article21 article11
Based on the two articles provided summarize and answer the following questions for each article for factors related to pay differences. Use APA format and cite the articles. One page per article
1. What were the main findings in the article?
2. To which of the above four factors (factors that are bolded) are the findings of the article providing support? Why?
3. How can the information from the article be helpful to management or job-seekers?
207
© Blackwell PublishingLtd and theDepartment ofEconomics,University ofOxford, 2010.PublishedbyBlackwell PublishingLtd,
9600 Garsington Road, Oxford OX4 2DQ, UK and 350 Main Street, Malden, MA02148, USA.
OXFORD BULLETIN OF ECONOMICSAND STATISTICS, 73, 2 (2011) 0305-9049
doi:10.1111/j.1468-0084.2010.00613.x
Preferences,ComparativeAdvantage, and
CompensatingWageDifferentials forJob
Routinization*
Climent Quintana-Domeque
Departament de Fonaments de l’Anàlisi Econòmica, Universitat d’Alacant, Campus de Sant
Vicent del Raspeig, 03690, Alacant, Spain (e-mail:climent@ua.es)
Abstract
Iattempt toexplainwhythere isnotmuchevidenceoncompensatingwagedifferentials for
job disamenities. I focus on the match between workers’preferences for routine jobs and
thevariability in tasksassociatedwith the job.UsingdatafromtheWisconsinLongitudinal
Study, I find that mismatched workers earn lower wages and that both male and female
workers in routinized jobs earn, on average, 5.5% and 7% less than their counterparts
in non-routinized jobs. However, once preferences and mismatch are accounted for, this
differencedecreases to2%formen,and4%forwomen,not statistically significant inboth
cases.
I. Introduction
Formore than30years, labour economists havebeen trying tofindevidenceofwagepre-
miumsfor jobs that involvesuchdisamenitiesasphysicaleffort, routinenatureof thework
or job insecurity.According to the theory of compensating wage differentials, which goes
back to Adam Smith and involves the framework of analysis outlined by Rosen (1974),
ÅThis article is a revised version of Chapter 1 of my Ph.D. dissertation at Princeton University. I would like to
thank my advisor,Alan Krueger, who has always been exceptionally generous with his advice. I am also extremely
grateful to Jesse Rothstein for his insights and suggestions. I am particularly indebted to Carlos Bozzoli and Marco
Gonzalez-Navarro for their many thoughtful remarks. Thanks also go to the Editor Beata Javorcik and the seminar
participants at Princeton University, Universitat d’Alacant, Universitat de les Illes Balears, SAE 2007 Meetings in
Granada and Universidad Pablo de Olavide. Special thanks go to an anonymous referee of the Oxford Bulletin of
EconomicsandStatisticswhosecomments, insightsandsuggestionsmade theconceptual frameworkshorter, clearer
and neater. I also want to thank Erik Plug for providing me with the codes used in his previous work. Financial
support from theRafael delPinoFoundation, theBankofSpain and theSpanishMinistryofScience and Innovation
(ECO2008-05721/ECON) is gratefully acknowledged. The usual disclaimers apply. Previous versions of this paper
are circulated as Industrial Relations Section Working Paper 525, Princeton University, and IVIE Working Paper
WP-AD 2010-06. This research uses data from the Wisconsin Longitudinal Study (WLS) of the University of Wis-
consin-Madison. Since 1991, theWLS has been supported principally by the National Institute onAging (AG-9775
andAG-21079), with additional support from the Vilas Estate Trust, the National Science Foundation, the Spencer
Foundationand theGraduateSchoolof theUniversityofWisconsin-Madison.Apublicusefileofdata fromtheWLS
is available from the University ofWisconsin-Madison, 1180 Observatory Drive, Madison,Wisconsin 53706 and at
http://www.ssc.wisc.edu/∼wls/data/. The opinions expressed herein are those of the authors.
JELClassification numbers: J3, J31.
208 Bulletin
workersmust receiveawagepremiumforsuffering fromjobdisamenities, ceterisparibus.
However, a surveyof theevidencehasconcluded that ‘testsof the theoryofcompensating
wage differentials are inconclusive with respect to every job characteristic except risk of
death’(Borjas, 2005, Ch. 6, p. 224).
It is obvious that on-the-job risk of death is an undesirable job characteristic, and the
availableempirical evidence indeedsuggests thatwagesarepositivelyassociatedwithon-
the-jobriskofdeath(ViscusiandAldy,2003).However,manyother jobcharacteristicsare
not regarded as intrinsically undesirable by all workers. Instead, the desirability of a large
number of job attributes depends crucially on individual workers’ tastes or personalities.
Smith(1979)notes that theheterogeneityofworkers’tastesmaketestingforcompensating
wage differentials difficult.
At first glance, preference heterogeneity may seem consistent with mixed results for
repetitivework.Forexample,Lucas (1977)findsevidenceof significant compensation for
repetitivework,whileBrown(1980) reportsanegativeestimate.Almost20years later, the
mixed results are even more striking. Daniel and Sofer (1998) present some such results
in their paper.
One straightforward way to account for preference heterogeneity when looking for
compensating wage differentials is to run separate wage regressions for workers with
different preferences. Still, as I show in section II, non-routine-preferring workers earn
lower wages in routinized jobs, which is contrary to what the theory of compensating
wage differentials would predict. Therefore, preference heterogeneity by itself does not
explain the puzzle of compensating wage differentials.
Why,evenafteraccountingforpreferenceheterogeneity,arecompensatingwagediffer-
entialsnotobservedor incorrectlysigned?What ifworkers’preferences forone typeof job
(or job attribute) are related to their productivity in performing that type of job?Workers’
tastes for a certain job attribute may correlate with their comparative advantage in such
jobs. This is not the same as saying that preferences can have a direct effect on wages,
independent of the type of job; i.e. workers with different preferences may have differ-
ent absolute advantages in performing any job. Rather, the key insight here is that when
workers’ preferences do not match job attributes, they are less productive. For example,
non-routine-preferring workers are likely to be more productive in non-routinized jobs
than routine-preferring workers. By the same token, routine-preferring workers are likely
to be more productive in routinized jobs than non-routine-preferring workers.
If matching were perfect and each worker was assigned to a job according to com-
parative advantage, then the marginal routine-preferring worker would be willing to pay
for working in a routinized job. Similarly, the marginal non-routine-preferring worker
would need to be compensated for working in a routinized job. This would be consistent
with the compensating wage differentials theory.
However, as Lang and Majumdar (2004) pointed out, both casual empiricism and
researchshowthatmatchingis imperfect.Morerecently,Shimer(2007)acknowledgesthat
skills and geographical location of workers are poorly matched with the skill requirement
and locationof jobs: unemployedworkers are attached to anoccupation and ageographic
location where jobs with their skills are currently scarce. Here, a similar point can be
made. As I will show, a mismatch between workers’ preferences and job attributes does
exist, and must be taken into account when looking for compensating wage differentials.
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
Compensating wage differentials for job routinization 209
Indeed, my findings indicate that not accounting for mismatch in wage equations could
bias compensating wage differentials estimates.
I propose a simple assignment model with Nash bargaining over wages for analysing
the role of mismatch when looking for compensating wage differentials. Assuming that
observed workers are not in long-run market equilibrium, all workers, no matter what
their preferences are, need to be compensated if working in the sector with a shortage of
workers in the absence of pay differentials. However, only mismatched workers, who are
less productive because their sectors do not match their preferences, are penalized.
This simple frameworkoffers a rationale for theexistenceofmixedestimates for com-
pensatingwagedifferentials. Indeed, in the literature, thestandardestimatesmayconfound
theeffectonwagesofthejobattributebeinganalysedwiththeoneattributabletomismatch.
Thispaper focusesonjobroutinization(i.e. jobs involvingrepetitiveandroutine tasks).
I consider this isan important jobattribute tostudybecauseestimates for it in the literature
aremixed (e.g.Lucas, 1977;Brown, 1980;Daniel andSofer, 1998). So, this analysismay
shednewlighton thesourcesof thesemixed results.Furthermore,Table1shows that29%
of male workers and 36% of female workers report that ‘being able to do different things
rather than thesamethingsoverandover’is ‘muchmore important thanhighpay’. Indeed,
the table indicates that variability of tasks is one of the most highly valued characteristics
on the job for workers. This suggests that it should be easier to find compensating wage
differentials for job routinization than for other job attributes.
Using data from the Wisconsin Longitudinal Study (WLS), I find that mismatched
workersearn lowerwages.Myresultsalso indicate thataccountingformismatch is impor-
tant in obtaining more reliable estimates of compensating wage differentials. On average,
male workers in routinized jobs are paid 5.5% less than workers in non-routinized jobs,
after accounting for: differences in completed years of education, IQ measured at high
school, highschool rank, adult cognition, tenure,occupationandfirmsize.Thisdifference
decreases to4.5%after accounting fordifferences in thepreference for routinework.Fur-
thermore, controlling formismatch reduces thedifference inaveragewagesbetweenmale
workers in routinizedvs.non-routinized jobs to2%,and this isnot statistically significant.
For female workers, the difference decreases from 7% to 4%.
This paper is laid out as follows. Section II briefly describes the puzzle. It presents a
brief review of the compensating wage differentials literature, offers a description of the
WLS dataset, and takes a first look at the data. Section III presents a model that sheds
TABLE 1
Percentage of currently employed individuals reporting that job characteristic is much
more important than high pay, WLS 1992–93
Job characteristic
Men Women
Being able to do different things rather than the same things over and over 29 36
Being able to work without frequent checking by a supervisor 22 27
Having the opportunity to get on-the-job training 18 25
Having a job that other people regard highly 7 11
Being able to avoid getting dirty on the job 2 6
Source: Table 2 inAndrew et al. (2006), page 51.
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
210 Bulletin
light on the puzzle. Section IV offers the empirical model. My results are in section V.
SectionVI offers some robustness checks. Finally, sectionVII concludes.
II. The puzzle
A brief review
More than twocenturies ago,AdamSmithnoted thatworkerswith the same level of com-
petence should be paid different wages if their working conditions are different. Rosen
(1974) formalizesAdam Smith’s ideas showing that, under perfect competition, identical
workers need to be compensated for job disamenities.
The standard method for testing the prediction of this theory is to estimate a wage
equation with characteristics of the job (z) and personal characteristics (p). In general, the
equation is of the form:
ln(w)= � + �z + �p+ �. (1)
The estimation of equation (1) using cross-sectional data identifies a market relationship
between ln(w) and z. If the market relationship is linear, then � measures the marginal
cost of the disamenity for any worker who is in his most preferred job in long-run market
equilibrium. For an undesirable job attribute, the theory predicts that �
>0.
This iswhat is predictedby the standard theoryof compensatingdifferentials:workers
have heterogeneous preferences, and firms are heterogeneous with respect to the costs
of providing good working conditions; in long-run equilibrium workers who value good
conditions most are matched with firms that have the lowest costs of providing them, and
in thesematchesconditionsaregoodandwagesare low.Conversely,highwagesandpoor
conditionsareobservedinmatchesofworkerswhocarelesswithfirmsthathavehighcosts.
This long-run relationship is the only thing that the standard theory gives us.1 However,
theempirical evidenceoncompensatingwagedifferentials ismixed for jobcharacteristics
other than the risk of death [see Rosen (1986) for a classical discussion on the theory of
equalizing differences].
There have been several previous attempts at solving this puzzle. First, omitted vari-
ables can lead to biased estimates because of the correlation between unobserved skills,
preferences for the job attribute under study, individual productivities and the quality of
workingconditions(e.g.Brown,1980;DuncanandHolmlund,1983;Garen,1988;Kostiuk,
1990; Hwang, Reed and Hubbard, 1992). Second, when working conditions are reported
by the workers themselves, the estimates are likely to suffer from simultaneity bias (e.g.
McNabb,1989).Further, ifanswerstosurveyquestionsaboutworkingconditionsaregiven
in subjective terms, then the estimates are likely to suffer from subjectivity biases (e.g.
McNabb, 1989). Finally, when worker conditions are defined using average occupation
(or industry)characteristicsand thenmatched to individualworkers,misclassificationbias
may arise.
From an empirical perspective, in this paper I take into account most of these biases.
First, I control for preferences for the job attribute under study, and use IQ measured at
1I thank an anonymous referee of the Oxford Bulletin of Economics and Statistics for clarifying this issue.
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
Compensating wage differentials for job routinization 211
high school and high school rank as proxies for unobserved skills and individual produc-
tivities, and occupation and size of firm dummy variables to account for characteristics
other than job routinization (the job attribute under study) that may be related to worker
productivities. Second, job routinization is measured by time spent doing monotone tasks
in order to circumvent the problem of subjectivity biases due to the use of answers given
in subjective terms. Last but not least, I measure working conditions at the worker level,
not at the occupation level, to avoid misclassification bias.
From a theoretical point of view, this paper can be thought of as looking at the conse-
quenceof thepossibility thatobservedworkersarenot inalong-runequilibrium,providing
anaccountof short-runwagedeterminationwhenworkers arenotperfectlymatched.This
is the gap that this study aims to fill. I present a very simple model: workers are randomly
assigned to jobs and wages are determined by Nash bargaining. The model highlights
the effect of mismatch on wages, which must be taken into account when looking for
compensating wage differentials.
Istartbypresentingtheimplicationsofpreferenceheterogeneity(about theattractiveor
unattractive features of performing a job task) for estimates of compensating wage differ-
entials. Suppose there are two types of workers: those who enjoy z (x =1) and those who
havedistasteforz (x =0).Inthatcase, totest thetheoryofcompensatingwagedifferentials,
the following regressions should be run:
ln(w)= �0 + �0z + �0p+ �0, if x =0 (2)
ln(w)= �1 + �1z + �1p+ �1, if x =1. (3)
If the theory is correct, I should find evidence on �0>0 and �1<0: workers who have distaste for z (x =0) are compensated for working in a job involving high levels of z, while workers who enjoy z (x =1) are willing to pay for working in a job involving high levels of z. With these predictions at hand, I can assess the existence of compensating wage differentials for job routinization depending on workers’preferences. Before taking a first look at the data, I provide a description of the dataset used in this paper.
Data
I use data from the WLS of the University of Wisconsin-Madison. The sample contains
information on 10,317 men and women who graduated from Wisconsin high schools in
1957,approximatelyone-thirdofall seniors inWisconsinhighschools in1957. It contains
a rich set of self-reported information fromsamplemembers, siblings andparents, aswell
asadministrativedata, collected inaseriesof surveys:1957(graduates),1964(graduates),
1975 (graduates), 1977 (siblings), 1992–93 (graduates), 1993–94 (siblings) and 2003–05
(graduates and spouses).
I focuson the1992–93waves,when respondentswere in their early50s.Thisdecision
is based on both informational requirements and sample (size and selectivity) consider-
ations.Firstofall, informationonworkers’preferencesisnotavailablepriortothe1992–93
waves.Second,participationinthelabourmarket ishigherforpeople intheir50s(1992–93
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
212 Bulletin
waves) than in their 60s (2003–05 waves): 92.4% of men were employed in 1992 while
only47.8%ofthemwereemployedin2004.Finally, thishelpsmetominimizenon-random
attrition problems.
