Content uploaded by Jan Marcus
Author content
All content in this area was uploaded by Jan Marcus on Sep 04, 2019
Content may be subject to copyright.
econstor
Make Your Publications Visible.
A Service of
zbw Leibniz-Informationszentrum
Wirtschaft
Leibniz Information Centre
for Economics
Marcus, Jan
Article — Manuscript Version (Preprint)
The Effect of Unemployment on the Mental Health of
Spouses – Evidence from plant closures in Germany
Journal of Health Economics
Provided in Cooperation with:
German Institute for Economic Research (DIW Berlin)
Suggested Citation: Marcus, Jan (2013) : The Effect of Unemployment on the Mental Health
of Spouses – Evidence from plant closures in Germany, Journal of Health Economics, ISSN
0167-6296, Elsevier, Amsterdam, Vol. 32, Iss. 3, pp. 546-558,
http://dx.doi.org/10.1016/j.jhealeco.2013.02.004 ,
http://www.sciencedirect.com/science/article/pii/S0167629613000118
This Version is available at:
http://hdl.handle.net/10419/74545
Standard-Nutzungsbedingungen:
Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen
Zwecken und zum Privatgebrauch gespeichert und kopiert werden.
Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle
Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich
machen, vertreiben oder anderweitig nutzen.
Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen
(insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten,
gelten abweichend von diesen Nutzungsbedingungen die in der dort
genannten Lizenz gewährten Nutzungsrechte.
Terms of use:
Documents in EconStor may be saved and copied for your
personal and scholarly purposes.
You are not to copy documents for public or commercial
purposes, to exhibit the documents publicly, to make them
publicly available on the internet, or to distribute or otherwise
use the documents in public.
If the documents have been made available under an Open
Content Licence (especially Creative Commons Licences), you
may exercise further usage rights as specified in the indicated
licence.
www.econstor.eu
The effect of unemployment on the
mental health of spouses -
Evidence from plant closures in Germany
Jan Marcus1
Studies on health effects of unemployment usually neglect spillover effects
on spouses. This study specifically investigates the effect of an individ-
ual’s unemployment on the mental health of their spouse. In order to al-
low for causal interpretation of the estimates, it focuses on plant closure as
entry into unemployment, and combines difference-in-difference and match-
ing based on entropy balancing to provide robustness against observable and
time-invariant unobservable heterogeneity. Using German Socio-Economic
Panel Study data the paper reveals that unemployment decreases the men-
tal health of spouses almost as much as for the directly affected individuals.
The findings highlight that previous studies underestimate the public health
costs of unemployment as they do not account for the potential consequences
for spouses.
JEL: I12, J65
Keywords: unemployment, mental health, plant closure, entropy balancing,
matching, job loss
1Jan Marcus, Mohrenstraße 58, DIW Berlin, 10117 Berlin, Germany; jmarcus@diw.de. Valuable
comments by Adam Lederer, Frauke Peter, Anika Rasner, Thomas Siedler, Michael Weinhardt, the
anonymous referee and the editor Richard Frank are gratefully acknowledged. I would also like to
thank seminar participants at the Annual Meeting of the German Economic Association, the Twenty
First European Workshop on Econometrics and Health Economics, the Leibniz Seminar on Labour
Research (Berlin), and the Warsaw International Economic Meeting for their helpful suggestions.
NOTICE: This is the author’s version of a work that was accepted for
publication in “Journal of Health Economics”. Changes resulting from
the publishing process, such as peer review, editing, corrections,
structural formatting, and other quality control mechanisms may not
be reflected in this document. Changes may have been made to this
work since it was submitted for publication. A definitive version
was subsequently published in Journal of Health Economics 32 (2013),
3, pp. 546–558 and is online available at http://dx.doi.org/10.1016/
j.jhealeco.2013.02.004.
1. Introduction
Apart from income, employment has many non-financial benefits, such as structured
time, social status and identity, social contact, collective purpose, as well as activity
(Jahoda 1979). Unemployment results in the loss of the pecuniary and non-pecuniary
work benefits, and these losses also impact other household members. Spouses of newly
unemployed individuals have to cope with reduced household income, a presumably more
depressed partner, the partner’s unfamiliar presence at home as well as a reduced social
status. For spouses, too, these negative consequences of unemployment might result in
depressive symptoms and other mental health issues.
Yet, while there is a whole branch of literature on the health implications of job loss
and unemployment for those individuals directly affected (see e.g. Browning et al. 2006;
Brand et al. 2008;Eliason & Storrie 2009;Kuhn et al. 2009;Salm 2009;Sullivan & von
Wachter 2009;Deb et al. 2011;Schmitz 2011;Browning & Heinesen 2012;Marcus 2012),
few studies address the impact on their spouses. Not considering the potential negative
externalities on spouses might result in underestimating the public health costs of job loss
(e.g. Kuhn et al. 2009). This study contributes to our understanding of spillover effects
of unemployment on other household members by estimating the effect of unemployment
on the spouse’s mental health. In order to give the estimates a causal interpretation,
this study applies a combination of matching and difference-in-difference that is robust
against selection on observables and selection on unobservables with time-invariant ef-
fects. The matching part of the estimator constitutes one of the first applications of
entropy balancing (Hainmueller 2012), which balances the conditioning variables more
effectively than common propensity score methods. Furthermore, this study considers
only unemployment resulting from plant closures. Other causes of unemployment might
result from mental health issues and, hence, might be endogenous.
Using German Socio-Economic Panel Study (SOEP) data from 2002 through 2010,
this paper finds that the unemployment of one spouse2similarly affects the mental
health of both spouses. About one year after plant closure, unemployment decreased
mental health by 27 % of a standard deviation for unemployed individuals themselves
and by 18 % of a standard deviation for their spouses. In general, the decreases in
mental health are larger when the male spouse enters unemployment. The results are
robust over various matching specifications. Furthermore, this paper shows that changes
in mental health do not differ between treated and matched controls before the plant
2The terms “partner” and “spouse” are used interchangeably in this paper.
1
closure, adding additional credibility to the identification assumption. Analyzing other
reasons for unemployment and not just that connected to plant closures confirms the
finding that unemployment decreases the mental health for spouses almost as much as
for directly affected individuals. However, the effects for other reasons of unemployment
are not larger than for unemployment due to plant closure, suggesting that selection
issues with respect to entering unemployment might be less important than previously
thought.
The structure of this paper is as follows. The next section discusses related literature
in greater detail, section 3 illustrates the estimation strategy, section 4 introduces the
data, describes the construction of treatment and control group, and provides descriptive
statistics. The main results are presented in section 5, while section 6 performs further
analyses. Section 7 concludes.
2. Related Literature
The literature on the consequences of unemployment is closely related to the literature on
the consequences of job loss - especially when the focus is on the identification of causal
effects. In the following, I discuss these two branches of literature together, bearing in
mind that not all individuals who lose their jobs are unemployed afterwards and not
every individual in the state of unemployment experienced the event of an involuntary
job loss.
Previous studies provide some evidence for spillover effects of job loss and unemploy-
ment on other household members. For instance, Winkelmann & Winkelmann (1995)
report decreases in subjective well-being following the partners’ unemployment (without
differencing between the reasons for unemployment). Stephens (2002), using U.S. Panel
Study of Income Dynamics (PSID) data, provides evidence for the “added worker effect”,
that women increase their labor supply due to their husbands’ job losses resulting from
lay-offs or plant closures. Again with PSID data, Charles & Stephens (2004) show that
job loss resulting from lay-off or plant closure increases the probability of divorce. Other
studies indicate that children are also affected by their parents’ loss of employment. For
instance, using SOEP data, Siedler (2011) finds that experiencing parental unemploy-
ment (for any reason) increases the probability of children to support extreme right-wing
parties. With data from the PSID, Lindo (2011) provides evidence for a reduction in
children’s birth weight following a father’s experience of plant closing. This already
indicates that the public health costs of job loss are underestimated if spillover effects
2
on other household members are not taken into account.3Taken together, the find-
ings on spillover effects on other household members lead to the expectation of negative
consequences of unemployment for the spouse’s mental health.
