The Causal Effect of Job Loss on
Mental Health in the Netherlands
Eline Jorritsma s2521075 June 5, 2019
Master’s Thesis Economics (EBM877A20) Supervisor: Prof. V. Angelini
Abstract
This thesis studies the causal effect of job loss on five subjective mental health measures in the Netherlands. Data from the LISS Core Study from 2017 and 2018 was used to estimate the causal effects. With the aid of Propensity Score Matching and a parametric model that is dependent on a set of covariates, this thesis does not find any significant evidence for adverse effects of job loss on any of the five subjective mental health measures. The parametric approach does find large significant negative effects on subjective mental health for
individuals who have difficulties making ends meet. Moreover, ageing and higher levels of education have a significant positive impact on subjective mental health. These outcomes are assumed to provide proof of the existence of selection bias that may be caused by the
presence of the confounding factor socioeconomic status.
2
1. INTRODUCTION
Job loss is an involuntary life event with a potentially far-reaching impact on a worker’s life. As we recently went through a recession, we saw a rise in unemployment in the Netherlands. The unemployment level peaked in the first quarter of 2014, when the Netherlands saw 694,000 unemployed individuals in the work force, opposed to 316,000 individuals five years later, in the first quarter of 20191. Having experienced high levels of unemployment in the recent past, attention has been drawn to the effects job loss and unemployment have on individuals and society. Brand (2015) has looked into the impact of job loss and
unemployment on the life of American individuals. She has found that there are not only economic effects of job loss, but research suggests that it is also associated with non-economic effects such as declines in psychological and physical well-being.
The effect of job loss on mental health has been examined in many strands of research. Studies from research papers in medicine, psychology, sociology and economics have investigated and shown over the years that job loss is associated with adverse mental health outcomes. Both cross-sectional and panel data sets and objective and subjective mental health measures have been used in the literature. Paul and Moser (2009) examined the effect of unemployment on mental health with meta-analyses across 237 cross-sectional and 87 longitudinal studies. The overall outcome was that for several indicator variables of mental health such as depression, anxiety and subjective well-being, the unemployed show more distress than employed individuals. The negative effect of unemployment was found to be stronger in countries with weak levels of economic development, unequal income
distributions and weak employment protection systems. Their outcome supports the
assumption that unemployment is not only correlated with mental distress, but also causes it. Bartley and Owen (1996) have examined the relationship between the socioeconomic status of men and their employment status. They found that healthy men in lower socioeconomic groups showed higher rates of unemployment. Hudson (2005) tested the causal effect of socioeconomic status on mental illness using a longitudinal database in Massachusetts. The study found that socioeconomic status has a direct impact on mental illness. Moreover, socioeconomic status indirectly affects mental health through the economic hardship of low and middle income groups. Hence, it is possible that a difference in mental health between
3 employed and unemployed individuals is due to confounding factors, such as an individual’s socioeconomic status. Therefore, in a study on the effect of job loss on mental health such confounding factors must be taken into account.
The aim of this thesis is to estimate and isolate the causal effect of job loss on mental health on individuals in the Netherlands. This will be attempted using data from the LISS Core Study from 2017 and 2018 and two statistical approaches. The data provides five different subjective mental health states and provides data on individual characteristics. The first estimation approach is the non-parametric Propensity Score Matching (PSM) method. The second approach is an estimation with Ordinary Least Squares (OLS), using a vector of covariates to control for individual characteristics between individuals. This thesis does not find any significant evidence for a (negative) causal relationship between job loss and mental health. Thus, this thesis suggests that the difference in mental health states between employed and unemployed individuals in the Netherlands is not caused by job loss, but possibly by confounding factors such as an individual’s socioeconomic status.
This thesis adds to the existing literature on the relationship between mental health and unemployment by attempting to capture the causal impact of job loss. The existing literature mostly focuses on the causal effects that arise due to job displacement such as income loss, expectations on finding a new job and self-esteem as has been done by Strandh (2000) and Baik et al. (1989). However, few scholars have attempted to isolate and estimate the causal effect of job loss through a PSM model which is done in this thesis. In the literature the PSM model has been used to estimate the causal effect of job loss on physical health, as has been done by Böckerman and Ilmakunnas (2009), who find fail to find a causal effect of
unemployment on self-assessed health status using panel data on Finnish individuals.
4
2. LITERATURE REVIEW
2.1 Mental Health and Unemployment
2.1.1 Future Job Prospects
To individuals, unemployment is supposed to be a temporary period in their life. Unemployment normally only occurs after graduation, or when individuals are in-between jobs. However, unemployment is open-ended, it is uncertain when and if there will be an end to it. Baik et al. (1989) use American data from a survey that has been carried out in four different branches of a state employment commission to estimate the relationship between involuntary job loss and psychological distress. They find significant evidence that the expectations on finding a new job help explain the negative effect of job loss on mental health. The individuals with expectations on a successful job search had a better mental health status than individuals with lower expectations. When comparing British individuals who were laid-off for seven weeks to workers who were permanently laid-off, i.e. who needed to find a new job, Fryer and McKenna (1987) found a similar result. Even though all individuals in both groups were unemployed, those who knew it would be temporary, and thus had good expectations about re-employment, had a better mental health status than the individuals who did not know if they would find a new job. It was suggested that this was due to the
predictable nature of the unemployment for the temporarily laid-off group. Knowing that they would be working again after seven weeks, and thus having a positive outlook on
re-employment, reduced the negative impact of unemployment on mental health.
2.1.2 Economic Dependence
5 et al. (1993) study the effects of unemployment on the mental health of Finnish individuals who received a questionnaire after the wood-factory that previously employed them had been closed down. Using employees from a similar Finnish wood-factory as a control group, they found that individuals who rate their income level as ‘low’ compared to ‘moderate’ or ‘good’ suffer significantly more in unemployment from mental health issues than their employed counterparts. Marcus (2013) investigates the effect of unemployment on mental health of spouses, using German Socio-Economic Panel Study data. His findings, slightly contrasting the findings of Viinamaki et al. (1993), show that the mental health of spouses of the
unemployed is not strongly affected by the drop in household income. However, Marcus (2013) does suggest that the effects might work through uncertainty about future income and therefore cannot conclude that income as a whole is irrelevant to mental health. Theodossiou (1998) investigates the effects of low-pay and unemployment on six different measures of mental health. Using data from the 1992 British Household Panel Study (BHPS) Theodossiou (1998) finds that the unemployed have a significantly higher probability of suffering from a rise in depression, loss of confidence and anxiety and a decrease in self-esteem and the level of general happiness. Compared to individuals who are in low-paid employment these effects are still stronger for the unemployed. Hence, these findings suggest that individuals value being employed highly and value having a low-paid job higher than being unemployed.
