Content uploaded by Eva Berger
Author content
All content in this area was uploaded by Eva Berger on Jan 06, 2017
Content may be subject to copyright.
Discussion
Papers
Locus of Control and
Mothers' Return to Employment
Eva M. Berger and Luke Haywood
1586
Deutsches Institut für Wirtschaftsforschung 2016
Opinions expressed in this paper are those of the author(s) and do not necessarily reflect views of the institute.
IMPRESSUM
© DIW Berlin, 2016
DIW Berlin
German Institute for Economic Research
Mohrenstr. 58
10117 Berlin
Tel. +49 (30) 897 89-0
Fax +49 (30) 897 89-200
http://www.diw.de
ISSN electronic edition 1619-4535
Papers can be downloaded free of charge from the DIW Berlin website:
http://www.diw.de/discussionpapers
Discussion Papers of DIW Berlin are indexed in RePEc and SSRN:
http://ideas.repec.org/s/diw/diwwpp.html
http://www.ssrn.com/link/DIW-Berlin-German-Inst-Econ-Res.html
Locus of Control and Mothers’ Return to Employment
Eva M. Berger∗Luke Haywood
Johannes Gutenberg University Mainz DIW Berlin
Abstract
This paper investigates the effect of locus of control (LOC) on the length of mothers’
employment break after childbirth. Using data from the German Socio-Economic Panel
Study (SOEP), duration data reveals that women with an internal LOC return to employ-
ment more quickly than women with an external LOC. We find evidence that this effect
is mainly related to differential appreciation of the career costs of longer maternity leave.
Given the high level of job protection enjoyed by mothers in Germany, economic con-
sequences of differences in this noncognitive skill can be expected to be larger in other
settings.
Keywords: Locus of Control, Noncognitive Skills, Personality, Maternal Employment,
Female Labor Supply, Survival Analysis
JEL-codes: J22, J24
∗Corresponding author: Johannes Gutenberg University Mainz, Jakob-Welder-Weg 4, 55128 Mainz, Germany.
Phone: +49(0)6131-39-27295. Fax.: +49(0)6131-39-27695. E-mail: eva.berger@uni-mainz.de.
1
1 Introduction
Recent research on so-called noncognitive skills1demonstrates their crucial role for educational
attainment, employment outcomes, and a variety of risky behaviors (Blomeyer et al. 2009, Cole-
man and DeLeire 2003, Heckman et al. 2006, Piatek and Pinger 2015). In the labor market, a
number of studies find that noncognitive skills affect earnings (Andrisani 1977, 1981, Cebi
2007, Flossmann et al. 2007, Heckman and Rubinstein 2001, Heckman et al. 2006, Heineck
and Anger 2010, Mueller and Plug 2006, Nyhus and Pons 2005, Osborne Groves 2005) and
unemployment duration (Uhlendorff2004, Uysal and Pohlmeier 2009).2In the present study,
we test whether and how locus of control (LOC) affects women’s decision how long to stay out
of employment after childbirth, a crucial determinant of female labor supply and related to the
female wage gap (Blundell et al. 2013, Adda et al. 2016).
While the effect of various noncognitive skills on employment outcomes has been stud-
ied, little is known about its association with labor supply decisions.3We focus on one such
noncognitive skill, LOC. It describes “individual differences in a generalized belief for internal
versus external control of reinforcement” (Rotter 1966, p. 1). LOC is then a measure of the
degree to which an individual perceives that success or failure in life follows from her own
behavior rather than being controlled by outside forces. If life events are perceived by a per-
son as being contingent upon her own behavior or her own characteristics, this characterizes
an internal LOC. If, on the other hand, life events are perceived by a person as the result of
luck, chance, fate, or under the control of powerful others, this is labeled an external LOC.4
Using a common factor of the Rotter LOC scale and the Rosenberg self-esteem scale, Heckman
et al. (2006) find that the impact of noncognitive skills on the employment probability is greater
1We use the term non-cognitive skills as it is used in a large part of the related literature (see, e.g., Cunha
and Heckman 2007, 2008). These skills refer to abilities and personality traits other than pure cognitive ability.
Nevertheless, we are aware that most of these skills or traits do have a cognitive component.
2Furthermore, a number of studies in psychology investigate the relationship between personality traits and
occupational attainment (see (Roberts et al. 2007)).
3Wichert and Pohlmeier (2010) find a relationship between women’s big five personality traits and their labor
force participation. For a detailed discussion of the research on noncognitive skills in economics and psychology,
see Borghans et al. (2008).
4Following the original concept of LOC by Rotter (1966), we use a single index of LOC where internal and
external are the opposite poles of a single dimension. This aspect is discussed in detail in appendix A.
1
than the effect of cognitive skills. Moreover, they find that this pattern is more pronounced
for women than for men, suggesting that noncognitive skills matter particularly for women’s
labor supply. Departing from this observation, the present paper focuses on a particularly im-
portant labor supply decision for women—when to re-enter the labor market after childbirth.
Since noncognitive skills—though largely stable throughout adulthood—are malleable during
childhood (Cunha and Heckman 2007, 2008), the current study adds to the evidence suggesting
that early childhood investments may have long-lasting effects (Carneiro et al. 2011, Cunha and
Heckman 2009).
Post-birth employment behavior matters because it affects human capital accumulation,
future employment chances and wages (Davies and Pierre 2005, Gutierrez-Domenech 2005,
Lefebvre et al. 2009). Furthermore, maternal employment plays a large role in counteracting
the risk of child poverty, an important policy issue. Most studies on mothers’ return to employ-
ment focus on institutional factors such as parental leave schemes or tax rules (e.g., Bergemann
and Riphahn 2009, 2010, Burgess et al. 2008, Gutierrez-Domenech 2005), or on individual
and family-related factors such as income, educational attainment, and labor market experience
(Kuhlenkasper and Kauermann 2010).
The present paper thus contributes to two strands of the literature: First, the economic con-
sequences of noncognitive skills; second, the determinants of mothers’ return to employment
after childbearing. We focus on LOC and test several channels through which LOC could
influence the length of employment breaks women take after a childbirth. First, LOC might
affect women’s labor market attachment, viz. their willingness to be available all day for their
children. Second, LOC may affect the relative importance given to uncertain material gains
associated with earlier return to the labor market.
The paper is organized as follows: Section 2 presents a simple model which provides us with
a framework to test various channels by which LOC may influence maternal return to employ-
2
ment. Section 3 gives details on the data from the German Socio-Economic Panel (SOEP) that
we use. Section 4 describes our reduced-form estimation strategy based on a semi-parametric
survival model. Section 5 presents and discusses our main results: women with more inter-
nal LOC return to the labor market quicker because they are more sensitive to the career costs
associated with long leave durations. Section 6 concludes.
2 A Model of Return to Employment After Childbirth
The economic consequences of mothers’ decision when to return to the labor market after giv-
ing birth are very large (Adda et al. 2016) since human capital accumulates in employment.
Work experience can then be seen as a form of investment in human capital (Blundell et al.
2013). Coleman and DeLeire (2003) find that persons with an internal LOC make higher in-
vestments in human capital than persons with an external LOC. They argue that the benefits of
human capital investment are appreciated more strongly by individuals with an internal LOC
compared to individuals with an external LOC since the latter believe that future earnings de-
pend more strongly on factors over which they have no influence. The same argument applies
to an earlier return to work since this also leads to more work experience and can thus be seen
as an investment in human capital.
Importantly, while a more internal LOC is associated with a higher job search intensity by
unemployed persons (Caliendo et al. 2015, Spinnewijn 2015), our setting allows us to highlight
other aspects. Mothers in Germany benefit from a legal right to return to their previous position
(at the same wage) within three years after childbirth. Hence, mothers can freely decide about
the length of their leave without fearing to lose their job position or have their wage level cut on
returning—they do not need to search for a new job.5In practice very few mothers are observed
to enter new jobs after maternity leave, so that career costs do not arise as a result of differ-
ences in job search during maternity leave. In particular, mothers on leave report extremely low
5Obviously, job protection is only helpful for mothers who have a permanent contract or whose temporary job
does not expire during maternity leave. We do a robustness test of our main estimations below including only those
women in the sample who have a permanent job prior to first childbirth. Our results are robust.
3
search intensity, significantly less than unemployed and even employed individuals: Only 4%
of women on leave in the first year after birth report active job search activities in the previous
month, rising to only 9% and 12% in the second and third year after birth. This is significantly
less than a comparable sample of unemployed individuals of whom 66% report to have been
actively searching for work in the last month. We believe our evidence about the effect of LOC
on return to work thus highlights a different channel: differential appreciation of future wage
growth.
Even though mothers have the legal right to return to their previous position at the same
wage, the duration of leave might have consequences for their future career prospects. For ex-
ample, the probability of future promotions might be reduced for mothers taking long leaves.
Expectations about these consequences might vary across individuals: Women with a more in-
ternal LOC might fear stronger negative consequences than women with a more external LOC.
In the following section we formalize this intuition, allowing mothers to determine the du-
ration of leave to maximize their expected net present value of lifetime utility, comparing the
value of leave to that of work. We use this model to determine the factors we investigate in a
framework of duration models.
2.1 The Model
To inform our empirical results, we now present a simple model of return to the labor market
based on the German institutions of job protection. Mothers’ instantaneous utility u(.) depends
on the material and non-material benefits of staying at home with the child b(d), a function of
leave duration d. Utility also depends on other household income m, which can strongly influ-
ence mothers’ net income given joint household taxation in Germany. On returning to the labor
market, mothers receive their pre-childbirth wage level w0, in line with German job protection,
4
which extends until 36 months of maternity leave.
We believe that future wage growth prospects vary significantly as a function of maternity
leave duration. We do not discuss what causes these differences. A possible mechanism is that
wage offers reflect the latent productivity of returning mothers. The reduction in wage offers
would thus reflect the depreciation of human capital during maternity leave. Another reason
may be discrimination by employers.
In line with the literature on job search, we model wage growth as workers receiving job of-
fers from a wage distribution F(.) at Poisson rate λ. Dechter (2014) and Schmieder et al. (2015)
find that periods out of the labor market are followed by a reduction in job offers. We thus
allow for mothers to receive fewer offers when returning later to the labor market. Assuming
mothers accept offers with a higher wage than in their current job, this then translates to a lower
wage growth rate for mothers with longer maternity leave spells. We also allow the job arrival
rate λto be a function of individual unobserved characteristics X(including LOC—as previous
evidence suggests that LOC influences job search intensity and this may likewise be the case
for job search intensity on-the-job6).