Few studies explicitly analyze the causal relationship between unemployment and the
partner’s mental health. Most of these studies focus on single plants or small geographic
areas (e.g. Liem & Liem 1988;Penkower et al. 1988;Dew et al. 1992).4To my knowledge
only two studies use nationally representative data in this context (Clark 2003;Siegel
et al. 2003). However, both studies do not take into account the endogeneity of the
treatment. Drawing on data from the British Household Panel Study, Clark (2003)
finds partner’s unemployment to reduce mental health, but less so if the respondent
is already unemployed. Yet, this study does not differentiate between the reasons for
unemployment. With data from the U.S. Health and Retirement Study (HRS) Siegel
et al. (2003) do not find a significant effect of husbands’ job loss on wives’ mental health
in general. However, this study also considers individuals who lost their job due to lay-
offs, which might result from mental health issues. Furthermore, due to the sampling
design of the HRS, this study only analyzes individuals over the age of 50. In contrast,
the present study looks at the entire range of working-age individuals and includes in
the treatment group only individuals who lost their job due to plant closure.
Previous studies document spillovers of other events in one spouse’s life on the other
spouse’s mental health. For instance, Bolger et al. (1989) find that stress experienced by
one spouse is associated with strain in the other spouse.5With a sample of the elderly
Dutch population, Lindeboom et al. (2002) provide evidence that a severe illness of the
partner significantly decreases own mental health. They also find that the death of the
spouse strongly increases the risk of developing a depression. Compared to the other
life events they analyze (e.g. disability, sever financial problems, death of child, death
of grandchild), conjugal bereavement has the largest effect on mental health.
3A further reason why the public health costs of unemployment might be underestimated is that
often the focus is on short-term effects, and long-term consequences like potential scarring effects of
unemployment (Clark et al. 2001) are not taken into account.
4Dew et al. (1992) provide an overview over earlier qualitative studies on this topic.
5Jones (1993) and ten Brummelhuis et al. (2010) come to similar conclusions. Though, it remains
unclear whether the associations presented in these studies can be interpreted as causal effects.
3
3. Estimation Strategy
In order to estimate the effect of unemployment on the mental health of couples, this
paper combines matching and difference-in-difference (DiD), which is regarded to be su-
perior to pure cross-sectional matching estimators (Heckman et al. 1997). This estimator
brings together the literature on selection on observables with the literature on selection
on unobservables. The idea of the estimator is rather simple. In the matching part of the
estimator, I take couples who are affected by unemployment and similar couples who do
not experience unemployment. In the DiD part, I compare changes in mental health of
these two groups. The DiD part eliminates time-invariant mental health differences be-
tween couples in the treatment and control group that result from unobserved variables
(like personality traits and differences in the reporting behavior). In all analyses I focus
on the average treatment effect on the treated (ATT), i.e. the unemployment induced
change in mental health of those couples who are actually affected by unemployment as
a result of plant closures.
The challenge is how to make treated couples (couples, where one spouse is affected
by unemployment resulting from plant closure) and control couples (couples without job
loss experiences in the period) similar. To increase similarity between the two groups,
propensity score methods are often applied (Caliendo & Kopeinig 2008),6where the
control group observations are reweighted either by weights that depend directly on
propensity score values (as in propensity score weighting) or by weights that depend
on propensity score distances to treatment observations (as, for example, in nearest
neighbor or kernel matching). However, I do not take the detour via the propensity score,
but instead implement a reweighting technique, entropy balancing (Hainmueller 2011,
2012), that focuses directly on the balancing of conditioning variables (see appendix A.1).
Entropy balancing is more effective in reducing covariate imbalance than propensity score
methods as it reweights the control group observations in such a way that the control
group satisfies pre-specified balancing requirements (here: same mean and variance of
conditioning variables as in the treatment group). In the section on sensitivity analyses
I also apply propensity score methods.
I perform entropy balancing separately for couples where the husband/wife enters
unemployment.7This is like exact matching on the gender of the directly affected
6The propensity score (Rosenbaum & Rubin 1983) is the probability to receive the treatment condi-
tional on the covariates.
7For convenience, I refer to the male and female in a couple as husband and wife - independent of
marital status.
4
spouse. Section 5 shows the results for all couples pooled and separately according to
whether husband or wife enters unemployment.
Obtaining the weights from entropy balancing constitutes the first step in implement-
ing the estimation strategy. This is the matching/reweighting step. The second step
is the regression step, where the change in mental health is regressed on the treatment
indicator with the sampling weights obtained in the first step. In the regression step, I
additionally control for all conditioning variables used in the matching step. This does
not alter the treatment effect as after weighting the treatment is mean-independent of
all conditioning variables. However, the regression-adjustment decreases the standard
errors of the treatment effect estimates because it reduces unexplained variance in the
outcome. This is similar to including control variables in randomized experiments.
Hence, the ATT of interest can be obtained from
b
β=(X’WX)−1X’∆y, (1)
where ∆yis the vector of mental health changes and Wa diagonal matrix with 1 in the
diagonal cells for couples of the treatment group and entropy balancing weights in the
diagonal cells for control group couples. Xis a n-by-(k+2) matrix, in which the first
column consists of a vector of 1s and the second column of the values of the treatment
group indicator, D, for each couple. Additionally, Xcontains one column for each of the
kconditioning variables. The estimator is similar to the regression-adjusted semipara-
metric difference-in-difference matching strategy proposed by Heckman et al. (1997).
It differs only with respect to the construction of the weights, which are computed by
propensity score methods in Heckman et al. (1997).
In order to give the estimates a causal interpretation, the estimator has to assume that
no unobserved variables exist that simultaneously influence changes in mental health and
the probability of entering unemployment due to plant closure, i.e. in the absence of
treatment (unemployment due to plant closure) the mental health of treated couples
and matched control couples follows the same trend:
E[Ya
0−Yb
0|EB(X), D = 1] = E[Ya
0−Yb
0|EB(X), D = 0],(2)
where ∆Y0=Ya
0−Yb
0refers to the change in mental health from before (b) to after
(a) the treatment in the absence of treatment and EB(X)refers to the weights from
entropy balancing.
5
4. Data
This paper makes use of data from the German Socio-Economic Panel (SOEP, v27)
from 2002 through 2010. The SOEP is among the largest and longest running house-
hold panel surveys in the world. Annually about 20,000 individuals participate in the
SOEP. It consists of several subsamples and is designed to be representative of the en-
tire population in Germany (Wagner et al. 2007). The SOEP hosts several features
that make it particularly attractive for the present analysis. Firstly, the longitudinal
nature of the data ensures that I can observe the mental health scores before and after
the treatment. Secondly, its large sample size facilitates analysis based on relatively
rare events like plant closures. Thirdly, the SOEP contains not only information on one
household member but also data on cohabiting spouses. These data are directly provided
by the spouses themselves. Fourthly, after household dissolutions the SOEP follows all
household members and not just the household head. This is of particular importance
for the present analysis as unemployment increases the probability of divorce (Charles
& Stephens 2004) and, hence, household dissolution. Only considering couples that still
live together after the treatment would result in a rather selective panel (though divorce
related panel attrition might still occur in the SOEP; see below). Fifthly, the SOEP pro-
vides a wide range of information at the individual and the household level, including
details about earnings, employment and living conditions. This enables me to include
almost all conditioning variables used in related studies.
4.1. Outcome
This study uses the Mental Component Summary Scale (MCS), which was developed as
a brief instrument to measure mental health in large scale surveys (Ware et al. 1996). It
is found to be reliable and valid (e.g. Salyers et al. 2000), with high values of sensitivity
and specifity compared to other brief scales of mental health (Gill et al. 2007). MCS
is not a disease-specific measure focusing on a particular condition or disease, but a
generic measure of mental health. MCS is used widely in the epidemiologic literature
and a broad literature in economics uses this summary measure of mental health as well
(e.g. Lechner 2009;Reichert & Tauchmann 2011;Schmitz 2011). It was included in the
SOEP as the measure of mental health. It is not a measure of general life satisfaction
although it correlates with life satisfaction.8
8The correlation in the used sample is r= 0.4.
6
The construction of MCS reverts to the SOEP version of the SF-12 questionnaire,
which was first adopted in the SOEP in 2002 and ever since is included every two years.
The SF-12 contains 12 health-related questions that encompass two dimensions: physical
and mental health. All items reflect the current health status of the respondents, as they
refer to the four weeks immediately prior to the interview (e.g. “How often did you feel
run-down and melancholy in the last four weeks?”). Appendix A.2 provides an overview
over the items of the SF-12 questionnaire and the answer categories. MCS is a weighted
combination of the 12 items and provided by the SOEP Group. It is computed by means
of explorative factor analysis and transformed to have mean 50 and standard deviation
10 in the 2004 SOEP sample (Andersen et al. 2007). Higher values indicate a better
mental health status.