2.1.3 Control of the Life Course
Moreover, Strandh (2000) suggests that the negative effects of unemployment on mental health are partially due to the loss of control over life. This is based on the theory brought up by Fryer (1986), who describes people as social actors who, to the best of their abilities, try to control a situation and reach a desirable outcome. When this fails, Fryer (1986) theorizes, mental health will deteriorate. In line with this theory, Strandh (2000), following Fryer’s agency perspective (Fryer, 1986; Fryer and McKenna 1987) suggests unemployment imposes unwanted restrictions on an individual’s ability to exercise power over their life, leading to a deterioration in mental health. The insecurity that comes with unemployment affects an individual’s ability to control their life course, to plan ahead, and to make
6 said that economic problems and restrictions on the ability to plan ahead lead to worse mental health in individuals who did not choose to go into unemployment voluntarily.
2.1.4 Age
Theodossiou (1998) finds that all measures of mental well-being are lower for the middle age group compared to younger and older age groups. Breslin and Mustard (2003) use Canadian data from the National Population Health Survey to examine the relationship between unemployment and mental health among young and older adults. They find that individuals between 31 and 55 years old who become unemployed show signs of increased distress and clinical depression at follow-up two years later. This effect of unemployment on mental health was not found for younger individuals between the ages of 18 and 30.
Suggested reasons for these findings are the possible family responsibilities and financial commitments that slightly older individuals have, compared to the younger individuals.
2.1.5 Gender
Several scholars have found that the effect of unemployment on mental health is not similar for men and women. Being a man is more traditionally associated with having a job. Women have a more widely accepted access to alternative roles in society, such as being a housewife or a stay-at-home mother. Although these traditional gender roles are slowly fading in western societies, it might be that effects of unemployment on mental health are not as strong for females than for males. Supporting this argument, meta-analyses by Paul and Moser (2009) show that the mental health of men is more strongly affected by unemployment than that of women.
2.2 Causation and the Selection Bias
7 any difference in mental health between unemployed and employed individuals is not caused by job loss. The bias that is associated with these phenomena is called the Selection Bias. The question rises whether there has been enough focus on this endogeneity problem in research for the causal effect of unemployment.
2.2.1 Plant Closures
Several scholars have tried to solve for this problem by using data on plant closures. As all workers are displaced from their jobs when a plant shuts down, their individual
characteristics or any confounding factors did not cause job loss and therefore it can be argued that it will greatly reduce the presence of selection bias.
Eliason and Storrie (2009) find, when using Swedish linked employer-employee data, that both males and females show an increase in alcohol-related mortality and suicide in the short-run after plant closure. Moreover, Kuhn, Rafael and Zweimüller (2009) use administrative data from Austria to study the short-run effect of plant closures on public expenditures for health care. They find that expenditures on medical treatments are not affected by the plant closures. However they do find evidence that job loss does significantly increase expenditures for antidepressants and related drugs and for hospitalization due to mental health problems for males. Viinamäki et al. (1993) find that men who were involved in a wood-processing factory closure in Finland, had worse mental health than individuals who were employed at a similar wood-processing factory. The effects of the factory closure on mental health are found to be associated with insufficient social support, subjectively poor health, low income and
uncertainty about the future.
2.2.2 Propensity Score Matching
8 treatment group and the synthetic control group is assumed to help remove the selection bias from the results and to identify the causal effect of unemployment on mental health. The PSM method is explained in more detail in the ‘Theoretical and Empirical Model’ section below. Browning et al. (2006) use a random ten percent sample of the adult male population of Denmark to investigate the causal effect of job loss on stress-related health outcomes. They use probit propensity scores and one-to-one matching to match every treated individual to a non-treated individual who had the closest propensity score. With this approach Browning, et al. (2006) do not find significant evidence that job displacement causes hospitalization for stress-related diseases.
Gebel and Voßemer (2014) not only use a PSM approach, but extend it with a Difference-in-Differences model. They use data from the German Socio-Economic Panel from 1995 until 2010 to estimate the effect of employment transitions on health. Similar to what Browning et al. (2006) have done, Gebel and Voßemer (2014) constructed the synthetic control group with the aid of PSM. They used a logit model to estimate the propensity scores and matched the individuals using a 5-Nearest Neighbour (NN) matching, without replacement, procedure. The Difference-in-Differences approach identifies the causal effect of job loss as the difference between the change in health of the treatment group between period 𝑡 and 𝑡 + 1 and the counterfactual change in health between the period 𝑡 and 𝑡 + 1 that the individuals would have experienced if they had not lost their job. This change in health by the counterfactual is estimated by the synthetic control group that was constructed with the aid of PSM. This is allowed as the individuals in this group are the ‘statistical twin’ of the individuals in the treatment group. The empirical results of Gebel and Voßemer (2014) illustrate that
unemployment negatively affects an individual’s mental health and re-employment improves an individual’s mental health with effects of similar size. It is stated, however, that a
comparison of these effects is not suitable to answer the question whether unemployment scars mental health beyond re-employment.
3. THEORETICAL AND EMPIRICAL MODEL
9 Rubin and Rosenbaum (1983). Additionally, the causal effect will attempted to be captured by estimating regressions, given some prespecified covariates.
3.1 Selection Bias
An important issue in any research into the causal effects of a certain type of treatment is selection bias, as this might overestimate or underestimate the causal effect of any type of treatment. In terms of this thesis, the selection bias prevails if the effect of job loss on mental health is overshadowed by confounding factors that might influence both the probability that an individual loses their job and an individual’s mental health. This could mean that
individuals who lose their job already suffered more from lower levels of mental health than employed individuals. As this is likely to hold in reality, it thus means that the difference found in mental health between employed and unemployed individuals might not completely be caused by job loss, but by a confounding factor, such as the socioeconomic status of individuals.
More formally, think about job loss as a binary random variable, 𝐽𝐿𝑖 which can either take up a value of zero in case the individual is employed, or a value of one when an individual loses their job. The mental health of individual 𝑖 is denoted by 𝑀𝐻𝑖. The central question in this thesis is whether 𝑀𝐻𝑖 is affected by job loss and by how much. It is assumed that there are
two potential mental health states, one state if someone loses their job, 𝑀𝐻1𝑖, and one if they remain employed, 𝑀𝐻0𝑖.
The observed outcome 𝑀𝐻𝑖 of individual 𝑖 can be written as follows:
𝑀𝐻𝑖 = 𝑀𝐻0𝑖+ (𝑀𝐻1𝑖− 𝑀𝐻0𝑖)𝐽𝐿𝑖 (1)
In this notation the causal effect of job loss to an individual is 𝑀𝐻1𝑖− 𝑀𝐻0𝑖. When taking the overall effect of the individuals in the sample it is likely that the treatment effect is different across individuals. Moreover, as both outcomes are never observed for one individual, when estimating the causal effect, the average mental health of individuals who have lost their job is compared to that of individuals who are employed.