TheWLSdatasetoffers anopportunity for exploring the roleofmismatch inobserving
compensating wage differentials. It contains a set of individual characteristics obtained
from the (graduate) respondents, such as IQ score measured at high school, high school
rank, adult cognition, education, tenure, preference for job routinization, hourly wages,
hours of work, number of hours performing different tasks on the job, etc. Moreover,
the sample is quite homogeneous (high school graduates from Wisconsin high schools in
1957), which makes any concerns about omitted variables less important.
Mysampleisrestrictedtoworkerswhowereemployedin1992.Unfortunately,employ-
mentstatusismissingfor1,824individuals.Thisimpliesadramaticdecreaseintheoriginal
sample size from 10,317 to 8,493. There are 7,196 individuals employed in 1992. After
restricting our sample size to those individuals having a positive hourly wage rate, the
number of observations decreases to 6,756. Focusing only on Wisconsin residents, the
sample decreases to 4,696.The sample also excludes individuals who were: working less
than20hoursperweek, self-employed, employeesof their owncompanyor familywork-
ers. Farm workers and members of the military also are excluded from my sample.After
applying these restrictions, my working sample is left with approximately 3,800 observa-
tions. The presence of extreme values in the wage distribution was detected accidentally
through the comparison of average wages for men and women. To avoid the estimates
being driven by extreme values in the wage distribution, I trim the tails of the log-wage
distribution at both the 3% bottom and the 3% top. Finally, after dealing with missing
observations for thevariablesused in theanalysis, theworkingsample size is about3,200.
The next subsection presents the definition of the main variables used in the empirical
analysis.
Definition of the main variables
The main variables in this paper are job routinization; worker’s preference for routine;
and mismatch, i.e. the discrepancy between job routinization and worker’s preference for
routine. In this subsection, I discuss how these variables are measured.
The job routinization indicator (z) – whether a job is classified as routinized or non-
routinized – is constructed using the fraction of working time doing the same things over
and over: job routinization is measured as 1 (routinized job) if the fraction of working
time doing the same things over and over is equal to or higher than 0.5. Sensitivity anal-
yses with alternative definitions of job routinization will be performed in the robustness
checks section. I compute this fraction as the ratio of the number of weekly hours doing
the same things over and over on the job to the total number of weekly working hours.
Note that the reportednumberofhours canbecomparedacross individuals; this addresses
standard subjectivity bias concernsdue toworkers’subjective assessments aboutworking
conditions.Moreover, the fact that thenumberofhoursworked is reportedby theworkers
themselves confronts the misclassification bias that is attributable to imprecise matching
of average job (occupationor industry) characteristics to individualswhose jobcharacter-
isticsmaydepart (byand large) fromtheaveragecharacteristicswithin theiroccupationor
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
Compensating wage differentials for job routinization 213
industry.Ofcourse, as inprevious studies, simultaneitybiasesmayexist:workerswhoare
unhappy with earnings that they receive may also respond negatively when asked about
job attributes (McNabb, 1989).
The worker’s preference for routine indicator (x) – whether a worker is classified as a
routine-preferringworkeroranon-routine-preferringworker–ismeasuredbytheresponse
to this question: ‘To what extent do you see yourself as someone who prefers work that
is routine and simple?’ The possible answers to this question are: agree strongly, agree
moderately, agree slightly, neither agree nor disagree, disagree slightly, disagree moder-
ately, disagree strongly.This is oneof thequestions asked in scoring thefive-factormodel
of personality structure, and it is included in the personality section of the 1992–93 ques-
tionnaire, separate from jobhistoryor current/last jobcharacteristics.Hence, thepotential
concerns about framing effects are minimized. For workers who agree strongly, moder-
ately or slightly, preferring work that is routine and simple, x =1. Sensitivity analyses
with alternative definitions of worker’s preference for routine will be performed in the
robustness checks section.
Finally, mismatch between job routinization and worker’s preference for routine and
simplework ismeasuredas theabsolutevalueof thedifferencebetween z and x,m(z,x)=
|z −x|. I adopt this approach because absolute value seems to be the most intuitive way
of thinking about the discrepancy between two variables. Note that for binary indicators,
the absolute-value deviation is equivalent to the quadratic deviation.
Descriptive statistics
Table2presents themaindescriptive statistics of theWLSsample for currently employed
individuals (1992–93). A first glance at the table shows that, on average, male
workers in non-routinized jobs earn $18.09 per hour, while male workers in routinized
jobs earn $15.21: a difference of approximately $3 in the hourly wage. Women in non-
routinized jobs earn $11.41 per hour, while women in routinized jobs earn $9.33.
Although these are unadjusted averages, workers do not seem to be compensated for job
routinization.
The tablealsoshowsthat themajorityofmen(52%)work innon-routinized jobs,while
the majority of women work in routinized jobs (64%). At the same time, the fraction of
workers who prefer routine and simple work is higher for women than for men: 0.24 vs.
0.18.Thefact thatworkers innon-routinizedjobsarenotcompensatedfor jobroutinization
is evenmore strikinggiven that the supplyof routine-preferringworkers seems tobevery
low(24%ofmaleworkers,18%of femaleworkers) incomparison to thedemandfor them
(48% of male workers, 64% of female workers).
Can mismatch explain the apparent lower wages in routinized jobs? The percentages
of well-matched workers (according to job routinization and preference for routine and
simple work) are 62% and 53% for men and women, respectively. Hence, mismatch is
higher forwomen (47%) than formen (38%).Forbothmenandwomen,mismatch isvery
high.Moreover,mismatchmayberesponsible for (partof) thedifference inaveragewages
between routinized and non-routinized jobs: mismatched men are paid $15.51 per hour
while those who are well-matched are paid $17.44 per hour. For women the difference is
smaller: $9.61 vs. $10.53.
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
214 Bulletin
TABLE 2
Descriptive statistics, WLS 1992–93
Men Women
Obs. Mean SD Obs. Mean SD
Hourly wage routinized jobs 800 15.21 4.98 1,111 9.33 3.46
Hourly wage non-routinized jobs 865 18.09 6.10 637 11.41 4.19
Job routinization (z =1 if fraction of weekly worked 1,665 0.48 0.50 1,748 0.64 0.48
hours doing the same things over and over is equal
or higher than 0.5, z =0 otherwise)
Routine-preferring worker (Preference for routine 1,656 0.18 0.38 1,743 0.24 0.42
and simple work: x =1 if strongly/moderately/slightly
agree, x =0 if strongly/moderately/slightly/disagree
or neither agree nor disagree)
Mismatch, |z −x| 1,656 0.38 0.49 1,743 0.47 0.50
Fraction of weekly worked hours doing the 1,665 0.48 0.38 1,748 0.61 0.37
same things over and over
Preferences for routine and simple work
Strongly agree 94 0.06 – 108 0.06 –
Moderately agree 162 0.10 – 238 0.14 –
Slightly agree 35 0.02 – 48 0.03 –
Neither agree nor disagree 6 0.00 – 18 0.01 –
Slightly disagree 59 0.04 – 68 0.04 –
Moderately disagree 447 0.27 – 503 0.29 –
Strongly disagree 853 0.52 – 760 0.44 –
Hourly wage mismatched workers 636 15.51 5.08 821 9.61 3.52
Hourly wage well-matched workers 1,020 17.44 6.04 922 10.53 4.12
Hourly wage 1,665 16.71 5.77 1,748 10.09 3.87
IQ (measured at high school) 1,665 98.95 14.35 1,748 100.10 13.89
High school rank 1,543 41.59 27.03 1,636 57.04 27.21
Education (years of completed education) 1,665 13.44 2.19 1,748 12.93 1.71
Adult cognition score (WAIS) 1,653 7.47 2.78 1,739 7.62 2.63
Tenure 1,659 19.34 11.00 1,744 12.09 9.00
Notes:Author’s calculations.
As expected, men are paid higher hourly wages than women: $16.71 vs. $10.09. Not
surprisingly, given the cohort under study, born around 1940, women on average are less
educated than men.
Table 3 shows the distribution of workers (by their preferences for routine and simple
work) across jobs (by routinization) and the average hourly wages by worker-job type.
Among men, 42% of non-routine-preferring workers are mismatched into routinized jobs
(567/1359×100),while thispercentage is57 forwomen(758/1331×100).Forbothmen
and women, the percentage of mismatched workers is lower in non-routinized jobs. This
is consistent with the fact that the majority of men and women are non-routine-preferring
workers (76%ofmen, and82%ofwomen).Regarding the averagehourlywage, the table
describes an interesting feature of my data: there are no differences in average wages
betweenmismatchedandwell-matched routineworkers. Indeed, thedifferencesare found
only for non-routine-preferring workers.
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
Compensating wage differentials for job routinization 215
TABLE 3
Distribution of workers across jobs and average hourly
wages by worker-job type, WLS 1992–93
(Number of observations) z =0 z =1
Male
x =0 48% 34%
18.4 15.7
(792) (567)
x =1 4% 14%
14.3 14.0
(69) (228)
Female
x =0 33% 43%
11.8 9.7
(573) (758)
x =1 4% 20%
8.4 8.5
(63) (349)
Notes:Author’s calculations.
TABLE 4
Fraction of weekly worked hours doing the same things over and over by
occupational category, WLS 1992–93
Occupational category Men Women
Professional and technical specialty occupations 0.27 0.46
Executive, administrative and managerial occupations 0.27 0.41
Sales occupations 0.54 0.70
Administrative support occupations (including clerical) 0.63 0.65
Precision production, craft, and repair occupations 0.44 0.83
Operators and fabricators 0.79 0.86
Service occupations 0.64 0.79
Handlers,equipmentcleaners,helpers, labourers, farmoperators 0.73 0.90
farm workers and related occupations
Source:Author’s calculations.
A first look at the data
I start bymeasuring job routinizationas the fractionof timeatworkdoing the same things
over and over. Routine-preferring workers (x =1) are defined as those individuals who
strongly agree, moderately agree, slightly agree or neither agree nor disagree, with the
statement ‘I see myself as someone who prefers work that is routine and simple’.
Table 4 reports the degree of job routinization by occupational category for men and
women, respectively. As the table makes clear, ‘Professional and Technical Specialty
Operations’, and ‘Executive,Administrative, and Managerial’occupational categories on
average involve less routinization, while occupations such as ‘Operators and Fabricators’
involve more routinization of tasks. Another interesting feature that emerges from this
table is that female workers tend to spend a higher fraction of time than male workers
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
216 Bulletin
TABLE 5
Job routinization and wages by workers’ preferences. OLS estimates for
men and women
Men Women
Workers’ preferences Workers’ preferences
Routine Non routine Routine Non routine
(1) (2) (3) (4)
Job routinization 0.035 −0.082 0.004 −0.096
(0.052) (0.024) (0.046) (0.016)
Completed years of education 0.044 0.032 0.042 0.016
(0.016) (0.005) (0.015) (0.006)
IQ measured at high school 0.002 0.002 0.002 0.004
(0.002) (0.001) (0.002) (0.001)
High school rank 0.000 0.000 −0.000 0.000
(0.001) (0.005) (0.001) (0.000)
Adult cognition score −0.005 0.004 −0.002 0.001
(0.007) (0.003) (0.006) (0.004)
Tenure 0.009 0.008 0.016 0.013
(0.002) (0.001) (0.002) (0.001)
R2 0.25 0.32 0.39 0.38
Number of observations 270 1,253 378 1,243
Notes: Dependent variable is log(hourly wage). Routine-preferring worker equals to 1 if
worker agrees strongly, moderately or slightly, preferring work that is routine and simple.
Job routinization is the fraction of weekly worked hours doing the same things over and
over. Heteroskedasticity robust standard errors are reported in parentheses.All regressions
include occupation dummy variables.
doing the same things over and over. In other words, women tend to do more routinized
tasks than men within occupational categories.
The results from Table 5 show evidence contrary to the theory of compensating wage
differentials:workerswithlowerpreferencesforroutineandsimpleworkearnlowerwages
in the routinized jobs.Columns(2)and(4)showthatbothnon-routine-preferringmaleand
female workers do not appear to be compensated for working in routinized jobs; rather,
if anything, they appear to be penalized. For routine-preferring workers, columns (1) and
(3), I find a positive but not statistically significant association between job routinization
and hourly wages.
The bottom line of Table 5 is that preference heterogeneity clearly matters, but in a
surprisingly opposite way to what one would have expected from a selection-bias explan-
ation: workers with lower preference for routine and simple work earn lower wages in
routinized jobs.This paper provides an explanation for such a finding.
Note that the implicit assumption behind the prediction of a positive association
betweenjobroutinizationandwagesfornon-routine-preferringworkersisthattheymustbe
compensatedbecauseof theirhigherdisutilitywhenworking in routinized jobs.However,
non-routine-preferringworkers are likely to be less productive in routinized jobs. In other
words, workers’preferences are likely to reflect two things that are equally important for
wage determination: their disutility from working, which will be higher as the discrep-
ancy between preferences and job attributes (characteristics or job tasks) increases; and
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
Compensating wage differentials for job routinization 217
their comparative advantage on the job, which will be lower as the discrepancy between
preferences and job attributes increases.
If matching were perfect, and each worker was assigned to a job according to her
comparative advantage, then the productivity effect of comparative advantage would not
play any role: productivity would be the same for every worker, because every worker
would be assigned to a job where her comparative advantage was maximized. However,
matching is far from perfect, and neglecting its influence on wages is likely to confound
thecompensatingwagedifferentialsestimates. Inotherwords,equations (2)and(3)would
be mis-specified if mismatch also matters.
Thus, a potential explanation for the puzzling results in Table 5 is that preferences for
performinga joband theworker’s comparativeadvantage inperforming it are (positively)
correlated. If this is the case, then workers with lower preference for routine and simple
work will earn lower wages in routinized jobs, not because they are not compensated for
taking such jobs but because they are less productive in performing them.
III. Conceptual framework
In this section, I present a simple assignment model with Nash bargaining to show the
effect of mismatch on the wage rate. The main purpose of the model is to highlight the
importance of the mismatch productivity effect on the wage rate, and its relevance for
understanding estimates of compensating wage differentials.
There are two types of workers x ∈ {0,1}, defined by their preferences for a job attri-
bute (x =0 fornon-routinepreferringworkers,x =1 for routine-preferringworkers) anda
continuumoffirms’types z∈[0,1], definedby the jobattribute (z =0 for completelynon-
routinized jobs, and z =1 for completely routinized jobs).Eachfirm is randomlymatched
witheachworker: (z,x) foreachfirm-workerpair.Then, thefirm zandtheworkerxbargain
over the division of the match surplus to decide the optimal wage.
