You are on page 1of 20

Labour Economics 8 2001.

223242
www.elsevier.nlrlocatereconbase

What really matters in a job? Hedonic


measurement using quit data
Andrew E. Clark )
CNRS and LEO, Uniersite dOrleans,
Rue de Blois, B.P. 6739, 45067 Orleans, Cedex 2, France

Accepted 5 February 2001

Abstract
This paper uses labour market spell data from the first seven waves of the British
Household Panel Survey BHPS. to model separations and quits. Three main results
emerge. First, job satisfaction data are powerful predictors of both separations and quits,
even controlling for wages, hours and standard demographic and job variables. Second, the
comparison of the power of seven domain job satisfaction measures in a quit equation
yields a ranking of job characteristics: job security and pay are the most important, followed
by use of initiative, the work itself, and hours of work. This ranking differs markedly across
different labour market groups. Last, union job dissatisfaction seems to be real: dissatisfied
union members are just as likely to quit as are dissatisfied union non-members. However,
union Afree-ridersB non-union members at establishments with union recognition. do seem
to behave differently from workers at establishments where unions are not recognised.
q 2001 Elsevier Science B.V. All rights reserved.
JEL classification: J28; J30; J60
Keywords: Job satisfaction; Job quality; Quits; Unions

1. Introduction
A large literature in economics has considered the determinants of labour
market mobility, with a particular emphasis on quits. This latter is typically
modelled in terms of discounted values of income streams, predicting that quits are
negatively correlated with current wages. A sub-strand of this literature has related
)

Tel.: q33-23841-7365.
E-mail address: Andrew.Clark@univ-orleans.fr A.E. Clark..

0927-5371r01r$ - see front matter q 2001 Elsevier Science B.V. All rights reserved.
PII: S 0 9 2 7 - 5 3 7 1 0 1 . 0 0 0 3 1 - 8

224

A.E. Clark r Labour Economics 8 (2001) 223242

quits not only to wages, but also to workers prior evaluations of their jobs
usually, a measure of overall job satisfaction.. Some studies e.g., GHS, 1973;
Shields and Wheatley Price, 1999. relate job satisfaction at time t to quit
intentions at time t. However, as this latter might be thought of as just another
way of expressing dissatisfaction, it is not clear that much is explained by
examining this correlation.
Other studies have used panel data to model realised quits in terms of preious
levels of reported job satisfaction. This relationship is typically negative and
significant. The current paper contributes to this literature using the first seven
waves of the British Household Panel Survey BHPS.. The finding that cross-section job satisfaction responses are indeed good predictors of future quits has two
major implications for labour economics. The first is that cross-section subjective
scores contain useful information predicting observable future labour market
behaviour: no relationship would be found if job satisfaction responses were
dominated by idiosyncratic individual effects. This finding thus validates the
analysis of such subjective data.
The second implication is that jobs are much richer than the typical empirical
article in labour economics would lead us to believe. Mobility is typically studied
in terms of wages, hours of work and tenure. However, these three variables may
not be the only important ones, or even the most important ones. The lack of
suitable data on other job characteristics does not usually allow this hypothesis to
be tested. Here, I will argue that domain job satisfaction measures are appropriate
indices of such job characteristics.
A ARoute OneB approach to finding out what makes a good job is to ask
workers. Clark 1997. uses data from the first wave of the BHPS, in which all
respondents in the labour force indicate the first and second most important
aspects of a job from a list of seven specific job attributes. Job Security is most
often cited as the first most important aspect of a job, followed by the work itself,
and then pay. The second most important aspect list is headed by pay, then job
security and the work itself. The least-cited aspects both first and second most
important. are hours and promotion opportunities. Splitting the sample by sex
shows that men seem more concerned with extrinsic aspects of work such as pay
and promotion. whereas women value the intrinsic returns to work.
An alternative approach, as taken by Ward and Sloane 2000., is to model the
relationship between overall job satisfaction and satisfaction with various job
domains pay, hours, work itself, etc... In somewhat the same vein, Clark 1998.
models overall job satisfaction as a function of 18 different job characteristics. The
characteristics which have the greatest effect on overall job satisfaction are argued
to be those which are most important.
One potential drawback of such studies, which typifies many economists
reactions to subjective data, is a suspicion that what workers say does not reveal
their preferences and thus their behaviour.. In the BHPS, in addition to a question
on overall job satisfaction, workers are asked to rate their satisfaction levels with

A.E. Clark r Labour Economics 8 (2001) 223242

225

seven facets of their job: promotion prospects, total pay, relations with supervisors, job security, ability to work on their own initiative, the actual work itself, and
hours of work. Instead of asking workers what is important, or analysing the
relationship between overall job satisfaction and various job characteristics, the
approach taken here is hedonic: panel data is used to pick out the types of job
satisfaction which best predict future quits. Job security and pay turn out to be the
most powerful predictors of quits the same characteristics as chosen by the
workers themselves when asked to choose the most important job aspects..
Sub-regressions reveal that these job characteristics differ notably by sex, age, and
hours of work. Last, dissatisfied union workers are shown to be just as likely as
dissatisfied union non-members to quit: in this sense union job dissatisfaction is
real.

2. Mobility and quits


Models of two-sided search have attracted a great deal of attention in labour
economics, leading to predictions about sub-optimal firm-worker matches see
Burdett, 1978; Jovanovic, 1979; Mortensen, 1988.. McLaughlin 1991. considers a
model with asymmetric information and wage revisions to derive a theoretical
distinction between quits the firm refuses the workers demand for a wage rise.
and layoffs the worker refuses the firms demand for a wage cut.. Sicherman and
Galor 1990. consider the optimal choice of education in the context of a career
path involving job switches.
Quits in search models occur when the appropriately discounted expected utility
stream at another job is higher than that in their current position, net of moving
costs. The theoretical and empirical literature on search and mobility almost
without exception identifies utility with wages, so that mobility depends on the
discounted present value of income streams. For example, Sicherman and Galor,
in their section on empirical implications, write that AWhile in some occupations
the returns to schooling are in the form of higher wages, in other occupations the
returns are in terms of higher probabilities of advancing to occupations with higher
wagesB p. 177.. Empirical estimation of quit functions, such as McLaughlin
1991. and Topel and Ward 1992., uses wage as the key job description variable.
Labour economists do pay a great deal of attention to wages, and this paper will
not argue that this attention is misguided: there is strong empirical evidence that
quits decline in the current jobs wage. Analogously, however, quits should also
decline in the level of all the other attractive features of the current job.
Considering V as a value function describing the utility stream in job i, an
individual will quit to job j if Vi - Vj y C, where C is any quitting costs. My
main goal in this paper is to put more structure into V than simply the wage rate.
In particular, I imagine that V s V w,h,Z ., where w is wages, h is hours of work,
and Z is a vector of other important job characteristics. For data reasons, Z will