10 status conditional on job loss is linked to the average causal effect of job loss on mental health and the selection bias:
𝐸[𝑀𝐻𝑖|𝐽𝐿𝑖 = 1] − 𝐸[𝑀𝐻𝑖|𝐽𝐿𝑖 = 0]
= 𝐸[𝑀𝐻1𝑖|𝐽𝐿𝑖 = 1] − 𝐸[𝑀𝐻0𝑖|𝐽𝐿𝑖 = 1]
+ 𝐸[𝑀𝐻0𝑖|𝐽𝐿𝑖 = 1] − 𝐸[𝑀𝐻0𝑖|𝐽𝐿𝑖 = 0].
(2)
First, the average causal effect of job loss on mental health is captured by
𝐸[𝑀𝐻1𝑖|𝐽𝐿𝑖 = 1] − 𝐸[𝑀𝐻0𝑖|𝐽𝐿𝑖 = 1] . (3) This is equal to the average treatment effect on the treated (ATT), where 𝐸[𝑀𝐻1𝑖|𝐽𝐿𝑖 = 1] is
the value of the mental health of an individual who lost their job and 𝐸[𝑀𝐻0𝑖|𝐽𝐿𝑖 = 1] is the
value the mental health would have been if the individual who lost their job had never lost their job. Second, the selection bias in the equation is captured by
𝐸[𝑀𝐻0𝑖|𝐽𝐿𝑖 = 1] − 𝐸[𝑀𝐻0𝑖|𝐽𝐿𝑖 = 0] . (4) Here 𝐸[𝑀𝐻0𝑖|𝐽𝐿𝑖 = 1] still has the same meaning as it had before and 𝐸[𝑀𝐻0𝑖|𝐽𝐿𝑖 = 0] is the mental health of an individual who is employed. Angrist and Pischke (2009) argue that randomized trials control for this selection bias. However, this thesis is not in the form of a randomized trial, individuals are not randomly selected to lose their job. In fact, it is unknown why these individuals lose their jobs and they might lose their jobs for several different reasons.
To control for the differences between individuals in the sample, vector 𝑥𝑖 is introduced as a vector of observed covariates of individual 𝑖 before job loss occurs. The vector of covariates is assumed to control for the selection bias as described by equation (4). This means that, if the set of covariates 𝑥𝑖 helps control for the differences between individuals sufficiently,
equation (2) is reduced to the following equation: 𝐸[𝑀𝐻𝑖|𝐽𝐿𝑖 = 1, 𝑥𝑖] − 𝐸[𝑀𝐻𝑖|𝐽𝐿𝑖 = 0, 𝑥𝑖]
= 𝐸[𝑀𝐻1𝑖|𝐽𝐿𝑖 = 1, 𝑥𝑖] − 𝐸[𝑀𝐻0𝑖|𝐽𝐿𝑖 = 1, 𝑥𝑖] + 0
(5)
11 Given the mental health states of individuals and the vector of covariates 𝑥𝑖, all that is left to estimate is 𝐸[𝑀𝐻0𝑖|𝐽𝐿𝑖 = 1, 𝑥𝑖]. In words, we must try and estimate the mental state an individual would have had if they had not lost their job. This outcome is always missing, hence, estimating the causal effect of a treatment is a missing data problem and leads to speculation about the effect of treatment. The assumption that there exists a unique value of 𝑀𝐻𝑟𝑖 for individual 𝑖 undergoing treatment 𝑟 is called the Stable Unit-Treatment Value Assumption (Rubin, 1980).
This thesis tries to solve for the missing data problem by using Propensity Score Matching to create a synthetic control group that is as similar as possible to the treatment group.
Additionally, this thesis will attempt to solve for the missing data problem with the aid of an Ordinary Least Squares (OLS) regression model that keeps the set of covariates 𝑥𝑖 as fixed.
3.2 Propensity Score Matching - Theory
The PSM approach is used because it holds some advantages over parametric approaches such as OLS. Firstly, as PSM is a non-parametric method, it does not rely on assumptions regarding the functional form of the model. Hence, it is not as sensitive to any issues of misspecification. Moreover, PSM allows for the construction of a counterfactual to be able to estimate the causal effects of treatment. PSM thus only takes individuals in to account that are similar to those in the treatment group.
Using the vector of covariates 𝑥𝑖 that has been introduced above, balancing scores can be estimated to group individuals from the treatment and control group in such a way that direct comparisons are more relevant. The balancing score, 𝑏(𝑥), is a function of the observed covariates 𝑥 such that the conditional distribution of 𝑥 given 𝑏(𝑥) is the same for both the individuals in the treatment group and those in the control group. That is, using the notation of Dawid (1979):
𝑥 ⫫ 𝐽𝐿|𝑏(𝑥). (6)
Let the conditional probability of assignment to job loss, given the covariates, be denoted by
𝑒(𝑥) = 𝑃(𝐽𝐿 = 1|𝑥), (7)
12 𝑃(𝐽𝐿1, … , 𝐽𝐿𝑛|𝑥1, … , 𝑥𝑛) = ∏ 𝑒(𝑥𝑖)𝐽𝐿𝑖{1 − 𝑒(𝑥𝑖)}1−𝐽𝐿𝑖.
𝑁
𝑖=1
(8) The function 𝑒(𝑥) is called the propensity score, which is defined as the propensity towards exposure to job loss, given the observed covariates 𝑥. The propensity score is the coarsest balancing score possible. The propensity score function 𝑒(𝑥) is unknown, however, 𝑒(𝑥) can be estimated from the observed data with the aid of a probit or logit model. Estimates of 𝑒(𝑥) are the probabilities that individual 𝑖 with covariate vector 𝑥 will experience treatment 𝐽𝐿𝑖 = 1, regardless of them actually experiencing job loss. Using these estimates of the propensity scores, it is possible to match individuals who experience job loss with individuals who have propensity scores that are as similar as possible and who remain employed. This matched group of individuals can then be considered as the ‘synthetic’ control group. Estimates of direct comparisons are more relevant when using these individuals as a control group, opposed to using all individuals in the dataset.
The quantity to be estimated after this matching procedure is the average treatment effect (ATE), defined as
𝐴𝑇𝐸 = 𝐸(𝑀𝐻1|𝑥) − 𝐸(𝑀𝐻0|𝑥), (9)
where 𝐸(𝑀𝐻1) is the expected mental health of those who suffered job loss, and 𝐸(𝑀𝐻0) is the expected mental health of the employed from the synthetic control group. Hence, the average treatment effect is the expected difference between the mental health of the
unemployed to that of the employed. Given the vector of covariates, this the causal effect of job loss on mental health since the individuals in the synthetic control group are assumed to have the mental health state individuals in the treatment group would have had if they had not lost their job. The ATE in this model is thus expected to be the same as the ATT.