The profit function of the firm is given by
� =A(m(z,x))−w, (4)
whereA isgross revenue(production),whichdependsnegativelyonmismatchm(z,x), and
w is the wage rate.The negative relationship between A and m is assumed on the grounds
that the worker’s taste for a job attribute (e.g. routine-preferring worker) is likely to be
positively correlated with his ability to perform well in a job with such an attribute (e.g.
routinized job). In other words, a routine-preferring worker will tend to have a compara-
tive advantage indoing repetitive things.Tinbergen (1975) sets a production function that
depends on the extent to which a person’s abilities match those required in the execution
of a job task.
The utility function of the worker is given by
u=w− v(z,m(z,x)), (5)
wherev is thedisutility fromwork,whichdependspositivelyonmismatchm(z,x)between
the job characteristic (z) and the worker’s preference for such a job characteristic (x), and
on the job characteristic (z).
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
218 Bulletin
This randomassignmentsettingcanbeunderstoodbyassumingthatdue tofrictions the
market is not in long-run equilibrium. This is a plausible assumption as the data suggest
thatmismatch is substantial:18%ofmaleworkersareclassifiedas routine-preferworkers,
while48%of themareworking in jobs involvinghalf (ormore)of theirweekly timedoing
the same thingsover andover.Hence, I assume that the routinized sector is the sectorwith
a shortage of workers in the absence of pay differentials,
∂v(z,m(z,x))
∂z
>0.
The solution to the Nash bargaining problem is obtained from
max
w
{��u1−�}, (6)
where 0<�<1 measures the firm bargaining power.
The FOC gives us the optimal wage rate:
wÅ(z,m(z,x))= �v(z,m(z,x)) +(1− �)A(m(z,x)). (7)
The marginal effect of z holding m constant, which is the ‘standard’ compensating
wage differential, is
∂wÅ(z,m(z,x))
∂z
= � ∂v(z,m(z,x))
∂z
, (8)
which is positive given my previous assumption.
However, the total effect of z holding x constant is
dwÅ(z,m(z,x))
dz
= ∂w
Å(z,m(z,x))
∂z
+ ∂w
Å(z,m(z,x))
∂m(z,x)
∂m(z,x)
∂z
, (9)
where ∂wÅ(z,m(z,x))/∂z>0 from equation (8), ∂m(z,x)/∂z>0 if x =0 (i.e. the higher
is job routinization, the higher is the mismatch for a non-routine preferring worker), and
∂m(z,x)/∂z<0 if x =1 (i.e. the higher is job routinization, the lower is the mismatch for
a routine preferring worker).
Equation (9)givesusprecisely theeffectsbeingestimatedas �0 and �1 inequations (2)
and (3), in which m is omitted. What is the sign of ∂wÅ(z,m(z,x))/∂m(z,x)? The answer
to this question is given by Proposition 1.
Proposition 1.When mismatch also affects gross revenue (output), it has an ambiguous
effect on the wage rate. If the productivity effect dominates the disutility effect, then
mismatchaffects thewage ratenegatively. If the reverse is thecase, thenmismatchaffects
thewage rate positively. If both effects cancel eachother out, thenmismatchhasnoeffect
on the wage rate.
Proof
∂wÅ(z,m(z,x))
∂m
= � ∂v(z,m(z,x))
∂m
︸︷︷︸
>0
+(1− �)∂A(m(z,x))
∂m
︸︷︷︸
<0
. (10)
�
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
Compensating wage differentials for job routinization 219
Hence, givenProposition1,weconclude that the total effect of z holding x constant is
ambiguous.2
IV. Empirical model
Mymodelyields threeparameters that are captured inequation (11): a routine sectormain
effect (the‘standard’compensatingwagedifferential, �); a routine-preferringworkermain
effect (the absolute advantage of this type of worker, �); and a negative wage effect for
workerswhoare ina sectorother than theone theyprefer (thenegativeproductivity effect
due to mismatch, ).
ln(w)= � + �z + �x + m(z,x)+ �. (11)
To identify the effects of m and z, I need to be aware of the possibility that the error
term � is correlated with m and/or z. First, mismatch (m) is likely to be correlated with
worker’s ability: workers with worse skills are likely to be paid lower wages and to end
up being mismatched. Second, the level of job routinization (z) could be correlated with
worker’s skills and skills requirements of the job: routine jobs are perhaps those requiring
unskilled workers.
Imeasure relevantworker’s characteristics thatmaybe related tobothwages andmis-
match by education (completed years of education), IQ score measured at high school,
high school rank and an adult cognition measure that is based on 8 of the 14 items from
the WeschlerAdult Intelligence Scale (WAIS). To account for the relevant characteristics
of the job thatmaybe related tobothwages and job routinization, I control for occupation
dummyvariables (the8occupationalcategoriesaredescribed inTable4).Notice thatonce
I control for occupation, the unique variation used to identify the wage premium/penalty
associated with job routinization is within-occupation variation. Further, I also control
for size of firm dummy variables. Given this rich set of control variables (C′), it seems
plausible to identify the effects of m and z by means of equation (12):
ln(w)= � + �z + �x + m(z,x)+C′� +u. (12)
Finally, although I have a rich set of control variables that helps me to identify the
effects of z and m, regression (12) contains worker’s preferences (x), which may well be
endogenously determined and thus may compromise the interpretation of my estimates:
workers’ preferences are likely to be affected by their labour market experience. More
specifically, an individual’s working experience on a particular job (tenure) is likely to
affect his preferences for such a job. Although I do not have suitable data for assessing
whetherworkers’preferenceschangeover time, I try toovercomethisshortcomingbycon-
trolling for tenure: keeping tenure constant, the effect of preferencesonwages is obtained
net of the effect of tenure on preferences. Hence, C′ will also include tenure.
2Borghans et al. (2006) show that the effect of people skills on wages (in the equilibrium assignment) can be
decomposed into two effects: first, workers with more people skills earn more because they generate higher (net)
revenue(productivityeffect);second,workerswithmorepeopleskillstakejobswherepeopletasksaremoreimportant
and these jobs pay less, all else equal (compensating wage differential effect).
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
220 Bulletin
V. Results
Empirical findings
Tables 6 and 7 present the results on the effect of job routinization on wages for men
and women, respectively. Column (1) in Table 6 shows that, on average, male workers in
routinized jobsearn11%less thanmaleworkers innon-routinized jobs.Once theworker’s
preference for routine work is accounted for, this penalty is reduced to 10% [see column
(2)]. Column (3) shows that routinized jobs on average pay 7% less than non-routinized
jobs when mismatch is controlled; on average, mismatched workers earn 4% less than
well-matched workers. Hence, if mismatch is not accounted for, the negative effect of
job routinization on wages is overestimated. Indeed, once mismatch is included as a new
variable in the wage regression, I can explain a substantial portion of the incorrectly-
signed estimate for job routinization.
While columns (1) to (3) control for worker heterogeneity, they do not account for
job heterogeneity. In columns (4)–(6), I add both occupation and size of firm dummy
variables into the previous specifications in an attempt to account for both kinds of
heterogeneity. Notice that controlling for occupation is crucial to account for different
skill requirements of the job. The results in columns (4)–(6), are qualitatively similar to
thoseincolumns(1)–(3):maleworkers inroutinizedjobsearn5.5%less thantheircounter-
TABLE 6
Mismatch and compensating wage differentials. OLS estimates for men
(1) (2) (3) (4) (5) (6)
Job routinization −0.107 −0.095 −0.068 −0.053 −0.045 −0.023
(0.016) (0.016) (0.022) (0.016) (0.016) (0.021)
Routine-preferring worker – −0.073 −0.091 – −0.056 −0.071
(0.020) (0.022) (0.019) (0.022)
Mismatch – – −0.037 – – −0.031
(0.022) (0.021)
Completed years of education 0.049 0.049 0.048 0.033 0.033 0.033
(0.003) (0.004) (0.004) (0.005) (0.005) (0.005)
IQ measured at high school 0.003 0.003 0.003 0.003 0.003 0.003
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
High school rank 0.000 0.000 0.000 0.000 −0.000 −0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Adult cognition score 0.006 0.005 0.005 0.003 0.002 0.002
(0.003) (0.003) (0.003) (0.003) (0.003) (0.003)
Tenure 0.008 0.008 0.008 0.007 0.007 0.007
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
Occupation dummy variables? No No No Yes Yes Yes
Firm size dummy variables? No No No Yes Yes Yes
R2 0.27 0.28 0.28 0.36 0.36 0.37
Adjusted R2 0.27 0.27 0.28 0.35 0.36 0.36
Number of observations 1,523 1,523 1,523 1,520 1,520 1,520
Notes: Dependent variable is log(hourly wage). Job routinization equals to 1 if fraction of weekly worked hours
doing the same things over and over is equal to or higher than 0.5. Heteroskedasticity robust standard errors are
reported in parentheses.
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
Compensating wage differentials for job routinization 221
TABLE 7
Mismatch and compensating wage differentials. OLS estimates for women
(1) (2) (3) (4) (5) (6)
Job routinization −0.100 −0.083 −0.034 −0.072 −0.064 −0.038
(0.018) (0.018) (0.021) (0.016) (0.016) (0.020)
Routine-preferring worker – −0.114 −0.157 – −0.076 −0.099
(0.018) (0.021) (0.017) (0.019)
Mismatch – – −0.071 – – −0.037
(0.021) (0.019)
Completed years of education 0.047 0.046 0.045 0.022 0.022 0.021
(0.006) (0.006) (0.006) (0.006) (0.006) (0.006)
IQ measured at high school 0.005 0.005 0.005 0.004 0.004 0.004
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
High school rank 0.000 −0.000 −0.000 0.000 0.000 0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Adult cognition score 0.004 0.004 0.004 0.001 0.001 0.001
(0.003) (0.003) (0.003) (0.003) (0.003) (0.003)
Tenure 0.015 0.015 0.015 0.012 0.012 0.012
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
Occupation dummy variables? No No No Yes Yes Yes
Firm size dummy variables? No No No Yes Yes Yes
R2 0.31 0.32 0.32 0.44 0.45 0.45
Adjusted R2 0.30 0.32 0.32 0.44 0.44 0.44
Number of observations 1,621 1,621 1,621 1,612 1,612 1,612
Notes: Dependent variable is log(hourly wage). Job routinization equals to 1 if fraction of weekly worked hours
doing the same things over and over is equal to or higher than 0.5. Heteroskedasticity robust standard errors are
reported in parentheses.
parts innon-routinized jobs [seecolumn(4)].Thispenaltydecreases to4.5%once I adjust
for differences in preferences [see column (5)]. Finally, once workers’ preferences and
mismatch are accounted for, this difference is reduced to 2% [see column (6)]. Moreover,
this is not statistically different from zero.
Table 7 reports similar results for women. Accounting for differences in preferences
slightly decreases the job-routinization wage penalty, from 10% to 8% [columns (1) and
(2)], or from 7% to 6.5% [columns (4) and (5)].Again, adding mismatch into the model
seems tobe important: theeffectof job routinizationdecreases from8%to3.5%[columns
(2) and (3)], or from 6.5% to 4% [columns (5) and (6)]. In none of the cases, the job
routinization effect on wages is statistically significant once both preferences for routin-
ization and mismatch are accounted for. Mismatched female workers earn 4% less than
well-matched female workers.
Overall, two features of the data stand out. First, mismatch is negatively related to
wages.ThisisconsistentwithbothmyassignmentmodelandBorghansetal.(2008):people
aremostproductive in jobs thatmatchtheirstyle,andtheyearn lesswhentheyhavetoshift
to other jobs. Indeed, I find a mismatch effect after accounting for worker type (worker’s
preference for routine work), job type (job routinization), and other observable character-
istics at the worker, occupation and firm levels. Second, once mismatch is accounted for,
the coefficient on job routinization is attenuated.The evident mismatch effect can explain
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
222 Bulletin
a substantial portion (but not all) of the incorrectly-signed compensating differential for
jobroutinization indicated inpreviousanalyses. Indeed, in themodelswithoccupationand
size of firm dummy variables, the compensating differential for job routinization cannot
be statistically distinguished from zero. In the next section, I perform several robustness
checks to the use of alternative measures and the presence of outliers. Before presenting
the results of my sensitivity analyses, it is important to discuss my results.
Discussion
Myresultsshowthataccountingformismatchexplainsasubstantialportion(butnotall)of
the incorrectly signedcompensatingdifferential for job routinization indicated inprevious
analyses. The fact that job routinization has still a negative sign could be reflecting that
workers in routine jobsare lessproductive thanworkers innon-routine jobs.However,we
control fordifferentproxies for individualproductivity suchaseducationand IQ.Further-
more, in the most complete empirical models, the coefficient on job routinization is not
statistically different from zero.
Regarding the estimated effect of mismatch on wages, it must be recognized that this
could be picking up two different kinds of effects. On the one hand, mismatch can have
a negative effect on productivity due to the discrepancy between worker’s preferences
for routine jobs and the variability in tasks associated with the job [Tinbergen (1975)
sets a production function that depends on the extent to which a person’s abilities match
those required in the execution of a job task]. On the other hand, mismatch may reflect
unobserved worker’s ability: mismatched workers could be less productive to start with.
Unfortunately, I cannot disentangle these two effects in my paper. Nonetheless, the fact
that mismatch must be accounted for in wage equations is an important one.
Future research could benefit from such a framework using new and better data that
may help to disentangle these two effects by using quasi-experimental variation in mis-
match.For example, plant closingcouldbeusedasan instrument formismatch to identify
theeffectofmismatchonwages for ‘workerswhohavebeendisplaced fromanon-routine
job to a routine one by plant closing’.
VI. Robustness checks
This section addresses some potential concerns about my previous estimates: the use of
alternativemeasuresof job routinization, routine-preferringworker andmismatchand the
sensitivity of OLS estimates to outliers.
Alternative measures
Thediscreteapproachtomeasuringjobroutinizationandworkers’preferencesisappealing
because it is neat and clear cut. Unfortunately, it does not take full advantage of all the
available informationcontained inmydata.Moreover, the thresholdsdefiningroutine jobs
and routine-preferring workers are arbitrary.
In this subsection, I start byexploiting thevariability inworkers’preferencesandmea-
sures of job routinization. Here, job routinization is measured as a continuous variable;
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
Compensating wage differentials for job routinization 223
workers’preferencesaremeasuredbyseveralbinary indicators;andmismatch ismeasured
as it is in the rest of the paper. More specifically, the new job routinization variable is the
fraction of working time doing the same things over and over on the job (as in Table 5).
Workers’preferenceforroutineiscapturedbyseveralbinaryindicators:Routine-Preferring
Worker 1 (equal to 1 for workers who disagree strongly or moderately with the statement
‘I see myself as someone who prefers work that is routine and simple’, zero otherwise);
Routine-Preferring Worker 2 (equal to 1 for workers who agree slightly, neither agree
nor disagree, or disagree slightly with the previous statement, zero otherwise); Routine-
Preferring Worker 3 (equal to 1 for those workers who agree moderately or strongly with
the previous statement, zero otherwise). Tables 8 and 9 present the new estimates using
thesealternativemeasuresof jobroutinizationandworkers’preferences,wheretheomitted
categoryisRoutine-PreferringWorker1.Thenewestimatesareverysimilar totheprevious
ones: thenegativeassociationbetweenwagesand jobroutinizationdecreasesdramatically
after accounting for worker’s preference and mismatch. The tables also reveal a negative
association between mismatch and wages for both men and women: on average, both
mismatched female and male workers earn 3% less than their well-matched counterparts.