226

A.E. Clark r Labour Economics 8 (2001) 223242

here include promotion prospects, relations with supervisors, job security, ability
to work on ones own initiative, and the actual work itself.
As I am suggesting that jobs are best described by a number of different
aspects, a summary measure of Ajob qualityB or utility at work would seem to be a
useful tool for the analysis of labour market behaviour. This paper considers what
still remains a fairly novel measure of match quality: overall job satisfaction, as
reported by workers themselves. If job utility depends on more than just wages
and hours, then overall job satisfaction should predict quits even when current
wages and hours are controlled for. This turns out to be the case.
This use of job satisfaction data remains relatively unusual in economics, but
widespread in other social science disciplines, such as psychology, sociology and
management. A substantial body of work in Psychology has looked at the job
satisfactionquits link. McEvoy and Cascio 1985. carry out a meta-analysis of
five American studies to show that job enrichment, here understood to be strongly
correlated with job satisfaction, is significantly negatively correlated with voluntary turnover. Carsten and Spector 1987. carry out a similar meta-analysis of 39
studies. Warr 1998. is a recent summary of work in this field. From an
economists point of view there are weaknesses in these meta-analyses, both in
terms of the sample sizes the average of the 39 summarised in Carsten and
Spector is only 300 individuals., heterogeneity in the other control variables, and
the non-representativeness of the samples which are often occupation-specific..
By way of contrast, the small literature in economics has used larger, random
samples. Flanagan et al. 1974., analyse 2275 white males from the NLS Older
Men Survey to show that quits are negatively correlated with job satisfaction
actually whether the worker likes their job or not., even controlling for wages.
Freeman 1978. uses panel data from three surveys: NLS Older Men; the
Michigan PSID; and NLS Younger Men. Job satisfaction scores, together with the
lagged wage, are entered directly into probit quit equations. As both variables are
treated as cardinal although the satisfaction scores are more properly rankings.,
Freeman is able to compare directly quit elasticities, finding that the elasticity with
respect to job satisfaction is, in two of the three datasets examined, at least as large
as that with respect to former wages. Akerlof et al. 1988. use data from the NLS
Older Men survey to reach the same conclusion. Lincoln and Kalleberg 1996.
analyse quits using American and Japanese data.
Two more recent papers analyse job satisfaction and quits in the German
Socio-Economic Panel GSOEP.. Clark et al. 1998. find, in the first 10 waves of
GSOEP, that workers reporting dissatisfaction with their jobs are statistically more
likely to quit than those with higher levels of satisfaction. The panel aspect of the
data allows controls to be introduced for unobserved individual heterogeneity, to
which the results are robust. Levy-Garboua
et al. 1999. analyse men aged 2540

in the first 11 waves 19841994. of the GSOEP. Their innovation is to consider


that job satisfaction reflects the whole experience of the job, whereas quits should
only be related to job satisfactions future component: past satisfaction or dissatis-

A.E. Clark r Labour Economics 8 (2001) 223242

227

faction in the job is therefore considered as a sunk cost. The part of job
satisfaction associated with past experience is thus subtracted from the overall
measure to calculate expected future job satisfaction; this latter is shown to be a
significant predictor of quits. 1
Job satisfaction thus seems to be a suitable index of a poor job match: someone
with low job satisfaction is significantly more likely to quit their job in the future
than is an otherwise identical person with high job satisfaction.2 The analysis of
observed behaviour in the labour market should reassure those who are squeamish
of results based uniquely on what people say.
The fact that job satisfaction predicts quits even when wages and hours of work
are controlled for suggests that utility at work depends on other, typically
unmeasured, variables. The data used in the current paper, which includes
information on satisfaction with seven job domains, will allow us to address this
issue. All of the above papers used only one summary measure of how much the
individual likes their job. This paper will show that satisfaction scores in five out
of seven work domains predict future quits. The relative power of the different job
satisfaction measures in explaining quits yields a ranking of Awhat really mattersB
in a job, not by asking workers what is important for them, but by seeing which
types of dissatisfaction are most likely to drive quits. This approach has some
similarities to Lockes 1976. definition of job satisfaction as the weighted sum of
satisfaction with work sub-domains, the weights being provided by the importance
which the worker attaches to each of them. The weights here are derived by the
contribution to the log-likelihood of different job satisfaction measures in a quit
equation.
3. Data and initial results
The BHPS covers a random sample of approximately 10,000 individuals in
5500 British households. This data set includes a wide range of information about
individual and household demographics, health, employment, values and finances.
The wave 1 data were collected in late 1991early 1992, the wave 2 data were
collected in late 1992early 1993, and so on. The BHPS data includes files
referring to individual labour market spells. The technical details regarding this
1
All of the above refers to match quality as reported by workers. Conversely, it would be of great
interest to have information on the firms evaluation of the match, and to relate that to future
separations. Such information is routinely collected by firms, as a result of internal performance
evaluations, but is typically not made available to outside researchers.
2
Cross-section job satisfaction data may also be validated by linking it to absenteeism negative:
Clegg, 1983. or productivity positive: Mangione and Quinn, 1975.. Patterson et al. 1997. analyse rich
panel data on over a hundred UK manufacturing companies to show that workers, job satisfaction
scores explain 5% and 16% of the variation in profitability and productivity, respectively. In both
cases, prior profits and productivity are controlled for.