4.2. Treatment and Control Group
Treatment and control group consist of couples who live together in the same household
before the treatment - irrespective of their marital status. Figure 1 provides an overview
of the construction of treatment and control group, which the following paragraph de-
scribes in more detail.
I only include couples where both spouses participate in the survey before (tb) and
after (ta) the treatment, and provide valid mental health information in both years. I
do not consider same-sex couples as there are none in the treatment group.
The treatment group comprises couples in which one spouse (the “directly affected
spouse”) enters unemployment due to plant closure between two survey waves with the
SF-12 questionnaire. I construct the treatment indicator by combining information from
the question on whether the respondent is officially registered as unemployed with infor-
mation on the reason why the individual left the last job. I only consider plant closures
since other reasons of unemployment might be endogenous, e.g. someone might be dis-
missed because of shrinking work productivity due to marital problems (and marital
problems might decrease mental health). Before the plant closure the directly affected
individual has to be employed either full-time or part-time in the private sector and
has to be between the ages of 18 and 62. I include couples in the treatment group
irrespective of age and employment status of the indirectly affected spouse (the partner
of the directly affected individual). However, I do not consider couples in which the
indirectly affected spouse experienced an involuntary job loss between tband tadue to
plant closure. That is done in order to prevent that own experiences deteriorate the
mental health effect of the indirectly affected spouse. The plant closure experience of
7
one spouse is likely to be correlated with a plant closure experience of the other spouse
since spouses might work in the same plant. Excluding couples in which indirectly af-
fected spouses experience plant closures themselves reduces the treatment group by 7
%.
Figure 1: Construction of treatment and control group
Before treatment
Spouse 1
Treatment and control group (same standard)
- Works full-time/part-time
- Employed in private sector
- Age: 18-62
Spouse 2
Treatment and control group (same standard)
- Cohabiting
After treatment
Spouse 1
Treatment group
- Unemployed due to plant closure
Control group
- No employer change
Spouse 2
Treatment and control group (same standard)
- Same spouse
- No job loss due to plant closure
after
t
before
t
Treatment period (~ 2 years)
Note: The figure presents an overview of the construction of treatment and control group. The boxes
show the requirements that couples have to meet to qualify for treatment and control group, respectively.
“Spouse 1” refers to the directly affected individual and “spouse 2” to the partner of “spouse 1”.
The control group consists of a potentially directly affected individual (who is of the
same sex as the directly affected individual in the treatment group) and a potentially
indirectly affected individual. For the potentially indirectly affected spouse the same
restrictions apply as in the treatment group construction, both before and after the
treatment. Similarly to the treatment group, at the pre-treatment interview the poten-
tially directly affected individual has to be employed either full-time or part-time in the
private sector and has to be between the ages of 18 and 62. However, couples qualify
only for the control group, if the potentially directly affected spouse did not leave the
previous employer during the treatment period (this excludes couples with any job loss
8
experiences; see figure 1). This leaves more than 14,000 couples for the control group,
compared to 109 couples in the treatment group: 70 couples in which the husband enters
unemployment and 39 couples in which the wife enters unemployment.
The plant closure can take place at any time between two survey waves that include the
mental health questions. Hence, there are four treatment periods: 2002-2004, 2004-2006,
2006-2008 and 2008-2010. I estimate treatment effects pooled over all four treatment
periods. On average, I observe the treatment group 11 months after the plant closure.
This time, however, varies between 0 and 23 months, with a rather uniform distribution.
Hence, the estimates are to be interpreted as averages over these different unemployment
durations; an interpretation inherent also to most applications of fixed effects panel
estimators.
4.3. Conditioning Variables
The set of conditioning variables, i.e. Xin equation (1), is selected following the screen-
ing of conditioning variables in other studies that analyze health effects of job loss and
unemployment on the directly affected individual (Browning et al. 2006;Böckerman &
Ilmakunnas 2009;Eliason & Storrie 2009;Salm 2009;Sullivan & von Wachter 2009;
Schmitz 2011). The richness of the SOEP data allows including almost all conditioning
variables used in these studies. Also the job security perceived by the individuals is
included, which can be seen as a variable that captures unobserved factors related to
the plant closure. The conditioning variables include demographic, labor market re-
lated, educational and health data of the directly affected spouse. I include the same
set of conditioning variables for the indirectly affected spouse as for the directly affected
spouse - except for those variables that make only sense for working individuals (e.g.
tenure with current firm, size of the company). The pre-treatment working status is
only included for indirectly affected spouses. Therefore, conditioning variables can be
divided into variables reported by the directly affected individual, variables reported by
the indirectly affected individual and variables on the couple (i.e. household) level. All
72 non-collinear conditioning variables originate from the pre-treatment interview. As
the conditioning variables also include the pre-treatment values of the mental health
score, the applied regression-adjusted matching DiD estimator resembles a regression-
adjusted matching estimator, where the outcome is the post-treatment mental health
score (Lechner 2010). I use the terminology matching DiD, in order to emphasize that
the estimator provides also some robustness against selection on unobservables. Table 1
provides an overview of the conditioning variables.
9
Table 1: Overview of the conditioning variables
Variable Definition D I
Individual information
Demographic
Age in years X X
Female 0=male, 1=female X X
Migrant 1=individual or parents moved to Germany, 0 otherwise X X
Non-German 0=German, 1=foreign citizenship X X
Health
Physical health based on SF12 questionnaire (see Andersen et al. 2007)X X
Mental health based on SF12 questionnaire X X
Often melancholic 0=never/almost never, 1=always/often/sometimes; part
of SF12
X X
Self-rated health 3 categories (very good/good, satisfactory, poor/bad) X X
Labor market
Labor earnings annual nominal labor earnings in 1000 Euro X X
Never unemployed 0=ever unemployed, 1=never unemployed X X
Tenure tenure with present employer (in years) X
Years in full time previous full-time experience in years X
Company size 4 categories (<20, 20-200, 200-2000, ≥2000 employees) X
Perceived job security 3 categories (big worries, some worries, no worries) X
Industry sector 10 categories X
Working status 3 categories (full time, part time, not employed) X
Educational
Secondary schooling 4 categories (no degree/basic school, intermediate/ X X
other school, academic school track (Abitur), technical
school)
University 0=no university degree, 1=university degree X X
Vocational training 0=no vocational training, 1=vocational training X X
Couple information
Children 1=children under 18 in household, 0 otherwise X
Regional unemployment yearly information on the state level X
Residential district 4 cat. (<2000, 2000-20 000, 20 000-100 000, ≥100 000
inhabitants)
X
Federal state 14 categories9X
Survey year 4 categories (2002, 2004, 2006, 2008) X
Note: The table describes the coding of the conditioning variables and indicates whether the variable is
included for the directly affected spouse (D), the indirectly affected spouse (I) or both. All conditioning
variables originate from the pre-treatment interview.
9I group Bremen with Lower Saxony and Hamburg with Schleswig-Holstein due to few cases.
10
4.4. Descriptive Statistics
Table 2 presents summary statistics of selected conditioning variables separately for
couples in the treatment and control group (before and after matching/reweighting).
The table in appendix A.3 provides the means of other conditioning variables.
Table 2: Summary statistics for selected variables before treatment
Variable Treated Controls (unmatched)
unmatched matched Difference
Directly affected spouse
Age 48.1 44.0 48.1−4.1∗∗∗
Female+35.8 43.6 35.8 7.9∗
Non-German+22.9 12.4 22.9−10.5∗∗∗
Mental health 49.9 50.7 49.9 0.8
Tenure 11.0 12.4 11.0 1.4
Never unemployed+62.4 70.6 62.4 8.2∗
Labor earnings 24.9 35.2 24.9 10.3∗∗∗
Big job worries+39.4 15.1 39.4−24.3∗∗∗
No job worries+22.0 39.0 22.0 16.9∗∗∗
University+8.3 23.3 8.3 15.1∗∗∗
Indirectly affected spouse
Mental health 49.3 50.4 49.3 1.1
Works full-time+49.5 52.2 49.5 2.7
Not working+31.2 22.7 31.2−8.5∗∗
N 109 14285 109
Note: The first three columns present means of selected variables before treatment for treated, controls
and matched controls, respectively. +indicates that the mean represents a percentage share. The last
column displays the difference in means between treatment and control group before matching; stars
indicate significant t-test differences between these two groups: * p < 0.1; ** p < 0.05; *** p < 0.01.