I also check the sensitivity of my estimates to the thresholds defining routine jobs and
routine-preferringworkers.Now, I classifya jobas routinized if the fractionof timedoing
TABLE 8
Mismatch and compensating wage differentials. OLS estimates for men. Alternative definitions
(1) (2) (3) (4) (5) (6)
Job routinization −0.144 −0.125 −0.086 −0.071 −0.058 −0.029
(0.021) (0.022) (0.027) (0.022) (0.022) (0.027)
Routine-preferring worker 2 – −0.052 −0.056 – −0.046 −0.049
(0.034) (0.034) (0.032) (0.032)
Routine-preferring worker 3 – −0.081 −0.104 – −0.061 −0.078
(0.022) (0.024) (0.021) (0.023)
Mismatch – – −0.046 – – −0.034
(0.020) (0.019)
Completed years of education 0.049 0.048 0.048 0.033 0.033 0.033
(0.004) (0.004) (0.004) (0.005) (0.005) (0.005)
IQ measured at high school 0.003 0.003 0.002 0.002 0.002 0.002
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
High school rank 0.000 0.000 0.000 0.000 −0.000 −0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Adult cognition score 0.006 0.004 0.004 0.003 0.002 0.002
(0.003) (0.003) (0.003) (0.003) (0.003) (0.003)
Tenure 0.008 0.009 0.009 0.007 0.007 0.007
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
Occupation dummy variables? No No No Yes Yes Yes
Firm size dummy variables? No No No Yes Yes Yes
R2 0.27 0.28 0.28 0.36 0.36 0.36
Adjusted R2 0.27 0.28 0.28 0.35 0.35 0.36
Number of observations 1,523 1,523 1,523 1,520 1,520 1,520
Notes:Dependentvariable is log(hourlywage). Job routinization is the fractionofweeklyworkedhoursdoing the
same things over and over. Heteroskedasticity robust standard errors are reported in parentheses.
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
224 Bulletin
TABLE 9
Mismatch and compensating wage differentials. OLS estimates for women.
Alternative definitions
(1) (2) (3) (4) (5) (6)
Job routinization −0.152 −0.126 −0.086 −0.106 −0.092 −0.068
(0.023) (0.023) (0.027) (0.022) (0.022) (0.025)
Routine-preferring worker 2 – −0.075 −0.086 – −0.047 −0.053
(0.030) (0.030) (0.028) (0.029)
Routine-preferring worker 3 – −0.110 −0.142 – −0.073 −0.092
(0.019) (0.022) (0.018) (0.020)
Mismatch – – −0.051 – – −0.029
(0.019) (0.018)
Completed years of education 0.045 0.045 0.044 0.021 0.022 0.021
(0.006) (0.006) (0.006) (0.006) (0.006) (0.006)
IQ measured at high school 0.005 0.005 0.005 0.004 0.004 0.004
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
High school rank −0.000 −0.000 −0.000 0.000 0.000 0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Adult cognition score 0.004 0.004 0.004 0.001 0.001 0.001
(0.003) (0.003) (0.003) (0.003) (0.003) (0.003)
Tenure 0.015 0.015 0.015 0.012 0.012 0.012
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
Occupation dummy variables? No No No Yes Yes Yes
Firm size dummy variables? No No No Yes Yes Yes
R2 0.31 0.32 0.33 0.44 0.45 0.45
Adjusted R2 0.31 0.32 0.32 0.44 0.44 0.44
Number of observations 1,621 1,621 1,621 1,612 1,612 1,612
Notes:Dependentvariable is log(hourlywage). Job routinization is the fractionofweeklyworkedhoursdoing the
same things over and over. Heteroskedasticity robust standard errors are reported in parentheses.
the same thingsover andover is above the third quartile on thedistributionof the fraction
of time. And, a worker is classified as routine-preferring if his score on the preference
for routine and simple work is above the third quartile on the distribution of preferences.
The new mismatch measure is the absolute value of the difference between these new
alternative measures. I provide new estimates with these alternative definitions for men
and women inTable 10.The new estimates are very similar.
Formen,column(1)showsthatworkers inroutinizedjobsonaverageearn7%less than
theircounterparts innon-routinizedjobs.Column(2)showsthataccountingfordifferences
in preferences makes the wage penalty lower: almost 6%. Finally, adding mismatch into
the model, column (3), decreases the wage penalty even further: 3%. Note too that being
mismatched is associatedwith awagepenalty of 7%.Similar qualitative results are found
for women in columns (4)–(6).
Sensitivity to outliers
The OLS estimates are known to be sensitive to outliers. In my analysis, I trimmed both
the bottom 3% and the top 3% of the wage distribution in order to avoid the influence
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
Compensating wage differentials for job routinization 225
TABLE 10
Mismatch and compensating wage differentials. OLS estimates for men and women.
Alternative thresholds defining routine jobs and routine-preferring workers
Men Women
(1) (2) (3) (4) (5) (6)
Job routinization −0.071 −0.056 −0.030 −0.089 −0.072 −0.054
(0.018) (0.018) (0.019) (0.017) (0.017) (0.018)
Routine-preferring worker – −0.086 −0.070 – −0.116 −0.114
(0.019) (0.020) (0.018) (0.018)
Mismatch – – −0.067 – – −0.038
(0.019) (0.017)
Completed years of education 0.050 0.050 0.049 0.050 0.048 0.048
(0.004) (0.004) (0.004) (0.005) (0.005) (0.005)
IQ measured at high school 0.003 0.003 0.003 0.006 0.005 0.005
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
High school rank 0.000 0.000 0.000 0.000 −0.000 −0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Adult cognition score 0.006 0.005 0.005 0.004 0.004 0.004
(0.003) (0.003) (0.003) (0.003) (0.003) (0.003)
Tenure 0.008 0.008 0.008 0.015 0.015 0.015
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
R2 0.26 0.27 0.27 0.30 0.32 0.32
Number of observations 1,523 1,523 1,523 1,621 1,621 1,621
Notes: The dependent variable is log(hourly wage). Job routinization equals to 1 if fraction of weekly worked
hours doing the same things over and over is above the third quartile on the distribution of time. Routine-preferring
worker equals to 1 if his score on the preference for routine and simple work is above the third quartile on the
distribution of preferences. Heteroskedasticity robust standard errors are reported in parentheses.
of extreme values. Here, I go one step further and perform a median Quantile regression
analysis tomake sure thatmypreviousOLSestimates arenot drivenbyextremevaluesof
the wage distribution.
The new (median) estimates reported in Tables 11 and 12 are robust to outliers and
very similar to my previous OLS estimates. In Table 11, column (1) shows that, at the
median, male workers in routinized jobs earn 11% less than male workers in non-
routinized jobs. Once the worker’s preference for routine work is accounted for, this
penalty is reducedto9%[column(2)].Column(3)showsthat routinizedjobsat themedian
pay 5% less than non-routinized jobs when mismatch is controlled. Mismatched work-
ers earn 6% less than well-matched workers. Table 12 shows similar results for women,
columns (1)–(3).
To sum up, my results appear to be robust. Moreover, the rich set of covariates I
consider in the WLS (education, IQ at high school, high school rank, cognition score,
preferences, tenure, occupation type and size of firm) helps me to control to some extent
forbothworkers’and job’sheterogeneity.Nonetheless, it shouldbenoted that theabsence
of comparable longitudinal information on job routinization and workers’preferences as
wellas theabsenceofanyvalid instrumentspreventsmefromarguingthat theassociations
I document are causal.
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
226 Bulletin
TABLE 11
Mismatch and compensating wage differentials. Quantile median estimates for men
(1) (2) (3) (4) (5) (6)
Job routinization −0.108 −0.088 −0.049 −0.044 −0.037 −0.016
(0.020) (0.021) (0.029) (0.021) (0.020) (0.027)
Routine-preferring worker – −0.074 −0.107 – −0.068 −0.087
(0.025) (0.030) (0.020) (0.026)
Mismatch – – −0.060 – – −0.029
(0.030) (0.025)
Completed years of education 0.053 0.052 0.052 0.034 0.034 0.035
(0.005) (0.005) (0.005) (0.005) (0.005) (0.005)
IQ measured at high school 0.003 0.003 0.002 0.003 0.003 0.003
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
High school rank 0.000 0.000 0.000 −0.000 −0.000 −0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Adult cognition score 0.007 0.005 0.006 0.001 0.003 0.003
(0.004) (0.004) (0.004) (0.004) (0.004) (0.003)
Tenure 0.008 0.008 0.008 0.006 0.006 0.006
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
Occupation dummy variables? No No No Yes Yes Yes
Firm size dummy variables? No No No Yes Yes Yes
Pseudo R2 0.16 0.17 0.17 0.22 0.22 0.22
Number of observations 1,523 1,523 1,523 1,520 1,520 1,520
Notes: Dependent variable is log(hourly wage). Job routinization equals to 1 if fraction of weekly worked hours
doing the same things over and over is equal to or higher than 0.5. Bootstrapped standard errors (1,000 replications)
are reported in parentheses.
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
Compensating wage differentials for job routinization 227
TABLE 12
Mismatch and compensating wage differentials. Quantile median estimates for women
(1) (2) (3) (4) (5) (6)
Job routinization −0.100 −0.089 −0.056 −0.070 −0.069 −0.037
(0.024) (0.022) (0.027) (0.022) (0.021) (0.024)
Routine-preferring worker – −0.134 −0.157 – −0.074 −0.111
(0.021) (0.027) (0.022) (0.025)
Mismatch – – −0.060 – – −0.052
(0.026) (0.022)
Completed years of education 0.064 0.063 0.063 0.032 0.032 0.030
(0.007) (0.007) (0.007) (0.007) (0.007) (0.007)
IQ measured at high school 0.006 0.005 0.006 0.004 0.004 0.004
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
High school rank 0.000 0.000 0.000 −0.000 −0.000 −0.000
(0.001) (0.001) (0.000) (0.001) (0.001) (0.001)
Adult cognition score 0.001 −0.003 −0.002 −0.001 −0.003 −0.002
(0.005) (0.006) (0.006) (0.004) (0.004) (0.004)
Tenure 0.016 0.017 0.017 0.013 0.014 0.014
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
Occupation dummy variables? No No No Yes Yes Yes
Firm size dummy variables? No No No Yes Yes Yes
Pseudo R2 0.20 0.21 0.21 0.30 0.30 0.30
Number of observations 1,621 1,621 1,621 1,612 1,612 1,612
Notes: Dependent variable is log(hourly wage). Job routinization equals to 1 if fraction of weekly worked hours
doing the same things over and over is equal to or higher than 0.5. Bootstrapped standard errors (1,000 replications)
are reported in parentheses.
VII. Conclusions
In this paper, my goal has been to argue that previous estimates of compensating wage
differentials are inconclusive because they do not account for the discrepancy between
workers’preferences and job attributes. Both casual empiricism and research results sug-
gest that this discrepancy indeed exists. In my sample, 38% of the men and 47% of the
women appear to be mismatched.
I propose a simple assignment model with Nash bargaining over wages for analysing
the role of mismatch when looking for compensating wage differentials. Assuming that
observedworkersarenot in long-runmarketequilibrium,allworkers,nomatterwhat their
preferences are, need to be compensated if working in the sector with a shortage of work-
ers in the absence of pay differentials. However, only mismatched workers, who are less
productivebecause their sectorsdonotmatch theirpreferences,arepenalized. Ifmismatch
isnot accounted, then theassociationbetweenwagesand jobattributesmaybepickingup
the correlation between job attributes, preferences and mismatch.
My empirical analysis uses the WLS and focuses on job routinization (the fraction of
working time spent doing the same things over and over). I report several findings. First,
mismatch is negatively related to wages, which is consistent with the negative mismatch
productivityeffectdominating thepositivecompensatingwagedifferential effect.Second,
for both men and women, I find that the negative relationship between wages and job
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
228 Bulletin
routinization is attenuatedoncemismatchandworkers’preferencesareaccounted for.The
evident mismatch effect can explain a substantial portion (but not all) of the incorrectly-
signed compensating wage differential for job routinization that previous analyses have
indicated.
In my view, this paper highlights the importance of accounting for mismatch when
looking for compensating wage differentials. Clearly, much more work needs to be done
on the theoretical front, for instance,byendogenizingmismatch.Nevertheless, I anticipate
that as long as there are search frictions that ensure that some workers remain in jobs that
arenotoptimalgiventheexistingwagerates, theresultsof theassignmentmodelpresented
here will generalize to a market setting. Given the substantial mismatch I find in the data,
these sorts of frictions seem realistic.
Final Manuscript Received: July 2010
References
Andrew, M., Bornstein, S., Carayon, P., Carr, D., Choi, H., Delamater, J., Doescher, H., Flynn, K., Frederick,
C., Fischer, D., Freese, J., Grol-Prokopczyk, H., Hauser, R., Hauser, T., Huang, R., Ho, J., Hoonakker, P.,
Khodyakov,D.,Merrill,D.,Piefer,L.,Piliavin, J., Pudrovska,T.,Rainwater,E.,Raymo, J.,Roan,C.,Ruel,
E.,Shinberg,D.,Sicinski,K.,Siegl,E.,Springer,K.,Warren,J.,Wicke,E.,Wright,A.andYonker, J. (2006).
‘The class of 1957 in their mid-60s: a first look’, in Hauser R. and Roan C. (eds), CDE Working Paper No.
2006-03, Center for Demography and Ecology, University ofWisconsin-Madison.
Borghans, L., Weel, B. and Weinberg, B. (2006). People People: Social Capital and the Labor-market
Outcomes of Underrepresented Groups, NBERWorking Paper No. 11985.
Borghans,L.,Weel,B. andWeinberg,B. (2008). ‘Interpersonal styles and labormarket outcomes’,Journal of
Human Resources,Vol. 43, pp. 815–858.
Borjas, G. (2005). Labor Economics, McGraw-Hill Irwin, NewYork.
Brown, C. (1980). ‘Equalizing differences in the labor market’, Quarterly Journal of Economics, Vol. 94,
pp. 113–134.
Daniel,C.andSofer,C.(1998).‘Bargaining,compensatingwagedifferentials,anddualismofthelabormarket:
theory and evidence for France’, Journal of Labor Economics,Vol. 16, pp. 546–575.
Duncan,G. andHolmlund,B. (1983). ‘WasAdamSmith right after all?Another test of the theoryof compen-
sating wage differentials’, Journal of Labor Economics,Vol. 1, pp. 366–379.