228

A.E. Clark r Labour Economics 8 (2001) 223242

spell data, including checks for recall bias using overlaps between retrospective
data at wave t and the same information from the individual interview at wave
t y 1, are described in Halpin 1997.. Using the start and end dates of each spell
the end date may be censored, i.e. the spell may be ongoing., spell information
can be matched back to individual records at the separate waves. This yields a rich
data set with multiple spells per individual, and multiple observations per spell. I
consider only those who are in employment omitting the self-employed,. and who
are aged between 16 and 65.
The original spell data include information on just under 15,000 ended job
spells both employed and self-employed: these are not distinguished in the spell
data., of which around 3500 ended with a quit. Of these spells, 7309 ended
employment spells of which 1771 ended in a quit. could be matched to individual
interview data3 employment can be distinguished from self-employment in the
wave records..
The spell data include start and end date, censoring status i.e., is the spell
ongoing?. and the reason for the end of terminated job spells. The distribution of
this latter is presented below. The first column refers to all job spells in the data,
while the second refers to all employment spells which were matched back to
individual interview data.
Reason for end of spell

All jobs
spells %.

Matched employment
Job spells %.

Promoted
Left Intentionally
Made Redundant
Dismissed
Temporary Job Ended
Took Retirement
Health Reasons
Left To Have Baby
Look After Family
Care Of Other Person
Other
Moved Away
Started College or University

16.61
23.58
12.37
2.50
12.18
3.43
3.75
2.14
2.02
0.24
18.90
1.48
0.81

24.37
24.58
13.05
1.94
5.49
4.34
3.69
2.00
2.04
0.15
16.30
1.48
0.57

The spell data include both internal and external job mobility, the first option
referring to a different job with the same firm. In the first column, the largest
3
Say that an individual had four completed employment spells in a given year. Obviously only one
of these could have been active on the day of his or her yearly interview. The remaining spells are
unmatched and will not be used in this analysis.

A.E. Clark r Labour Economics 8 (2001) 223242

229

percentage of spells just under a quarter. ended in a quit, followed by internal


mobility and other reasons, then the end of a temporary job and being made
redundant. No other reason accounts for more than 4% of all spell ends. In column
2, which shows the spells which can be matched to interview dates, the pattern of
responses is similar, although Atemporary job endedB is less prevalent and
promotion more prevalent. The shorter durations of temporary jobs make them less
likely to be matched with an individual interview; in addition, BHPS interviews
take place in the Autumn, when far fewer temporary Asummer jobsB will be
observed.
Quitting will be indicated by the second answer in the above list, ALeft
IntentionallyB. The distinction between quits and layoffs, in particular, will always
be somewhat hazy: someone who is fired, or who left for other reasons, may well
report that they themselves instigated the separation. One shred of evidence that
quits are different from other spell ends in this data comes from looking at the
next reported labour status in the spell data. For those who said that they left
intentionally, the overwhelming majority 94%. moved into employment, with
only 4% moving into unemployment. For all employment spell ends, these figures
are 63% and 21%, respectively.4
The key explanatory variable is job satisfaction. In the BHPS, respondents in
employment initially rate their satisfaction levels with seven specific facets of their
job promotion prospects, total pay, relations with supervisors, job security, ability
to work on ones own initiative, the actual work itself, and hours of work., each of
which was to be given a number from one to seven, where a value of one
corresponded to Anot satisfied at allB and a value of seven corresponded to
Acompletely satisfiedB. After they had rated their levels of contentment with the
list of topics, individuals were asked a final question, worded as:
All things considered, how satisfied or dissatisfied are you with your present
job overall using the same 17 scale?
Table 1 takes a first look at job mobility as a function of job satisfaction. For
each satisfaction measure, the logrank x 2 statistic tests for the equality of the
survivor function by the seven satisfaction levels. All of the correlations are
significant at better than the 0.1% level. For separations, satisfaction with job
security is the most powerful predictor. It is not clear how much economic
intuition this correlation provides, however, as those whose job will soon end for
example, temporary workers. will naturally be dissatisfied with their job security.
Satisfaction with pay, the work itself, and initiative are the next strongest
correlates.
With respect to quits, the strongest correlation is with overall satisfaction, as
might be hoped. Pay and the work itself are the most important domain satisfac4

For employment spells which end for AotherB reasons, 62% are followed by employment, and 23%
by unemployment.

230

A.E. Clark r Labour Economics 8 (2001) 223242

Table 1
Separations and quits: logrank tests of equality of the survivor functions

Overall job satisfaction


Satisfaction with
Promotion Opportunities
Total Pay
Relations With Supervisors
Security
Use Of Initiative
Work Itself
Hours Worked

Separation rate
x 2 6.

Quit rate
x 2 6.

259.0

100.2

74.4
166.3
46.5
364.8
120.6
167.3
104.9

21.5
97.0
24.2
48.1
53.9
79.1
54.2

Critical values for the x 2 6. distribution: 22.5s 0.1% level; 16.8s1% level; 12.6 s 5% level.

tion variables. This ranking of job characteristics will be partially preserved in the
multivariate analysis below.
4. Mobility, quits and job satisfaction: regression results
Table 1 presented some bivariate evidence that job mobility is correlated with
various job satisfaction measures. However, mobility likely depends on many
different factors, such as wages, hours of work, marital status, number of children
and so on. Multivariate regression analysis allows us to check the predictive power
of job satisfaction, even after standard controls are introduced.
The relationship between job satisfaction and quitting will be modelling using a
duration model. Specifically, the regression results come from a discrete time Cox
regression.5 The Cox specification assumes proportional hazards, with a non-parametric baseline hazard. All of the papers key results can be reproduced using a
Weibull model advanced failure time specification., where the assumption of
proportional hazards is dropped, but the baseline hazard is parameterised. The
BHPS panel data contain both entries and exits, and as such is unbalanced. Again,
the papers main results can all be obtained using only balanced panel data.
The data contain both multiple observations on individual spells, and multiple
observations on individuals. Separations and quits are therefore modelled using
time-varying covariates where appropriate.. As the same subjects appear repeatedly in the risk pools, a robust variance estimator is used with clustering on
individuals.
The first two columns of Table 2 concern separations. The data here refer to
31,791 observations of 13,875 spells, of which 7309 ended in separation. The
5
A discrete time model is used as dates in the spell data are only recorded to the nearest month. Ties
in the data are dealt with using the Breslow method. A nice presentation of the Cox model can be
found in a recent paper by Bratberg and Vaage 2000..