Since benefits of earlier return to the labor market are in the future and thus uncertain,
these might be appreciated differentially by individuals depending on their LOC. LOC is a
noncognitive skill related to beliefs, it plays a role in situations of uncertainty; i.e., individuals
form expectations about the net present value of labor market re-entry according to their LOC.
We operationalize this using an expectations operator ˜
ELOC
λ, indicating that differences in LOC
translate to differences in beliefs about λ. We can then give the net present value of being in
6Though we assume that mothers do not search during leave, we do allow for job search on-the-job, both prior
to childbirth as well as after re-entry. This is how we model wage growth.
5
maternity leave, V(.), and that of being employed, W(.), discounted by the interest rate ras
V(b,m,d,w0)=u(b(d),m)
r+1
rmax [V(b,m,d),W(w0,m,d)]−V(b,m,d)
W(w,m,d)=u(w,m)
r+1
r˜
ELOC
λ"λ(X,d)Zmax W(w0,m,d),W(w,m,d)−W(w,m,d)dF(w0)#
These continuous-time Bellman equations are often presented in terms of asset value formula-
tions
r V(b,m,d)=u(b(d),m)+max [W(w0,m,d),V(b,m,d)]−V(b,m,d) (1)
r W(w,m,d)=u(w,m)+˜
ELOC
λ"λ(X,d)Zmax W(w0,m,d),W(w,m,d)−W(w,m,d)dF(w0)#.
(2)
The structure of these expressions closely follows the “fundamental equation of search” for
the unemployed (Pissarides 2001). First, the flow utility u(b(d),m) depends on other household
income mand enjoyment of staying at home during maternity leave, b(d). We expect both the
material and non-material components of b(d) to be decreasing with the duration of maternity
leave d. Second, at any point in time during the first 36 months of maternity leave, mothers
may decide to move back into employment. We expect this to occur when W(.) exceeds V(.),
explaining the maximum operator in equation (1).
Equation (2) gives the value of being employed. This has a similar structure to that of being
in maternity leave: First, flow utility is given by the instantaneous utility of work depending on
wage wand other household income. Second, with some probability λ(.), individuals receive
job offers from a wage distribution with the cumulative distribution function F(.). When an
offer w0is accepted, the employment value increases to W(w0,m,d)—the maximum operator
indicates that only offers are accepted which increase the employment value. This is the case
for any offer that pays higher wages—and this is the way we model wage growth. Note that
we allow the rate of new job offers after return to the labor market to depend on LOC (as part
of X) and maternity duration d. Additionally, we allow individuals with different LOC to have
different beliefs about the probability of job offers arising. In particular, the literature reviewed
6
above suggests that individuals with internal LOC may expect stronger consequences of own
behavior on wage growth than individuals with external LOC. Thus, individuals with an internal
LOC may expect stronger consequences of maternity leave duration on wage growth (beliefs
about ∂λ/∂d).
A young mother then chooses her optimal duration of maternity leave, d∗, such that V(b,m,d∗)=
W(w0,m,d∗), which implies
u(b(d∗),m)=u(w0,m)+˜
ELOC
λλ(d∗,LOC)Zmax W(w0,m,d∗),W(w0,m,d∗)−W(w0,m,d∗)dF(w0).
(3)
Expression (3) implicitly defines the optimal time of return to the labor market (d∗) as a
function of the level of pre-birth earnings w0(as this is equal to wage at re-entry), other house-
hold earnings m, and the option value related to future wage growth—beliefs about which may
be moderated by LOC. In order to gain an estimable expression, we note that the expression
in the expectations-term is related only to wage gains in the future, thus is a function of wage
growth. Therefore we can write
u(b(d∗),m)=u(w0,m)+˜
ELOC
∆wT∆w(d∗,w0,m,X),(4)
where T(.) is a function ensuring that (3) is satisfied. The above formulation implies the
following paths by which LOC may determine return to the labor market:
1. Pre-birth wage rate w0—moderated by other household income m. The pre-birth wage
rate may also be a function of characteristics Xand thus of LOC, e.g., as a result of
differences in eliciting outside offers due to differences in job search intensity on-the-job
before childbirth (Spinnewijn 2015).7
7We do not explicitly write w0(X) in order to keep the notation short.
7
2. Wage growth ∆w(.). This may be a function of LOC (independently of leave duration),
again as a result of differences in job search intensity on-the-job.
3. Expectations about the monetary benefits of returning earlier (conditional on all other
effects, via ˜
ELOC
λ). As noted above, by changing perceived control, LOC may influence
the relevant expectations (especially for risk-averse individuals). We test two channels
for this: (i) wage growth might be appreciated differentially by people according to their
LOC; (ii) the effect of don wage growth post-return (i.e., the wage growth penalty) might
be appreciated differentially by people according to their LOC.
4. Utility of staying at home vis-a-vis working. Individuals with different LOC may have
related (unobserved) preferences related to staying at home with their child (cf., e.g.,
Noor 2002).
Section 4 presents our empirical framework to test these channels, especially the channels
(3.i) and (3.ii). Channel (4) constitutes a residual explanation in case none of the previous
channels (fully) explain the effect of LOC on d. However, based on our empirical results below
(presented in section 5) we conclude that the residual channel can be neglected as we find
channels (1) and (3ii) fully explain the effect of LOC on leave duration.
3 Data
The empirical analysis in this paper is based on data from the German Socio-Economic Panel
Study (SOEP) 1984–2013, an annual household panel study that is representative of the pop-
ulation in Germany (Wagner et al. 2007). We use data from women who have their first child
between 1992 and 2012.
We focus on the return decision of mothers, i.e., the decision to re-enter employment after
childbirth and therefore sample women who were employed prior to first childbirth. This allows
us to take into account employment characteristics prior to first childbirth and thus to analyze
the channels of pre-birth wage level and wage growth. Certainly, the population of employed
8
women prior to childbirth is selective. Women who are not employed prior to first childbirth are
partly still in education or have some other reason (e.g., health problems) not to participate in
the labor force. For this group the decision to enter employment after childbirth is presumably
different and our determinants may not be as relevant.
Further, we concentrate on first births because transitions into employment after higher order
births are more complex to model. For example, later transitions into employment are related to
the space between births and to previous career interruptions due to previous births. Focusing
on first birth allows us to control for employment characteristics of the job prior to first birth.
We later show that our association of LOC and labor market return is not related to timing of
second childbirth.
Women’s employment status is observed in the data on a monthly basis from the fourth
month after childbirth until they return to employment or until they are censored.8. The first
three months after childbirth are ignored because new mothers in Germany are not allowed to
be employed during the first eight weeks after childbirth; this period extents to twelve weeks
for mothers of multiple births or preterm births. Censoring occurs when an individual exits the
survey or when the most recent month observed (December 2012) is reached and no transition
into employment is observed until this point. Individuals for whom we observe a transition
into unemployment9or education are discarded from the sample. In total, the sample contains
18,967 person-month observations from 966 individuals observed between a minimum of one
and a maximum of 215 monthly spells (corresponding to 18 years). Average maternity leave
duration is 19.6 months (std. dev. 27.3). For 78.8% of individuals a transition into employment
is observed with a mean number of spells of 18.2 (std. dev. 21.9). The remaining 21.2% are
censored and have a mean spell duration of 25.1 (std. dev. 41.3). Figure (1) illustrates the
density of spells.
8The monthly employment status is surveyed retrospectively in each wave; therefore, the latest SOEP wave of
2013 contains information on the monthly employment status until December 2012.
9A person is defined as unemployed if she is registered as unemployed with the Employment Office.
9
Figure 1: Density of maternity leave spells
Our measure of LOC is based on five items each surveyed in the SOEP in waves 1999,
2005, and 2010. We use the earliest measure available for each individual, i.e., preferably that
of 1999.10 The LOC items from year 1999 were answered on a 4-point scale (1 “disagree com-
pletely” to 4 “agree completely”), while in the 2005 and 2010 surveys the items were answered
on 7-point Likert scales (1 “disagree completely” to 7 “agree completely”). To make the scales
compatible, we standardize the score of each item and take the average over the five standard-
ized scores.11 The higher the LOC index, the more a woman believes she has (internal) control
over life events.12 Since the LOC items are not collected yearly in the SOEP, we assume that
LOC is stable over time. Given previous evidence this seems reasonable.13 In section 5.1 we
10For 68% of individuals we have information from 1999, for 27% from 2005, and for 5% our LOC measure
stems from 2010. Section (5.1) shows that all results hold using only individuals for whom the 1999 LOC measure
is available.
11This way of construction the LOC index is also used by Coleman and DeLeire (2003) with the LOC items from
the National Educational Longitudinal Study (NELS) and by Cebi (2007) with the LOC items from the National
Longitudinal Survey of Youth (NLSY).
12Appendix A provides details about the construction of the LOC index as well as about the LOC items.
13Cobb-Clark and Schurer (2013) use panel data to show that LOC is relatively stable even during major life
events. Personality psychologists widely agree upon mean and rank-order stability of noncognitive skills (or per-
sonality traits as they call it) in adulthood (Caspi et al. 2005, Costa Jr and McCrae 1994, Fraley and Roberts 2005,
McCrae and Costa 1994, McCrae and Costa Jr 1996, 2003, Roberts and DelVecchio 2000). Heckman et al. (2006)
and Coleman and DeLeire (2003) make similar assumptions for data availability reasons.
10
test whether changes in LOC may be driving our results.
We use a number of socio-economic and demographic control variables: In the main es-
timations we include the first child’s birth cohort (in three-year groups), log of return wage
w0—equal to the wage prior to childbirth (we use gross hourly wage, inflation-adjusted to the
base year 2011, in Euros), log of other household income m(we use inflation-adjusted net
household income net of own labor earnings, in Euros per month), predicted wage growth ten-
dency ˆ
∆wT RE ND (see section 4.1), and the estimated wage growth penalty ˆg(see section 4.2).
In extensions of our main specification we include educational attainment (university degree,
vocational degree, no professional degree), ISCED categories of educational level, labor market
experience prior to first birth in years, mothers’ birth cohorts (in groups of five years), partner
status (cohabiting), an indicator for the presence of other adults present in the household, re-
gional dummies (16 federal states in Germany), a dummy for East versus West Germany, a
dummy variable indicating whether the woman has a second child within three years after the
first birth,14 the Big Five personality traits, job type (civil servant, self-employed, white-collar,
blue-collar), and employment hours (full-time, part-time, marginal employment). Summary
statistics of these variables are reported in table 1.