Garen,J. (1988). ‘Compensatingwagedifferentialsandtheendogeneityof jobriskiness’,ReviewofEconomics
and Statistics,Vol. 70, pp. 9–16.
Hwang,H.,Reed,W.andHubbard,C.(1992).‘Compensatingwagedifferentialsandunobservedproductivity’,
Journal of Political Economy,Vol. 100, pp. 835–858.
Kostiuk, P. (1990). ‘Compensating differentials for shift work’, Journal of Political Economy, Vol. 98,
pp. 1054–1075.
Lang, K. and Majumdar, S. (2004). ‘The pricing of job market characteristics when markets do not clear’,
International Economic Review,Vol. 45, pp. 1111–1128.
Lucas,R.(1977). ‘Hedonicwageequationsandpsychicwagesinthereturnstoschooling’,AmericanEconomic
Review,Vol. 67, pp. 549–558.
McNabb,R. (1989). ‘Compensatingwagedifferentials: someevidence forBritain’,OxfordEconomicPapers,
Vol. 41, pp. 327–338.
Rosen, S. (1974). ‘Hedonic prices and implicit markets’, Journal of Political Economy,Vol. 82, pp. 34–55.
Rosen, S. (1986). ‘The theory of equalizing differences’, in Ashenfelter O. and Layard R. (eds), Handbook
of Labor Economics,Vol. 1, Elsevier Science,Amsterdam,The Netherlands.
Shimer, R. (2007). ‘Mismatch’, American Economic Review,Vol. 97, pp. 1074–1101.
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
Compensating wage differentials for job routinization 229
Smith, R. (1979). ‘Compensating wage differentials and public policy: a review’, Industrial and Labor Re-
lations Review,Vol. 32, pp. 339–352.
Tinbergen, J. (1975). Income Distribution: Analysis and Policies, North Holland Co, Amsterdam, The
Netherlands.
Viscusi,K. andAldy, J. (2003). ‘Thevalueofa statistical life: a critical reviewofmarket estimates throughout
the world’, Journal of Risk and Uncertainty,Vol. 27, pp. 5–76.
© Blackwell Publishing Ltd and the Department of Economics, University of Oxford 2010
Copyright of Oxford Bulletin of Economics & Statistics is the property of Wiley-Blackwell and its content may
not be copied or emailed to multiple sites or posted to a listserv without the copyright holder’s express written
permission. However, users may print, download, or email articles for individual use.
WORKER LEARNING AND
COMPENSATING DIFFERENTIALS
W. KIP VISCUSI and MICHAEL J. MOORE*
The authors hypothesize that in industries with relatively high levels
of job-related injury risk, workers with longer job tenure will more
clearly appreciate the degree of job risk than will newly hired workers,
and will thus be more willing to accept lower wages in return for higher
workers’ compensation benefits. This hypothesis is confirmed by an
analysis of quit behavior using 1981-83 data from the Michigan Panel
Study of Income Dynamics and 1981-85 data from the National
Institute of Occupational Safety and Health.
IN the standard compensating wagedifferential model, workers value their
wage and workers’ compensation compo-
nents hased on full job risk information.
Market forces generate positive wage
differentials as ex ante compensation f
or
exposure to relatively high risk. Similarly,
market forces generate wage offsets for
the increases in ex post risk compensation
embodied in workers’ compensation bene-
fits.
These predictions can be modified to
take into account potential imperfections
in worker information, as in Viscusi
(1979a,b, 19
80
a,b,d), where the role of
learning is incorporated into the worker’s
decision model. The potential for learning
about risks introduces a new market
* W. Kip Viscusi is George G. Allen Professor of
Economics, Department of Economics, and Michael
J. Moore is Associate Professor of Economics, Fuqu
a
School of Business, Duke University. The first
author’s research was supported by the endowment
of the George G. Allen chair at Duke University, and
the second author’s research received partial support
from the Business Associates Fund at the Fuqua
School of Business.
The data and programs used in this study are
available on request to Michael J. Moore, Fuqua
School of Business, Duke University, Durham, NC
27706.
response through worker quitting after
the acquisition of adverse risk informa-
tion. In a full information world, after
controlling for health status, no unex-
pected job risk-quit relationship will be
observed. In the more realistic sequential
decision model in which there is an
opportunity for learning, the acquisition
of adverse new information by the worker
on the job may lead the worker to quit.
With the exception of the experimental
results reported in Viscusi and O’Connor
(1984), in which worker responses to
alternative chemical labels were moni-
tored, tests of the standard compensating
differential model and of the learning
models have been distinct, as each focuses
on a different aspect of labor market
behavior. The empirical evidence support-
ing compensating risk differentials is
substantial: greater job risks boost worker
wages, and workers are willing to accept a
wage cut in return for higher workers’
compensation benefits.’ These results are
‘ See, for example. Smith (1979) and Viscusi
(1979a) for analysis of wage-risk tradeoffs. Estimates
of wage—workers’ compensation tradeoffs appear in
Arnould and Nichols (1983), Butler (1983), Dorsey
and Walzer (1983), and Viscusi and Moore (1987). In
Industrial and Labor Relations Review, Vol. 45, No. 1 (October 1991). © by Cornell University.
0019-7939/91/4501 $01.00
80
JOB TENURE AND WORKERS’ COMPENSATION 81
the main predictions of the standard
compensating differential theory, and
they continue to hold if learning is
introduced. Market tests of the role of
worker learning, on the other hand, have
focused on two other empirical issues—the
effect of injury experiences on workers’
risk perceptions and the positive effect of
job risks on worker quitting.^
The focus of this paper is broader than
that of separate analyses of the wage and
quit effects of job risks because we use the
relationships typically estimated and tested
in the standard compensating differential
theory to examine the job risk-learning
model as well. In particular, using a large
data set on workers in the early 1980s, we
evaluate the tradeoffs between wages and
workers’ compensation benefits and be-
tween wages and risks implied by worker
quit behavior, and compare these tradeoffs
across worker tenure groups.
Younger workers will assess job risk less
precisely than more senior workers, since
their inf”ormational base for making these
judgments is smaller. In multi-period
models that incorporate learning and
experimentation with risky jobs, workers’
reservation wage rates will be less for jobs
posing less precisely understood risks, for
any given mean level of risk. The empiri-
cal prediction is that more senior workers
will demand greater compensation for risk
because of their more precise judgments.
In addition, workers who are on the quit
margin (those who would need only a
small inducement to quit their jobs) will
have greater subjective risk perceptions
than other workers. These workers conse-
‘ quently will demand greater wage com-
pensation for higher risk levels, since they
will be comprised disproportionately of
workers whose risk beliefs have been
adversely affected by on-the-job experi-
ences. These workers will also assess a
greater chance of receiving workers’ com-
pensation benefits, so their expected value
Moore and Viscusi (1990) we provide a literature
review of estimates of the effects of fatality risks and
workers’ compensation on wages.
2 See Viscusi (1979a,b, 1980a,b,d, and 1983) and
Viscusi and O’Connor (1984).
of higher benefits will be greater. We
therefore expect to observe greater wage-
risk and wage-workers’ compensation
tradeoffs for these workers.
Theoretical Eramework
The learning model that we apply here
to the wage-workers’ compensation trade-
off was first introduced in Viscusi (1979a,b,
1980a,b,d).3 Our overall objective is to ex-
plore the relationships between the wage-
risk tradeoff and the wage-workers’ com-
pensation tradeoff for new workers and for
senior workers on the quit-no quit margin.
The essential ideas can be captured in a
two-period model. Let there be two health
states: healthy and injured. In the good
health state the worker receives a wage
rate w, from which he derives utility U^{w).
In the injured state the worker receives
workers’ compensation benefits b, where w
> b. This assumption refiects the structure
of workers’ compensation programs in
virtually every state, since benefits are
typically two-thirds of the wage or less,
except for workers with very low wages.
We assume that the worker would rather
be healthy than not (that is, f/'(x) > U^{x)),
has a higher marginal utility of income
when healthy (Ul > U^), and is either
risk-averse or risk-neutral (t/^, U%; ^ 0).
The worker values payoffs over time using
a discount factor p that equals the recipro-
cal of 1 plus the interest rate.
Suppose that there are two possible jobs,
a risky job and a safe job. We can assume
with no loss of generality that the safe job
poses no risk of injury.* The safe job
offers a payoff WQ forever. The risky job
offers the worker an initial perceived
probability of not being injured equal to p
and a 1-;!? chance of suffering an injury
that lasts a single period. If the worker is
not injured in period 1, he revises his
‘ A variant of this analysis without learning
appears in Diamond (1977).
One could adopt the assumption that the
alternative job poses a known risk of injury without
altering the model structure, even in the n-period
case. If both jobs are uncertain and there are more
than two periods, the model structure becomes more
complex. See Viscusi (1979a) for these extensions.
82
INDUSTRIAL AND LABOR RELATIONS REVIEW
assessed probability of not being injured
upward to p’^. If he is injured, the
assessed probability is p~. It is also
possible, as noted by Viscusi (1979a), that
workers revise their expectations based on
observations of other workers’ injury
experiences. The revision of workers’
probabilistic beliefs follows a standard
Bayesian learning process, where
(3) = 0 = pU\w) + i l –
(1) >p>p-
The workers’ initial job decision involves
a choice between two periods of work on
the safe job or initial work on the risky job,
after which he can quit if he is injured in
period I. As first noted by Viscusi (1979a),
the worker’s problem mirrors the classic
two-armed bandit problem, which de-
scribes the optimal sequence of plays on
two slot machines. On one machine the
probability of success is known, and on the
other it is uncertain. Thus, a payoff on the
uncertain machine yields information in
addition to monetary rewards. Eor this
class of two-armed bandit problems, it is
shown in Viscusi (1979a,b, 1980a,b,d) that
the stay-on-a-winner rule is always opti-
mal. The worker will not leave the risky
job after a favorable experience in period
I. The worker also will not leave the safe
job once he starts on it.
The wage package for the marginal new
worker attracted to the firm must satisfy
the condition that expected lifetime utility,
V, is equal between the two jobs, given the
opportunity the worker has to switch from
the risky job following an unfavorable
period 1 outcome:
(2) V =
= pU\w) + {\-p)U\b)
p-U\w) + {\-p-)U\b)].
If we set U^{WQ) equal to zero, with no loss
of generality, we have
Not all workers will quit their jobs prior
to period 2 after an unfavorable period 1
job experience. The focus here, however, is
on the wage-benefit tradeoff of the mar-
ginal senior worker relative to the pay pack-
age that will attract the worker to the job
initially. The marginal senior worker will
quit after an adverse experience in period
1, so the wage package {•w,b) sufficient to
attract the worker initially must satisfy
(4) y = 0 = pu\w) + (1 – ^
since the last term in equation 5 equals zero
after an unfavorable period 1 experience.
The first issue analyzed is the wage-
workers’ compensation tradeoff that will
be reflected in the {w,b) package for new
hires. Implicit differentiation of equation
3 yields
(5)
dw _ –
Bb V Ul[p +
The value of dwidb represents the wage
offset in response to b for the new hire in a
two-period job choice problem. The initial
wage package {w,b) will be adequate to
retain the worker if his on-the-job experi-
ences are favorable.
The worker on the margin at the start of
period 2 has an expected utility Z equal to
(6) Z = 0 = p-U\w) + {\-p-)U\b),
since he is indifferent between leaving
(where U^{WQ) = 0) and staying on the
risky job. The wage-workers’ compensa-
tion tradeoff for this worker equals
(7) Tr =
One issue that we investigate empirically
JOB TENURE AND WORKERS’ COMPENSATION 83
is the relative magnitude of the wage-
workers’ compensation tradeoffs in equa-
tions 5 and 7. Workers who have experi-
enced or observed an on-the-job injury
should value workers’ compensation more
highly, since they will assess a higher
probability of receiving such benefits
than will other workers. For workers with
an adverse job experience, the expected
amount of workers’ compensation bene-
fits will be {l-p’*’)b. For new hires who
plan to quit if their initial period job
experience is unfavorable, the discounted
expected benefit amount is {l-p)b +
^{l -p){\ -p”)b, where these benefits are
provided over (l-p) + ^{l-p) periods.
The lower expected amount of benefits
per period of work for new hires is
reflected in a lower expected utility as
well, which will influence the compensat-
ing differential they are willing to accept
for these benefits. In particular, one
would expect
Quit New
Margin Hire
Tradeoff Tradeoff
(11)
U\w) =
-{\-p)U\b) –
db Z,.
or
(9)
ul[p
After some algebraic manipulation, equa-
tion 9 reduces to
(10) P’
Given the restrictions on probabilities
outlined in equation 1 above, equation 10
always holds.
The second empirical concern will be
the effect of worker learning on the
character of the wage premiums com-
manded by job risks. From the develop-
ment above, equation 4 implies that, for
new hires.
Similarly, solving for U^{iu) from equation
6 for workers on the quit margin yields
the condition that
(12)
The utility metric defined by setting
U^{WQ) equal to zero leads to a negative
value of U^{b), implying that the right
sides of equations 11 and 12 are positive.
The condition that must be met for the
risk premium required by senior workers
on the margin to be higher than that of
new hires is that C/*(w) be greater for this
group. Thus, the right side of equation 12
must exceed the expression on the right
side of equation 11. This requirement is
always satisfied if inequality 1() holds, as is
assumed.
In addition to compensation for risk, the
overall wage structure of the firm will, of
course, also include returns to worker ex-
perience and performance. This wage struc-
ture can be viewed as defining the pecuni-
ary returns to the worker over time. As is
standard in agency theory models of wages,
it is the new hires and other workers on the
margin of leaving the firm who are of great-
est concern. By altering the entering wage
level, the firm ensures a flow of new work-
ers to the firm. Higher wages also diminish
the tendency to quit, but since quitters tend
to be workers with particularly adverse job
experiences, learning-induced quits will
continue to occur. Because of the differ-
ence in risk perceptions of the potential
quitters, both safety and workers’ compen-
sation should be more highly valued by this
group than by the new worker group, which
has a lower assessment of the job risk both
for the initial period of work and for their
expected duration of work at the firm.
The principal predictions that we ex-
plore below stem from the effect of worker
experiences on risk perceptions. For senior
workers on the margin, the wage—workers’
compensation tradeoff should be more neg-
84 INDUSTRIAL AND LABOR RELATIONS REVIEW
ative (and less than zero), as indicated by
equation 8. This relationship arises be-
cause the higher probability these workers
attach to receiving benefits as a result of
their on-the-job learning about the risks
should be reflected in the preferences cap-
tured in quit equations. In addition, the
compensating differential experienced
workers demand for risk will also rise for
any given level of perceived risk, to the
extent that more experienced workers place
less value on future benefits of job exper-
imentation because of their more precise
evaluations of job risk. Experienced work-
ers on the quit margin will also consist dis-
proportionately of workers who have had
or observed adverse job experiences, boost-
ing the required wage-risk tradeoff.