The first two columns of table 2 display means of the conditioning variables for the
treated and the control group couples, respectively. The last column displays differences
in means between treatment and control group before matching and tests for the signifi-
cance of these differences. Directly affected spouses differ in many respects significantly
from their control group counterparts. For instance, they are older (48.1 vs. 44.0 years),
are less likely to be female, more often do not have the German citizenship and earn
annually on average 10,000 Euro less. Individuals in the control group are almost three
times more likely to have a university degree than individuals who experience unemploy-
ment due to plant closure (23.3 % vs. 8.3 %). However, before the treatment there are
11
no significant differences with respect to tenure and mental health. The mental health
score is only about 0.8 points (about 8 % of a standard deviation) lower for the directly
affected spouses in the treatment group and the difference is not statistically significant
from zero.
Similarly, the indirectly affected spouses do not differ significantly from their control
group counterparts with respect to mental health (49.3 vs 50.4), but with respect to
age, German citizenship, education and earnings (see appendix A.3). Of these indirectly
affected spouses, 31.2 % of them do not work, compared to 22.7 % in the control group.
The third column of table 2 reports means for the matched controls. After the
reweighting based on entropy balancing the means in the control group equal the means
in the treatment group. The applied entropy balancing scheme not only balances the
means but also the variances of the conditioning variables.
The table in appendix A.3 not only reports the means for other conditioning variables,
but it also compares the matched control groups resulting a) from entropy balancing and
b) the propensity score based kernel matching (Heckman et al. 1997). Kernel matching
results rely on an Epanechnikov kernel and a bandwidth of 0.06, which improves the
covariate balance particularly well compared to other propensity score based specifica-
tions.10 The table shows that also kernel matching works quite well. For all but one
variable (technical college for the directly affected spouse) the standardized bias is below
the value of 5, which is regarded to be low (Caliendo & Kopeinig 2008).11 However, the
table in appendix A.3 also depicts that entropy balancing clearly outperforms kernel
matching as it better improves covariate balance between treatment and control group.
Kernel matching even increases the standardized bias for some variables (e.g. for living
in a small city).
5. Results
Table 3 shows the results for the effect of unemployment on the mental health of couples.
The table shows the findings pooled for all couples and separately according to the sex
of the directly affected spouse. It starts with a simple specification and then gradually
incorporates more sophisticated procedures. The first specification provides a simple
10For propensity score methods I also include the squared terms of all cardinal conditioning variables
to improve balance on the second moments.
11The standardized bias is a measure of the matching quality, and defined as the difference between
the means of treated and controls as a percentage share of the square root of the average of the
variances in the two groups (see also appendix A.3).
12
comparison of the average mental health of treated and (all) controls after treatment.
Since there might be fundamental differences between the treatment and control group
(e.g. with respect to reporting behavior or the general mental health level), specification
(2) uses the change in mental health as outcome. Hence, this specification resembles an
ordinary difference-in-difference estimator without control variables. It might be that
the mental health of treated couples follows a different trend than the mental health of
other couples, i.e. even in the absence of treatment the mental health of the treated
couples would change in a different way. The previous section shows that treated and
control couples differ indeed in many respects. This makes it more likely that the treated
are on a different mental health track. Matching ensures that only comparable couples
are compared. Specification (3) displays the results for the matching DiD estimator.
Specification (4) is the preferred specification as, in addition to specification (3), it
includes the covariates in the outcome equation as well. This leaves the estimates of the
ATT unchanged, because by construction of the entropy balancing scheme, the treatment
indicator is mean-independent of all conditioning variables. However, including the
covariates in the outcome regression reduces the variance in the outcome and, hence,
makes the estimates more precise.
Specification (1), the simple mean comparison, shows in the first cell that the mental
health score of spouses who entered unemployment due to plant closure is on average
about 2.68 units lower than the mental health score of control group individuals. This
implies a difference of about 26.8 % of a standard deviation since the mental health score
is normed to have a standard deviation of 10. For the spouses of the directly affected
individuals, mental health is on average 2.59 points lower. The second and third panel
show that the mental health is worse for both spouses irrespective of the sex of the
directly affected spouse. However, the difference between treated and controls is more
pronounced for couples where the husband became unemployed, and not significant for
couples where the wife entered unemployment. In general, the differences are of similar
magnitude for own mental health and the spouse’s mental health, and not statistically
significant from each other (as indicated by the p-values). This finding suggests that
unemployment of one spouse similarly affects the mental health of both spouses.
The results for specification (2) are similar to the previous results, although the coef-
ficient estimates decrease somewhat in magnitude (especially in the last panel). In the
matching DiD estimator in specification (3) both standard errors and coefficient esti-
mates slightly increase, t-statistics (not shown) marginally increase for most estimates.
However, incorporating matching does not change the overall picture much.
13
Table 3: The effect of unemployment on mental health - main results
Mean Main
difference DiD Match specification
Outcome (1) (2) (3) (4)
All couples
Own mental health −2.68 ∗ ∗ −1.84 −2.68 ∗ ∗ −2.68∗∗∗
(1.04) (1.12) (1.16) (0.72)
Partner’s mental health −2.59 ∗ ∗ −1.47∗ −1.81 ∗ ∗ −1.81∗∗∗
(1.01) (0.84) (0.88) (0.66)
p-value of difference 0.94 0.78 0.53 0.32
Husband’s unemployment
Own mental health −3.22 ∗ ∗ −2.70∗ −3.13 ∗ ∗ −3.13∗∗∗
(1.35) (1.53) (1.56) (0.88)
Partner’s mental health −2.66 ∗ ∗ −2.04∗ −2.01∗ −2.01∗∗∗
(1.34) (1.04) (1.09) (0.77)
p-value of difference 0.73 0.71 0.53 0.27
Wife’s unemployment
Own mental health −2.02 −0.28 −1.86 −1.86∗∗∗
(1.62) (1.49) (1.58) (0.63)
Partner’s mental health −2.11 −0.46 −1.46 −1.46∗
(1.45) (1.40) (1.50) (0.77)
p-value of difference 0.96 0.93 0.85 0.63
Note: The table presents the effect of one spouse’s entry into unemployment on the mental health of
both spouses. Each cell displays the ATT from a separate regression and its robust standard error in
parentheses. Additionally, the table provides p-values for the t-tests whether unemployment differently
influences directly and indirectly affected spouses. The upper panel considers all couples, while the
two lower panels display results separately according to the sex of the directly affected spouse. The
results rely on 109 couples in the treatment group (including 70 couples where the husband enters
unemployment) and more than 14,000 couples in the control group. * p < 0.1; ** p < 0.05; ***
p < 0.01. The first column refers to the mean difference in mental health after the treatment and “DiD”
to the simple difference-in-difference estimator without matching and “Match” to the DiD results after
entropy balancing. The last column presents results for the matching DiD estimator that includes the
covariates in the outcome equation.
14
The last column displays the results for the preferred specification, which includes
regression-adjustment. In specification (4) unemployment decreases mental health for
the directly affected individuals by 2.68 points on average, or about 26.8 % of a standard
deviation. This decrease is stronger when the husband enters unemployment (3.13 vs.
1.86 points). The impact of the spouse’s unemployment is only slightly smaller for the
indirectly affected spouse (1.81 points on average) and the difference in the effects is far
from being statistically significant. Again, the effect is stronger if the husband enters
unemployment (2.01 vs. 1.46 points). However, the effects of the wife’s unemployment
become significant in this last specification. These effects exhibit negative signs in all
specifications but were not estimated precisely enough before. The results also suggest
that the mental health decrease for wives whose husbands enter unemployment and
for wives who enter unemployment themselves is very similar, and the difference is far
from being significant. It has to be considered that the two effects are based on two
rather different groups of women (working wives vs. wives of working men). Hence,
one can not interfere that women are more/less affected by the spouse’s unemployment
than by their own unemployment. Similarly, one cannot say for sure that men are more
affected by their own unemployment than by their spouse’s unemployment. In general,
the coefficient estimates of the preferred specification closely resemble the results of
specification (1), the simple mean comparison. This implies that selection into treatment
is not strong with respect to mental health, which might provide additional credibility for
studies that evaluate the impact of plant closures but do not observe the pre-treatment
outcome.