Table 2
Separations, quits and overall job satisfaction: duration models
Separations

All

All

y0.118 0.008 .

y0.113 0.008 .
y0.037 0.033 .
y0.118 0.009 .
1.204 0.116 .
y0.114 0.034 .
0.181 0.048 .
0.875 0.045 .
y0.073 0.035 .
y0.123 0.034 .
0.173 0.042 .
0.203 0.090 .
0.217 0.062 .
0.107 0.151 .
0.115 0.044 .
0.015 0.040 .
y0.018 0.036 .
y0.013 0.031 .
0.150 0.033 .
0.135 0.044 .
0.070 0.045 .
0.061 0.023 .
0.104 0.029 .
0.055 0.063 .
Yes
Yes
Yes
28,036
11,674
6253
y50,888.4
y51,650.5

y0.145 0.017 .

y0.114 0.019 .
0.085 0.076 .
y0.038 0.022 .
y0.408 0.295 .
y0.381 0.069 .
0.330 0.097 .
0.909 0.097 .
y0.201 0.090 .
y0.336 0.081 .
0.159 0.090 .
0.168 0.195 .
0.273 0.149 .
y0.042 0.398 .
0.305 0.096 .
0.062 0.087 .
0.085 0.081 .
0.081 0.073 .
0.198 0.072 .
0.100 0.101 .
y0.149 0.101 .
0.092 0.051 .
0.134 0.065 .
0.195 0.134 .
Yes
Yes
Yes
21,779
9211
1459
y11,046.4
y11,690.1

31,461
13,670
7217
y60,471.3
y60,577.1

24,402
10,784
1752
y14,211.2
y14,250.2

Men

Women

y0.126 0.027 .

y0.103 0.028 .

y0.028 0.031 .
y0.437 0.407 .
y0.353 0.102 .
0.282 0.124 .
0.888 0.123 .
y0.187 0.120 .
y0.301 0.107 .
y0.092 0.136 .
0.025 0.254 .
0.204 0.200 .
y0.273 0.449 .
0.296 0.133 .
0.068 0.117 .
0.042 0.109 .
0.041 0.103 .
0.206 0.100 .
0.085 0.161 .
y0.123 0.164 .
0.175 0.067 .
0.305 0.084 .
0.262 0.169 .
Yes
Yes
Yes
11,102
4722
753
y5177.6
y5514.6

y0.077 0.033 .
0.022 0.435 .
y0.370 0.096 .
0.462 0.201 .
0.944 0.153 .
y0.144 0.136 .
y0.373 0.124 .
0.424 0.124 .
0.279 0.314 .
0.112 0.241 .
0.102 0.913 .
0.335 0.145 .
0.087 0.131 .
0.095 0.123 .
0.110 0.104 .
0.223 0.106 .
0.054 0.132 .
y0.163 0.130 .
0.005 0.081 .
y0.141 0.109 .
0.130 0.205 .
Yes
Yes
Yes
10,677
4489
706
y4816.3
y5164.5

231

Standard errors are in parentheses.

All

A.E. Clark r Labour Economics 8 (2001) 223242

Job satisfaction
Male
Age
Age-squaredr1000
Log wage
Log hours
TemporaryrFixed-Term Contract worker
Union member
Union recognised at workplace
Married
Separated
Divorced
Widowed
Education: High
Education: ArOrNursing
Firm Size - 25
Firm Size 25 199
House Renter
Number of Children aged 0 2
Number of Children aged 3 4
Number of Children aged 5 11
Number of Children aged 12 15
Number of Children aged 16 18
Regional Dummies 18 .
Industry Dummies 10 .
Occupation Dummies 9 .
Number of observations
Number of individuals
Number of separationsrquits
Log Likelihood
Log Likelihood at zero

Quits

All

232

A.E. Clark r Labour Economics 8 (2001) 223242

remaining columns refer to quits. Here, I have omitted the last observation in all
spells which ended for a non-quit reason, transforming them from non-censored
non-quit ended. into censored spells. Similar results are obtained by modelling
quit-ended spells against all other spells, which increases the sample size, but
designates all other reasons for spell termination including missing values and
AotherB . as non-quits. The initial quit sample consists of 24,654 observations on
10,944 spells, of which 1771 ended in a quit.
The variable of most interest in Table 2 is job satisfaction. Many different
specifications of this variable were tested, including dummies, dummies for the
minimum satisfaction expressed over the spell, and dummies for job satisfaction at
the beginning of the spell. For expositional purposes, I here present results using
the simple job satisfaction scores which assumes cardinality.. This makes the
tables easier to read; similar results are obtained by using more elaborate specifications. Other control variables in these regressions include log wages deflated by
the CPI. and log hours of work, sex, age, age-squared, marital status, education,
house renter, a dummy for temporary or fixed-term contract work, number of
children by their ages., union membership, union recognition at the workplace,
and firm size, industry, occupation and region dummies.
The first two columns of Table 2 model all separations, first only as a function
of job satisfaction, and then including the other controls outlined above. A
negative coefficient in this model implies that separation happens later, rather than
sooner, so that job durations are longer. Job satisfaction is a very strong predictor
of separations: the more satisfied is the individual, the less likely he or she is to
separate from their current job. The other estimated coefficients reveal that job
durations are longer for those with higher wages, union members, and in workplaces with union recognition. Conversely, job durations are shorter for those
working longer hours, temporaryrfixed-term contract workers, the married, those
with high education,6 those who rent their accommodation, and those with more
children. Separations are hump-shaped in age, reaching a maximum in the late
40s.
The remainder of Table 2 refers to quits. Column 3 models quits as a function
of job satisfaction alone, while column 4 adds the full set of control variables.
Columns 5 and 6 repeat column 4s analysis for men and women, respectively. Job
satisfaction predicts quits in all four columns, with t-statistics of 8.7, 5.9, 4.6 and
3.7, respectively. This is the first main conclusion of the paper.