The linkage of the theoretical predic-
tions to the empirical model will be fairly
direct. Wage and workers’ compensation
levels are directly observable. We do not
have direct measures of workers’ subjec-
tive probability of being injured for the
data set that we will use below, but we do
have relatively good data on the fatality
rate for the worker’s industry. The under-
lying assumption is that workers will have
higher risk perceptions and will be more
likely to acquire adverse job information
that will generate quit behavior in indus-
tries with high objective measures of risk.
The results in Viscusi (1979a) and Viscusi
and O’Connor (1984) indicate that the job
risk-quit relationship is similar whether
one uses objective industry risk variables
or subjective risk assessments. We use the
objective measure here because our focus
on state differences in workers’ compensa-
tion requires that we use a large national
data set for which there are no subjective
risk data.
Sample and
Empirical Results
Data
The main requirement with respect to
the survey data for the study is that they
include information on wage rates, quit
behavior, the worker’s state of residence
(to establish matchups to workers’ com-
pensation benefit formulas), and the
worker’s state of residence and industry
(to establish matchups with risk data).^
The principal data source we use is the
1981, 1982, and 1983 waves of the
University of Michigan Panel Study of
Income Dynamics (PSID). The PSID is a
broad longitudinal survey data set that
contains information pertaining to the
characteristics of individuals and their
jobs. These data are matched to informa-
tion on the risk of an on-the-job fatality
provided by the National Institute for
Occupational Safety and Health (NIOSH)
as part of its National Traumatic Occupa-
tional Fatality Project (NTOF). Unlike
comparable risk data available from other
sources, the NIOSH data represent a
census of all occupational fatalities, aver-
aged over the years 1981-85. As such,
these data are not subject to the sampling
error that is present in risk survey data.
Furthermore, the NIOSH data vary by
both state and industry, making them
more comparable to the workers’ compen-
sation benefit data, which vary primarily
by state. This feature provides a better
matchup than the matchup possible using
other available national survey data, which
vary only by industry. The NIOSH data
yield over 400 distinct observations of job
risk, thus providing one of the most
detailed breakdowns of injury risk cur-
rently available.^
The workers’ compensation data, which
are described in detail in Viscusi and
Moore (1987), are based on benefits for
temporary total disabilities, the injury
category under which approximately 65%
of total claims fall. Furthermore, in recent
years benefit ceilings for temporary total
^ An exact matchup of benefits and risk levels with
the worker’s state of employment, although desir-
able, is not possible with our data, since sample
members report only their state of residence. For the
majority of cases, however, the state of residence and
the state of employment are the same,
** The properties of the NIOSH data are explored
in relation to the BLS data in Moore and Viscusi
(1988, 1990),
Since the model is based on perceived changes in
actual risk, the state-industry risk data will not allow
us to distinguish among changes in perceptions,
actual risk changes, and changes in risk preferences.
We assume that the latter two change slowly, if at all.
JOB TENURE AND WORKERS’ COMPENSATION 85
disabilities and for permanent total dis-
abilities have become equal in practically
all states.•̂ Fatality benefits also usually
equal benefits for the above two catego-
ries, although there are some exceptions.
This standardization of benefit ceilings
allows a more representative measure of
ex post accident compensation than was
available in most earlier studies. On the
other hand, it ignores some other impor-
tant aspects of benefit structure, such as
differences between ceilings for partial
and total disability, waiting periods, and
duration. These differences, although im-
portant, cannot be captured in a refined
fashion, and researchers typically resort to
using temporary total disability ceilings or
payments as proxies for each state’s
benefit level.
The workers’ compensation benefit lev-
els are matched to workers in the PSID by
state. They are then used in conjunction
with information on the worker’s weekly
wage, marital status, and family size to
determine the weekly benefit for wbich
tbe worker qualifies. The benefit variable
is computed using the formula b = (2/3
weekly wage) x (I – D) + (benefit
maximum) X D, where D = 1 if the
worker qualifies for the maximum benefit
level, and 0 otherwise. Tbe variable b is
tben divided by the worker’s after-tax
weekly wage to construct tbe wage re-
placement rate measure tbat is used in tbe
empirical analysis.^
For purposes of estimation, workers
wbo report tbeir occupation as farming
are excluded from tbe sample, since tbe
agricultural sector is excluded from tbe
NIOSH data. Also excluded are workers
wbose reported hourly wage is below tbe
statutory minimum wage, non-beads of
‘ A detailed exploration of the differences in
disability benefits and their interrelationship is
provided by Burton and Krueger (1986) and
Krueger and Burton (1983), A detailed analysis of
the important permanent partial disability compo-
nent is provided by Burton (1983), More generally,
see Berkowitz and Burton (1987) for an analysis of
permanent disability,
^ We assume that benefits are computed based on
full-time weekly earnings. This assumption reflects
the law in many states, and the actual work week for
the majority of the sample.
bousebolds, blacks, workers wbo are over
65 years old or are not in tbe labor force,
and cases witb missing data. Tbe sample
tbat remains consists of 2,571 observations
on 857 workers. The mix of tbe workers
in tbe sample follows tbe expected pat-
terns. Fifteen percent of tbe sample
members are women (FEMALE), 11% are
single (SINGLE), and tbe mean number of
dependent cbildren is close to 1 (DEPEN-
DENTS).
Tbe human capital variables include tbe
standard measures. Tbe workers have an
average of about 13 years of schooling
(EDUCATION), 10 years of experience at
tbeir current firm (TENURE), and 18 years
of job experience overall (EXPERIENCE).
Because of tbe interrelationsbips among
tbe various buman capital variables, a
worker age measure is not included in tbe
analysis. Tbe sample consists primarily of
workers in industries in wbicb job hazards
are likely to be of consequence. In
particular, tbe sample restrictions we bave
imposed yield a sample tbat is over 30%
unionized (UNION).
Tbe average after-tax real bourly wage
equals approximately $7 in 1982 dollars.
Tbe tax component of wages was calcu-
lated using information on marginal tax
rates provided by tbe PSID sample mem-
bers. Although most wage equation stud-
ies utilize tbe pre-tax wage for simplicity,
tbe inclusion of workers’ compensation
benefits in tbe equation increases tbe
importance of using after-tax wages, since
tbese benefits bave favorable tax status.
Moreover, tbe extent of one’s tax savings
varies with one’s tax bracket and state tax
level. Failure to make an adjustment for
taxes would tbus distort tbe tradeoff rate
between wages and workers’ compensa-
tion, wbicb is a central empirical concern
in tbis paper.
Tbe deatb risk measure (RISK), wbicb was
described earlier, implies an average deatb
risk of 7.6/100,000 for sample members.
Tbis risk level is only sligbtly different from
tbe national average NTOF risk measure
of 9/100,000, so tbe sample is representa-
tive of tbe industry mix captured in tbe
NTOF data set. Tbe rougbly 1/10,000 deatb
risk level sbould be viewed as a typical risk
86 INDUSTRIAL AND LABOR RELATIONS REVIEW
sample rather than a high-risk sample. We
will use the death risk variable as a proxy
for the overall job risk, since this measure
is available on a state-specific basis, whereas
published nonfatal injury data are not
readily available.^
The basic workers’s compensation vari-
able in Table 1 is the real dollar value of
the state weekly workers’ compensation
maximum benefit, WCMAX, which averages
$207. Benefits are computed using infor-
mation on the worker’s state of residence,
martial status, number of dependents, and
the formula given above. We also compute
a benefit variable following Moore and Vis-
cusi (1989), using the maximum benefit
level in each state as a proxy for expected
benefits. Since changes in the maximum
influence the distribution of benefits for
which each worker can potentially qualify,
this variable measures changes in expected
benefits when there is wage rate uncer-
tainty. To control for the fact that in-
creases in the maximum are more highly
valued by workers whose wage places them
at or above the maximum, we estimate the
effect of this variable separately for each
class of worker. The relative sizes of the
estimated effects provide a check on the
plausibility of our results. These results also
provide a check on the robustness of the
results derived using the actual benefit level
in the replacement rate.
As shown in Viscusi and Moore (1987),
workers only value accident insurance at
positive risk levels. Thus, the appropriate
measure of workers’ compensation, the
weighted weekly benefit level, involves an
interaction of the death risk and the
benefit level. Insurance benefits are cap-
tured by the risk-weighted replacement
rate. The benefit level for which the
worker qualifies is interacted with the risk
variable to create the weighted benefit
measure and then divided by the worker’s
after-tax weekly wage to create the
weighted replacement rate variable. This
^ Published single-digit (SIC) injury rates for fatal
and nonfatal injuries exhibit a correlation of about
,70 for 1986, significant at the ,10 level. See U,S,
Department of Commerce (1989), Tables 681 and
682,
formulation recognizes the fact that the
value of the benefit varies with the risk
and the fact that workers’ compensation
benefits are tax exempt. Use of the
weighted replacement rate variable is
consistent with much of the previous
research on the wage effects of workers’
compensation, although some studies have
entered wages and benefits separately.i°
As Table 1 indicates, the average re-
placement rate for the workers in each
tenure group is 0.70. This rate is slightly
higher than the nominal replacement rate
of 0.66 used by most states. Two compet-
ing influences lead to this divergence.
Since our replacement rate is computed
on an after-tax basis, it will tend to be
higher than the before-tax nominal rate of
0.66. On the other hand, since many
workers’ wages put them above the maxi-
mum, the observed replacement rate will
tend to be lower. The net effect is to yield
a rate slightly above the state-mandated
nominal rate.
Empirical Results
The empirical tests of the worker
learning hypothesis compare the wage-
benefit and wage-risk tradeoffs estimated
in quit equations across two tenure
groups—workers with at least three years
of tenure, and those with less than three
years. We focus on the quit equation
rather than a wage equation because the
primary matter of interest is how the
wage-risk and wage-benefit preferences of
different groups of workers are altered,
not how market contracts respond. The
cutoff point at three years of tenure is the
division used in Kahn’s (1987) analysis of
the preferences of marginal workers.
Furthermore, restricting the newly hired
worker sample to two years of tenure or
less, as in Viscusi (1980c), yielded very
small samples.
Estimates of the parameters of the quit
equations for the senior worker group
provide information on the behavior of
senior workers on the margin, whereas
estimates of the parameters of the quit
10 See Worral and Butler (1985),
J O B T E N U R E AND WORKERS’ COMPENSATION 87
Table 1. Variable Definitions and Sample Characteristics,
Variable
Name
Mean (Standard Deviation)
Variable Definition
EDUCATION Y e a r s of e d u c a t i o n
FEMALE 1 if w o r k e r is f e m a l e , 0 o t h e r w i s e
HEALTH 1 if w o r k e r r e p o r t s t h e p r e s e n c e of a h e a l t h i m p a i r m e n t
LIMITATION t h a t limits t h e a m o u n t of work he can d o , 0
otherwise
DEPENDENTS N u m b e r of d e p e n d e n t c h i l d r e n
SINGLE 1 if w o r k e r has n e v e r b e e n m a r r i e d , 0 otherwise
EXPERIENCE Years w o r k e d for pay since a g e 16
UNION STATUS 1 if worker’s j o b is covered by a collective b a r g a i n i n g
a g r e e m e n t , 0 otherwise
WAGE W o r k e r ‘ s a f t e r – t a x h o u r l y w a g e in 1982 d o l l a r s
( G N P d e f l a t o r )
RISK N T O F fatality r a t e v a r i a b l e : n u m b e r of fatalities p e r
1 0 0 , 0 0 0 w o r k e r s , by s t a t e a n d o n e – d i g i t (SIC) i n d u s t r y
WCMAX M a x i m u m b e n e f i t level f o r t e m p o r a r y disability in t h e
w o r k e r ‘ s s t a t e
REPLACEMENT P o r t i o n o f w e e k l y a f t e r – t a x w a g e r e p l a c e d by w o r k e r s ‘
RATE compensation
d 1 if 2/3 of worker’s weekly wage exceeds WCMAX, 0
otherwise
STRONG QUIT 1 if Worker is looking very h a r d for a new j o b ,
INTENTIONS 0 OTHERWISE
WEAK QUIT 1 if w o r k e r is looking at least s o m e w h a t h a r d for a
INTENTIONS new j o b , 0 otherwise
ACTUAL QUITS 1 if w o r k e r c h a n g e d j o b s in t h e past year, 0 otherwise
SAMPLE SIZE
Tenure s 3
12,93
(2,61) .
0,11
(0,31)
0,07
(0,26)
1,05
(1,13)
0,09
(0,28)
21,17
(12,15)
0,34
(0,48)
7,11
(2,33)
7,54
(9,81)
210,81
(69,89)
0,70
(0,20)
0,74
(0,44)
0,08
(0,27)
0,14
(0,34)
0,03
2007
Tenure < 3
13,23
(2,35)
0,21
(0,41)
0,05
(0,22)
0,97
(1,09)
0,20
(0,40)
13,61
(10,08)
0,23
(0,42)
6,31
(2,45)
8,43
(9,50)
195,35
(65,50)
0,70
(0,18)
0,66
(0,47)
0,15
(0,36)
0,22
(0,42)
0,10
564
equations for the junior worker group
provide information on the preferences of
the newly hired worker. Our principal
hypothesis is that the wage-benefit trade-
offs for the newly hired worker group will
be less negative than those of the marginal
senior worker group, which will include
many workers who have acquired unfa-
vorable information about risks on the job.
We expect the risk effect to be larger in
the senior worker group for the same
reason. The self-selection of workers with
more extensive learning and more precise
risk perceptions into the junior worker
group will tend to work against our
principal hypotheses.
The quit equations are estimated using
three measures of quit behavior. Table 1
defines these variables. If a worker an-
swers “yes” to a question asking whether
he is considering looking for a new job,
the weak quit intention variable equals
one; if “no,” it equals zero. A similarly
constructed measure of strong quit inten-
tions equals one if the worker reports that
he is seriously considering a new job and
zero otherwise. Finally, the actual quit
variable equals one if the worker quit
during the year and zero otherwise. This
variable pertains only to the 1981 and
1982 data. The quit variables reflect
aggregate quit behavior fairly closely, as
the average quit rate in manufacturing
industries equaled about 1.5% per month
in the late 1970s.•’ The average value of
the weak quit intention variable of 14%, or
‘ See U,S, Department of Labor (1977),
INDUSTRIAL AND LABOR RELATIONS REVIEW
1.2% per month, roughly equals the
observed rate. The actual quit rate in our
sample, 4.5% per year, is lower than the
aggregate manufacturing rate, as ex-
pected, given the broader mix of indus-
tries represented in our sample.
The quit equations estimated are of the
form
(13) Quit, = [1 + exp – (a’X,
-I- 8^Weighted
Replacement Rate,)]
Due to the binary nature of the depen-
dent variable in these equations and the pres-
ence of the endogenous wage and replace-
ment rate variables on the right-hand side
of equation 13, nonlinear two stage least
squares is used to estimate the parameters
of the model.’2 As shown by Amemiya
(1985), these estimates are both consistent
and asymptotically normal. Instrumental
variables include all of the exogenous ex-
planatory variables in the quit equations and
state dummy variables.’^
Higher worker wages should reduce
quitting by increasing the attractiveness of
the worker’s current job. Quit rates should
increase with risk levels if there are
learning-induced quits, and higher work-
ers’ compensation benefits should dimin-
ish quitting. The coefficients of interest
are 4>q> 7o. and 8̂ , which we will use to
calculate the wage—workers’ compensation
‘^ The SAS procedure SYSNLIN, with the Gauss-
Newton Minimization method, is used to estimate the
model.