3.3 Propensity Score Matching - Procedure
13 programme by Leuven and Sianesi (2003). The procedure to estimate the causal effect of job loss on mental health using PSM consists of the five following five steps:
Firstly, by the aid of a logit model, the propensity scores are estimated. These propensity scores represent the probability than an individual will lose their job, irrespective of them actually losing their job. The covariates that are used in the propensity score estimating are described in the ‘Data’ section, moreover table 3 provides some summary statistics on these covariates.
Secondly, these estimated propensity scores are used to create a counterfactual based on the similarity between the individuals. In the matching procedure an unemployed individual with propensity score 𝑒(𝑥)1𝑖 ∈ [0, 1] is matched to a similar employed individual with propensity score 𝑒(𝑥)𝑖0 ∈ [0, 1]. In other words, using the estimated propensity scores, individuals who are unemployed are matched with another individual who is not unemployed. If executed accurately, apart from being employed, this control group is as similar to the treatment group as possible. This matching can be executed by different methods, the procedure that is used in this thesis is Nearest Neighbour matching (NN), without replacement. This is the most
straightforward matching method, an individual in the treatment group is matched with an individual in the control group that is closest in terms of propensity score. The choice has been made to match without replacement, which means that an individual can only be
matched with another individual who has not been used in another match. However, a risk of the nearest neighbour matching method is that there could be some bad matches in case the nearest neighbour is far away in terms of propensity scores. Other methods of matching are described and tested below in the ‘PSM Quality Check Results’ section.
14 Fourthly, after finding common support, the match quality needs to be assessed. It is
important that it is checked that the matching procedure balances the distribution of the relevant variables in both the treatment and control group. This will be done by comparing the situation before and after matching and checking whether there is any significant bias left after matching. A formal way of testing the distance in marginal distributions of the covariates is by using the standardised bias (SB) as introduced by Rosenbaum and Rubin (1985). For each covariate the SB before and after matching is given by the following equations:
𝑆𝐵𝑏𝑒𝑓𝑜𝑟𝑒 = 100 ∗ 𝑋1 ̅̅̅ − 𝑋̅̅̅0 √0.5 ∗ (𝑉1(𝑋) + 𝑉0(𝑋)) (10) 𝑆𝐵𝑎𝑓𝑡𝑒𝑟= 100 ∗ 𝑋1𝑀 ̅̅̅̅̅ − 𝑋̅̅̅̅̅0𝑀 √0.5 ∗ (𝑉1𝑀(𝑋) + 𝑉0𝑀(𝑋)) (11) In these equations 𝑋1, 𝑋0, 𝑉1 and 𝑉0 are the mean and the variance, respectively, of the covariate in the treatment and control group before matching. Similarly, 𝑋1𝑀, 𝑋0𝑀, 𝑉1𝑀 and 𝑉0𝑀 are the mean and variance of the covariate for the treatment and control group for the
matched samples. The SB approach is a common approach, however in terms of standardized bias a formal cut-off point of what the indication of a successful match should be has never been set. In most of the literature it is commonly agreed upon that a bias reduction to a level below 3 or 5 percent is as sufficiently low.
Another way of testing the matching quality is by using a two-sample t-test to check whether there is significant difference between covariate means for the control and treatment group. It is expected that there are differences before matching, however, after matching there should be a balance between both groups and thus there should not be any significant differences between the two groups. In other words, insignificant results of a two-sample t-test indicates that there are no differences between the control group and treatment group after matching has occurred. A shortcoming of testing this with a two-sample t-test is that the bias reduction before and after matching is not clearly visible.
Finally, the causal impact of job loss on mental health is estimated as the average treatment effect on the unemployed, compared to their similar employed counterparts. As the
15
3.4 Ordinary Least Squares Regression Model
Another method to find the causal effect of job loss on mental health is with the aid of a regression model. In this model we will again use a vector of fixed covariates 𝑥𝑖 which will control for the confounding factor and the differences between individuals. This leads to the Conditional Independence Assumption (CIA), which is assumed to hold so the estimation will have a causal interpretation. In a formal way,
𝐸[𝑀𝐻1𝑖− 𝑀𝐻0𝑖|𝑥𝑖] = 𝐸[𝑀𝐻𝑖|𝑥𝑖, 𝐽𝐿𝑖 = 1] − 𝐸[𝑀𝐻𝑖|𝑥𝑖, 𝐽𝐿𝑖 = 0] . (12)
The CIA assures that the difference between the mental health of individuals who lost their job and that of those who did not, given the vector of covariates 𝑥𝑖, is equal to the causal effect of job loss on mental health.
Following the theory as described above which is in line with Angrist and Pischke (2009), we end up with the following regression model to estimate the causal effect of job loss on mental health:
𝑀𝐻𝑖 = 𝛼 + 𝛿 ∙ 𝐽𝐿𝑖 + 𝑥𝑖′ ∙ 𝛾 + 𝑢𝑖, (13)
where 𝑀𝐻 is the mental health status of the individuals in the control group and treatment group, 𝛿 is the causal effect of job loss on mental health, 𝐽𝐿 is the dummy variable denoting job loss, 𝑥 is the vector of covariates keeping the differences between individuals fixed and 𝑢 is the error term which is assumed to be uncorrelated with 𝐽𝐿, 𝛿 and covariates 𝑥. The
covariates used in this model are the same as those used in the propensity score matching model and can be found in table 3 below in the ‘Data’ section.
16
4. DATA
4.1 Data Collection
This thesis uses data from the LISS panel (Longitudinal Internet Studies for Social sciences), which is provided by CentERdata (Tilburg University, the Netherlands). The LISS panel consists of 4500 representative households in the Netherlands. The panel members complete online questionnaires every month, some of which are part of the LISS Core Study. The LISS Core Study is a longitudinal study which is repeated annually and follows changes in life course and living conditions of the panel members. The LISS panel has been collecting data since October 2007. This thesis uses merged data from wave 10 and wave 11 of the LISS Core Study panels, which correspond to the years 2017 and 2018. Of the LISS Core Study panels the surveys on ‘Income’, ‘Work and Schooling’ and ‘Health’ are used. Additional information is collected from the ‘Background Variables’, which contains periodically updated general individual information. The data from the questionnaires ‘Work and
Schooling’, ‘Income’ and ‘Health’, are collected in April, June and November, respectively. As information about employment status is available in both the ‘Work and Schooling’ survey and the ‘Background Variables’ data-file, this thesis uses the ‘Background Variables’ that are collected in April 2017 and April 2018, which are the same months as when the ‘Work and Schooling’ data is collected. This is done to make sure both files can complement each other and there are no differences in employment status between the two datasets. Important is that the data on mental health, which is from the ‘Health’ questionnaire, is collected after the data on employment status, this way the effect of unemployment can truly be measured. In this thesis, however, there is no information on the employment status of individuals in
November. Therefore, in this thesis it is assumed that the individuals who were unemployed in April are still unemployed in November, which is a data flaw as this assumption might very well not hold in reality.
