
292)
90-
500.4
855
0.37
450.
1110
0.61
900.
5032
0.11
58(0
.017
8)
(0.0
232)
(0.0
293)
(0.0
190)
(0.0
262)
(0.0
295)
Rew
eighte
ddis
trib
uti
ons
usi
ng
the
DF
Lappro
ach
.T
he
pro
pen
sity
score
use
sa
ver
yflex
ible
model
.C
on-
trols
incl
ude
educa
tion,
aquadra
tic
poly
nom
ial
inex
per
ience
and
tenure
;a
whit
e/nonw
hit
eand
mar-
ried
/si
ngle
dum
my;
occ
upati
on,
indust
ry,
yea
rand
regio
nfixed
effec
ts;
and
inte
ract
ions
bet
wee
nsc
hooling
and
exp
erie
nce
.B
oots
trapp
edst
andard
erro
rsare
obta
ined
by
runnin
g1,0
00
iter
ati
ons
usi
ng
bala
nce
dsa
mple
sof
1,0
00
per
form
ance
and
non-p
erfo
rmance
pay
jobs
each
(i.e
.su
bsa
mple
sco
nta
in2,0
00
obse
rva-
tions)
.
35
Tab
le11
:D
ecom
pos
itio
nre
sult
s:B
HP
S
90–10logwagegap
90–50logwagegap
50–10logwagegap
Variance
Act
ual
Rew
eigh
ted
Act
ual
Rew
eigh
ted
Act
ual
Rew
eigh
ted
Act
ual
Rew
eigh
ted
1991–1993
Un
adju
sted
Ch
ange
0.0
872
0.08
720.
0548
0.05
480.
0324
0.03
240.
0369
0.03
69(0
.0558
)(0
.055
8)(0
.045
4)(0
.045
4)(0
.035
5)(0
.035
5)(0
.016
2)(0
.016
2)
Com
posi
tion
effec
t0.0
402
0.04
380.
0029
0.00
240.
0373
0.04
130.
0120
0.01
50(0
.0308
)(0
.050
9)(0
.027
4)(0
.044
2)(0
.021
2)(0
.034
3)(0
.009
4)(0
.015
9)W
age
Str
uct
ure
0.04
71
0.04
960.
0519
0.03
01-0
.004
90.
0194
0.02
500.
0094
(0.0
575
)(0
.073
1)(0
.048
7)(0
.065
3)(0
.040
1)(0
.048
9)(0
.016
7)(0
.021
5)
2006–2008
Un
adju
sted
Ch
ange
0.0
604
0.06
040.
0862
0.08
62-0
.025
8-0
.025
80.
0243
0.02
43(0
.0457
)(0
.045
7)(0
.033
9)(0
.033
9)(0
.037
0)(0
.037
0)(0
.018
4)(0
.018
4)
Com
posi
tion
effec
t0.0
046
-0.0
176
0.00
240.
0016
0.00
22-0
.019
30.
0066
0.00
40(0
.0380
)(0
.133
5)(0
.029
6)(0
.081
0)(0
.030
5)(0
.092
7)(0
.013
1)(0
.039
0)W
age
Str
uct
ure
0.05
88
0.09
630.
0838
0.00
66-0
.027
90.
0168
0.01
770.
0269
(0.0
574
)(0
.264
0)(0
.033
9)(0
.023
0)(0
.051
0)(0
.138
0)(0
.021
5)(0
.061
4)
Rew
eighte
ddis
trib
uti
ons
usi
ng
the
DF
Lappro
ach
.T
he
pro
pen
sity
score
use
sa
ver
yflex
ible
model
.C
ontr
ols
incl
ude
educa
tion,
aquadra
tic
poly
nom
ial
inex
per
ience
and
tenure
;a
whit
e/nonw
hit
eand
marr
ied/si
ngle
dum
my;
occ
upati
on,
indust
ry,
yea
rand
regio
nfixed
effec
ts;
and
inte
ract
ions
bet
wee
nsc
hooling
and
exp
erie
nce
.R
ecen
tere
dIn
fluen
ceR
egre
ssio
ndec
om
posi
tions
dec
om
pose
educa
tion,
aquadra
tic
poly
nom
ial
inex
per
ience
and
tenure
;a
whit
e/nonw
hit
eand
marr
ied/si
ngle
dum
my;
occ
upati
on,
indust
ry,
yea
rand
regio
nfixed
effec
ts.
Boots
trapp
edst
andard
erro
rsare
obta
ined
by
runnin
g1,0
00
iter
ati
ons
usi
ng
bala
nce
dsa
mple
sof
1,0
00
per
form
ance
and
non-p
erfo
rmance
pay
jobs
each
(i.e
.su
bsa
mple
sco
nta
in2,0
00
obse
rvati
ons)
.