It would be interesting to investigate whether the effect of unemployment differs be-
tween subgroups. However, due to the rather small number of couples in the treatment
group - especially when only considering couples in which the wife enters unemployment
- I refrain from more detailed inspections of potential mechanisms and treatment effect
heterogeneity.12
When including the post-treatment household net income as additional control vari-
able in the regression step, in order to analyze potential channels, the estimated effects
12Tentative analyses indicate that effects tend to be larger when - before the plant closure - the directly
affected individual provided a higher share of household income and when the indirectly affected
spouse did not work full-time. Further tentative analyses do not find supportive evidence for the
hypothesis that unemployment hurts less if individuals have more previous unemployment expe-
riences (habituation hypothesis; Clark et al. 2001). On the contrary, some of these specifications
suggest that unemployment has a more negative impact on mental health for those with previous
unemployment times, particularly for those couples where the indirectly affected spouse has been
unemployed before.
15
of unemployment on mental health only slightly decrease.13 This suggests that the drop
in current household income is not the main channel for the estimated effects. However,
the effects might work through the uncertainty about future income levels. Hence, it is
not possible to conclude that income is irrelevant as potential channel.14
6. Further Results
This section consists of two parts. The first part performs sensitivity analyses. It
applies different matching procedures (propensity score weighting, kernel matching),
analyzes the sensitivity of the results to a redefinition of the treatment and runs a
placebo regression to check the plausibility of the identifying assumption. The second
part investigates the potential endogeneity of other reasons for unemployment. Table
4 presents the results for the various sensitivity analyses, table 5 for other reasons for
unemployment.
6.1. Sensitivity Analyses
For the first two sensitivity checks I rely on propensity score methods instead of entropy
balancing. Propensity score weighting (PSW) and kernel matching differ from entropy
balancing with respect to the weighting matrix Win equation (1). Propensity score
weighting assigns to each control observation a weight that equals 1/(1 −P(X)), where
P(X)is the propensity score. Kernel matching matches to each treatment observation
control observations that are close in terms of the estimated propensity score. However,
it does not assign equal weights to all the matched neighbors but instead assigns weights
according to distances in the propensity score.15 Specifications (5) and (6) present the
results for propensity score weighting and kernel matching, respectively. The effects are
similar to the results in the main specification. However, the effects for wife’s unem-
ployment slightly decrease, while the effects for husband’s unemployment on husband’s
13These results are available from the author upon request. The average monthly net household in-
come is estimated to decrease by about 500 Euro due to one spouse’s unemployment, applying the
previously outlined estimation strategy to household income. This decrease accounts for less than
20 % of the pre-treatment household income.
14Couples tend to have stronger decreases in mental health due to unemployment, when the income of
the directly affected spouse was more important for the couple’s income. Also this finding from the
tentative analyses underlines that income should not be excluded as moderator variable.
15As in appendix A.3, for kernel matching I use the linear index of the propensity score and an Epanech-
nikov kernel with a bandwidth of 0.06.
16
Table 4: The effect of unemployment on mental health - sensitivity analyses
Main PS- PS- All plant Placebo
specification weighting matching closures regress.
Outcome (4) (5) (6) (7) (8)
All couples
Own mental health −2.68∗∗∗ −2.60∗∗∗ −2.73∗∗∗ −0.76 −0.32
(0.72) (0.70) (0.74) (0.50) (0.64)
Partner’s mental health −1.81∗∗∗ −1.83∗∗∗ −1.64 ∗ ∗ −1.22 ∗ ∗ 0.13
(0.66) (0.61) (0.67) (0.48) (0.63)
p-value of difference 0.32 0.36 0.23 0.44 0.53
NT reated 109 109 109 288 62
Husband’s unemployment
Own mental health −3.13∗∗∗ −3.03∗∗∗ −3.52∗∗∗ −0.59 −0.91
(0.88) (0.84) (0.92) (0.60) (0.57)
Partner’s mental health −2.01∗∗∗ −2.11∗∗∗ −1.91 ∗ ∗ −0.96∗ −0.05
(0.77) (0.72) (0.79) (0.58) (0.52)
p-value of difference 0.27 0.34 0.14 0.62 0.14
NT reated 70 70 70 194 33
Wife’s unemployment
Own mental health −1.86∗∗∗ −1.67∗∗∗ −1.39 ∗ ∗ −1.12∗0.36
(0.63) (0.58) (0.63) (0.67) (0.69)
Partner’s mental health −1.46∗ −1.30 ∗ ∗ −1.09∗ −1.76∗∗∗ 0.35
(0.77) (0.63) (0.64) (0.65) (0.75)
p-value of difference 0.63 0.61 0.69 0.40 0.99
NT reated 39 39 39 94 29
Note: The table presents the effect of one spouse’s entry into unemployment on the mental health of
both spouses. Each cell displays the ATT from a separate regression and its robust standard error in
parentheses. Additionally, the table provides the number of treated couples (NT reated ) and p-values for
the t-tests whether unemployment differently influences directly and indirectly affected spouses. The
upper panel considers all couples, while the two lower panels display results separately according to the
sex of the directly affected spouse. Specification (4) is the main estimation specification as in table 3,
specifications (5) and (6) display the results for propensity score weighting and propensity score (kernel)
matching, respectively. In specification (7) the treatment group comprises all couples that experienced
a plant closure and specification (8) performs a placebo regression that pretends that the treatment
takes place two years earlier. * p < 0.1; ** p < 0.05; *** p < 0.01.
17
and wife’s mental health get more similar for propensity score weighting and more dif-
ferent for kernel matching. Also these specifications do not reject the null-hypothesis
that unemployment influences directly and indirectly affected spouses in the same way,
as indicated by the p-values.
As outlined in section 2, research on the consequences of unemployment is closely
intertwined with research on the consequences of job loss. In order to make the results
comparable to other studies that solely analyze job losses (e.g. Browning et al. 2006;
Brand et al. 2008;Kuhn et al. 2009;Salm 2009), the treatment group in specification
(7) comprises all couples that experienced a plant closure (irrespective of unemploy-
ment experiences).16 The results in table 4 indicate that plant closures per se do not
have such negative impact on mental health. Rather, it is the unemployment experi-
ence that decreases mental health. Hence, failing to differentiate between job loss and
unemployment might be a potential reason why existent studies do not find an effect of
job loss on own mental health (e.g. Browning et al. 2006;Brand et al. 2008;Salm 2009).
Own plant closure experiences significantly decrease own mental health for wifes, but
no longer for husbands. The decrease in the spouses’ mental health remains significant.
Also the effects on the spouses’ mental health are smaller than in the main specification
indicating that those with unemployment following a plant closure have larger drops in
mental health.17 In specification (7), the mental health consequences are slightly larger
for couples in which the wife experienced a plant closure.
To identify causal effects, all matching procedures assume that the conditioning vari-
ables include all variables simultaneously influencing changes in mental health and the
probability of becoming unemployed due to plant closure. This assumption cannot be
directly tested. In order to add additional credibility to this assumption, I perform a
placebo regression. For this purpose, I pretend that the treatment takes place two years
earlier. Accordingly, for the first step I compute the weights based on conditioning vari-
ables obtained in the last year with health data before the placebo job loss. Specification
(8) shows that the placebo treatment does not influence changes in mental health. All
estimated effects are insignificant and close to zero. This specification adds plausibility
to the assumption that the mental health of treated and matched controls follows a
16Apart from this difference, the same rules for the selection of the treatment group apply as before
(see figure 1 and section 4.2), in particular couples where both spouses experienced a plant closure
are disregarded.
17Considering only couples that are included in specification (7) but not in specifications (1)-(6) indi-
cates that only the effect of the wive’s plant closure on the mental health of the spouse is significant
(results are not shown, but are available upon request).
18
similar trend before the treatment.
Despite the robustness of the results to different matching procedures, the ATT es-
timates might be downward-biased for several reasons. First, there might be selective
panel attrition. Couples experiencing a greater negative impact from unemployment
(e.g. with respect to finances or identity) might be more likely to drop out of the sam-
ple.18 These couples might also be more likely to experience greater decreases in mental
health. Related to this argument for selective panel attrition is the finding that job loss
increases the risk to commit suicide (Eliason & Storrie 2009), which is an extreme form of
mental health problems. Similarly, also divorce increases the chances of panel attrition,19
and job loss increases the probability of divorce (Charles & Stephens 2004). Second, the
expectation of the plant closure might already decrease mental health. Hence, for treated
couples the pre-treatment mental health score would be lower than their “normal” men-
tal health score. Using German SOEP data, Reichert & Tauchmann (2011) provide
evidence that already the fear of becoming unemployed decreases mental health. Yet,
specification (8) indicates that treated and matched controls do not differ with respect
to their mental health trend before the treatment. One reason for this might be that the
conditioning variables include also the perceived job security and that I, hence, compare
only couples having similar risks of job loss.