4 Empirical Strategy
We want to test whether and how LOC influences return to the labor market. We now present a
strategy to estimate the relative importance of our key determinants of maternal return decisions
and in particular to assess the channels by which LOC might influence maternal return. The key
determinants and channels are deducted from our model above, as derived in section 2.
14We tried a number of different specifications to account for having another child within a short period of
time; we controlled for having a second or third child within different periods of time and for the spacing between
children (in months) and the spacing squared. The main results are always largely the same. See columns (5) and
(6) of table 4, for example.
11
Table 1: Summary statistics of covariates
Mean /Percent std. dev. Min. Max.
LOC (higher =more internal) 0.053 0.568 -1.900 1.525
Spell duration
4-6 months 25%
7-9 months 9 %
10-12 months 13%
13-18 months 18%
19-24 months 8%
25-30 months 6%
31-36 months 5%
37-48 months 5%
49-60 months 4%
60+months 7%
Child birth cohorts
1992-1994 10%
1995-1997 13%
1998-2000 16%
2001-2003 16%
2004-2006 17%
2007-2009 14%
2010-2012 14%
Return wage w054.428 22.591 4.783 205.808
Other household income m2410.929 1308.446 1 13929.91
Wage growth tendency ˆ
∆wT EN D 0.102 0.189 -0.02 1.704
ˆγ0.026 0.013 0.008 0.050
Educational degree
University degree 23%
Vocational degree 67%
No professional degree 10%
ISCED educational levels
general elementary 9.6%
middle vocational 45.5%
vocational +high school degr Abitur 11.3%
higher vocational 10.5%
higher education 23.1%
Experience 7.647 4.307 0 24.6
Mother birth cohorts
1955-1959 1%
1960-1964 5.3%
1965-1969 22.5%
1970-1974 28.3%
1975-1979 26.6%
1980-1989 16.4%
Partner in HH 94.5%
Other adults in HH 2.7%
East Germany 21.8 %
2nd child w/in 3 yrs 28.9%
Big Five
Neuroticism -0.036 0.746 -2.013 1.797
Openness 0.007 0.765 -2.371 1.546
Conscientiousness 0.009 0.757 -3.155 0.941
Extraversion 0.039 0.775 -2.668 1.359
Agreeableness 0.016 0.693 -2.277 1.114
Employment hours
Full-time employment 84.5%
Part-time employment 8.5%
Marginal employment 1.3%
Job type
Civil servant 8.7%
Self-employed 3.7%
White-collar 69.4%
Blue-collar 11.2%
Note: N =966 individuals.
12
First, we have information about return wages as these are equal to pre-birth wages. Second,
in section 4.1 we show how we estimate wage growth tendencies from our pre-birth observa-
tions. Third, section 4.2 shows how we assess the influence that maternity leave duration has on
post-return wage growth (the wage growth penalty) in different occupations. Fourth, in section
4.3 we explain how we test the role of LOC on return durations via the different channels—i.e.,
via return wage, wage growth tendency, and differential appreciation of wage growth and wage
growth penalty by LOC—in a simple estimation framework. Finally, in subsection 4.4, we give
details of the survival model used for estimation.
4.1 Individual Wage Growth Tendency
Following the model presented above, mothers receive wage w0on returning to the labor market.
Wage growth after re-entry is modeled in a search framework and the amount of job offers
arriving may depend on a number of characteristics including LOC as well as d. The model
implies that with probability λ(.) mothers receive a wage offer from distribution F(.), such that
wage growth after return is given by
wd+1−wd=λ(d,X)Zw0
w0−w0dF(w0),(5)
Wage growth depends on factors Xincluding LOC. We assume that these factors are time-
constant and affect wage growth similarly before and after childbirth. We call this the wage
growth tendency of a woman. Women with a high wage growth tendency have incentives to
return to the labor market quicker than women with a low wage growth tendency because they
have higher opportunity costs of staying out of the labor market (see section 2). To empirically
assess the influence of wage growth tendency on the return decision, we would not wish to use
actual observed wage growth after return (subject to selection issues since λ(.) also depends on
d). We thus follow a different strategy to estimate the wage growth tendency of young mothers.
We could use observationally similar men or women who have no children; however this group
may be specific in other (unobserved) dimensions. In our main specification we thus estimate
13
wage growth for women based on their pre-birth wage dynamics. We then predict wage growth
tendencies ( ˆ
∆wT RE ND ) taking into account both observable individual and job characteristics
(education, occupation, labor market experience, tenure with the same firm, and region) xi,tas
well as an unobserved individual component ηi; in addition, we control for macroeconomic
trends νt(year fixed effects). Hence, we estimate women’s pre-childbirth earnings dynamics as
∆wi,t=ηi+x0
i,tα+νt+εi,t,(6)
where ∆wi,t=log(wi,t+1)−log(wi,t). We estimate equation (6) using all periods tprior to
pregnancy (at least nine months prior to first childbirth). Note that if LOC influences on-the-
job search intensity, this will be part of the individual time-invariant component (ηi) of wage
growth.15 Appendix B provides further details and presents results of the estimation of equation
(6). We can now use our estimated parameters ˆη, ˆα, and ˆνtto predict young mothers’ wage
growth tendency at the time of childbirth ( ˆ
∆wi
T RE ND ) and argue that this is the wage growth
that women could expect on returning immediately to the labor market—the following section
considers how these wage growth prospects evolve as a function of the duration of maternity
leave. Note that this prediction takes into account the fact that women may change their jobs
and occupations over the course of their careers.16
Our identifying assumption is that the only difference between post- and pre-birth wage
growth arises as a result of different maternity leave durations d. Then, our estimable equation
(6) is a linear (first-order) approximation of wage growth given by expression (5) with d=0.
Since young mothers do not know the precise rate of their future wage growth, they will base
15This is true if the role of LOC is constant before and after birth. If, however there was some effect of LOC on
wage growth that is specific to post-birth periods, this would not be taken into account.
16In predicting counterfactual levels of wage growth for mothers out of the labor market, we could focus only
on an individual fixed effect with no other observables. However, pre-birth characteristics may be informative of
future wage growth. Being in occupation jat time ton wage growth in time t+1 might be associated with a high
probability of moving to occupation k, where earnings may be higher. For a young mother in occupation jbefore
leaving the labor market to maternity, this may thus be a valuable predictor. The time-varying characteristics we
use are subsumed in the vector xin equation (6).
14
their decisions about don predictions similar to ours.
Table 1 shows predicted yearly wage growth at the time of first childbirth ˆ
∆wi
T RE ND . Mean
wage growth is 10% with considerable variance.17 Differences in wage growth across occupa-
tions is largely consistent with previous findings—e.g., higher for women in technical occupa-
tions (14.8%) than in health service occupations (8.8%).
We also test an on-the-job search channel: if LOC importantly influences job search in em-
ployment, it should show up in a higher predicted wage growth tendency (as we include the
individual fixed effect in our prediction). However, we find that the predicted wage growth
tendency does not depend on LOC (the correlation of ˆηwith LOC is insignificant). Hence, em-
pirically, on-the-job search does not constitute an important channel by which LOC influences
mothers’ return decisions.
4.2 Wage Growth Penalty
We think that the predicted wage growth tendency reflects important individual prospects in
wage growth. We now consider by how much wage growth post return is reduced depending
on maternity leave duration d. Recall that the wage on returning to the labor market is fixed at
w0due to German job protection legislation; however, wage growth at return might be reduced
depending on leave duration d. We denote the extent of wage growth reduction the wage growth
penalty. The wage growth penalty varies across occupations. Our hypothesis is that in occupa-
tions with a high wage growth penalty, LOC has a larger effect on dthan in occupations with a
low wage growth penalty. In an extreme case, in an occupation with no wage growth penalty at
all, there is no reason for LOC to affect dbecause there is no cost (no penalty) that needs to be
attributed to internally or externally controlled factors (according to internal or external LOC).
17Expectations of strongly negative wage growth tendencies appear unrealistic, thus we exclude as measurement
error rates below -2%.
15
To test this channel, we first assess the average wage growth penalty in different occupa-
tion categories j. We estimate wage growth after return in different occupations as a function
of the leave duration din combination with the predicted wage growth tendency in the same
occupation j, i.e., we estimate
∆wi,d=1−gj(d)ˆ
∆wT RE ND
i+ιi,(7)
where ∆wi,d=log(wi,d+1)−log(wi,d). The wage growth penalty gj(d) is thus the extent by
which wage growth at return is reduced compared to the wage growth tendency. In our estima-
tion we specify a linear penalty function gj(d)=γjd. This means that the wage growth penalty
for each month of leave is constant. In an occupation with a zero wage growth penalty, we
would find γ=0; i.e., in this occupation, wage growth after return (∆wi,d) corresponds to the
predicted wage growth tendency ( ˆ
∆wT RE ND
i) plus noise (ιi).
Since wage growth post return is estimated as a function of dand dis a choice variable,
issues of selection may arise. Most obviously, women with certain unobservable characteristics
choose higher dand at the same time have lower wage growth. Our estimation procedure will
not be affected if these unobservable characteristics are constant before and after birth and affect
wage growth before and after birth equally; this follows from the fact that we estimate wage
growth post return as a function of individual wage growth tendency ( ˆ
∆wT RE ND
i). Since the wage
growth tendency (see the preceding section) is a prediction based on estimated wage growth in
pre-birth periods (including individual fixed effects, thus including differences in LOC), it is
generated independently of d. Only if the unobservable characteristics affect wage growth be-
fore birth differently than wage growth after birth, the γ-coefficients might be estimated with
bias. Does this bias our results? Not necessarily, since we are not primarily interested in the
predicted effect of don wage growth post return, ˆγ; rather our aim is to compare wage growth
penalties across occupations in order to test whether the effect of LOC on the return decision
is moderated by the industry’s wage growth penalty—this will depend more on the quality of
16
the ranking, less on the absolute values. Therefore, if the distribution of the relevant unobserv-
able characteristics and thus selection into different dwas similar in all occupation categories
in the way that ˆγjrelative to ˆγkis close to the relation between the true wage growth penalties
in occupations jand kfor all j, selection would not lead to false conclusions in our context. By
contrast, if the bias varies strongly between occupation categories, such that the order of wage
growth across occupations is affected, we face a problem. We are reassured by the fact that the
ranking across different occupations’ wage growth penalties seem very plausible (see below,
full results are relegated to appendix C).