4.2 Data Description
17 that were collected from the LISS Core Study, but from these individuals not everyone
participated in every survey. Moreover, not all these individuals answered every survey completely, which lead to missing data. Where possible, the missing information on for example age, gender, education and unemployment was reconstructed by using
complementary data from other questionnaires in the study. Unfortunately however, for some variables, such as ‘difficulty making ends meet’ and ‘living in an urban area’, this was not possible as information on these variables was not available in other questionnaires. Due to non-response, missing information in relevant questions and individuals not being relevant for this study, the data set was greatly reduced. For this thesis, relevant individuals are considered to be people who are between the age of 18 and 70, they are employed in 2017 and either employed or unemployed after job loss in 2018 and they answered all the relevant questions for this thesis completely.
As has been described in the previous section, the propensity score matching is done by using characteristics of all individuals that are observed before treatment occurs. This means that all individuals in the sample had to be employed in 2017 to be considered. This results in a dataset, consisting of in total 1,010 individuals of which 15 individuals are unemployed in 2018. This corresponds to an unemployment rate of 1.49 percent, which is lower than the population unemployment rate in the Netherlands in this period, which was 4.1 percent2 in April 2018.
4.3 Mental Health
This thesis uses five different measures of mental well-being to estimate the causal effects of unemployment on mental health. These are subjective self-assessed measures of an individual’s mental health status. Participants were asked to rate on a scale from one to six for five different mental states how they felt during the past month The six possible answers range from feeling this way ‘never’ to ‘continuously’. The five mental states are described as ‘feeling depressed and gloomy’, ‘feeling very anxious’, ‘feeling so down that nothing could cheer me up’, ‘feeling calm and peaceful’, and ‘feeling happy’. A description of the way these questions were proposed can be found in Appendix B. Below, Table 1 shows summary
statistics of the dependent variable for the used dataset. The table includes the results of a
18 test, testing whether the mental health scores were significantly higher or lower in the
treatment group compared to all other employed individuals in the data.
Table 1: Mental health status before matching, (N=1,010).
Variable Min. Max. Mean Mean (JL=0) Mean (JL=1)
Depressed and gloomy 1 6 1.890 1.886 2.133
Very anxious 1 6 1.950 1.951 1.933
Down, unable to cheer up 1 6 1.529 1.525 1.8
Calm and peaceful 1 6 4.312 4.314 4.2
Happy 1 6 4.306 4.313 3.867**
Note: */**/*** indicates that there are significant differences at the 10%/5%/1% level for the mean differences between JL=0 and JL=1, as assessed by a one-sided t-test.
A value of ‘1’ for these variables means that the respondent never felt the way as described by the variable, a value of ‘6’ meant that these feelings were continuously experienced. The table shows that the unemployed, on average, have a higher score for the ‘depressed and gloomy’ and ‘down, unable to cheer up’ variables, which might indicate that they have a lower mental health than others, however these differences are not found to be significantly different. Moreover, on average, the unemployed rate lower on the ‘very anxious’ variable. The more positive mental health states ‘calm and peaceful’ and ‘happy’ both have a lower average for the unemployed individuals compared to the employed. However, the only significantly lower rated state, according to a two-sample t-test, is ‘happy’.
19
Table 2: Correlation coefficients between mental health measures, (N=1,010)
Depressed and gloomy Very anxious Down, unable to cheer up Calm and peaceful Very anxious 0.543 - - -
Down, unable to cheer up 0.696 0.573 - -
Calm and peaceful -0.424 -0.463 -0.431 -
Happy -0.544 -0.366 -0.515 0.535
Note: All coefficients are significant at a 1% level.
4.4 Job Loss
The employment status of the individuals is denoted by the variable job loss, which is the treatment variable in the models. The dataset only includes individuals who report to be employed in April 2017. In April 2018 some of these individuals report to be seeking a job after experiencing job loss, these are the people that are considered to have experienced the treatment ‘job loss’. All other individuals report to be employed in 2018. In 2018 15
individuals state to have lost their job, 995 individuals state to be employed. In Appendix B a description is given on the possible answers as individuals are asked to describe their
employment situation.
4.5 Covariates
A set of covariates is used to estimate the propensity score matching model. Some of the variables that are used are similar to the covariates used by Gebel and Voßemer (2014) and Marcus (2013) in their propensity score matching procedures. Moreover, this thesis adds extra information on financial well-being and living characteristics.
20 into six categories, where the highest level of education with diploma is either ‘1’: primary school, ‘2’: intermediate secondary education (US: junior high school), ‘3’: higher secondary education and preparatory university education (US: senior high school), ‘4’: intermediate vocational education (US: junior college), ‘5’: higher vocational education (US: college), or finally ‘6’: university. The variable single has a value of ‘0’ when an individuals is married or co-habiting, it has a value of ‘1’ if an individual is single, divorced or a widower. The
migration background variable non-western has a value of ‘0’ when the individual has a Dutch background, or is a first or second generation foreigner with a western background. It has a value of ‘1’ when the individual is a first or second generation foreigner with a non-western background.
The household characteristics in the models are given by the amount of children in the household, whether they have difficulty making ends meet, if they are living in an urban area and if they inhabit a self-owned dwelling. Children in household has a value of ‘0’ if there are no children living in the household, and a value larger than zero indicating the number of children living in the household, where the maximum amount of children in the household in this sample is five. The variable difficulty making ends meet states whether an individual experienced difficulties making ends meet at the time of the questionnaire. A value of ‘0’ indicates that the individual has no difficulty making ends meet, a value of ‘1’ indicates the opposite. The variable living in an urban area demonstrates the urban character of the place of residence of the individual’s household. A value of ‘1’ illustrates an extremely urban character where surrounding address density per square kilometre is 2,500 or more. A
household lives in a very urban character area when the surrounding address density is 1,500 – 2,500 per square kilometre, which corresponds to a value of ‘2’ in the data. A moderately urban character (density between 1,000 – 1,500 per square kilometre) has a value of ‘3’ and a slightly urban character (density between 500 – 1,000 per square kilometre) has a value of ‘4’. Lastly, a value of ‘5’ indicates a non-urban character where the density is less than 500 per square kilometre. The covariate home-owner has a value of ‘1’ when the dwelling the
household inhabits is self-owned and has a ‘0’ if it is another type of dwelling, such as a rental dwelling, sub-rented dwelling, or a cost-free dwelling.
21
Table 3: Covariates (N=1,010)
Covariate Min. Max. Mean Mean
(JL=0) Mean (JL=1) Age 21 69 46.608 46.569 49.2 Male 0 1 0.525 0.522 0.733* Education level 1 6 4.146 4.158 3.333*** Single 0 1 0.323 0.318 0.667*** Non-western 0 1 0.067 0.067 0.067 Children in household 0 5 0.829 0.835 0.4*
Difficulty making ends meet 0 1 0.086 0.085 0.133
Living in an urban area 1 5 2.842 2.849 2.333*
Home-owner 0 1 0.767 0.771 0.533**
Note: */**/*** indicates that there are significant differences at the 10%/5%/1% level for the mean differences between JL=0 and JL=1, as assessed by a one-sided t-test.