The two education dummies are respectively for a college degree and advanced high school
A-level approximately., or intermediate high school O-level approximately.. The omitted category is
for no or low qualifications. These categories are for achieved paper certificates and not merely for
years of schooling.

A.E. Clark r Labour Economics 8 (2001) 223242

233

With respect to the other explanatory variables, quits are less likely for
high-wage workers, and for union workers both members and at union-covered
establishments.; they are more likely for workers with long hours, temporary
contracts, higher education, and for those who rent their accommodation. In the
last two columns the estimated quit equations for men and women are fairly
similar. The only notable differences are that quits by women fall with age,
married women are more likely to quit, whereas married men are not, and that
male quitting is related to the number of younger school-age children aged
between 5 and 15..
The next part of the paper uses the same methodology to establish a ranking of
what matters in a job. It does this by replacing the overall job satisfaction score in
Table 2s quit equations by a series of seven domain job satisfaction variables. The
sample in these regressions is restricted to those for whom all seven measures are
available. The log-likelihoods can thus be compared directly across the different
regressions: descriptive statistics for this regression sample are contained in
Appendix A. The estimated coefficients on the non-job satisfaction explanatory
variables are not reported, as these do not change in any important way from those
in Table 2. The firm size, industry and age-squared variables are dropped from
these equations, as they proved to be insignificant in all of the estimated equations.
Table 3 has eight columns. Each shows the estimate on the job satisfaction
variable and the log-likelihood from the estimation. It is worth repeating that a full
set of other demographic and job controls are used in these regressions. The least
negative log-likelihood i.e., the regression with the greatest explanatory power. is
found for the equation including overall job satisfaction, as might be hoped if this
is to be considered a useful summary measure. With respect to the seven domain
satisfaction variables, the most powerful is satisfaction with job security. In other
words, if we had to choose between the seven domain satisfaction measures in
order to predict quits, then satisfaction with job security would allow us to make
the most accurate prediction.7 After job security, satisfaction with pay 8 , the work
itself, use of initiative and hours of work come in a group, the regressions having
similar log-likelihoods. The least powerful predictors of quits are satisfaction with
relations and with promotion opportunities, which are insignificant. This is the
papers second main finding: different types of job satisfaction predict quitting
differently, allowing a ranking of job characteristics to be established. This shows
that pay and hours of work are indeed important, as the empirical literature has
emphasised. However, the use of initiative and the work itself are just as
7

Deery et al. 1999. use organisational records from 600 employees at an Australian automotive
plant to show that perceived job security is the most important work setting variable predicting
absenteeism.
8
Note that pay satisfaction predicts quits een when wages are controlled for: wages by themselves
seem an insufficient statistic for predicting the effect of pay on quitting. This is consistent with utility
being at least partly relative with respect to pay see Clark, 1996, 1999; Clark and Oswald, 1996..

234

Type of job satisfaction


Overall

Promotion

Pay

Relations

Job security

Initiative

Work itself

Hours

All, Controlling for wages and hours


Satisfaction coefficient y0.108 0.020. y0.013 0.017. y0.058 0.018. y0.005 0.019. y0.073 0.018. y0.063 0.020. y0.052 0.021. y0.064 0.019.
Log likelihood
y9744.29
y9759.20
y9753.60
y9759.53
y9750.60
y9754.37
y9756.04
y9753.35
Log likelihood at zero sy10,277.8; number of observationss18,707; number of individualss8461; number of quitss1308
All, Not controlling for wages and hours
Satisfaction coefficient y0.113 0.020. y0.014 0.017. y0.076 0.017. y0.001 0.019. y0.072 0.018. y0.068 0.020. y0.057 0.020. y0.064 0.019.
Log likelihood
y9896.83
y9913.29
y9902.55
y9913.70
y9904.66
y9907.66
y9909.48
y9907.31
Log likelihood at zero sy10,422.7; number of observationss18,954; number of individualss8553; number of quitss1324
All regressions include the same control variables as in Table 2 above, apart from age-squared, firm size, and industry, all of which are insignificant across the
eight equations. Standard errors are in parentheses.

A.E. Clark r Labour Economics 8 (2001) 223242

Table 3
Quits and domain job satisfaction: duration models

Table 4
Quits and domain job satisfaction. results by sex, age, and hours of work
Type of job satisfaction
Overall

Promotion

Pay

Relations

Job security

Initiative

Work itself

Hours

Women
Satisfaction coefficient y0.119 0.029 . y0.015 0.023 . y0.047 0.025 . y0.015 0.026 . y0.057 0.026 . y0.111 0.027 . y0.080 0.030 . y0.078 0.027 .
Log likelihood
y4456.28
y4464.74
y4462.94
y4464.81
y4462.45
y4456.73
y4460.98
y4460.60
Log Likelihood at zero s y4725.7; number of observations s 9150; number of individuals s 4263; number of quits s 663
30 or Under
Satisfaction coefficient y0.078 0.028 . y0.004 0.023 . y0.032 0.024 . y0.018 0.025 . y0.068 0.024 . y0.085 0.027 . y0.022 0.028 . y0.067 0.026 .
Log likelihood
y4723.26
y4727.50
y4726.55
y4727.24
y4723.46
y4722.26
y4727.16
y4723.95
Log likelihood at zero s y4862.9; number of observations s 5741; number of individuals s 3448; number of quits s 736
Oer 30
Satisfaction coefficient y0.157 0.029 . y0.036 0.024 . y0.089 0.026 . 0.005 0.030 .
y0.075 0.026 . y0.054 0.031 . y0.098 0.030 . y0.059 0.028 .
Log likelihood
y4054.71
y4067.64
y4062.58
y4068.86
y4064.41
y4067.25
y4063.43
y4066.41
Log likelihood at zero s y4301.7; number of observations s 12,966; number of individuals s 5355; number of quits s 572
Part-time workers
Satisfaction coefficient y0.205 0.050 . y0.122 0.037 . y0.098 0.043 . 0.024 0.052 .
y0.120 0.044 . y0.091 0.047 . y0.133 0.051 . y0.164 0.045 .
Log likelihood
y1300.53
y1303.17
y1305.70
y1308.57
y1304.82
y1306.82
y1304.76
y1301.69
Log likelihood at zero s y1427.1; number of observations s 3277; number of individuals s 1698; number of quits s 234