‘^ We experimented with the use of age, experi-
ence, and tenure variables as instruments. Since
these variables are important predictors of the wage,
they would serve as useful instruments if they were
independent of the error term €^. When experience
and tenure variables are added to he vector of
instruments, there is no change in point estimates of
the coefficients. The estimated standard errors are
larger when the age variable is used in the weak quit
equation. The main results, particularly those in the
actual quit equation, are unaffected. We report
results using the age, experience, and tenure
variables as instruments.
and wage-risk tradeoffs implied by work-
ers’ quit behavior.
Table 2 presents estimates of equation
13, using the actual quit variable, for
workers in each tenure group. The actual
quit variable provides a strong measure of
the job satisfaction of workers, since
workers are less likely to quit than to
merely seek a new job. This measure
should consequently reflect most strongly
the role of worker learning in affecting
the wage-risk and the wage-workers’
compensation tradeoffs.
The Table 2 results indicate that wages
and job risk characteristics are the most
important determinants of workers’ quit
behavior. For workers with more than
three years of tenure, increases in the
wage exert significant downward pressure
Table 2. Determinants of Quits: 857 Workers,
1981-1983.
(Standard Errors in Parentheses)
Independent
Variable”
EDUCATION
FEMALE
HEALTH LIMITATION
DEPENDENTS
SINGLE
EXPERIENCE
UNION STATUS
In (Weekly Wage)
RISK
WEIGHTED REPLACEMENT
RATE
CONSTANT
Tenure s 3
0.366**
(0.167)
-0.832
(0.795)
0.408
(0.641)
0.418**
(0.227)
0.534
(0.671)
-0.038*
(0.029)
-0.050
(0.568)
-4.111***
(1.430)
0.590***
(0.215)
-1.673***
(0.659)
18.029
(7.098)
Tenure £ 5
-0.525***
(0.211)
-0.549
(0.659)
-0.591
(1.632)
-0.306
(0.293)
0.398
(0.670)
-0.073*
(0.046)
-0.361
(0.557)
3.076**
(1.472)
-0.028
(0.144)
-0.141
(0.251)
-12.826
(6.472)
Sources: Michigan Panel Study of Income Dynam-
ics, 1981, 1982, 1983; National Institute for Occupa-
tional Safety and Health data for 1981-85.
^ Also included as explanatory variables are a
Southeast regional dummy variable, a city size
variable, and a year dummy variable.
* Statistically significant at the .10 level: ** at the
.05 level; *** at the .01 level (one-tailed tests).
JOB TENURE AND WORKERS’ COMPENSATION 89
on quits. Similarly, increases in workers’
compensation benefits (as captured by tbe
weighted replacement variable) decrease
quits by workers who have been on the job
for three years or more. Furthermore,
among senior workers, increases in the
risk level have a significant positive effect
on quit behavior, as on-the-job experience
makes workers more aware of adverse
working conditions and also increases the
precision of their estimates of the proba-
bility of a job-related injury or health
problem. This risk effect is the same as
that found by Viscusi (1979a). The wage
effect, too, mirrors the earnings effect in
Viscusi’s quit intention equation. The
additional effect shown by the workers’
compensation variable provides further
support for the model of rational worker
learning: as workers perceive their jobs to
be more dangerous, workers’ compensa-
tion benefits become more valuable to
them and serve to dampen the influence
of risk on turnover by mitigating the
financial losses associated with an accident.
The remaining variables in the quit
equation measure the effect of the work-
er’s characteristics on quits, holding con-
stant the wage, the job risk, and the
weighted replacement rate. In most cases,
the signs of these variables are theoreti-
cally indeterminate. Significant effects are
found, however, for education, number of
dependents, and work experience in the
more senior worker group.
The equation for workers with less than
three years of tenure indicates that
roughly the same control variables exert a
significant effect on worker quits as in the
senior group. In one case, however, the
effect is opposite that for the older group:
education increases quits by the older
workers but decreases quits by younger
workers. This result could reflect the net
effects of a number of influences. Educa-
tional attainment may affect one’s pros-
pects for external mobility in a manner
that varies with the extent of one’s
job-specific experience. Education may
also complement specific training, which is
accumulated in the early years of a job.
This complementarity would tend to re-
duce quits initially. The negative educa-
tion effect for junior workers could also
reflect firms’ greater commitment to pre-
serving the match for more educated
young workers. As workers acquire more
work history, the job-signaling informa-
tion content of the education becomes less
important to the employer. For senior
workers, greater education increases job
mobility, with little connection between a
worker’s education and the firm’s desire to
retain the worker.
Consistent with our theoretical predic-
tions, none of the job risk characteristics
variables are statistically significant in the
equation for junior workers, because their
perceptions of risks and, therefore, their
valuations of risk insurance are very
imprecise. Indeed, in the early stages of
the employment process, workers have
been shown by Viscusi (1979a) to show a
systematic preference for jobs with charac-
teristics that are only poorly understood.
This preference is reflected in the lack of
an effect of the risk variable and the
weighted replacement rate variable.
The wage variable exerts a significant
positive effect for the junior workers. This
unexpected result could be due to a
number of influences. The most obvious
explanation, that the causality between
quits and wages runs in the opposite
direction, not only would have to survive
the instrumentation but also would have
to provide an explanation of why reverse
causality only matters for junior workers.
Alternatively, wages could be acting as a
proxy for skills that junior workers are
offering to different employers. These
skills could be observable to both the
worker and the firm, unlike those in the
signaling context discussed above, but not
captured in our data. More able workers
and firms would be willing to invest more
in the job matching process. Furthermore,
in the job matching model of Mortensen
(1978), wages are not necessarily nega-
tively related to turnover. Rather, the
wage acts as a proxy for the match-specific
capital; and inclusion of these components
of the capital as regressors will eliminate
any wage effect. Finally, it could be the
case that the expected search costs are
lower for high-wage young workers than
90 INDUSTRIAL AND LABOR RELATIONS REVIEW
for those bound by minimum wages, who
face large queues for available jobs, thus
making the high-wage junior workers
more likely to quit.
For the actual quit equation, we thus
have reasonably precise estimates of the
parameters used to calculate the wage-
workers’ compensation tradeoff for work-
ers with more than three years of tenure.
The wage, risk, and workers’ compensa-
tion variables are all significant at the 1%
level or lower for these workers. As
expected, variation in benefits does not
cause any significant variation in quits for
new hires.
Further support for the learning model is
found by comparing the effects of RISK on
quit intentions and on actual quits across
tenure groups for alternative specifications.
Table 3 presents the risk and workers’ com-
pensation effects for all three measures of
quit behavior, as well as the effects of dif-
ferent measures of the wage variable (the
after-tax weekly wage is always used in the
replacement rate variable). In addition to
the actual quit variable, we also estimate
Table 3. Wage-Risk and Wage-Replacement Rate Tradeoffs: Summary of Coefficient Estimates.
(Standard Errors in Parentheses)
Dependent
Variable
Actual Quits
(i) ln (Weekly Wage)
TENURE S: 3
TENURE < 3
(ii) Weekly
Wage
TENURE a S
TENURE < 3
Weak Quit Intentions
(i) ln (Weekly Wage)
TENURE a 3 .
TENURE < 3
(ii) Weekly Wage
TENURE a 3
TENURE < 3
Strong Quit Intentions
(i) ln (Weekly Wage)
TENURE a 3
TENURE < 3 (ii) Weekly Wage TENURE a 3 TENURE < 3
Sources: See notes to Table 2.
‘ Estimates would not converge.
* Statistically significant at the .10 level;
Wage
-4.111***
(1.430)
3.076**
(1.472)
-0.012**
(0.005)
0.004**
(0.002)
-2.246***
(0.573)
0.458
(0.681)
-0.005**
(0.002)
0.04E-3
(1.58E-3)
-2.104**
(0.646)
3.077**
(1.472)
-0.005**
(0.002)
a
** at the .05 level;
Risk
0.590***
(0.215)
– 0 . 0 2 8
(0.144)
0.650***
(0.245)
0.003
(0.115)
0.077*
(0.051)
– 0 . 0 3 1
(0.057)
0.072*
(0.048)
– 0 . 0 0 9
(0.053)
0.085**
(0.052)
-0.028
(0.144)
0.087**
(0.050)
a
Weighted
Replacement Rate
-1.673***
(0.659)
-0.141
(0.251)
-1.819***
(0.690)
-0.130
(0.221)
-0.122*
(0.083)
0.042
(0.076)
– 0 . 1 1 5 *
(0.079)
0.014
(0.072)
– 0 . 1 1 1 * *
(0.087)
– 0 . 1 4 1
(0.251)
-0.112*
(0.083)
a
*** at the .01 level (one-tailed testsV
JOB TENURE AND WORKERS’ COMPENSATION 91
equation 13 using measures of weak and
strong quit intentions as dependent vari-
ables. As the results in Table 3 indicate, the
findings in Table 2 are quite robust over the
different specifications. Both higher wages
and increases in the weighted replacement
rate significantly reduce quits and quit in-
tentions for workers in the more senior ten-
ure groups. Eor workers with less than three
years’ tenure, actual quits are positively and
significantly related (as before) to wages; nei-
ther risk nor the weighted replacement rate
shows any systematic effects on quits or quit
tendencies for these workers.
To evaluate the robustness of our
results, particularly with respect to the
specifications used in some of” our previ-
ous work, we also estimated the quit
equations using the benefit measure of
Moore and Viscusi (1989), which analyzed
the effects of workers’ compensation on
job fatalities. As noted above, the numera-
tor of the weighted replacement rate
variable in this case uses the maximum
benefit payment for the worker’s state as
the numerator of the replacement rate.
To account for the fact that changes in the
benefit maximum will be more highly
valued by the workers whose wage exceeds
the maximum, we estimate 8̂ separately
for each worker group (that is, for those
workers whose current weekly wage places
them above or below the maximum). This
comparison is accomplished using the
dummy variable D, defined earlier. The
variable D is also treated as endogenous in
estimating the quit equations. An addi-
tional test of the plausibility of our results
is the prediction that the estimate of 8̂ ,
given D = 1, should be more negative
than its estimated value when D = 0.
Use of the benefit maximum as the
numerator could introduce some error
into our measure of the replacement rate,
since it overstates the replacement rate for
workers whose wages are low enough to
place them below the maximum in their
state. In Moore and Viscusi (1989), how-
ever, we argue that changes in the
maximum will be valued by all workers
because weekly wages are not known with
certainty a priori. An increase in the
maximum will therefore increase expected
benefits for all workers, with the extent of
the increase being felt most strongly by
workers whose wage places them above
the maximum. Einally, since benefit ceil-
ings are one of the primary policy instru-
ments available for altering benefit levels,
direct estimates of the effect of changes in
the maximum are most relevant for policy
purposes.'”*
The results of this estimation, summa-
rized in Table 4, essentially replicate those
reported in Table 3 in terms of sign and
statistical significance. The quit behavior of
workers with more than three years of ten-
ure is systematically related to both the risk
level and the risk-weighted value of work-
ers’ compensation benefits. Increases in risk
lead to significantly more quits, and to in-
creased intentions to quit. This effect is once
again reduced by insurance for financial
losses and medical costs associated with an
injury that are embodied in the workers’
compensation program. Quits and quit in-
tentions of workers with less than three
years’ tenure, on the other hand, are not
significantly related to either of these forces.
A further test of the plausibility of the
model compares the coefficients on the
weighted replacement rate variables for
workers whose wage places them above or
below the benefit maximum. If workers
are uncertain about their future wage,
then each worker will attach some likeli-.
hood to the possibility that the wage at the
time of an injury will exceed the benefit
maximum. Workers whose wages cur-
rently exceed the maximum will attach a
greater probability to this outcome and
will, therefore, value changes in the
benefit maximum more highly. As a
consequence, the estimated replacement
rate effects should be more negative for
workers whose current weekly wage ex-
ceeds the benefit maximum.
Our results support this prediction for the
‘̂ A further reason for use of the maximum is that
it allows direct comparisons with our other published
studies on this subject (Moore and Viscusi 1990), An
earlier version of this paper used the benefit maxima
rather than the replacement rate as the benefit
measure. The signs, significance levels, and magni-
tudes of the key coefficients in the wage and quit
equations mirror those reported here.
92 INDUSTRIAL AND LABOR RELATIONS REVIEW
Table 4. Wage-Risk and Wage-Benefit Tradeoffs: Summary of Coefficient Estimates
Using Alternative Benefit Variable.
(Standard Errors in Parentheses)
Dependent
Variable
Weak Quit Intentions
TENURE a 3
TENURE < 3 Strong Quit Intentions TENURE a 3 TENURE < 3
Actual Quits”
TENURE a 3
TENURE < 3
In (Wage)
-2.022***
(0.680)
1.193
(1.025)
-2.200***
(0.750)
-0.422
(0.904)
-0.070**
(0.042)
0.151
(0.106)
Independent
Risk
0.266**
(0.115)
-0.062
(0.087)
0.411***
(0.159)
0.042
(0.086)
5.73E-3***
(2.32E-3)
-2.28E-3
(5.58E-3)
Variable
Weighted
Replacement Rate
X d
-0.579**
(0.290)
0.091
(0.124)
-1.028**
(0.464)
– 0 . 1 2 9
(0.176)
-8.45E-3***
(2.94E-3)
-3.09E-3
(6.82E-3)
X (l-d)
-0.205**
(0.101)
– 0 . 1 0 1
(0.094)
-0.290**
(0.139)
– 0 . 0 2 6
(0.072)
a
a
Sources: See notes to Table 2.
” Not significantly different from adjacent estimates.
* Statistically significant at the .10 level: ** at the .05 level; *** at the .01 level (one-tailed tests).
senior worker group in two of three cases.
For both quit intention variables, the
weighted replacement rate effects are neg-
ative and significant whether the wage ex-
ceeds or falls below the maximum. Further-
more, the estimated coefficient is always
more negative than the corresponding esti-
mate for low-wage workers. In the actual
quit equation estimated for more senior
workers, the two estimates are not signifi-
cantly different from each other. When re-
stricted to equality, they are negative and
significant. Consistent with our principal hy-
pothesis, the estimated wage-benefit trade-
offs are less negative for the junior worker
group. This pattern obtains for both low-
and high-wage workers, regardless of the
quit variable used. These results, and the
general robustness of the results in Table 3,
indicate that the hypothesis withstands a va-
riety of changes in specification.