Finally, another selection problem might occur if gj(d) is importantly non-linear and thus
the penalty for a further month of leave depends on the level of d. However, we find the simple
linear form g(d)=γdperforms as well as a quadratic form. We also tested further polynomials
and other functional forms such as piece-wise linear transformations. If this function is indeed
linear, then the penalty for one more month of leave is constant across d. Furthermore, the dis-
tribution of the leave duration dseems to be very broad in all occupation categories (cf. figure
A2 in appendix C).
Table A5 in appendix C gives the results of the estimation of equation (7); the distribution
of individuals in the eight occupation categories at return is provided in table A6. The estimated
wage growth penalties vary between 0.8 and 5 percent per month of maternity leave. The mean
wage penalty in our sample is 2.6%.18 This implies that a significant proportion of the wage
growth tendency is forgone as a result of maternity leave. We now test how these different leave
penalties ˆγjinfluence maternal return decisions, especially as moderated by LOC.
18The rank order of occupations’ estimated γseems to be plausible: The occupation category with the highest
estimated γ(5%) is cat. 4 (“merchants of services or associated occupations”). The occupation category with the
lowest estimated γ(0.8%) is cat. 8 (“occupations in transport, regulation, security, writing, translation, librarians,
artists, and other service occupations”). The occupation category with the second lowest estimated γ(1.2%) is cat.
6 (“occupations in health services”.
17
4.3 Determinants of Returning to Work
We now estimate determinants of young mothers’ return to work as a function of individuals’
wage level w0,19 other household earnings m, wage growth tendency ˆ
∆wT RE ND , occupation-
specific wage growth penalty ˆg, and LOC.
LOC influences mothers’ maternity leave controlling for the wage level and the wage growth
tendency. This means that the effect of LOC on return decisions is not merely an artefact of
selection (by LOC) into particular jobs with higher wages or higher wage growth.20 If LOC
influences maternity leave duration conditional on wage level and wage growth, this may be
because mothers with different LOC have different perceptions of human capital depreciation
when they are out of employment. By contrast, if we find no evidence for any of these human
capital stories, we may conclude that the evidence points towards differences in preference for
time spent with young children. The condition for women to return, including multiple channels
of LOC influencing return, is
u(w0,m)+˜
ELOC
∆wT∆w(d,w0,m,LOC)>u(b(d),m).(8)
Expression (8) implies the following candidates for influencing return decisions:
w0,i,ˆ
∆wi
T RE ND ,LOCi,ˆ
∆wi
T RE ND ×LOCi,ˆγj,ˆγj×LOCi(9)
where the interaction terms show that LOC may influence maternal return decisions by
moderating the weight of other determinants.
19This is the pre-childbirth wage level, but also the wage at which most women can be expected to return to the
labor market.
20Note that since dmay influence wage growth, we use wage growth predictions from pre-maternity periods as
explained above in subsection (4.1).
18
4.4 The Duration Model
Since return to the labor market is a binary choice, a linear model would not be appropriate;
therefore, we focus on the hazard rate of returning to work, θ(.), i.e., the probability at any
duration diof returning to work given Xi. We use a mixed proportional hazard framework, such
that
θ(d,X, η)=θ0(d)exp(β0X) exp(ξ),(10)
where θ0(d)is the baseline hazard which depends on maternity leave duration dbut not on
the covariates X.βdescribes the parameter vector to be estimated and ξis unobserved hetero-
geneity.
Note three important features of this empirical specification: First, in common with a large
literature using survival models, we assume a proportional hazards framework: conditional on
ξ, absolute differences in the covariates imply proportionate differences in the hazards at each
d.21
Second, we allow for a flexible semi-parametric form of duration dependence (i.e., the func-
tional form of θ0(.)) by including interval-specific intercepts. Given that each time interval in
the data is of unit length (one month)22 the discrete time hazard can be written as
h(d,X, ξ)=1−exp[−exp(δkDk+β0X+ξ)]
21We find descriptive evidence in favor of this assumption by looking at the hazard by different covariates.
Merely one determinant, w0, appears to change its influence on the hazard. However, conditional on our estimated
ˆηthis is no longer the case, consistent with the mixed proportional hazard formulation.
22This requires a complementary log-log transformation, see Jenkins (2005, p.41–42).
19
where dis the number of months after childbirth and Dkis an indicator equal to one if month
dlies within the kth interval of maternity duration.23 The parameter vector δis estimated by the
model along with the parameter vector β.
Third, we include unobserved heterogeneity in a semi-parametric form also, without as-
suming a specific distribution for the random effect. We assume a discrete distribution of types
following Heckman-Singer:24 the likelihood contribution of a person with spell length dmonths
is thus
L=
M
X
m
π(ξm)Lξ(d,X, ξm)
where Mis the number of groups of unobserved types25 and πis the probability of belonging
to type ξ. Conditional on the belonging to type ξ, the likelihood is then
Lξ(d,X, ξ)= h(d,X, ξ)
1−h(d,X, ξ)!cd
Y
d0=11−h(d0,X, ξ),(11)
where cis the censoring indicator. The parameters πand ξare estimated by the model
together with βand δ.
5 Estimation Results
The main estimation results are reported in table 2. We first show the basic impact of LOC
on the timing of labor market return not controlling for the other determinants. Column (1)
shows that LOC significantly influences the decision about maternity leave duration. Mothers
23That is, the model contains one dummy variable for each group of duration: 4-6 months (0-3 are omitted for
reasons outlined in section (3)); 7-9 months; 10-15 months; 16-21 months; 22-27 months; 28-33 months, 46-57
months and 58+months.
24The model has also been estimated with a parametrically specified random effect. Gamma distributed and
normally distributed ηwere tested, as well as a specification without any random effect. The sign and significance
levels of paremter estimates do not change compared to the main results discussed in section (5) below.
25We use two mass points in estimation, we also tested specifications with three and four mass points. However
Bayesian and Akaike information criteria point to the model with two mass points being suited best.
20
with more internal LOC have a higher hazard rate of return to employment; i.e., women with
a high belief in internal control return to employment significantly earlier than women with
a high belief in external control. To illustrate this result, figure 2 plots the hazard functions
for two individuals who are equal in all characteristics but LOC.26 The solid line in the graph
represents a woman with high internal LOC (75th percentile of the LOC score), the dashed
line a woman with more external LOC (25th percentile of the LOC score). The solid line runs
above the dashed line, illustrating that a person with internal LOC is more likely to return to
employment in each period. These differences in hazard rates create differences in the proba-
bility of returning to employment. For example, 36% of women with a high internal LOC score
(at the 75th percentile) return to employment twelve months after childbirth; this is more than
16% higher than for women with a low internal LOC score (at the 25th percentile). Moving
from one standard deviation below to one standard deviation above the mean of LOC increases
median return time by two months, i.e., by over 13%. The magnitude may seem modest but cu-
mulative knock-on career effects must also be taken into account. Finally, the estimation results
that we present next show that this LOC effect masks important effect heterogeneity depending
on mothers’ occupation: LOC really influences return decisions in occupations with high wage
growth penalties.
The step pattern of the graph in figure 2 is a result of our modeling the duration dependency
(see section 4). The first clear peak of the hazard function is just after 12 months, suggesting
that there is a relatively high probability of returning to employment around the time the child
turns one year. The second clear peak is observed in the graph during months 33 to 37, that is,
around the third year after childbirth. This is likely to be a result of the legal right to return to
the previous job, which expires after three years. We discuss whether this specific institutional
aspect may be driving our results in the following subsection. A smaller peak of the hazard
function is observed when the child is six years old, which is the usual school entry age in
26The graph uses the results of column (4) of table 2 (see below). For the graph all non-binary covariates other
than LOC are set to their mean and the child birth cohort is set to the most recent category (i.e., 2010-2012).
Considering women with different values in these characteristics would not change the pattern of the graph but
only rescale it—due to the proportional hazards feature of the model.
21
Germany. The irregular pattern of the hazard function over time confirms the importance of
allowing for non-parametric duration dependency.
Note: The estimated parameters of model (4) of table 2 are used to calculate the hazard functions. All non-binary
covariates other than LOC are set to their mean. We observe the child birth cohort of 2010-2012.
Figure 2: Hazard function by level of LOC
22
Table 2: Determinants of return to employment after first childbirth (dependent variable: hazard rate of return), discrete semi-parametric hazard
estimation (Heckman-Singer model)
(1) (2) (3) (4) (5) (6) (7) (8)
Locus of control LOC 0.300*** 0.287* 0.215*** 0.302*** 0.298*** 0.256*** 0.244** -0.112
(0.100) (0.160) (0.081) (0.098) (0.110) (0.098) (0.098) (0.204)
Return wage log(w0) 0.203 0.174* 0.389** 0.388** 0.284* 0.306* 0.303*
(0.212) (0.100) (0.159) (0.160) (0.152) (0.162) (0.163)
Other household income log(m) -0.337*** -0.436*** -0.435*** -0.316*** -0.313*** -0.312***
(0.034) (0.072) (0.076) (0.055) (0.052) (0.052)
W growth tendency ˆ
∆wT EN D -0.005 -0.012 0.259 0.300 0.280
(0.253) (0.278) (0.261) (0.261) (0.261)
Interaction ˆ
∆wT EN D ×LOC 0.038
(0.481)
W growth penalty ˆγj-6.493 -6.635
(4.381) (4.171)
Interaction ˆγj×LOC 13.444**
(6.461)
Duration dependency Yes Yes Yes Yes Yes Yes Yes Yes
Child birth cohort Yes Yes Yes Yes Yes Yes Yes Yes
Constant -3.352*** -3.983*** -2.115*** -2.103*** -2.106*** -1.473*** -1.318*** -1.215**
z0.929** 0.967 0.198 0.841*** 0.839*** 0.489** 0.434** 0.435**
p(pct) 0.496 0.622 2.191 0.590 0.590 0.632 0.691 0.692
Number of person-spells 18967 18967 18967 18967 18967 12262 12262 12262
Number of persons 966 966 966 966 966 748 748 748
Note: Columns 6-8 are based on a smaller sample since information necessary to estimate wage-growth depreciation terms is not available for all individuals. Standard
errors are in parentheses. * p<0.10, ** p<0.05, *** p<0.01. Data source: SOEP 1992–2013, authors’ calculations.
23
We now test whether the effect of LOC on return is related to material factors. First, column
(2) of table 2 introduces return wages w0and finds that these do not have an effect on return as
long as other household income mis not included. As soon as the latter is included, we find
a positive effect of the wage level on return. This is consistent with the idea that the marginal
utility of income (and, given the German tax system, net income) depends on other household
income.27 Further, we find that the coefficient related to LOC is slightly reduced from .300 in
column (1) to .215 in column (3). This suggests that one reason why individuals with more
internal LOC return earlier is that they have higher wages. However, even controlling for the
wage level, LOC still has a significantly positive effect on leave duration.