A.E. Clark r Labour Economics 8 (2001) 223242

Men
Satisfaction coefficient y0.099 0.030 . y0.006 0.025 . y0.073 0.027 . 0.006 0.029 .
y0.089 0.025 . y0.016 0.030 . y0.022 0.030 . y0.053 0.028 .
Log likelihood
y4346.80
y4353.20
y4348.90
y4353.20
y4346.27
y4353.07
y4352.91
y4351.07
Log likelihood at zero s y4644.3; number of observations s 9557; number of individuals s 4198; number of quits s 645

Full-time workers
Satisfaction coefficient y0.091 0.022 . 0.014 0.019 .
y0.052 0.020 . y0.010 0.021 . y0.065 0.020 . y0.060 0.022 . y0.039 0.023 . y0.038 0.021 .
Log likelihood
y7788.38
y7797.11
y7793.63
y7797.27
y7791.58
y7793.53
y7795.78
y7795.67
Log likelihood at zero s y8237.3; number of observations s 15,430; number of individuals s 6993; number of quits s 1074

235

All regressions include the same control variables as in Table 2 above, apart from age-squared, firm size, and industry. Standard errors are in parentheses.

236

A.E. Clark r Labour Economics 8 (2001) 223242

important, and job security is the most important of the seven characteristics
examined here.
One pertinent criticism with respect to the first half of Table 3 is that we are
stacking the odds rather against satisfaction with pay and satisfaction with hours of
work, as log wages and log hours already feature amongst the right-hand side
variables and attract significant estimates.. As a check, the quit regressions were
re-run without log wages and log hours. The results are shown in the second half
of Table 3. Pay does indeed move up the ranking, becoming the most important
characteristic. The relative position of the other job characteristics remains unchanged.
Table 4 continues this analysis for different sub-groups of workers: men;
women; the young 30 or under.; the over 30s; part-time workers under 30 hours
per week.; and full-time workers. Table 5 presents an overview of the subsequent
rankings of job characteristics. Job security is an important factor for all of the
groups considered; relations with supervisors, on the other hand, are not significantly correlated with quitting for any of the groups.
Table 5 shows that, according to the quit equations, different groups of workers
do indeed seem to value different things in their jobs. Satisfaction with the ability
to use initiative is the most powerful cause of quits for both women and younger
workers; it is, however, not correlated with quits by men or workers over the age
of 30. Part-timers are the only group for whom promotion opportunities appear to
be an important facet of the job.
Table 4s ranking is consistent with that found by simply asking workers about
the most important aspects of a job. A question regarding the first and second most
Table 5
Ranking of job characteristics, as revealed by subsequent quit behaviour
Most important

Next most important

Insignificant

All

Job Security

Promotion
Relations

Men

Job Security

Women

Initiative

30 or Under

Initiative

Over 30

Pay
Work Itself
Job Security
Promotion
Hours
Job Security

Pay
Initiative
Work Itself
Hours
Pay
Hours
Hours
Work Itself
Job Security
Hours
Hours

Part-time
Full-time

Job Security
Work Itself
Pay
Initiative

Everything Else
Promotion
Relations
Everything Else
Promotion
Relations
Initiative
Relations
Everything Else

A.E. Clark r Labour Economics 8 (2001) 223242

237

important aspect of a job appeared in wave one of the BHPS. Workers chose the
work itself and job security as the two most important aspects of a job, followed
by pay and initiative. The least important aspects chosen by the fewest workers.
were hours of work and promotion opportunities. Similar results are reported in
Ward and Sloane 2000. and Clark 1998.. Not only do workers tell us that certain
aspects of a job are important in their work, and not only do these characteristics
correlate the most strongly with their reported job satisfaction, but dissatisfaction
with these same characteristics is the most powerful predictor of workers future
quit behaviour. This paper thus finds that workers do not seem to misrepresent
either their preferences or their job satisfaction in cross-section data. Workers
consistently say, and their actions bear out their statements, that job security, pay,
the work itself and use of initiative are important in their jobs. To the extent that
much academic research has focussed exclusively on pay and hours, labour
economics may be missing out some of the most important variables in explaining
labour market behaviour.
5. Is union job dissatisfaction real?
An ongoing literature has considered the ArealityB of union members lower job
satisfaction see Bender and Sloane, 1998; Hersch and Stone, 1990; Leigh, 1986..
Are union members encouraged to voice their dissatisfaction more than are union
non-members, or do union members really have more dissatisfying jobs? This
papers modelling approach suggests an innovative answer to this question: if
union members are being encouraged to voice their dissatisfaction, we would
expect low job satisfaction amongst union members to be less correlated with quits
than low job satisfaction amongst union non-members. If, however, union members and union non-members express their job satisfaction in exactly the same
way, then there should be no difference by union status in quitting as a function of
job satisfaction.
Table 6 presents some results. I examine overall job satisfaction, pay satisfaction, and satisfaction with job security the two most important domains in Table
Table 6
Is union job dissatisfaction real?
Type of job satisfaction
Overall

Pay

Job security

Job Satisfaction
Union Member=Job Satisfaction

y0.116 0.023.
0.032 0.044.

y0.056 0.020.
y0.009 0.038.

y0.085 0.021.
0.042 0.037.

Job Satisfaction
Union Recognised=Job Satisfaction

y0.141 0.025.
0.083 0.040.

y0.086 0.023.
0.068 0.034.

y0.109 0.023.
0.083 0.033.