5. RESULTS
5.1 Propensity Score Matching
Below, Table 4 reports the average treatment effect for all five measures of subjective mental health. Job loss increases the likelihood of individuals feeling depressed and gloomy with 0.133 points. In contrast, the unemployed rate both feelings of feeling anxious and feeling so down, nothing could cheer them up, 0.133 points lower than their employed counterparts. This indicates that employed individuals suffer more from these feelings than those who have lost their jobs, which is not in line with the hypothesis that the unemployed have lower mental well-being. However, all these results are insignificant, so it cannot be concluded that there is a difference in mental health states that is caused by job loss. The two positive mental health states, feeling calm and peaceful and feeling happy show ambiguous results as well. Unemployed individuals are more likely to rate their calm and peaceful feeling 0.2 points higher than the employed people do. Moreover, individuals rate feeling happy 0.2 points lower after job loss than their employed counterparts, which
22 well, so it cannot be concluded that individuals rate their positive feelings higher or lower after job loss than employed individuals do.
Table 4: Average Treatment Effects of Job Loss – Nearest Neighbour Matching
(1) Depressed (2) Anxious (3) Down (4) Peaceful (5) Happy 0.133 (0.363) -0.133 (0.377) -0.133 (0.410) 0.2 (0.393) -0.2 (0.406) Note: */**/*** indicate significance at the 10%/5%/1% level. Brackets indicate standard errors.
5.2 Regression Model
A White test to check for heteroskedasticity was performed and found proof of heteroskedasticity in the data. To solve for this problem White standard errors are applied in all five models.
23 The estimated value reports that the unemployed experience feeling very anxious 0.035 points lower than employed individuals, however as is the case with the other estimates of the causal effects, this result is not statistically significant. Similar to findings in model (1), ageing individuals have 0.015 points lower reported levels of anxiousness, which is significant at the one percent level. Additionally, males report 0.113 points lower levels and having children decreases the level of anxiety by 0.057 for every additional child, which is significant at the five percent level for both covariates. Moreover, individuals who experience difficulties making ends meet report 0.271 points higher levels of feeling anxious at a five percent significance level than those who do not experience any financial difficulties.
The output estimate of the causal effect for the self-reported state of feeling so down nothing
could cheer me up is 0.209, implicating a decrease in mental health in case of job loss. This
result however is insignificant and thus does not provide proof that job loss affects mental health in any way. Similar to the findings from model (1), ageing lowers the self-reported level of feeling down significantly by 0.012 points at the one percent level. This implies that an additional year of ageing leads to less suffering from feeling down than it does for younger individuals. A higher level of education significantly lowers the reported level of feeling
down by 0.049 points at the five percent level. Individuals of non-western descent, however,
report significantly higher levels of feeling down by 0.255 points at the five percent level. Moreover, individuals experiencing difficulties making ends meet report feeling more down by 0.380 points at a one percent significance level.
Table 6 shows the estimation results of model (4) and (5), that estimate the causal effect of job loss on feeling calm and peaceful and feeling happy. The treatment effect of job loss results in insignificant results, where the estimates report that unemployed individuals have 0.080 and 0.282 points lower levels of feeling calm and peaceful and happy, respectively. Model (4) does, however, show that as individuals age and as they have completed a higher level of education, they significantly report 0.016 and 0.069 points higher levels of feeling
calm and peaceful, respectively, at the one percent level. Moreover, males report 0.125 points
24 What these five regression models have in common is that individuals experiencing
25
Table 5: OLS Regression Output – Negative Mental Health States
(1) Depressed (2) Anxious (3) Down Treatment effect Job loss 0.184 (0.269) -0.035 (0.213) 0.209 (0.267) Covariates Age -0.012*** (0.003) -0.015*** (0.003) -0.012*** (0.003) Male -0.042 (0.057) -0.113** (0.059) -0.073 (0.053) Education -0.060** (0.024) -0.029 (0.026) -0.049** (0.023)
Children living at home -0.029
(0.028) -0.057** (0.026) -0.010 (0.025) Single -0.002 (0.066) 0.018 (0.068) 0.075 (0.060) Non-western 0.248* (0.129) 0.092 (0.134) 0.255** (0.123) Difficulty making ends meet 0.421***
(0.133)
0.271** (0.121)
0.380*** (0.122)
Living in urban area -0.023
(0.023) -0.020 (0.024) -0.024 (0.020) Home-owner -0.040 (0.075) -0.015 (0.077) -0.048 (0.071) Constant 2.785*** (0.199) 2.926*** (0.215) 2.355*** (0.195) R2 0.055 0.052 0.065 Number of observations 1,010 1,010 1,010
26
Table 6: OLS Regression Output – Positive Mental Health States
(4)
Calm and peaceful
(5) Happy Treatment effect Job loss -0.080 (0.246) -0.282 (0.267) Covariates Age 0.016*** (0.003) 0.005 (0.003) Male 0.125* (0.066) -0.075 (0.062) Education 0.069*** (0.027) 0.042* (0.025)
Children living at home 0.026
(0.030) 0.001 (0.030) Single -0.025 (0.076) -0.222*** (0.075) Non-western -0.137 (0.142) -0.183 (0.145) Difficulty making ends meet -0.475***
(0.143)
-0.704*** (0.135)
Living in an urban area -0.038
(0.028) -0.007 (0.026) Home-owner -0.096 (0.085) 0.078 (0.084) Constant 3.282*** (0.230) 4.057*** (0.219) R2 0.059 0.075 Number of observations 1,010 1,010
27
6. PSM Quality Check Results
The Theoretical and Empirical Model section above describes the necessary steps to build a PSM model that provides unbiased results. In this section these steps will be reviewed. Moreover, as there exist several methods to match individuals with the aid of propensity scores, some of these methods are reviewed to make sure the best method for this research was used to estimate the causal effects of job loss on subjective mental health states.
6.1 Step 1: Propensity Score Estimation
As mentioned above, the propensity scores are estimated using a logit model. The results of this estimating can be found below in Table 7. It shows that the only covariates that significantly increase the probability of losing one’s job in the used dataset are being male and being single, moreover a higher education level lowers the probability of becoming
unemployed. These variables are all significant at the ten percent level.
6.2 Step 2: Matching Procedure
After having estimated the propensity scores, as described in the first step, the propensity scores were used to match individuals in the treatment group with individuals in the control group to construct a counterfactual. The matching procedure of choice in this thesis was NN matching, however other methods such as Caliper Matching, Radius Matching, Kernel Matching and Local Linear Matching (LLM) could have been used as well. Caliper and Radius Matching are similar procedures, matches are made based on the maximum allowed distance in propensity scores between matches. The difference between the two is that with Caliper Matching an individual in the treatment group gets one single match with an individual from the control group that is closest in terms of propensity scores. Radius
28 individual. This linear term is advantageous whenever the observations of the control group are distributed asymmetrically around the treated individuals.