In a second step, we introduce workers’ predicted wage growth tendency ˆ
∆wT RE ND (from
section 4.1) in column (4). It does not appear to influence return decisions; i.e., wage growth
tendencies seem not to be a channel by which LOC influences return decisions. Further, the in-
teraction between wage growth tendency and LOC in column (5) is not a significant determinant
of return decisions either; put differently, the effect of wage growth tendency is not differentially
appreciated depending on LOC.
Third, column (7) shows that the average wage penalty in an occupation does not signifi-
cantly affect the return hazard. This suggests that women do not self-select into occupations
with a certain ˆgjaccording to their planned d. However, introducing the interaction between
ˆgjand LOC reveals that wage growth penalties do play a role for mothers’ return decision
for individuals with high LOC scores. In fact, this appears to be the main channel by which
LOC influences maternal return decisions: the direct, unconditional measure of LOC is now no
longer significant. Given that we have fewer data-points on occupational categories for these
analyses, we also test our benchmark results using this smaller sample to ensure that any results
are related to sample selection. Column (6) shows that estimates are similar in this sample.
27We tested a number of functional forms for this relationship, including own wage in levels and logs and relative
to total household wage and find the results of column (4) to be largely robust.
24
We next turn to non-wage dimensions of human capital. We find no evidence that the in-
troduction of education variables influences the effect of LOC on return decisions, as both the
introduction of simple dummies for formal education (in column (2) of table 3) as well as more
detailed ISCED education categories (in column (3) of table 3) confirm. While there may be a
small effect of experience on return (controlling for wage and wage growth), the effect is only
significant at 10% (column (4) of table 3). When we introduce maternal birth cohorts (thus im-
plicitly controlling for mothers’ age at first childbirth) on top of child birth cohorts , the results
change very little, as can be seen from column (5) of table 3.
We include a range of further covariates to test the sensitivity of our results. Among those
covariates we find regional dummies to have most significance (see column (2)-(4) of table 4):
living in East Germany is related to a quicker return—a finding that may relate both to dif-
ferences in child-care provision and social norms. Cohabiting with a partner and the presence
of another adult living in the same household do not appear to affect return to employment
(columns (6) and (7) of table 3) conditional on household income.
25
Table 3: Determinants of return to employment after first childbirth (dependent variable: hazard rate of return), discrete semi-parametric hazard
estimation (Heckman-Singer model) — with further covariates (I)
(Benchmark) (2) (3) (4) (5) (6) (7)
Locus of control LOC 0.302*** 0.298*** 0.302*** 0.304*** 0.287** 0.261** 0.313***
(0.098) (0.103) (0.101) (0.100) (0.121) (0.102) (0.097)
Return wage log(w0) 0.389** 0.203 0.207 0.336** 0.331 0.379*** 0.401**
(0.159) (0.159) (0.154) (0.164) (0.202) (0.147) (0.159)
W growth tendency ˆ
∆wT EN D -0.005 -0.178 -0.134 0.084 0.074 -0.050 0.004
(0.253) (0.265) (0.270) (0.286) (0.270) (0.242) (0.253)
Other household income log(m) -0.436*** -0.451*** -0.463*** -0.432*** -0.450*** -0.468*** -0.437***
(0.072) (0.079) (0.081) (0.079) (0.072) (0.067) (0.069)
University degree 0.633***
(0.137)
No professional degree -0.155
(0.249)
ISCED: middle vocational 0.247
(0.257)
ISCED: vocational +high school degr (Abitur) 0.298
(0.330)
ISCED: higher vocational 0.415*
(0.250)
ISCED: higher education 0.926***
(0.286)
Experience 0.072
(0.049)
Experience2-0.004*
(0.002)
Maternal cohorts 6 cat.
Cohabiting 0.522
(0.342)
Other adult in household 0.316
(0.368)
Duration dependency Yes Yes Yes Yes Yes Yes Yes
Child birth cohort Yes Yes Yes Yes Yes Yes Yes
Constant -2.103*** -1.223 -1.371* -2.140*** -2.745** -2.323*** -2.171***
z0.841*** 0.847*** 0.862*** 0.809*** 0.842*** 0.867*** 0.823***
p(pct) 0.590 0.589 0.599 0.595 0.602 0.587 0.597
Number of person-spells 18967 18946 18857 18967 18967 18967 18967
Number of persons 966 963 956 966 966 966 966
Note: Standard errors are in parentheses. * p<0.10, ** p<0.05, *** p<0.01. Data source: SOEP 1992–2013, authors’
calculations.
26
Table 4: Determinants of return to employment after first childbirth (dependent variable: hazard rate of return), discrete semi-parametric hazard
estimation (Heckman-Singer model) — with further covariates (II)
(Benchmark) (2) (3) (4) (5) (6)
Locus of control LOC 0.302*** 0.195*** 0.195*** 0.194*** 0.305*** 0.294***
(0.098) (0.072) (0.072) (0.073) (0.098) (0.083)
Return wage log(w0) 0.389** 0.237** 0.238** 0.242** 0.391** 0.180*
(0.159) (0.105) (0.103) (0.100) (0.161) (0.097)
W growth tendency ˆ
∆wT EN D -0.005 -0.346 -0.335 -0.243 -0.004 -0.262
(0.253) (0.238) (0.237) (0.233) (0.253) (0.263)
Other household income log(m) -0.436*** -0.342*** -0.341*** -0.334*** -0.433*** -0.334***
(0.072) (0.035) (0.039) (0.034) (0.075) (0.032)
Regional dummies 16 cat. 16 cat.
Other adult in household 0.674* 0.542
(0.371) (0.342)
Cohabiting 0.282 0.223
(0.275) (0.206)
East Germany 0.243**
(0.100)
Second child within 3 yrs -0.076 -0.179
(0.137) (0.112)
Interaction second child ×LOC -0.252
(0.161)
Duration dependency Yes Yes Yes Yes Yes Yes
Child birth cohort Yes Yes Yes Yes Yes Yes
Constant -2.103*** -1.855*** -2.164*** -2.626*** -2.115*** -2.023***
z0.841*** 0.175 0.164*** 0.161*** 0.849*** 0.169***
p(pct) 0.590 0.920 0.934 0.928 0.581 0.926
Number of person-spells 18967 18967 18967 18967 18929 18929
Number of persons 966 966 966 966 960 960
Note: Standard errors are in parentheses. * p<0.10, ** p<0.05, *** p<0.01. Data source: SOEP 1992–2013, authors’
calculations.
27
Finally, we find evidence of significant unobserved heterogeneity in most specifications.
Since our estimation is based on random effects, these are not treated as individual effects to
be estimated but as nuisance parameters that are integrated out. To gain an understanding of
the dimensions which the random effects are covering,28 we recover estimated probabilities by
considering the posterior distribution of types after convergence in our benchmark specification
(see table 2). The only significant correlate of type probabilities we find is the East dummy,
viz., some regional dummies if we include one for each region. Most obviously, this may be
related to differences in child-care provision across different regions.
5.1 Robustness Tests
In this subsection, we discuss a number of challenges to our findings. A first concern refers to
the endogeneity of LOC with childbirth. We assume that LOC is constant over time for each
individual, in line with previous research. Our LOC measure comes from different years (1999,
2005, and 2010), for some women this is prior to having the first child, while for others it is
after having the first child. If LOC evolves over time, certain characteristics of a newborn child
(e.g., poor health) might cause a mother to adopt a more external LOC and, at the same time,
extend her leave duration. To test the robustness of our results, we first create a dummy for
women who had their first child after their LOC was measured. We find this indicator not to
be significantly correlated with LOC; the same is true if we regress LOC on this indicator in-
cluding further covariates such as children’s birth cohorts. This provides first evidence for the
robustness of our results. In a second step we carry out the following robustness tests of our
main estimations: First we restrict the sample to mothers of children born after LOC is surveyed
(column (2) of table 5). To make the sample even more homogeneous, second, we restrict the
sample to mothers for whom LOC is measured in 1999 (column (3) of table 5). Third, we re-
28We are cautious about interpreting the predicted probabilities. Heckman and Singer (1984) note that their
estimation method “recovers the structural parameters of the underlying models very well but does not accurately
estimate the distribution of unobservables even in very large samples.” If we had multiple spells per individual, we
might have more confidence in the identification of these parameters (for a discussion on the benefits of multi-spell
data, see van den Berg (2001)).
28
strict the sample to mothers for whom LOC is measured in 1999 and who have their first child
after LOC is measured (i.e., after 1999) (column (4) of table 5). The results are robust to all
these restrictions. This suggests that our main result is unlikely to be affected by LOC evolving
as a result of childbirth.
A second concern relates to LOC picking up other noncognitive skills that might be the true
reason for a delayed or quicker return to work. We thus now control for important other noncog-
nitive skills in order to minimize the probability of LOC picking up other skills or traits. We
focus on the Big Five personality traits identified in personality psychology: neuroticism,open-
ness to experience,conscientiousness,extraversion, and agreeableness (John and Srivastava
1999, McCrae and Costa Jr 1996, 1999). The Big Five concept is among the best-established
models in personality psychology and widely used in empirical research (Caspi et al. 2005).
Column (5) of table 5 reports the results of including these five dimensions in the benchmark
specification.29 We find that extraversion is associated with faster return to the labor market. We
note that the effect of LOC on the outcome persists—in fact, the magnitude has increased. The
effect of LOC on the outcome seems not to be due to LOC picking up other personality traits.
Finally, LOC could pick up the effect of cognitive ability, which is likely correlated. However,
the most important channel of cognitive skills affecting mothers’ return decision is predicted to
go through their reward on the labor market, and should be reflected in wage levels and growth.
We have demonstrated that LOC influences return conditional on wage level and growth, argu-
ing that LOC operates by different appreciation of wage growth penalties.
29The SOEP wave 2005 provides a set of fifteen items of the big five—three for each of the five dimensions—
that were answered on a 7-point Likert scale. We use standardized mean answer scores as indicators for the five
variables. For more information on the implementation of the Big Five traits in the SOEP survey as well as on the
reliability and validity, see Dehne and Schupp (2007).