All regressions include the same control variables as in Table 2 above, apart from age-squared, firm
size, and industry. Standard errors are in parentheses.

238

A.E. Clark r Labour Economics 8 (2001) 223242

3.. In the first half of the table, union membership, which has been the focus of
the existing literature, is interacted with the job satisfaction measures. These
interactions attract only insignificant estimates in the quit equation. While average
job satisfaction is negatively correlated with quits, there is no difference in this
relation by union membership. This result therefore suggests that union dissatisfaction is real. Rather than being an artefact of institutional structures that make union
members more likely to express dissatisfaction, a dissatisfied union member and a
dissatisfied union non-member have a statistically equal probability of quitting. It
seems that when union members say that their jobs are dissatisfying, they mean it.
The second half of Table 6 sounds a note of caution, however. Here, I interact
job satisfaction scores with working at an establishment where trade unions are
recognised. Just over half of the sample work in establishments where trade unions
are recognised, and these individuals account for around 40% of all recorded quits
in this data as the regression results show, unionism is negatively correlated with
quitting.. The interaction between working at an establishment with union recognition and job satisfaction is positive and significant. The main effect of job
satisfaction remains negative and significant. A worker with low job satisfaction at
a Aunion-recognisedB workplace is therefore less likely to quit than an identical
worker with low job satisfaction at a workplace where a union is not recognised.
As the first half of Table 6 suggests, this difference does not result from union
members, but rather from union non-members at establishments with union
recognition. Comparison of the members and non-members at workplaces with
union recognition reveals a number of differences between the two samples: the
non-members are younger, have lower wages, are three times as likely to be
temporary workers, and are lower-educated. All of these variables are controlled
for in the regressions, of course, but there may well be other demographic or job
variables which differ between the two groups which could explain their difference in quitting behaviour as a function of job satisfaction. The main conclusion
from this part of the paper is that, according to the test proffered, dissatisfaction
amongst union members is just as real as that amongst non-union members.
However, union Afree-ridersB non-union members at establishments with union
recognition. do seem to behave differently from workers at establishments where
unions are not recognised.
6. Conclusion
This paper has used labour market spell data from the first seven waves of the
BHPS to model the effect of eight different job satisfaction measures on separation
and quit probabilities. Three main conclusions result. The first is that cross-section
job satisfaction responses are significantly correlated with future separations and
quits. This finding validates the cross-section analysis of job satisfaction: those
who say that they are less happy with their job are indeed more likely to both
separate from it and to quit it.

A.E. Clark r Labour Economics 8 (2001) 223242

239

The second conclusion concerns the make-up of a Agood jobB. Instead of


directly asking workers what job aspects are important, this paper has taken a
hedonic approach, by seeing which aspects of a job are most likely to induce quits.
Seven different aspects of the job are considered promotion, pay, relations with
management, job security, use of initiative, the work itself, and hours worked., as
measured by the job satisfaction reported by workers. Satisfaction with pay and
job security are shown to be the most important factors; the use of initiative, the
work itself and hours of work are also significant predictors of future quits. The
relative importance of different job characteristics differs markedly across different groups in the labour market. Standard analyses of labour market mobility,
which concentrate only on pay and hours, may thus omit potentially very
important variables. It is not that economists have been barking up the wrong tree
with the emphasis on wages and hours, but rather that they have not been barking
up enough of them.
The last conclusion concerns the reality of union workers lower job satisfaction. An interaction between union membership and job satisfaction in a quit
equation is insignificant: job dissatisfaction amongst union members therefore
seems to be real. However, there is a significant difference in quitting behaviour
according to union recognition at the workplace. The relationship between quitting
and union Afree-ridersB non-members at covered workplaces. would seem worthy
of further attention.
The most general conclusion concerns directions for future work. This paper
has shown that domain satisfaction scores do have some kind of objective reality,
in that they predict workers future quits. As these domain job satisfaction
measures often cover aspects of a job which seem to be important, but which
would be near impossible to measure objectively such as use of initiative, or the
work itself., they would seem to be a useful complement to the standard toolbox
that economists use to analyse labour market behaviour.

Acknowledgements
I thank Liliane Bonnal, Matthieu Bunel, Kevin Clarke, Christophe Daniel, Bill
Gould, Brendan Halpin, David Margolis, Jan van Ours, Antoine Terracol and
seminar participants at the 1st EALErSOLE conference, the 16th Journees
de
Microeconomie
Appliquee
Paris 1, and the TSER STT Confer
Lyon., Orleans,

ence Orleans
.. I am also grateful to two referees for their thoughtful comments.
The BHPS data were made available through the ESRC Data Archive. The data
were originally collected by the ESRC Research Centre on Micro-social Change at
the University of Essex. Neither the original collectors of the data nor the Archive
bear any responsibility for the analyses or interpretations presented here. The
Stata. programmes used to analyse the BHPS data are available on request.
Financial support from the European Union is gratefully acknowledged.

240

A.E. Clark r Labour Economics 8 (2001) 223242

Appendix A
Variable

Job satisfaction: overall


Job satisfaction:
promotion prospects
Job satisfaction: total pay
Job satisfaction: relations
with supervisors
Job satisfaction: job security
Job satisfaction: use of initiative
Job satisfaction: work itself
Job satisfaction: hours worked
Male
Age
Age-squaredr1000
Log wage
Log hours
TemporaryrFixed-Term
Contract worker
Union member
Union recognised at workplace
Married
Separated
Divorced
Widowed
Education: High
Education: ArOrNursing
Firm Size - 25
Firm Size 25199
House Renter
Number of Children aged 02
Number of Children aged 34
Number of Children aged 511
Number of Children aged 1215
Number of Children aged 1618

Mean Standard Min


error

Max

% of spells in
which variable
varies

5.427 0.0099
4.484 0.0133

1
1

7
7

33.6
37.8

4.615 0.0124
5.529 0.0112

1
1

7
7

37.4
34.9

5.192
5.745
5.562
5.214
0.511
37.481
1.524
6.879
3.507
0.044

0.0122
0.0101
0.0101
0.0112
0.0037
0.0800
0.0062
0.0051
0.0028
0.0015

0.416
0.586
0.624
0.021
0.072
0.011
0.363
0.385
0.294
0.367
0.183
0.071
0.076
0.280
0.173
0.039

0.0036
0.0036
0.0035
0.0010
0.0019
0.0008
0.0035
0.0036
0.0033
0.0035
0.0028
0.0019
0.0020
0.0045
0.0033
0.0015

1
1
1
1
0
16
0.256
1.675
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

7
7
7
7
1
65
4.225
10.192
4.595
1
1
1
1
1
1
1
1
1
1
1
1
2
2
4
3
2

Number of observationss 18,707, except for firm size 18,669..