Table 7: Logit Propensity Score Estimation
Covariates Job Loss
Age 0.006 (0.025) Male 1.044* (0.605) Education Level -0.384* (0.197) Children in Household -0.241 (0.338) Single 1.136* (0.589) Non-Western -0.978 (1.112)
Difficulty Making Ends Meet -0.006
(0.813)
Living in an Urban Area -0.299
(0.242) Home-Owner -0.566 (0.591) Constant -2.914 (1.850) N 1,010 Pseudo R2 0.125
29
6.3 Step 3: Test for Common Support and Overlap
The next step after matching is to test for common support and overlap between the two matched groups. Below, Figures 1-6 show the density functions of the propensity scores of the control and treatment group before and after all five matching procedures. From these figures it can be concluded that NN Matching and Caliper Matching show common support and that the other matching methods fail to provide it.
Figure 1: Density Function Before Matching Figure 2: Density Function After Nearest Neighbour Matching
Figure 3: Density Function After Caliper Matching Figure 4: Density Function After Radius Matching
30
6.4 Step 4: Testing the Match Quality
Testing for common support is not enough to conclude that the matching procedure resulted in a dataset with good matches. Therefore the match quality is assessed with the aid of a Standardised Bias (SB) test and a two-sample t-test.
6.4.1 Standardised Bias
To test the matching quality of the matching procedures, the standardised bias (SB) was estimated for all covariates in the models. As was mentioned above in section three, a match is considered a good match if the standardised bias is below five percent. Table 7 shows that before matching, the standardised bias for almost all covariates was (much) higher than five percent. The only covariate with a low enough SB is ‘non-western’, with an absolute value of 0.3. After NN matching, however, for all covariates except ‘non-western’ the SB has become much smaller. Despite the improvement in SB, the SB for the variables ‘age’, ‘male’, ‘non-western’, ‘living in an urban area’ and ‘home-owner’ are still above the five percent cut-off point. The only covariates with a low enough SB are ‘education’, ‘children living at home’, ‘single’ and ‘difficulty making ends meet’, they all have a SB of 0.0. The other four matching procedures show even less covariates with a SB below five percent. It thus seems as though the matching procedure did not result in a sample with a good quality control group.
Table 7: Standardised Bias
Covariate Before NN Caliper Radius Kernel LLR
Age 20.8 12.1 30.4 13.6 7.8 13.2
Male 44.2 13.9 16.1 18.1 21.1 27.8
Education -56.6 0.0* 5.3 5.9 -23.4 9.2
Children living at home -47.7 0.0* 0.0* -15.8 -24.6 0.0*
Single 73.2 0.0* 0.0* 13.9 36.7 0.0*
Non-western -0.3* 26.2 30.3 3.7* -2.0* 26.2
Difficulty making ends meet 15.1 0.0* 0.0* -7.1 7.7 0.0* Living in urban area -42.8 -27.6 -38.3 -12.5 -20.7 -27.6
Home owner -50.4 42.5 49.0 -1.8* -21.1 42.5
Mean Bias 39.0 13.6 18.8 10.3 18.3 16.3
N 1,016 30 30 950 1,010 29
31
6.4.2 Two-Sample t-test
Another way of testing the match quality is by the two-sample t-test of which the results are shown in Table 8. In this case, insignificant results imply that a match is a good match. The tests show that for nearest neighbour matching, caliper matching and local linear regression matching there are no significant differences between the control groups and the treatment groups. As table 3 in the Data section showed, before matching there were significant differences between all covariates except ‘age’, ‘non-western’ and ‘difficulty making ends meet’. This means that the matching procedure has increased the quality of the control groups as opposed to using the entire sample before matching as a control group.
Table 8: Two-Sample t-test
Covariates NN Caliper Radius Kernel LLR
Age C 47.667 45.846 46.448 46.569 47.071 T 49.200 49.692 49.692 49.200 49.200 Education C 3.333 3.538 4.185* 4.158*** 3.357 T 3.333 3.615 3.615* 3.333*** 3.333 Male C 0.667 0.615 0.517 0.521* 0.643 T 0.733 0.692 0.692 0.733* 0.733 Children at home C 0.400 0.462 0.844* 0.835* 0.429 T 0.400 0.462 0.462* 0.400* 0.400 Single C 0.667 0.615 0.310*** 0.318*** 0.643 T 0.667 0.615 0.615*** 0.667*** 0.667 Non-Western C 0.000 0.000 0.068 0.067 0.000 T 0.067 0.077 0.077 0.067 0.067
Difficulty Making ends meet C 0.133 0.077 0.084 0.085 0.143
T 0.133 0.077 0.077 0.133 0.133
Living in an urban area C 2.667 2.769 2.862* 2.849* 2.786 T 2.333 2.308 2.308* 2.333* 2.333
Home-owner C 0.333 0.385 0.779* 0.771** 0.357
T 0.533 0.615 0.615* 0.533** 0.533
N C 15 13 983 995 14
T 15 13 13 15 15
32 It has become clear that the NN-matching procedure increases common support. Moreover, the two-sample t-test provides evidence that there is no significant difference between the covariates in the control and treatment group. However, even though matching has decreased the SB for all covariates except non-western, the SB is not below the required five percent level for all covariates. It can thus not be concluded that PSM has increased the quality of the data enough to be able to draw unbiased conclusions from the estimations. The causal output of all matching procedures can be found in the Appendix in Table B.
7. CONCLUSION AND DISCUSSION
This thesis studied the causal effects of job loss on five different subjective mental health states. Confounding factors such as an individual’s socioeconomic status might affect the causal results. With the aid of the non-parametric Propensity Score Matching approach and a parametric Ordinary Least Squares approach this thesis attempts to isolate the causal effect of job loss while controlling for confounding factors and individual heterogeneity. The PSM model is based on the theory as introduced by Rubin and Rosenbaum (1983) and is estimated with the aid of guidelines provided by Caliendo and Kopeinig (2008). The parametric
approach is based on theory as provided by Angrist and Pischke (2009). With the aid of a vector of covariates that controls for individual heterogeneity, the causal effect of job loss was estimated on five different measures of mental health. These measures are described as
feeling depressed and gloomy, anxious, so down nothing could cheer me up, calm and
peaceful and happy.
The results of both approaches have failed to show any significant evidence that job loss causes any effects on the mental health of individuals. The absence of any statistically significant causal effect on job loss is considered to be evidence of the presence of selection bias. This selection bias may result from health selection and confounding factors. The health selection exists when individuals who suffered from lower mental health went into
unemployment and therefore lower the level of mental health among the unemployed.