29
Table 5: Determinants of return to employment after first childbirth (dependent variable: hazard rate of return), discrete semi-parametric hazard
estimation (Heckman-Singer model) — robustness tests
(Benchmark) (2) (3) (4) (5) (6) (7)
Locus of control LOC 0.302*** 0.265* 0.396*** 0.283** 0.270**
(0.098) (0.142) (0.115) (0.111) (0.106)
Locus of control LOC from 1999 0.343*** 0.365**
(0.102) (0.145)
Return wage log(w0) 0.389** 0.386 0.428** 0.375 0.365** 0.409** 0.127
(0.159) (0.249) (0.175) (0.257) (0.165) (0.170) (0.149)
W growth tendency ˆ
∆wT EN D -0.005 -0.294 0.173 -0.247 0.182 -0.168 -0.533
(0.253) (0.386) (0.276) (0.342) (0.321) (0.299) (0.401)
Other household income log(m) -0.436*** -0.615*** -0.412*** -0.592** -0.447*** -0.455*** -0.329***
(0.072) (0.236) (0.073) (0.238) (0.064) (0.069) (0.050)
Neuroticism 0.109
(0.088)
Openness -0.004
(0.105)
Conscientiousness 0.042
(0.090)
Extraversion 0.241***
(0.087)
Agreeableness -0.033
(0.098)
Part-time employment -0.282
(0.235)
Marginal employment 1.010
(0.814)
Civil servant 0.392**
(0.184)
Self-employed 1.591***
(0.419)
Blue-collar worker 0.100
(0.200)
Duration dependency Yes Yes Yes Yes Yes Yes Yes
Child birth cohort Yes Yes Yes Yes Yes Yes Yes
Constant -2.103*** 2.192 -2.346*** 2.246 -1.736** -2.054*** -2.428***
z0.841*** 1.215*** 0.937*** 1.496*** 0.652*** 0.914*** 0.171***
p(pct) 0.590 0.631 0.548 0.586 0.701 0.582 0.942
Number of person-spells 18967 7394 15309 5614 16547 18967 6944
Note: Standard errors are in parentheses. * p<0.10, ** p<0.05, *** p<0.01. Data source: SOEP 1992–2013, authors’
calculations.
30
A third major concern could be that women with certain LOC select themselves into spe-
cific job types—e.g., civil service or self-employment—or choose especially family friendly
employment hours—like part-time schedules—and that these job types and employment hours
are, at the same time, more compatible with a longer or shorter leave. If this is the case, then
LOC would not directly affect the decision to return to employment but rather indirectly through
the choice of the job type and employment hours prior to childbirth. To check this channel, we
introduce a set of variables characterizing the job type occupied prior to childbirth: civil servant,
self-employed, blue collar worker, and white collar worker (reference category). Additionally,
working hours categories are introduced: we differentiate part-time and marginal employment
(tax-free minijobs) from full-time employment. The results with the additional covariates are
reported in column (6) of table 5. Although it emerges that some of the variables have additional
predictive power—not surprisingly, self-employed women appear to return to employment more
quickly—the main results remain stable. If anything, the magnitude of the coefficient on LOC
has again increased, suggesting that it is not via choice of job type that LOC influences mater-
nity leave durations.
Fourth, we may be interested to assess to what extent the result is driven by mothers’ deci-
sion to have a second child. Differences in LOC may relate to different preferences about child
spacing rather than labor supply. If individuals (with different LOC levels) have the same total
number of children, our results could be consistent with the same life-time labor supply despite
later returns of mothers who have their children with shorter intervals. We address this concern
in two ways: First, we focus on women who are observed to only have one child in the sample,
yielding a sample of 6944 spells from 481 women.30 Column (7) in table 5 shows that the effect
of LOC remains largely unchanged in this smaller sample. Second, we find similar patterns
when we censor women at the moment they become pregnant with their second child.31 Finally,
note that average child spacing in Germany is over 3 years, whereas median return to the labor
market occurs after around 10 months in our data. These all point against a strong influence of
30We attempted to restrict the sample to women who have completed their fertility (age 40 plus), but many
women are younger at the last observation in the sample, yielding a too small sample for estimation.
31Results are available on request.
31
second children on return decisions after the first child.
Fifth, our model is based on labor supply decisions driving return. This makes sense for
returns to employment within the institutional framework of job protection. Within this period,
we are confident that the most important effects will be labor-supply driven. The loss of job
protection at 36 months of maternity leave is an important labor demand effect: Mothers are
no longer guaranteed their previous wage but may need to search for a job. If this institutional
feature is an important determinant of maternal return, we should observe a significant effect of
this threshold on maternal returns. Figure (1) shows the density of (completed) maternity leave
durations. We find, first, that only a small fraction—less than 15%—of maternal returns occur
after 36 months. Second, there appears to be some bunching at the threshold, i.e., a positive
mass of individuals choose exactly 36 months. We thus test the importance of the other deter-
minants of the model in explaining durations by adding a specific shifter for return at around
36 months.32 While the hazard at this specific date is indeed discretely higher than for other
periods controlling for all the other determinants,33 controlling for this specific group does not
alter the magnitude or significance of LOC in return decisions.34 Furthermore, it is not the
case that mothers react to the institutional set-up differentially according to their LOC. Finally,
conditional on this institutional threshold, we still find that the main effect of LOC runs by a
differential appreciation of the wage growth penalty. We thus believe that while labor demand
constraints resulting from differential job protection rules are an important part of return deci-
sions, our results are robust to taking this aspect into account.
32After inspection of the data, we include individuals who report 37 months of maternity leave in this category.
33We believe this bunching is not a function of a sudden change in underlying preferences of mothers. Next to
the labor demand effect discussed here, the difference could also stem from effects of child-care provision at age
three.
34The results of this robustness check are available on request.
32
6 Conclusion
The present study investigates the effect of locus of control (LOC) on the duration of mothers’
leave after first childbirth. Using data from the German Socio-Economic Panel Study (SOEP)
we find that women with a more internal LOC return to employment faster. We show that this
is not mainly a result of differences in wage levels or wage growth rates. Rather, we find that
the main reason that individuals with more external LOC return later is that they react less to
what we call wage growth penalties. These are estimated reductions in mothers’ wage growth
prospects—reducing wage growth more the longer women stay out of the labor market after
childbirth. The paper contributes to understanding the impact that LOC has on important eco-
nomic outcomes.
First, if mothers with external LOC do not fully take into account the effects of long labor
market breaks, they will stay out of the labor market for too long. This may cause large career
costs for mothers with external LOC.
Second, our results indicate that changes in financial incentives influence LOC-external
mothers’ return decisions less strongly if they accrue in the future and are seen as uncertain.
This insensitivity is relevant to policy: For example, subsidizing child-care costs financed by
higher taxation after return to work could be welfare-enhancing. Also, this lack of sensitivity
means there will be less negative moral hazard issues associated with transfers later in life. Con-
sider two reforms that were instituted in Germany: (i) providing pensioners a minimum level
of state pension; (ii) increasing pension payments to mothers with long periods out of the labor
market (M¨
utterrente). Both reduce incentives to work for individuals with low contributions. If
pension payments are seen as uncertain, women with external LOC will reduce their labor sup-
ply less as a reaction to the more generous new pension rules. This may then suggest an overall
smaller reaction of female labor supply to pension changes—consistent with early evidence of
the M¨
utterrente.
33
Third, noncognitive skills such as LOC—whilst relatively stable across adult life—are not
predetermined but develop during childhood and adolescence (Cunha and Heckman 2007, 2008).
Family and education can importantly determine non-cognitive skills. Policies which foster in-
ternal LOC may especially benefit girls by fostering their long-term career prospects.
34
Appendix A
The Construction of the LOC Variable and Factor Analysis
LOC is typically conceived as a unidimensional concept with “internal” and “external” as op-
posite poles of the same dimension (see, e.g., Piatek and Pinger 2015, Caliendo et al. 2015,
Coleman and DeLeire 2003, Cebi 2007). LOC describes the extent to which individuals believe
they can determine events in their own lives (internal) as opposed to feeling dependent on fac-
tors that are outside of their control (external). The unidimensional LOC index we use in our
estimations is based on five items surveyed in the SOEP; the items are given in table A1. We
standardize the score of each item and take the average over the five standardized scores. All
items but the first are reverse coded such that a higher LOC index reflects a more internal LOC.
Running an exploratory factor analysis over the five items in our sample confirms the as-
sumption of one single latent factor behind the items: only the eigenvalue of factor 1 is larger
than one; the others are clearly below. The relative signs of the factor loadings correspond to
the way the items are included in the LOC score (i.e., item 1 is reverse compared to the other
items). The results of the factor analysis are given in table A1.
Discussion of Alternative Versions of Constructing a LOC Variable
In the SOEP, the five LOC items are part of a 10-item battery that are related to LOC but that
do not all reflect the original concept of LOC by Rotter (1966). Table A2 gives the 10 items
(the five items used in our LOC score are in bold). We excluded five items from our main
LOC index for the following reasons: Item Q2 (Compared to other people, I have not achieved
what I deserve) does not reflect internal versus external LOC but rather measures an individuals
sense of fairness (Weinhardt and Schupp 2011, Piatek and Pinger 2015). Item Q4 (If a person
is socially or politically active, he/she can have an effect on social conditions) rather measures
an individuals degree of social involvement (Weinhardt and Schupp 2011, Piatek and Pinger
35
Table A1: Factor analysis with five LOC items
Factor1 Factor2
Q1 How my life goes depends on me (Wie mein Leben verl¨
auft,
h¨
angt von mir selbst ab).
-.2598271 .0660362
Q3 What a person achieves in life is above all a question of fate or
luck (Was man im Leben erreicht, ist in erster Linie eine Frage von
Schicksal oder Gl¨
uck).
.407111 .0884543
Q5 I frequently have the experience that other people have a con-
trolling influence over my life (Ich mache h¨
aufig die Erfahrung,
dass andere ¨
uber mein Leben bestimmen).
.575017 -.0324929
Q8 The opportunities that I have in life are determined by the social
conditions (Welche M ¨
oglichkeiten ich im Leben habe, wird von
den sozialen Umst¨
anden bestimmt).
.3125508 .1449504
Q10 I have little control over the things that happen in my life
(Ich habe wenig Kontrolle ¨
uber die Dinge, die in meinem Leben
passieren).
.6047107 -.0751981
Eigenvalue
Factor 1 1.027257
Factor 2 .0399061
Factor 3 -.0181629
Factor 4 -.0878149
Factor 5 -.2522844
N=966 individuals.