Additionally, the effects might be biased since the analyses do not include couples
where both spouses experience a plant closing. The direction of this bias is not clear. If
both partners experience a plant closure, the drop in household income is even larger.
This suggests larger decreases in mental health. However, findings in Clark (2003)
suggest that unemployment hurts less when the partner is unemployed as well. When
including the seven couples in the treatment group where both partners experienced a
plant closure, the results do not change meaningfully.
As the present analysis focuses on the short-term impact of unemployment on mental
health, it does not consider the full mental health consequences of unemployment. Long-
term consequences like potential scarring effects (past unemployment has a negative
impact even when the individuals have become reemployed; Clark et al. 2001) and the
effects of future job insecurity (Stevens 1997;Knabe & Rätzel 2011) are not taken into
account.
18For instance, Dorsett (2010) finds unemployment to be related to panel attrition when comparing
survey and register data.
19Although the SOEP following rules should mitigate divorce related panel attrition, see section 4.
19
6.2. Other Reasons for Unemployment
In order to assess the endogeneity of other reasons for unemployment, table 5 presents
the effects of different reasons for entering unemployment on the mental health of both
spouses.20 Similar to Kassenboehmer & Haisken-DeNew (2009), I differentiate between
unemployment due to plant closure, dismissal, and further reasons.21 I also examine the
effect of all reasons for unemployment jointly.22
Many studies that analyze health consequences of job loss and unemployment do
not differentiate between the reasons for unemployment/job loss (e.g. Browning et al.
2006;Eliason & Storrie 2009;Kuhn et al. 2009;Sullivan & von Wachter 2009;Deb
et al. 2011;Browning & Heinesen 2012). Those studies that differentiate, provide mixed
evidence on the differential effects. Salm (2009) does not find any decreases in health
for neither job loss due to lay-off nor for job loss due to plant closure in the U.S.; he
observes a decrease only for those who left the job for health reasons. Schmitz (2011)
does not find any effects of unemployment due to plant closure on health for Germany.
However, he provides evidence that unemployment due to other reasons decreases health
satisfaction, increases the number of hospital visits and reduces mental health. Marcus
(2012) shows that job loss increases smoking initiation, irrespective of whether job loss
due to plant closure or due to dismissal is considered. A study on subjective well-being
(Kassenboehmer & Haisken-DeNew 2009) demonstrates that, in Germany, the subjective
well-being decrease is even larger from unemployment due to plant closures than from
unemployment due to dismissal.
Comparing the effects of all unemployment (first column in table 5) with the effect
of unemployment due to plant closure (second column) shows that unemployment due
to plant closure seems to hurt more than unemployment in general. Potentially the all
unemployment category includes also cases of “voluntary” unemployment, which might
be less detrimental for both spouses’ mental health. Therefore, the next column looks at
unemployment due to dismissal, which is often regarded as involuntary unemployment
(e.g. Kassenboehmer & Haisken-DeNew 2009;Marcus 2012). Two arguments support
the idea that unemployment due to dismissals is more detrimental for mental health
20All specifications in table 5 base upon the procedure of the main specification (specification 4 in table
3), with the only difference that the treatment group varies between the columns.
21Further reasons include own resignation, mutual agreement, end of temporary contract and leave of
absence/sabbatical.
22I would like to thank the anonymous referee for drawing my interest in investigating the effect of
other reasons for unemployment.
20
Table 5: Changes in mental health after different entries into unemployment
Reason for unemployment
plant
Outcome all closure dismissal further
All couples
Own mental health −1.42∗∗∗ −2.68∗∗∗ −1.83∗∗∗ −0.65
(0.38) (0.72) (0.53) (0.53)
Partner’s mental health −0.91∗∗∗ −1.81∗∗∗ −0.83 −0.59
(0.35) (0.66) (0.51) (0.48)
p-value of difference 0.25 0.32 0.11 0.92
NT reated 761 109 323 329
Husband’s unemployment
Own mental health −1.66∗∗∗ −3.13∗∗∗ −1.61∗∗∗ −1.20 ∗ ∗
(0.46) (0.88) (0.61) (0.61)
Partner’s mental health −0.98 ∗ ∗ −2.01∗∗∗ −0.54 −0.92
(0.46) (0.77) (0.64) (0.63)
p-value of difference 0.26 0.27 0.18 0.72
NT reated 456 70 205 181
Wife’s unemployment
Own mental health −1.08∗ −1.86∗∗∗ −2.22∗∗∗ 0.02
(0.60) (0.63) (0.79) (0.75)
Partner’s mental health −0.81 −1.46∗ −1.33∗ −0.18
(0.50) (0.77) (0.70) (0.58)
p-value of difference 0.68 0.63 0.29 0.80
NT reated 305 39 118 148
Note: The table presents the effect of different reasons for entering unemployment on the mental health
of both spouses. Each cell displays the ATT from a separate regression and its robust standard error in
parentheses. Additionally, the table provides p-values for the t-tests whether unemployment differently
influences directly and indirectly affected spouses. The upper panel considers all couples, while the two
lower panels display results separately according to the sex of the directly affected spouse. * p < 0.1;
** p < 0.05; *** p < 0.01.
21
than due to plant closures. First, many researchers argue that poor health might cause
unemployment (e.g. Kuhn et al. 2009;Schmitz 2011). In line with this argument, some
dismissals might result from deteriorating health.23 Second, dismissals rather than plant
closures might be attributable to own behavior. Therefore, dismissals might hurt more.
Contrary, there are also arguments that unemployment due to plant closures is more
detrimental for mental health than due to dismissals. Plant closures might make it more
difficult to find a new job as others in the region (with similar qualifications) also lost
their jobs due to the plant closure. Additionally, individuals staying with the firm until it
finally closes, might be a group with the worst labor market prospects outside that firm,
low flexibility and/or a high identification with the firm. Furthermore, some dismissals
might actually be voluntary quittings in order to get compensation (Kassenboehmer &
Haisken-DeNew 2009: 453-454).
Table 5 shows that for wife’s entry into unemployment the decreases in mental health
are similar for plant closure and dismissal. However, for husband’s entry into unem-
ployment, plant closures are more detrimental than dismissals for the mental health of
both spouses. The effect on the indirectly affected spouse is even insignificant in this
specification. This comparison suggests that the mental health consequences of unem-
ployment are not overestimated when considering dismissals as well. Either endogeneity
issues with respect to the effect of dismissals on mental health are rather weak in Ger-
many or the effects of different endogeneity mechanisms cancel out. When looking at
the effect of unemployment that results from other reasons than dismissal and plant
closure in the last column, the effects of wife’s unemployment on mental health do not
only turn insignificant but become virtually zero. For husband’s own unemployment the
decrease in mental health is smaller than before but still significant, while the effect on
his partner is only slightly smaller but turns insignificant. However, this category might
also include some cases of chosen unemployment.
Table 5 highlights that for each specification, the coefficients for the indirectly affected
spouse is similar to that for the directly affected spouse. The two coefficients are not
significantly different in any of the specifications. This confirms the conclusion that
unemployment decreases the mental health of spouses almost as much as for the directly
affected individuals.
23Even controlling for pre-treatment mental health may not rule out reverse causality, as mental health
might decrease between the last pre-treatment interview and the dismissal. Researchers could not
distinguish this decrease before the dismissal from a decrease due to the dismissal.
22
7. Conclusion
This paper analyzes spillover effects of unemployment on other household members by
estimating the causal effect of unemployment on spouse’s mental health. Using data
from the German Socio-Economic Panel Study (SOEP) from 2002 through 2010, this
paper finds that unemployment of one spouse similarly affects the mental health of both
spouses. About one year after the plant closure, unemployment decreases mental health
by 27 % of a standard deviation for the unemployed individuals themselves and by 18
% of a standard deviation for their spouses. In general, the decreases in mental health
are larger when the husband enters unemployment. The findings are robust over various
specifications.
In order to give the estimates a causal interpretation, this study focuses on an ex-
ogenous entry into unemployment (i.e. plant closure), and applies a combination of
matching and difference-in-difference estimation that is robust against selection on ob-
servables and selection on unobservables with time-invariant effects. The matching part
of the estimator constitutes one of the first applications of entropy balancing (Hain-
mueller 2012), which balances the conditioning variables more effectively than common
propensity score methods. The estimation strategy assumes that no unobserved vari-
ables exist that simultaneously influence changes in mental health and the probability
of entering unemployment due to plant closure. This paper provides an indirect test to
show that this identifying assumption is not violated, as mental health does not follow
a different trend for treated and matched controls before the plant closure.