Confounding factors such as an individual’s socioeconomic status or low job satisfaction may cause job loss and lower levels of mental health. The results of the parametric approach and the logit model that estimated the propensity scores cautiously prove the existence of the confounding factor socioeconomic status. The logit propensity score estimation model
33 percent level show that males and singles have a higher chance of becoming unemployed and education decreases the chance of becoming unemployed. The parametric approach shows that education positively and significantly affects an individual’s mental health for all models except model (2). Moreover, individuals are significantly worse off in terms of mental health status if they experience difficulties in making ends meet. Lastly, individuals of non-western descent significantly show lower levels of mental health in model (1) and (3). Whereas being male is not a factor affecting an individual’s socioeconomic status, education levels, marital status, experiencing difficulties in making ends meet and being of non-western descent are factors that decide an individual’s socioeconomic status. Hence, this thesis finds proof that the confounding factor socioeconomic status affects the probability of becoming unemployed and the mental health status of individuals.
Recommendations on further research in this topic are to use data with more (unemployed) individuals and more information on employment characteristics, such as duration of (un)employment and general job descriptions so the outcome of the estimations are more representative of the population outcome. Moreover, to increase the power of the statistical method used in this thesis, one could attempt to estimate the causal effects by using a similar modal as the Difference-in-Differences Propensity Score Matching method as used by Gebel and Voßemer (2014). Lastly, as age is a significant variable affecting mental health, it may be interesting to estimate the models again by age categories.
8. REFERENCES
Angrist, J. D., & Pischke, J. S. (2009). Mostly harmless econometrics: An empiricist's companion. Princeton university press.
Baik, K., Hosseini, M., & Priesmeyer, H. R. (1989). Correlates of psychological distress in involuntary job loss. Psychological Reports, 65, 1227-1233.
34 Böckerman, P., & Ilmakunnas, P. (2009). Unemployment and self‐assessed health: evidence from panel data. Health economics, 18(2), 161-179.
Brand, J. E. (2015). The far-reaching impact of job loss and unemployment. Annual review of
sociology, 41, 359-375.
Breslin, F.C., Mustard, C. (2003). Factors influencing the impact of unemployment on mental health among young and older adults in a longitudinal, population-based survey.
Scandinavian Journal of Work, Environment & Health, 19(1), 5-14.
Browning, M., Moller Dano, A., & Heinesen, E. (2006). Job displacement and stress‐related health outcomes. Health economics, 15(10), 1061-1075.
Caliendo, M., & Kopeinig, S. (2008). Some practical guidance for the implementation of propensity score matching. Journal of economic surveys, 22(1), 31-72.
Dawid, A. P. (1979). Conditional independence in statistical theory. Journal of the Royal
Statistical Society: Series B (Methodological), 41(1), 1-15.
Eliason, M., & Storrie, D. (2009). Does job loss shorten life? Journal of Human
Resources, 44(2), 277-302.
Fryer, D. (1986) Employment Deprivation and Personal Agency during Unemployment – A Critical discussion of Jahoda’s Explanation of the Psychological Effects of Unemployment.
Social Behaviour, 1, 3-23.
Fryer, D. and McKenna, S. (1987) The Laying Off of Hands – Unemployment and the
Experience of Time. In S. Fineman (ed.), Unemployment – personal and social consequences, London: Tavistock Publications.
Gebel, M., & Voßemer, J. (2014). The impact of employment transitions on health in
Germany. A difference-in-differences propensity score matching approach. Social science &
medicine, 108, 128-136.
Hudson, C.G. (2005). Socioeconomic status and mental illness: tests of the social causation and selection hypotheses. American Journal of Orthopsychiatry, 75(1), 3-18.
Kuhn, A., Lalive, R., & Zweimüller, J. (2009). The public health costs of job loss. Journal of
35 Lechner, M. (2001). A note on the common support problem in applied evaluation
studies. Univ. of St. Gallen Economics, Disc. Paper, 1.
Leuven, E., Sianesi, B. (2003). “PSMATCH2: Stata Module to Perform Full Mahalanobis and Propensity Score Matching, Common Support Graphing, and Covariate Imbalance Testing”. Software: http://ideas.repec.org/c/boc/bocode/s432001.html.
Marcus, J. (2013). The effect of unemployment on the mental health of spouses–Evidence from plant closures in Germany. Journal of health economics, 32(3), 546-558.
Paul, K.I., Moser, K. (2009). Unemployment impairs mental health: Meta-analyses. Journal
of Vocational Behaviour, 74, 264-282.
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.
Rosenbaum, P., Rubin, D. (1985). Constructing a control group using multivariate matched sampling methods that incorporate the propensity score. The American Statistician, 39, 33-38. Rubin, D.B. (1980). Discussion of the paper by D. Basu. Journal of the American Statistical
Association 75, 591-593.
Strandh, M. (2000). Different exit routes from unemployment and their impact on mental well-being: the role of the economic situation and the predictability of the life course. Work,
Employment and Society, 14(3), 459-479.
Theodossiou, I. (1998). The effects of low-pay and unemployment on psychological well-being: a logistic regression approach. Journal of Health Economics, 17, 85-104.
Viinamäki, H., Koskela, K., Niskanen, L., Arnkill, R., & Tikkanen, J. (1993). Unemployment and mental wellbeing: a factory closure study in Finland. Acta Psychiatrica
36
9. APPENDIX
Appendix A: Tables
Table A: Frequencies of the Measures of Mental Health States in Percentages (N=1,010)
Mental Health Never Seldom Sometimes Often Mostly Continuously Depressed and gloomy 40.16 37.21 17.52 4.13 0.59 0.39
Very anxious 37.70 36.12 21.36 3.64 0.79 0.39
Down, unable to cheer up 64.57 22.15 10.43 1.967 0.49 0.39
Calm and peaceful 2.56 2.76 14.67 27.56 45.87 6.59
Happy 1.18 2.76 17.32 28.45 43.70 6.59
Table B: Average Treatment Effects by Matching Procedure
37
Appendix B: Survey questions
Dependent variables: from “Health – LISS Core Study”:
For every question, please choose the answer that best describes how you felt during this past month. This past month …
ch00x011: I felt very anxious
ch00x012: I felt so down that nothing could cheer me up ch00x013: I felt calm and peaceful
ch00x014: I felt depressed and gloomy ch00x015: I felt happy 1: never 2: seldom 3: sometimes 4: often 5: mostly 6: continuously
Treatment variable: from “Background Variables” as belbezig and “Work and Schooling – LISS Core Study” as cw00x525:
Primary occupation: 1: Paid employment
2: Works or assists in family business
3: Autonomous professional, freelancer, or self-employed 4: Job seeker following job loss
5: First-time job seeker
6: Exempted from job seeking following job loss 7: Attends school or is studying
8: Takes care of housekeeping
9: Is pensioner ([voluntary] early retirement, old age pension scheme) 10: Has (partial) work disability
11: Performs unpaid work while retaining unemployment benefits 12: Performs voluntary work
13: Does something else