36
Tab
le12
:D
ecom
pos
itio
nre
sult
s:S
OE
P
90–10logwagegap
90–50logwagegap
50–10logwagegap
Variance
Act
ual
Rew
eigh
ted
Act
ual
Rew
eigh
ted
Act
ual
Rew
eigh
ted
Act
ual
Rew
eigh
ted
1985–1989
Un
adju
sted
Ch
ange
0.1
452
0.14
520.
1290
0.12
900.
0162
0.01
620.
0302
0.03
02(0
.0382
)(0
.038
2)(0
.030
4)(0
.030
4)(0
.025
6)(0
.025
6)(0
.010
1)(0
.010
1)
Com
posi
tion
effec
t0.0
731
0.10
260.
0082
0.03
550.
0649
0.06
720.
0245
0.03
22(0
.0333
)(0
.087
6)(0
.029
9)(0
.068
1)(0
.025
7)(0
.068
9)(0
.010
1)(0
.023
8)W
age
Str
uct
ure
0.07
21
0.06
960.
1208
0.10
29-0
.048
7-0
.033
30.
0057
0.00
69(0
.0487
)(0
.127
3)(0
.039
3)(0
.105
8)(0
.037
5)(0
.090
8)(0
.011
0)(0
.027
2)
2004–2008
Un
adju
sted
Ch
ange
0.0
331
0.03
310.
0143
0.01
430.
0188
0.01
880.
0066
0.00
66(0
.0000
)(0
.050
7)(0
.039
9)(0
.039
9)(0
.037
0)(0
.037
0)(0
.014
2)(0
.014
2)
Com
posi
tion
effec
t-0
.1513
-0.1
544
-0.1
633
-0.1
635
0.01
200.
0091
-0.0
335
-0.0
353
(0.0
559
)(0
.141
6)(0
.031
2)(0
.059
6)(0
.053
4)(0
.121
5)(0
.013
5)(0
.034
0)W
age
Str
uct
ure
0.18
44
0.18
630.
1776
0.14
550.
0068
0.04
080.
0401
0.04
79(0
.0782
)(0
.152
0)(0
.047
7)(0
.090
7)(0
.070
9)(0
.126
6)(0
.019
4)(0
.036
1)
Rew
eighte
ddis
trib
uti
ons
usi
ng
the
DF
Lappro
ach
.T
he
pro
pen
sity
score
use
sa
ver
yflex
ible
model
.C
ontr
ols
incl
ude
educa
tion,
aquadra
tic
poly
nom
ial
inex
per
ience
and
tenure
;a
whit
e/nonw
hit
eand
marr
ied/si
ngle
dum
my;
occ
upati
on,
indust
ry,
yea
rand
regio
nfixed
effec
ts;
and
inte
ract
ions
bet
wee
nsc
hooling
and
exp
erie
nce
.R
ecen
tere
dIn
fluen
ceR
egre
ssio
ndec
om
posi
tions
dec
om
pose
educa
tion,
aquadra
tic
poly
nom
ial
inex
per
ience
and
tenure
;a
whit
e/nonw
hit
eand
marr
ied/si
ngle
dum
my;
occ
upati
on,
indust
ry,
yea
rand
regio
nfixed
effec
ts.
Boots
trapp
edta
ndard
erro
rsare
obta
ined
by
runnin
g500
iter
ati
ons
usi
ng
bala
nce
dsa
mple
sof
1,0
00
per
form
ance
and
non-p
erfo
rmance
pay
jobs
each
(i.e
.su
bsa
mple
sco
nta
in2,0
00
obse
rvati
ons)
.
37
Figure 1: Incidence of Performance Pay
(a) BHPS 1991–2008 (b) SOEP 1985–2010
(c) BHPS 1991–2008 (d) SOEP 1985–2010
38
Figure 2: Wage Inequality in Great Britain and Germany
(a) Distribution of Wages (Great Britain) (b) Distribution of Wages (Germany)
(c) St.Dev. log hourly wages (Great Britain) (d) St.Dev. log hourly wages (Germany)
39
Figure 3: Reweighting procedure: Kernel densities
(a) BHPS 1991–1993 (b) SOEP 1985–1989
(c) BHPS 2003–2008 (d) SOEP 2004–2008
40
Figure 4: Quantile RIF regressions: BHPS
(a) High (First or Higher Degree) (b) Medium (HND, HNC or Teaching)
(c) Low (A Level) (d) O Level
41
Figure 5: Quantile RIF regressions: SOEP
(a) High (University) (b) Medium (Technical)
(c) Low (Vocational)
42