The findings highlight that unemployment has severe consequences not just for the
directly affected individuals, but also for their spouses. Hence, previous studies under-
estimate the public health costs of job loss as they do not consider the consequences
for spouses. When comparing costs and benefits of labor market policies to prevent un-
employment, policy-makers should take into account that employment has non-financial
benefits not only for the employed individuals themselves but also for their spouses.
23
References
Andersen, H. H., Mühlbacher, A., Nübling, M., Schupp, J., & Wagner, G. G. (2007).
Computation of standard values for physical and mental health scale scores using the
SOEP version of SF-12v2. Schmollers Jahrbuch, 127(1), 171–182.
Böckerman, P. & Ilmakunnas, P. (2009). Unemployment and self-assessed health: Evi-
dence from panel data. Health Economics, 18(2), 161–179.
Bolger, N., DeLongis, A., & Kessler, R. (1989). The contagion of stress across multiple
roles. Journal of Marriage and the Family, 51(1), 175–183.
Brand, J. E., Levy, B. R., & Gallo, W. T. (2008). Effects of layoffs and plant closings
on depression among older workers. Research on Aging, 30(6), 701–721.
Browning, M., Dano, A. M., & Heinesen, E. (2006). Job displacement and stress-related
health outcomes. Health Economics, 15(10), 1061–1075.
Browning, M. & Heinesen, E. (2012). Effect of job loss due to plant closure on mortality
and hospitalization. Journal of Health Economics, 31(4), 599–616.
Caliendo, M. & Kopeinig, S. (2008). Some practical guidance for the implementation of
propensity score matching. Journal of Economic Surveys, 22(1), 31–72.
Charles, K. K. & Stephens, M. (2004). Job displacement, disability, and divorce. Journal
of Labor Economics, 22(2), 489–522.
Clark, A. E. (2003). Unemployment as a social norm: Psychological evidence from panel
data. Journal of Labor Economics, 21(2), 323–351.
Clark, A. E., Georgellis, Y., & Sanfey, P. (2001). Scarring: The psychological impact of
past unemployment. Economica, 68(270), 221–241.
Deb, P., Gallo, W. T., Ayyagari, P., Fletcher, J. M., & Sindelar, J. L. (2011). The effect
of job loss on overweight and drinking. Journal of Health Economics, 30(2), 317–327.
Dew, M. A., Bromet, E. J., & Penkower, L. (1992). Mental health effects of job loss in
women. Psychological Medicine, 22(3), 751–764.
Dorsett, R. (2010). Adjusting for nonignorable sample attrition using survey substitutes
identified by propensity score matching: An empirical investigation using labour mar-
ket data. Journal of Official Statistics, 26(1), 105–125.
Eliason, M. & Storrie, D. (2009). Does job loss shorten life? Journal of Human Re-
sources, 44(2), 277–302.
Gill, S. C., Butterworth, P., Rodgers, B., & Mackinnon, A. (2007). Validity of the
mental health component scale of the 12-item Short-Form Health Survey (MCS-12) as
measure of common mental disorders in the general population. Psychiatry Research,
152(1), 63–71.
24
Hainmueller, J. (2011). Ebalance: A Stata package for entropy balancing. MIT Political
Science Department Research Paper, 24.
Hainmueller, J. (2012). Entropy balancing for causal effects: A multivariate reweighting
method to produce balanced samples in observational studies. Political Analysis, 20,
25–46.
Heckman, J., Ichimura, H., & Todd, P. E. (1997). Matching as an econometric evaluation
estimator: Evidence from evaluating a job training programme. Review of Economic
Studies, 64(4), 605–654.
Jahoda, M. (1979). The impact of unemployment in the 1930s and the 1970s. Bulletin
of the British Psychological Society, 32(2), 309–314.
Jones, F. (1993). An empirical study of occupational stress transmission in working
couples. Human Relations, 46(7), 881–903.
Kassenboehmer, S. C. & Haisken-DeNew, J. P. (2009). You’re fired! The causal negative
effect of entry unemployment on life satisfaction. Economic Journal, 119(536), 448–
462.
Knabe, A. & Rätzel, S. (2011). Scarring or scaring? The psychological impact of past
unemployment and future unemployment risk. Economica, 78(310), 283–293.
Kuhn, A., Lalive, R., & Zweimüller, J. (2009). The public health costs of job loss.
Journal of Health Economics, 28(6), 1099–1115.
Kullback, S. (1959). Information theory and statistics. New York.
Lechner, M. (2009). Long-run labour market and health effects of individual sports
activities. Journal of Health Economics, 28(4), 839–854.
Lechner, M. (2010). The estimation of causal effects by difference-in-difference methods.
University of St. Gallen Department of Economics Working Paper Series, 28.
Liem, R. & Liem, J. H. (1988). Psychological effects of unemployment on workers and
their families. Journal of Social Issues, 44(4), 87–105.
Lindeboom, M., Portrait, F., & van den Berg, G. J. (2002). An econometric analysis
of the mental-health effects of major events in the life of older individuals. Health
Economics, 11(6), 505–520.
Lindo, J. (2011). Parental job loss and infant health. Journal of Health Economics,
20(5), 869–879.
Marcus, J. (2012). Does job loss make you smoke and gain weight? SOEPpapers on
Multidisciplinary Panel Data Research, 432.
25
Penkower, L., Bromet, E. J., & Dew, M. A. (1988). Husbands’ layoff and wives’ mental
health: A prospective analysis. Archives of General Psychiatry, 45(11), 994–1000.
Reichert, A. & Tauchmann, H. (2011). The causal impact of fear of unemployment on
psychological health. Ruhr Economic Papers, 266.
Rosenbaum, P. R. & Rubin, D. B. (1983). The central role of the propensity score in
observational studies for causal effects. Biometrika, 70(1), 41–55.
Salm, M. (2009). Does job loss cause ill health? Health Economics, 18(9), 1075–1089.
Salyers, M. P., Bosworth, H. B., Swanson, J. W., Lamb, J., & Osher, F. C. (2000).
Reliability and validity of the SF-12 health survey among people with severe mental
illness. Medical Care, 38(11), 1141–1150.
Schmitz, H. (2011). Why are the unemployed in worse health? The causal effect of
unemployment on health. Labour Economics, 18(1), 71–78.
Siedler, T. (2011). Parental unemployment and young people’s extreme right-wing party
affinity: Evidence from panel data. Journal of the Royal Statistical Society: Series A
(Statistics in Society), 174(3), 737–758.
Siegel, M. J., Bradley, E. H., Gallo, W. T., & Kasl, S. V. (2003). Impact of husbands’
involuntary job loss on wives’ mental health, among older adults. The Journals of
Gerontology. Series B, Psychological Sciences and Social Sciences, 58(1), S30–S37.
Stephens, M. (2002). Worker displacement and the added worker effect. Journal of
Labor Economics, 20(3), 504–537.
Stevens, A. H. (1997). Persistent effects of job displacement: The importance of multiple
job losses. Journal of Labor Economics, 15(1), 165–188.
Sullivan, D. G. & von Wachter, T. (2009). Job displacement and mortality: An analysis
using administrative data. Quarterly Journal of Economics, 124(3), 1265–1306.
ten Brummelhuis, L. L., Haar, J. M., & van der Lippe, T. (2010). Crossover of distress
due to work and family demands in dual-earner couples: A dyadic analysis. Work &
Stress, 24(4), 324–341.
Wagner, G., Frick, J., & Schupp, J. (2007). The German Socio-Economic Panel Study
(SOEP) - Scope, evolution and enhancements. Schmollers Jahrbuch, 127(1), 139–169.
Ware, J. E., Kosinski, M., & Keller, S. D. (1996). A 12-item short-form health survey
of scales and preliminary construction tests of reliability and validity. Medical Care,
34(3), 220–233.
Winkelmann, L. & Winkelmann, R. (1995). Happiness and unemployment: A panel
data analysis for Germany. Konjunkturpolitik, 41(4), 293–307.