36.2
32.6
33.5
36.5
0.0
45.5
45.5
46.2
25.2
1.9
3.2
3.8
4.3
1.6
1.8
0.3
1.9
1.8
7.9
12.2
3.3
5.8
5.8
8.3
7.6
4.5

A.E. Clark r Labour Economics 8 (2001) 223242

241

References
Akerlof, G.A., Rose, A.K., Yellen, J.L., 1988. Job switching and job satisfaction in the U.S. labor
market. Brookings Papers on Economic Activity 2, 495582.
Bender, K., Sloane, P., 1998. Job satisfaction, trade unions and exit-voice revisited. Industrial and
Labor Relations Review 51, 222240.
Bratberg, E., Vaage, K., 2000. Spell durations with long unemployment insurance periods. Labour
Economics 7, 153180.
Burdett, K., 1978. A theory of employee job search and quit rates. American Economic Review 68,
212220.
Carsten, J.M., Spector, P.E., 1987. Unemployment, job satisfaction, and employee turnover: a
meta-analytic test of the Muchinsky model. Journal of Applied Psychology 72, 374381.
Clark, A.E., 1996. Lutilite est-elle relative? Analyse a` laide de donnees
Economie et
sur les menages.

Prevision
121, 151164.

Clark, A.E., 1997. Job satisfaction and gender: why are women so happy at work? Labour Economics
4, 341372.
Clark, A.E., 1998. What Makes a Good Job? Evidence from OECD Countries, LEO, University of
Orleans,
Discussion Paper No. 98-26.

Clark, A.E., 1999. Are wages habit-forming? Evidence from micro data. Journal of Economic Behavior
and Organization 39, 179200.
Clark, A.E., Oswald, A.J., 1996. Satisfaction and comparison income. Journal of Public Economics 61,
359381.
Clark, A.E., Georgellis, Y., Sanfey, P., 1998. Job satisfaction, wage changes and quits: evidence from
Germany. Research in Labor Economics 17, 95121.
Clegg, C.W., 1983. Psychology of employee lateness, absence and turnover: a methodological critique
and an empirical study. Journal of Applied Psychology 68, 88101.
Deery, S., Erwin, P., Iverson, R., 1999. Industrial relations climate, attendance behaviour and the role
of trade unions. British Journal of Industrial Relations 37, 533559.
Flanagan, R.J., Strauss, G., Ulman, L., 1974. Worker discontent and work place behavior. Industrial
Relations 13, 101123.
Freeman, R.B., 1978. Job satisfaction as an economic variable. American Economic Review 68,
135141.
General Household Survey GHS., 1973. GHS Introductory Report. OPCS, London.
Halpin, B., 1997. Unified BHPS work-life histories: combining multiple sources into a user-friendly
format. ESRC Research Centre on Micro-social Change, University of Essex, mimeo.
Hersch, J., Stone, J., 1990. Is union job dissatisfaction real? Journal of Human Resources 25, 736751.
Jovanovic, B., 1979. Job matching and the theory of turnover. Journal of Political Economy 87,
972990.
Leigh, D., 1986. Union preferences, job satisfaction, and the union voice hypothesis. Industrial
Relations 25, 6571.
Levy-Garboua,
L., Montmarquette, C., Simonnet, V., 1999. Job Satisfaction and Quits: Theory and

Evidence from the German Socio-Economic Panel. TEAM, University of Paris I, mimeo.
Lincoln, J., Kalleberg, A., 1996. Commitment, quits and work organization in Japanese and U.S.
plants. Industrial and Labor Relations Review 50, 3959.
Locke, E.A., 1976. The nature and causes of job satisfaction. In: Dunnette, M.D. Ed.., Handbook of
Industrial and Organizational Psychology. Rand-McNally, Chicago.
Mangione, T.W., Quinn, R.P., 1975. Job satisfaction, counterproductive behavior, and drug use at
work. Journal of Applied Psychology 60, 114116.
McEvoy, G.M., Cascio, W.F., 1985. Strategies for reducing employee turnover: a meta-analysis.
Journal of Applied Psychology 70, 342353.

242

A.E. Clark r Labour Economics 8 (2001) 223242

McLaughlin, K., 1991. A theory of quits and layoffs with efficient turnover. Journal of Political
Economy 99, 129.
Mortensen, D., 1988. Wage, separation, and job tenure: on-the-job specific training or matching?
Journal of Labor Economics 6, 445471.
Patterson, M., West, M., Lawthorn, R., Nickell, S., 1997. Impact of People Management Strategies on
Business Performance, Institute of Personnel and Development. Issues in People Management No.
22.
Shields, M., Wheatley Price, S., 1999. Racial Harassment, Job Satisfaction and Intentions to Quit.
University of Leicester, mimeo.
Sicherman, N., Galor, O., 1990. A theory of career mobility. Journal of Political Economy 98,
169192.
Topel, R., Ward, M., 1992. Job mobility and the careers of young men. Quarterly Journal of
Economics 107, 439479.
Ward, M., Sloane, P., 2000. Non-pecuniary advantages versus pecuniary disadvantages: job satisfaction
among male and female academics in Scottish Universities. Scottish Journal of Political Economy
47, 273303.
Warr, P.B., 1998. Well-being and the workplace. In: Kahneman, D., Diener, E., Schwartz, N. Eds..,
Understanding Quality of Life: Scientific Perspectives on Enjoyment and Suffering. Russell Sage,
New York.

You might also like