36
2015). Item Q6 (One has to work hard in order to succeed) measures the degree to which
someone attributes success to effort, which would be part of an internal LOC. However, for
respondents who disagree with item Q6’s statement it is unclear whether they attribute success
to stable skills (which would also be part of an internal LOC) or to forces outside of themselves
(which would be part of an external LOC). Therefore, this item is inappropriate for measuring
internal versus external LOC. Item Q7 (If I run up against difficulties in life, I often doubt my
own abilities) displays a similar problem as item Q6; for respondents who disagree with the
item’s statement it is unclear whether they attribute difficulties to the lack of own effort (which
would also be part of an internal LOC) or whether they attribute difficulties rather to forces
outside of themselves (which would be part of an external LOC). Item Q9 (Innate abilities are
more important than any efforts one can make) explicitly measures the weight a person places
on effort versus skills—while both would be subsumed under an internal LOC according to Rot-
ter (1966). For these reasons, we exclude these five items from our main LOC index. The five
items we do include in our LOC index (the items printed bold) are appropriate for measuring
the unidimensional trait of internal versus external LOC. These items correspond to the LOC
items in the National Educational Longitudinal Study (NELS), which has been used e.g., by
Cebi (2007), Coleman and DeLeire (2003) and Heckman et al. (2006). The five inappropriate
SOEP items have no corresponding items in the NELS.
Caliendo et al. (2015) construct a similar LOC index as we do but they include nine out of
the ten SOEP items, dropping only item Q4 (and reverse coding items Q2, Q3, Q5, Q7, Q8,
and Q10). We run a robustness test with a LOC index using their items; the main results of the
paper do not change (see column (2) of table A3). Piatek and Pinger (2015) use only items Q3,
Q5, Q7, Q8, and Q10, i.e., they use similar items as we do but substitute Q1 by Q7. We run
a robustness test with a LOC index using the items selected by Piatek and Pinger (2015); the
results of our paper do not change (see column (3) of table A3).
37
Running a factor analysis over the ten items confirms the strength of the first latent factor.
The results of the factor analysis are given in table A2. The factor loadings suggest dropping
items Q4, Q6, and Q9 which do not load on the latent factor. A similar result of a factor analy-
sis of the LOC-items from the SOEP is found for example by Specht et al. (2013).35 Since the
factor analysis does not imply to exclude items Q2 and Q7, we do another robustness test: even
though items Q2 and Q7 do not fit into the original one-dimensional concept of LOC as argued
above, we include these items in the LOC index (reverse coding items Q2, Q3, Q5, Q7, Q8, and
Q10); our main results are robust to this modification (see column (4) of table A3).
A further robustness test is conducted employing a factor score of LOC weighting the items
according to the factor analysis reported in table A1 (similar to Piatek and Pinger (2015)) in-
stead of weighting all component items of the LOC index equally. The results are reported in
column (5) of table A3. Our main results are robust to using this alternative LOC-score.
35They also drop items Q4, Q6, and Q9 ex-ante, using only seven of the ten items.
38
Table A2: Factor analysis with ten LOC items
Factor 1 Factor 2 Factor 3
Q1 How my life goes depends on me (Wie mein Leben
verl¨
auft, h¨
angt von mir selbst ab).
-.2244663 .3317205 .1397904
Q2 Compared to other people, I have not achieved what
I deserve (Im Vergleich mit anderen habe ich nicht das
erreicht, was ich verdient habe).
.4855427 .0709979 -.0730634
Q3 What a person achieves in life is above all a ques-
tion of fate or luck (Was man im Leben erreicht, ist in
erster Linie eine Frage von Schicksal oder Gl¨
uck).
.405968 .0981938 -.1910283
Q4 If a person is socially or politically active, he/she
can have an effect on social conditions (Wenn man sich
sozial oder politisch engagiert, kann man die sozialen
Verh¨
altnisse beeinflussen).
.0963083 .0461371 .1972745
Q5 I frequently have the experience that other people
have a controlling influence over my life (Ich mache
h¨
aufig die Erfahrung, dass andere ¨
uber mein Leben
bestimmen).
.6344261 .0101093 .1099268
Q6 One has to work hard in order to succeed (Erfolg muss
man sich hart erarbeiten).
.0317271 .3904663 -.0564504
Q7 If I run up against difficulties in life, I often doubt my
own abilities (Wenn ich im Leben auf Schwierigkeiten
stoße, zweifle ich oft an meinen F¨
ahigkeiten).
.4587426 -.0118316 .1196966
Q8 The opportunities that I have in life are deter-
mined by the social conditions (Welche M¨
oglichkeiten
ich im Leben habe, wird von den sozialen Umst¨
anden
bestimmt).
.3267092 .1440986 -.0532071
Q9 Innate abilities are more important than any efforts
one can make (Wichtiger als alle Anstrengungen sind die
F¨
ahigkeiten, die man mitbringt).
.1013728 .3544741 .0105248
Q10 I have little control over the things that happen
in my life (Ich habe wenig Kontrolle ¨
uber die Dinge,
die in meinem Leben passieren).
.6018275 -.1673338 .0313426
Eigenvalue
Factor 1 1.553382
Factor 2 .4539728
Factor 3 .1338106
Factor 4 .0469835
Factor 5 .0194704
Factor 6 -.088547
Factor 7 -.1033458
Factor 8 -.1480346
Factor 9 -.2053569
Factor 10 -.2317325
N=964 individuals.
39
Table A3: Determinants of return to employment after first childbirth (dependent variable: hazard rate of return), discrete semi-parametric
hazard estimation (Heckman-Singer model) — robustness tests w.r.t. the construction of the LOC variable
(Benchmark) (2) (3) (4) (5)
Locus of control LOC 0.302***
(0.098)
LOC Caliendo et al. (2015) 0.047***
(0.013)
LOC Piatek and Pinger (2015) 0.039***
(0.014)
LOC incl Q2 +Q7 0.260***
(0.072)
LOC factor score -0.137***
(0.045)
Return wage log(w0) 0.389** 0.414*** 0.166* 0.387** 0.382**
(0.159) (0.153) (0.098) (0.167) (0.164)
W growth tendency ˆ
∆wT EN D -0.005 0.018 -0.195 0.011 -0.008
(0.253) (0.255) (0.253) (0.253) (0.251)
Other household income log(m) -0.436*** -0.439*** -0.339*** -0.434*** -0.435***
(0.072) (0.069) (0.032) (0.074) (0.075)
Duration dependency Yes Yes Yes Yes Yes
Child birth cohort Yes Yes Yes Yes Yes
Constant -2.103*** -2.137*** -2.070*** -2.117*** -2.091***
z0.841*** 0.849*** 0.190*** 0.820*** 0.849***
p(pct) 0.590 0.597 0.908 0.603 0.584
Number of person-spells 18967 18966 18967 18967 18967
Note: Standard errors are in parentheses. * p<0.10, ** p<0.05, *** p<0.01. Data source: SOEP 1992–2013, authors’
calculations.
40
Appendix B
We estimate wage growth prior to first childbirth as a function of occupation, education, labor
market experience (in several polynomials), tenure with the same firm (in several polynomials),
region (East versus West Germany), year fixed effects and interactions between year fixed ef-
fects and region. Wage growth of person iin year tis defined as log(waget+1)−log(waget),
where wage means gross hourly wage. We choose the covariates in the estimation to maximize
the model fit relying on the Akaike information criterion. Occupation is split into eight groups
based on the classification of occupations (Klassifikation der Berufe 1992 (KldB 92)) by the
Germany Statistical Office (Statistisches Bundesamt). Though the classification is available in
the SOEP on a 4-digit level, we are obliged to use relatively wide categories in order to have
a reasonable number of observations in each category. The categories are as follows: cat. 1 =
occupations in agriculture, animal husbandry, forestry, horticulture, manufacturing, and other
occupations; cat. 2 =technical occupations; cat. 3 =merchants of goods; cat. 4 =merchants
of services or associated occupations; cat. 5 =occupations in organizing, administration, or
office; cat. 6 =occupations in health services; cat. 7 =occupations in social services and ed-
ucation, other humanistic and natural scientific occupations; cat. 8 =occupations in transport,
regulation, security, writing, translation, librarians, artists, and other service occupations. The
results of the estimation of wage growth prior to first childbirth are reported in table A4. The
overall R2(including the individual effect) of the estimation is 23.6%, i.e., we explain 23.6% of
the variation of wage growth. This might seem low if we consider that we include an individual
fixed effect. Note, however, that the periodicity of wage adjustments may matter: Assume we
have different types of individuals in the sample, with differing wage growth trajectories, with
all individuals receiving a promotion every three years. While the individual effects would be
perfect predictors of latent wage growth, the explained variation in wage changes from any one
year to the next would only be one third.
41
Table A4: Estimation of wage growth prior to first childbirth (dependent variable:
log(waget+1)−log(waget))
Occupation
Cat. 2 -0.015
(0.082)
Cat. 3 -0.136**
(0.067)
Cat. 4 -0.069
(0.089)
Cat. 5 -0.082
(0.061)
Cat. 6 0.101
(0.111)
Cat. 7 0.018
(0.080)
Cat. 8 -0.128**
(0.060)
ISCED educational level
General elementary 0.103
(0.128)
Middle vocational -0.089
(0.125)
Vocational +high school degr (Abitur) -0.208
(0.146)
Higher vocational -0.174
(0.129)
Higher education -0.219
(0.147)
Experience -0.341***
(0.033)
Experience20.053***
(0.008)
Experience3-0.005***
(0.001)
Experience40.000***
(0.000)
Experience5-0.000***
(0.000)
Tenure 0.105***
(0.017)
Tenure2-0.022***
(0.004)
Tenure30.002***
(0.000)
Tenure4-0.000***
(0.000)
East Germany -0.618
(0.791)
Year fixed effects Yes
Year fixed effects X East Germany Yes
Constant -0.755***
(0.207)
No. of person-year obs 6059
No. of persons 1295
Note: The reference category for occupation is cat. 1 (de-
tails about the occupation categories can be found in the
text in appendix B.); the reference category for ISCED
educational level is “inadequately”. Standard errors are
in parentheses. * p<0.10, ** p<0.05, *** p<0.01. Data
source: SOEP 1984–2013, authors’ calculations.
42
Appendix C
Table A5: Estimation of the leave-related wage growth penalty (dependent variable: wage
growth at return)
g1(occ. cat. 1) 0.042**
(0.017)
g2(occ. cat. 2) 0.020
(0.032)
g3(occ. cat. 3) 0.028***
(0.011)
g4(occ. cat. 4) 0.050*
(0.029)
g5(occ. cat. 5) 0.022
(0.016)
g6(occ. cat. 6) 0.012
(0.014)
g7(occ. cat. 7) 0.039***
(0.012)
g8(occ. cat. 8) 0.008
(0.019)
No. of persons 541
Note: Details about the occupation categories can be
found in appendix B. Standard errors are in parentheses.