26
A. Appendix
A.1. Entropy balancing
Entropy balancing reweights the control group observations in such a way that the con-
trol group satisfies pre-specified balancing requirements (here: same mean and variance
of conditioning variables as in the treatment group). Among the possible sets of weights
that fulfill these balancing requirements, entropy balancing choses the set of weights
that deviates as little as possible from uniform weights (Hainmueller 2012).24 Entropy
balancing spares the need to check for covariate balance since balance according to the
pre-specified balancing requirements is fulfilled by construction. This makes the burden-
some procedure of propensity score methods unnecessary, where “researchers ’manually’
iterate between propensity score modeling, matching, and balance checking until they
attain a satisfactory balancing solution” (Hainmueller 2012: 25). Entropy balancing is
more effective as it improves the balance reached by common propensity score methods
for all covariates. Furthermore, while propensity score methods often decrease balance
on some covariates, entropy balancing improves balance for all conditioning variables.
In contrast to propensity score methods, entropy balancing is fully non-parametric.
In the present analysis, I require the control group to have the same mean and the same
variance as the treatment group for all conditioning variables after entropy balancing. In
applications of propensity score methods usually only the balance of the first moments is
checked. Entropy balancing is implemented using the program “ebalance” (Hainmueller
2011) in Stata 11.2.
Besides the unconfoundedness assumption (see eq. 2), a further requirement of match-
ing estimators is the overlap condition (P(D= 1|X)<1). It ensures that for any given
X, there are not just treated observations but also control observations. It is not clear
how to impose common support conditions for entropy balancing. Most studies relying
on propensity score methods exclude treated observations whose propensity score ex-
ceeds the maximum of the propensity score in the control group in order to implement
the common support for the ATT (Caliendo & Kopeinig 2008). However, this common
support condition is not binding in the present analysis since members of the control
group have the highest propensity score values. Hence, the overlap condition seems to
be satisfied in the present case. Furthermore, entropy balancing requires that there are
more observations in the control group than moment restrictions.
24To compare deviations from uniform weights, entropy weighting uses the entropy divergence (Kullback
1959) as distance measure. Hence the name entropy balancing.
27
A.2. Overview of the SOEP version of the SF-12 questions
28
A.3. Descriptive statistics - means and standardized bias
Means Means
treated controls Standard. Bias (%)
Variable raw EB kernel raw EB kernel
Directly affected spouse
Age 48.1 44.0 48.1 48.2 45.4 0.0−1.7
Female+35.8 43.6 35.8 35.8−16.1 0.0 0.0
Migrant+23.9 15.3 23.9 22.4 21.7 0.0 3.4
Non-German+22.9 12.4 22.9 21.8 27.8 0.0 2.8
Physical health 49.8 52.2 49.8 50.0−28.0 0.0−2.1
Mental health 49.9 50.7 49.9 50.0−9.6 0.0−1.2
Often melancholic+51.4 46.2 51.4 50.4 10.4 0.0 2.0
Bad health+14.7 9.3 14.7 13.2 16.4 0.0 4.2
Medium health+34.9 32.5 34.9 35.5 5.0 0.0−1.4
Good health+50.5 58.1 50.5 51.2−15.4 0.0−1.5
Labor earnings 24.9 35.2 24.9 25.2−46.3 0.0−2.3
Never unemployed+62.4 70.6 62.4 62.5−17.4 0.0−0.3
Tenure 11.0 12.4 11.0 10.9−14.6 0.0 0.6
Small company+33.9 19.9 33.9 32.1 32.0 0.0 3.9
Medium-small company+45.9 30.5 45.9 47.4 31.9 0.0−3.0
Medium company+11.9 24.0 11.9 12.3−31.8 0.0−1.2
Large company+8.3 23.5 8.3 8.2−42.5 0.0 0.2
Big job worries+39.4 15.1 39.4 39.9 56.6 0.0−0.9
Some job worries+38.5 45.9 38.5 38.8−15.0 0.0−0.5
No job worries+22.0 39.0 22.0 21.3−37.4 0.0 1.8
Years full-time 21.2 18.0 21.2 21.3 29.3 0.0−0.4
Primary sector+0.0 1.3 0.0 0.0−16.4 0.0 0.0
Manufacturing+39.4 30.0 39.4 38.8 19.9 0.0 1.4
Energy and water+0.0 1.3 0.0 0.0−16.0 0.0 0.0
Construction+13.8 5.5 13.8 14.0 28.0 0.0−0.6
Wholesale and retail+23.9 11.8 23.9 24.7 31.9 0.0−1.9
Hotel and restaurants+0.9 1.4 0.9 1.1−4.4 0.0−1.5
Transport+1.8 5.3 1.8 1.7−18.5 0.0 0.9
Banking and insurance+2.8 5.1 2.8 2.4−12.3 0.0 2.0
Health services+2.8 11.7 2.8 3.0−35.1 0.0−1.7
Other services+11.0 23.3 11.0 11.2−33.0 0.0−0.7
Basic school+45.0 29.1 45.0 45.4 33.2 0.0−0.8
Intermediate school+45.0 44.9 45.0 42.9 0.1 0.0 4.1
Technical college+2.8 6.8 2.8 3.7−19.1 0.0−5.3
Highest secondary+7.3 19.2 7.3 8.1−35.4 0.0−2.7
University+8.3 23.3 8.3 9.6−42.2 0.0−4.8
Vocational training+78.9 76.8 78.9 77.9 4.9 0.0 2.5
29
Means Means
treated controls Standard. Bias (%)
Variable raw EB kernel raw EB kernel
Indirectly affected spouse
Age 47.2 44.0 47.2 47.7 31.4 0.0−4.1
Migrant+26.6 15.4 26.6 24.4 27.7 0.0 5.1
Non-German+22.9 12.4 22.9 21.2 27.7 0.0 4.1
Physical health 48.9 51.4 48.9 49.0−29.1 0.0−1.0
Mental health 49.3 50.4 49.3 49.4−11.7 0.0−0.7
Often melancholic+57.8 47.9 57.8 55.8 19.8 0.0 4.0
Bad health+17.4 11.7 17.4 17.6 16.1 0.0−0.4
Medium health+38.5 32.0 38.5 37.0 13.6 0.0 3.2
Good health+44.0 56.2 44.0 45.5−24.5 0.0−2.9
Labor earnings 21.0 24.6 21.0 20.7−14.4 0.0 1.3
Never unemployed+56.9 65.1 56.9 57.9−17.0 0.0−2.1
Works full-time+49.5 52.2 49.5 48.0−5.4 0.0 3.0
Works part-time+19.3 25.1 19.3 20.4−14.1 0.0−2.9
Not working+31.2 22.7 31.2 31.6 19.3 0.0−0.8
Basic school+36.7 28.1 36.7 38.3 18.5 0.0−3.2
Intermediate school+48.6 46.1 48.6 47.6 5.0 0.0 2.0
Technical college+6.4 6.4 6.4 5.4 0.1 0.0 4.2
Highest secondary+8.3 19.4 8.3 8.7−32.7 0.0−1.5
University+11.9 22.4 11.9 11.8−28.0 0.0 0.5
Vocational training+74.3 75.4 74.3 73.5−2.4 0.0 1.9
Couple information
Children+37.6 49.8 37.6 36.9−24.7 0.0 1.4
Regional unemployment 9.6 10.2 9.6 9.7−13.7 0.0−1.9
Village+11.0 9.1 11.0 10.8 6.5 0.0 0.6
Small town+37.6 35.8 37.6 39.1 3.7 0.0−3.0
Small city+27.5 28.1 27.5 25.6−1.3 0.0 4.3
Big city+23.9 27.0 23.9 24.5−7.2 0.0−1.5
Year 2002+40.4 28.4 40.4 39.4 25.3 0.0 2.0
Year 2004+23.9 25.6 23.9 24.4−4.1 0.0−1.2
Year 2006+11.9 24.6 11.9 10.8−33.1 0.0 3.6
Year 2008+23.9 21.4 23.9 25.5 5.9 0.0−3.7
N 109 14285
Note: Summary statistics for treated couples, all control couples and matched control couples. The first
two columns present the means of selected variables before treatment for treated and controls. Third
and fourth column show the means for the reweighted control group according to entropy balancing
(EB) and kernel matching, a propensity score method. The last three columns display a measure for
the quality of the matching process. The standardized bias is defined for each conditioning variable s
as SBs= 100 ·s1−s0
√1
2(σ2
s1+σ2
s0), where s1and s0are the means of treated and controls, respectively, and
σ2
s1and σ2
s0the corresponding variances. +indicates that the mean represents a percentage share.
30