Tests of Worker Learning-Wage
Benefit Tradeoffs
The model predicts that the wage-
benefit tradeoff should be more negative
for senior workers on the quit-no quit
margin than for the new hires. In an
important sense, the detailed calculations
are unnecessary, since none of the trade-
offs for the junior worker groups are
based on coefficients that differ signifi-
cantly from zero. Nonetheless, point esti-
mates will indicate whether the hypotheses
hold for the data in our sample. Using the
results in Table 3, the tradeoffs can be
calculated directly.
In the quit equations, the wage-benefit
tradeoff is computed as the negative of the
ratio of the partial effect of a dollar
increase in workers’ compensation bene-
fits on quit intentions, -(dQ/db), to the
effect of a dollar increase in wages on
the same dependent variables, {dQ/dw),
where b and w denote the benefit and the
wage:
(14)
ola; _ -dQ/
db
db ~ dQ/dw •
To evaluate this expression, the quit
equation given by equation 13 above is
rewritten as
JOB TENURE AND WORKERS’ COMPENSATION 93
(15) Q =
[1 + exp – {a’X +
+ -ij, + \pR)],
where w denotes the wage, p the death
risk, H the hours worked per week, and
R the replacement rate.’^ Letting P ( 0
denote the right-hand side of this equa-
tion, the partial effect of an increase in
the wage on quits and quit intentions
equals
= p(Q) (l-P(Q)) (4),// +
(20)
dw
The replacement rate can be written as a
function of the weekly wage, the dummy
variable D, the benefit ceiling WCMAX, and
the tax rate:
(2/3) {\-D)wH
Differentiating this expression with re-
spect to the wage then yields the expres-
sion
(18)
dw
-(m)wH]— – DWCMAX]
dw
The first term in this expression will
always equal zero, since variation in the
weekly wage will affect the variable D only
when the worker just qualifies for the
maximum, at which point the first part of
this term equals zero. Inserting the non-
zero portion of equation 18 into equation
16 yields the expression
(19) a Q ^
dw
Again using equation 15 and the defini-
tion of the replacement rate R = bl
w{\—t)H, the effect of an increase in
benefits on quit behavior equals
‘^ Note that we now use p to denote the risk of an
injury, whereas the theoretical model used p to
denote the probability of no injury.
db
The wage-benefit tradeoff then equals the
negative of the ratio of equation 20 to
equation 19,
db
Similar calculations yield the expression
for the wage-risk tradeoff
(22) ^ =
dp
Table 5 summarizes the median wage-
benefit tradeoffs, and also the median wage-
risk tradeoffs computed using the coeffi-
cient estimates in Table 3. We report medians
to control for outlier problems associates with
computing the sample means.
In every case considered, as predicted
by the theory, the median wage-benefit
tradeoffs are more negative (and less than
zero) for the senior workers. For the
median senior worker in the quit equa-
tion, the wage-benefit tradeoff equals
Table 5. Wage-Benefit and Wage-Risk
Tradeoffs: Median Estimated Effects.
Quit and
Wage Variables
a
3 years 3 years
Wage-Benefit Tradeoffs
Actual Quits
ln (Weekly Wage) (x 10-‘)
Weekly Wage (x 10-‘)
Weak Quit Intentions
ln (Weekly Wage) (X 10″^)
Weekly Wage (x 10″‘)
Strong Quit Intentions
ln (Weekly Wage) (X 10″‘)
Weekly Wage (x 10″‘)
– 3 . 5
– 3 . 6
– 6 . 6
– 7 . 6
– 6 . 4
– 7 . 4
Wage-Risk Tradeoffs
Actual Quits
In (Weekly Wage) (x 10″^)
Weekly Wage (x 10 “2)
Weak Quit Intentions
ln (Weekly Wage) (X 10″^)
Weekly Wage (x 10 “2)
Strong Quit Intentions
ln (Weekly Wage) (X 10-^)
Weekly Wage (x 10 “2)
22.5
12.2
– 4 . 4
– 6 . 3
0.4
1.7
5.3
12.0
– 0 . 2
48.9
5.3
a
25.5
37.9
– 2 . 6
-20.8
25.5
a
Sources: See notes to Table 2.
94 INDUSTRIAL AND LABOR RELATIONS REVIEW
– . 3 5 cents per dollar of henefits, com-
pared to a positive wage-benefit tradeoff
of .53 cents per dollar of benefits for
junior workers. In addition to the absolute
differences observed in the quit equations
and the similar differences in the weak
and strong quit intentions equations, the
wage-benefit tradeoffs for the senior
workers are all based on coefficient esti-
mates that are statistically significant,
whereas those tradeoffs estimated for the
junior workers represent a combination of
coefficients, none of which are signifi-
cantly different from zero. Finally, the
wage-benefit tradeoffs for the senior
workers are quite stable across all six
equations. Thus, in addition to the ob-
served differences in the tradeoffs be-
tween the more and less senior workers,
the tradeoffs estimated for the senior
worker group are also much more precise.
We conclude that the predictions of the
model are supported and appear to be
robust across a variety of specifications of
the dependent variables.
The wage-risk tradeoff results summa-
rized in the second panel of Table 5 are
less consistent with the learning model in
which experienced workers have higher
risk assessments. If workers were fully
informed and all workers’ perceptions
were identical, the wage-risk tradeoffs
would not vary by tenure group. In almost
every instance, however, there is evidence
of a larger positive wage premium for risk
in the case of less experienced workers on
the quit margin. Once again, however, few
of the coefficient estimates in the junior
worker group are significant, so this result
is not particularly troublesome. Moreover,
since the risk variable enters the equation
directly and through the expected replace-
ment rate variable, it may be difficult to
reliably estimate the effects of these
correlated variables.
Due to the insignificance of the coeffi-
cients in the junior worker equations, the
most important information to be gained
from these calculations lies not in the com-
parison of effects, but in the policy implica-
tions of the estimates for the senior worker
group. Most important, the results support
the view of the labor market as an efficient
sorter of workers that compensates workers
for exposure to risk and subsidizes injury
insurance through wage reductions.
We can also use these results to estimate
whether benefit levels are adequate, using
the procedure in Viscusi and Moore
(1987). In this framework, the observed
tradeoff will equal the ratio —pl(\ —p),
where p is the probability of an accident.
For the fatality risk data, this ratio equals
about – 1/10,000. The estimated tradeoffs
between benefits and wages in the actual
quit equations, when both wages and
benefits are put on a weekly basis, are
many times greater than this figure.’^
Thus, since the observed tradeoff exceeds
the optimal tradeoff, it would appear that
benefit levels are too low. Since we do not
include a nonfatal risk variable in the quit
equations, however, the observed rate of
tradeoff might reflect instead the rate that
is optimal for risks of all types of injuries,
which would be much closer to the
observed tradeoff.
The same coefficients can also be used to
calculate the value of life and the degree of
risk aversion implied by our estimates. Us-
ing the estimates in the strong quit equa-
tion, the estimated value of life implied by
the estimates ranges between 1 and 4,25
million dollars.’^ This range is lower than
the estimates in the wage equations esti-
mated by Moore and Viscusi (1989) using
the NTOF data, and is similar to many es-
timates found in the literature. Since the
wage-risk tradeoffs are not as stable as the
wage-benefit tradeoffs across equations
within the senior worker group, the value-
of-life calculations are not as meaningful as
the calculations for assessing benefit opti-
mality.
Finally, since we know that workers will
give up approximately .004 dollars in
weekly wages for a one-dollar increase in
benefits, we can compute the degree of
“‘The estimated tradeoff of ,0035 dollars per
hour multiplied by 40 hours yields a weekly wage
decrease of 14 cents corresponding to a $1 benefit
increase,
‘ ‘ For example, the wage-risk tradeoff of ,017,
when multiplied by 2,000 hours, an inflation factor
of 1,25, and 100,000 (the risk scaling factor that
yields one statistical life), equals 4,25 million.
JOB TENURE AND WORKERS’ COMPENSATION 95
risk aversion exhibited by the workers in
our sample. If benefits were to rise hy two
dollars per week, wages would fall by
about $.01, for an annual decline of 50
cents. The discounted expected value of
this two-dollar benefit increase, at an
annual risk level of 1/10,000 and with 52
weeks per year and a real discount^ate of
5%, is 21 cents. Thus, the implication is
that workers are risk averse, and are
willing to sacrifice one dollar of wages for
about 42 cents in expected insurance
benefits.
Conclusion
The results in this paper and in other
recent studies suggest that compensating
differentials for risk should be viewed as a
broader issue than the standard wage-risk
tradeoff literature implies. Wages and
workers’ compensation serve as comple-
mentary compensation mechanisms, with
wages providing ex ante risk compensa-
tion and workers’ compensation providing
ex post earnings replacement. Each of
these earnings components reduces
worker quitting, and workers accept a
lower wage in return for higher workers’
compensation benefits. This tradeoff re-
flects worker preferences for different
forms of risk compensation.
The results presented here utilize these
wage and quit relationships to explore the
differences in the wage-workers’ compen-
sation tradeoff for workers on the quit
margin. Estimated tradeoffs indicate that
workers on the quit margin place a much
higher relative value on workers’ compen-
sation than do new hires. Moreover,
greater job risks lead to significant in-
creases in quits and quit intentions for
workers on the quit margin but have no
effect on the quit behavior of junior
workers. These fmdings are consistent
with a model in which worker quits are
induced in part by learning about risks on
the job.
REFERENCES
Amemiya, Takeshi, 1985, Advanced Econometrics.
Cambridge, Mass,: Harvard University Press,
Arnould, Richard J,, and Len M, Nichols, 1983,
“Wage-Risk Premiums and Workers’ Compensa-
tion: A Refinement of Estimates of Compensating
Wage Differential,” Joumal of Political Economy,
Vol, 91, pp, 332-40,
Berkowitz, Monroe, and John F, Burton, Jr, 1987,
Permanent Disability Benefits in Workers’ Compensa-
tion. Kalamazoo, Mich,: LIpjohn Institute,
Burton, John F,, Jr, 1983, “Compensation for
Permanent Partial Disabilities,” In John D. Worral,
ed,. Safety and the Work Force: Incentives and
Disincentives in Compensation. Ithaca, N,Y,: ILR
Press, pp, 18-60,
Burton, John F,, Jr,, and Alan B, Krueger, 1986,
“Interstate Variations in the Employers’ Cost of
Workers’ Compensation, with Particular Reference
to Connecticut, New Jersey, and New York,” In
James R, Chelius, ed,. Current Issues in Workers’
Compensation. Kalamazoo, Mich,: LIpjohn Institute,
Butler, Richard J, 1983, “Wage and Injury Rate
Response to Shifting Levels of Workers’ Compen-
sation,” In John D, Worral, ed,. Safety and the Work
Force: Incentives and Disincentives in Compensation.
Ithaca, N,Y,: ILR Press, pp, 61-86,
Butler, Richard J,, and John D, Worral, 1983,
“Workers’ Compensation: Benefit and Injury
Claims Rates in the Seventies,” Review of Economics
and Statistics, Vol, 65, No, 4, pp, 580-99,
Diamond, Peter, 1977, “Insurance Theoretic Aspects
of Workers’ Compensation,” In Peter Diamond,
ed,. Natural Resources, Uncertainty, and General
Equilibrium Systems. New York: Academic Press, pp,
67-89,
Dorsey, Stuart, and Norman Walzer, 1983, “Work-
ers’ Compensation, Job Hazards, and Wages,”
Industrial and Labor Relations Review, Vol, 36, No. 4,
pp, 642-54,
Ehrenberg, Ronald G, 1989, “Workers’ Compensa-
tion, Wages, and the Risk of Injury,” In John F,
Burton, Jr,, ed,. New Perspectives on Workers’
Compensation. Ithaca, N,Y,: ILR Press, pp, 71-96,
Kahn, Shulamit, 1987, “Occupational Safety and
Worker Preferences: Is There a Marginal Worker?”
Review of Economics and Statistics, Vol. 69, No, 2, p p ,
262-68,
Krueger, Alan B,, and John F, Burton, Jr, 1990,
“The Employers’ Cost of Workers’ Compensation
Insurance: Magnitudes, Determinants, and Public
Policy,” Review of Economics and Statistics, Vol. 72,
No, 2, pp, 228-40,
Moore, Michael J,, and W, Kip Viscusi, 1988.
“Doubling the Estimated Value of Life: Results
Using New Occupational Fatality Data,” Joumal of
Policy Analysis and Management, Vol, 7, No, 3, p p ,
476-90,
1989, “Promoting Safety Through Workers’
Compensation: The Efficacy and Net Wage Costs
96 INDUSTRIAL AND LABOR RELATIONS REVIEW
of Injury Insurance,” Rand Journal of Economics,
pp, 499-515,
_, 1990, Compensation Mechanisms for Job Risks:
Wages, Workers’ Compensation, and Product Liability.
Princeton, N,J,: Princeton University Press,
Mortensen, Dale, 1978, “Specific Capital and Labor
Turnover,” Bell Journal of Economics, Vol, 9, No, 2,
pp, 572-86,
Smith, Robert S, 1979, “Compensating Differentials
and Public Policy: A Review,” Industrial and Labor
Relations Review, Vol, 32, pp, 339-52,
U,S, Department of Labor, 1977, Handbook of Labor
Statistics. Bulletin No, 1966, Washington, D,C,:
U,S, GPO,
tJ,S, Department of Commerce, 1990, Statistical
Abstract of the United States. Washington, D,C,: U,S,
GPO,
Viscusi, W, Kip, 1979a, Employment Hazards: An
Investigation of Market Performance. Cambridge,
Mass,: Harvard University Press,
1979b, “Job Hazards and Worker Quit Rates:
An Analysis of Adaptive Worker Behavior,”
Intemationai Economic Review, Vol, 20, No, 1, pp,
29-58,
1980a, “A Theory of Job Shopping: A
Bayesian Perspective,” Quarterly Journal of Econom-
ics, Vol, 94, pp, 609-14,
_, 1980b, “Imperfect Job Risk Information and
Optimal Workman’s Compensation Benefits,”
Journal of Public Economics, Vol, 14, No, 3, pp,
319-37,
_, 1980c, “Sex Differences in Worker Quitting,”
Review of Economics and Statistics, Vol, 62, No, 3, pp,
388-98,
_, 1980d, “Self-Selection, Learning Induced
Quits, and the Optimal Wage Structure,” Interna-
tional Economic Review, Vol, 21, No, 3, pp, 529-46,
Viscusi, W, Kip, and Michael J, Moore, 1987,
“Workers’ Compensation: Wage Effects, Benefit
Inadequacies, and the Value of Health Losses,”
Review of Economics and Statistics, Vol, 69, No, 2, pp,
249-61,
Viscusi, W, Kip, and Charles O’Connor, 1984,
“Adaptive Responses to Chemical Labeling: Are
Workers Bayesian Decision Makers?” American
Economic Review, Vol, 74, pp, 942-56,
Worrall, John D,, and Richard J, Butler, 1985,
“Work Injury Compensation and the Duration of
Nonwork Spells,” Economic Journal, Vol, 95, pp,
714-24,