* p<0.10, ** p<0.05, *** p<0.01. Data source: SOEP
1992–2013, authors’ calculations.
43
Table A6: Occupation categories
Item Number Percent
Cat. 1 43 8
Cat. 2 24 4
Cat. 3 68 13
Cat. 4 46 8
Cat. 5 144 27
Cat. 6 81 15
Cat. 7 79 15
Cat. 8 58 11
Total 543 100
Note: Details about the occupation categories can be
found in in appendix B.. Data source: SOEP 1992–2013,
authors’ calculations.
Note: Details about the occupation categories can be found in appendix B.
Figure A1: Distribution of LOC by occupation category
44
Note: Details about the occupation categories can be found in appendix B.
Figure A2: Distribution of leave duration by occupation category
45
References
Adda, J., C. Dustmann, and K. Stevens (2016). The career costs of children. Journal of Political
Economy forthcoming.
Andrisani, P. J. (1977). Internal-external attitudes, personal initiative, and the labor market
experience of black and white men. Journal of Human Resources 12(3), 308–328.
Andrisani, P. J. (1981). Internal-external attitudes, sense of efficacy, and labor market experi-
ence: A reply to Duncan and Morgan. Journal of Human Resources 16(4), 658–666.
Bergemann, A. and R. T. Riphahn (2009). Female labor supply and parental leave benefits –
The causal effect of paying higher transfers for a shorter period of time. SOEPpapers 161,
DIW, Berlin, Germany.
Bergemann, A. and R. T. Riphahn (2010). The introduction of a short-term earnings-related
parental leave benefit system and differential employment effects. SOEPpapers 315, DIW,
Berlin, Germany.
Blomeyer, D., M. Laucht, K. Coneus, and F. Pfeiffer (2009). Initial risk matrix, home resources,
ability development, and children’s achievement. Journal of the European Economic Associ-
ation 7(2-3), 638648.
Blundell, R., M. C. Dias, C. Meghir, and J. Shaw (2013). Female labour supply, human capital
and welfare reform. mimeo.
Borghans, L., A. L. Duckworth, J. J. Heckman, and B. ter Weel (2008). The economics and
psychology of personality traits. Journal of Human Resources 43(4), 972–1059.
Burgess, S., P. Gregg, C. Propper, and E. Washbrook (2008). Maternity rights and mothers’
return to work. Labour Economics 15, 168–201.
Caliendo, M., D. Cobb-Clark, and A. Uhlendorff(2015). Locus of control and job search
strategies. Review of Economics and Statistics 97(1), 88–103.
46
Carneiro, P., C. Meghir, and M. Parey (2011). Maternal education, home environments and
the development of children and adolescents. Journal of the European Economic Associa-
tion forthcoming.
Caspi, A., B. W. Roberts, and R. L. Shiner (2005). Personality development: Stability and
change. Annual Review of Psychology 56, 453–484.
Cebi, M. (2007). Locus of control and human capital investment revisited. Journal of Human
Resources 42(4), 920–932.
Cobb-Clark, D. and S. Schurer (2013). Two economists musings on the stability of locus of
control. Economic Journal 123(570), F358F400.
Coleman, M. and T. DeLeire (2003). An economic model of locus of control and the human
capital investment decision. Journal of Human Resources 38(3), 701–721.
Costa Jr, P. T. and R. R. McCrae (1994). Set like plaster? Evidence for the stability of adult
personality. In T. F. Heatherton and J. L. Weinberger (Eds.), Can Personality Change?, pp.
21–40. Washington, DC: American Psychological Association.
Cunha, F. and J. J. Heckman (2007). The technology of skill formation. American Economic
Review Papers and Proceedings 97(2), 31–47.
Cunha, F. and J. J. Heckman (2008). Formulating, identifying and estimating the technology of
cognitive and noncognitive skill formation. Journal of Human Resources 43(4), 738–782.
Cunha, F. and J. J. Heckman (2009). The economics and psychology of inequality and human
development. Journal of the European Economic Association 7(2-3), 320–364.
Davies, R. and G. Pierre (2005). The family gap in pay in Europe: A cross-country study.
Labour Economics 12, 469–486.
Dechter, E. K. (2014). Maternity leave, effort allocation, and postmotherhood earnings. Journal
of Human Capital 8(2), 97–125.
47
Dehne, M. and J. Schupp (2007). Pers¨
onlichkeitsmerkmale im Sozio-oekonomischen Panel
(SOEP) — Konzept, Umsetzung und empirische Eigenschaften. Research Notes 26, DIW,
Berlin, Germany.
Flossmann, A. L., R. Piatek, and L. Wichert (2007). Going beyond returns to education: The
role of noncognitive skills on wages in Germany. Working paper, University of Konstanz.
Fraley, C. R. and B. W. Roberts (2005). Patterns of continuity: A dynamic model for con-
ceptualizing the stability of individual differences in psychological constructs across the life
course. Psychological Review 112, 60–74.
Gutierrez-Domenech, M. (2005). Employment after motherhood: A European comparison.
Labour Economics 12, 99–123.
Heckman, J. J. and Y. Rubinstein (2001). The importance of noncognitive skills: Lessons from
the GED testing program. American Economic Review Papers and Proceedings 91(2), 145–
149.
Heckman, J. J. and B. Singer (1984). A method for minimizing the impact of distributional
assumptions in econometric models for duration data. Econometrica 52(2), 271–320.
Heckman, J. J., J. Stixrud, and S. Urzua (2006). The effects of cognitive and noncognitive
abilities on labor market outcomes and social behavior. Journal of Labor Economics 24(3),
411–482.
Heineck, G. and S. Anger (2010). The returns to cognitive abilities and personality traits in
Germany. Labour Economics 3, 535–546.
Jenkins, S. P. (2005). Survival analysis. Lecture notes manuscript, University of Es-
sex. Online available at http://www.iser.essex.ac.uk/files/teaching/stephenj/ec968/pdfs/-
ec968lnotesv6.pdf.
48
John, O. P. and S. Srivastava (1999). Big Five trait taxonomy. History, measurement, and
theoretical perspectives. In L. Pervin and O. P. John (Eds.), Handbook of Personality. Theory
and Research, pp. 102–139. New York, NY: Guilford.
Kuhlenkasper, T. and G. Kauermann (2010). Duration of maternity leave in Germany: A case
study of nonparametric hazard models and penalized splines. Labour Economics 17, 466–
473.
Lefebvre, P., P. Merrigan, and M. Verstraete (2009). Dynamic labour supply effects of childcare
subsidies: Evidence from a Canadian natural experiment on low-fee universal child care.
Labour Economics 16(5), 490–502.
McCrae, R. R. and P. T. Costa (1994). The stability of personality: Observation and evaluations.
Current Directions in Psychological Science 3(6), 205–220.
McCrae, R. R. and P. T. Costa Jr (1996). Toward a new generation of personality theories:
Theoretical contexts for the Five-Factor Model. In J. S. Wiggins (Ed.), The Five-Factor
Model of Personality: Theoretical Perspectives, pp. 5187. New York, NY: Guilford.
McCrae, R. R. and P. T. Costa Jr (1999). A five-factor theory of personality. In L. A. Pervin and
O. P. John (Eds.), Handbook of Personality: Theory and Research, pp. 139153. New York,
NY: Guilford.
McCrae, R. R. and P. T. Costa Jr (2003). Personality in Adulthood: A Five-Factor Theory
Perspective. New York, NY: Guilford Press.
Mueller, G. and E. J. S. Plug (2006). Estimating the effect of personality on male and female
earnings. Industrial &Labor Relations Review 60(1), 1–20.
Noor, N. M. (2002). Work-family conflict, locus of control, and women’s well-being: Tests of
alternative pathways. Journal of Social Psychology 142(5), 645–662.
Nyhus, E. K. and E. Pons (2005). The effect of personality on earnings. Journal of Economic
Psychology 26, 363–384.
49
Osborne Groves, M. (2005). How important is your personality? Labor market returns to
personality for women in the US and UK. Journal of Economic Psychology 26, 827–841.
Piatek, R. and P. Pinger (2015). Maintaining (locus of) control? data combination for the iden-
tification and inference of factor structure models. Journal of Applied Econometrics forth-
coming.
Pissarides, C. (2001). Economics of search. In N. Smelser and P. Baltes (Eds.), International
Encyclopedia of the Social and Behavioral Sciences, pp. 13760–13768. Elsevier.
Roberts, B. W. and W. F. DelVecchio (2000). The rank-order consistency of personality traits
from childhood to old age: A quantitative review of longitudinal studies. Psychological
Bulletin 126, 3–25.
Roberts, B. W., N. R. Kuncel, R. Shiner, A. Caspi, and L. R. Goldberg (2007). The power of per-
sonality: The comparative validity of personality traits, socioeconomic status, and cognitive
ability for predicting important life outcomes. Perspectives on Psychological Science 2(4),
313–345.
Rotter, J. B. (1966). Generalized expectancies for internal versus external control of reinforce-
ment. Psychological Monographs 80(1).
Schmieder, J. F., T. von Wachter, and S. Bender (2015). The effect of unemployment benets and
nonemployment durations on wages. mimeo.
Spinnewijn, J. (2015). Unemployed but optimistic: Optimal insurance design with biased be-
liefs. Journal of the European Economic Association 13, 130–167.
Uhlendorff, A. (2004). Der Einfluss von Pers¨
onlichkeitseigenschaften und sozialen Ressourcen
auf die Arbeitslosigkeitsdauer. Zeitschrift f¨ur Soziologie und Sozialpsychologie 56(2), 279–
303.
Uysal, S. D. and W. Pohlmeier (2009). Unemployment duration and noncognitive skills. Work-
ing paper, University of Konstanz.
50
van den Berg, G. (2001). Duration models: Specification, identification, and multiple durations.
In J. Heckman and E. Leamer (Eds.), Handbook of Econometrics, Volume 5. Elsevier.
Wagner, G. G., J. R. Frick, and J. Schupp (2007). The German Socio-Economic Panel Study
(SOEP) — scope, evolution and enhancements. Schmollers Jahrbuch 127, 139–169.
Weinhardt, M. and J. Schupp (2011). Multi-itemskalen im soep jugendfragebogen. Data Doc-
umentation 60, DIW, Berlin, Germany.
Wichert, L. and W. Pohlmeier (2010). Female labor force participation and the Big Five. Dis-
cussion Paper 10-003, ZEW, Mannheim, Germany.
51
