SEARCH

SEARCH BY CITATION

Abstract

  1. Top of page
  2. Abstract
  3. Introduction
  4. 1 Related Literature
  5. 2 Empirical Strategy
  6. 3 Data and Variables
  7. 4 Matching Quality
  8. 5 Results
  9. 6 Robustness
  10. 7 Heterogeneity Analysis
  11. 8 Conclusion
  12. Acknowledgments
  13. References
  14. Supporting Information

This paper estimates the effect of involuntary job loss on smoking behaviour and body weight using German SOEP data. Baseline non-smokers are more likely to start smoking due to job loss, while smokers do not intensify smoking. In particular, single individuals and those with lower health or socioeconomic status prior to job loss exhibit high rates of smoking initiation. Job loss increases body weight slightly, but significantly. The applied regression-adjusted semiparametric difference-in-difference matching strategy is robust against selection on observables and time-invariant unobservables. This paper provides an indirect test that the identifying assumption is not violated.


Introduction

  1. Top of page
  2. Abstract
  3. Introduction
  4. 1 Related Literature
  5. 2 Empirical Strategy
  6. 3 Data and Variables
  7. 4 Matching Quality
  8. 5 Results
  9. 6 Robustness
  10. 7 Heterogeneity Analysis
  11. 8 Conclusion
  12. Acknowledgments
  13. References
  14. Supporting Information

The loss of employment has a deep impact on the individual affected. It not only reduces income in both the short and long run (Jacobson et al. 1993), but also results in a plethora of other negative issues, such as decreased life satisfaction (Knabe and Rätzel 2011), increased risk of divorce (Charles and Stephens 2004) and further job losses (Stevens 1997), as well as negative consequences for the children of those affected (Lindo 2011). This paper contributes to the literature on the effects of job loss by looking at the impact of involuntary job loss on smoking behaviour and body weight. Knowing more about the overall consequences of job loss is crucial for policymakers who are comparing the costs and benefits of labour market policies to prevent job losses. Furthermore, analysing the effects of job loss on health behaviours may help us to better understand two scientific puzzles: why job loss increases mortality at the individual level (Eliason and Storrie 2009; Sullivan and von Wachter 2009) and why—at an aggregated level—health behaviours improve when the unemployment rate is high (Ruhm and Black 2002; Ruhm 2005). The topic is also relevant from a public health perspective, as smoking and excess body weight are among the major causes of preventable deaths.

There is ample evidence that unemployed individuals engage in unhealthy behaviours more often than those in employment (see Henkel 2011; Roelfs et al. 2011). However, it is unclear whether this represents the causal effect of unemployment, the reverse causal effect, or a spurious correlation. Most studies on the effect of job loss on health behaviours suffer from various shortcomings. The studies focus on correlations, concentrate on case studies, do not deal with the endogeneity of job loss, or do not draw on longitudinal data (see Roelfs et al. 2011). This paper addresses these shortcomings by reverting to population-representative survey data, by looking at involuntary job loss and by applying a regression-adjusted semiparametric difference-in-difference matching strategy. This strategy is robust not only against selection on observables (like conventional matching estimators), but also against selection on unobservables with time invariant effects (e.g. ability, childhood conditions). To interpret the estimated effects as causal, the estimation strategy assumes that no unobservable variables exist that simultaneously influence the probability of job loss and changes in health behaviours. This paper provides an indirect test to show that this identifying assumption is not violated in the present case.

Empirical evidence on the effects of job loss on health behaviours is mixed, although most studies find that job loss changes health behaviours for the worse (see Henkel 2011; Roelfs et al. 2011). It seems likely that these contradictory findings are at least partially due to methodological weaknesses. Only a few studies rely on longitudinal data and simultaneously do not concentrate on single plants. Among these studies, Morris et al. (1992) find job loss-related increases in body weight but no increases in either smoking or drinking. Another British study (Montgomery et al. 1998) finds an increased risk of smoking and low body weight among male participants with experience of unemployment—even when controlling for health behaviours at age 16.

Two studies using the US Health and Retirement Survey (HRS) provide the most sophisticated strategies for the identification of causal effects, as they rely on longitudinal data and focus on involuntary job loss. Falba et al. (2005) show that job loss increases the probability of smoking relapse and increases the daily number of cigarettes smoked by existing smokers. Deb et al. (2011) find an increase in drinking and in the probability of being overweight, but only for individuals who already had poor health behaviours prior to job loss. This paper supplements these two studies by applying a different estimation technique, and by looking at the effect of job loss in a different welfare regime. While Falba et al. (2005) apply logit/OLS regressions and Deb et al. (2011) apply finite mixture models, this paper resorts to matching, an intuitive approach that does not rely as much on linearity assumptions. Due to the HRS sampling design, these two studies can look at only individuals over 50, while the data of the German Socio-Economic Panel Study (SOEP) allow for an analysis of the entire range of working-age individuals. By focusing on Germany, this paper looks at a country with more generous unemployment assistance than the USA.

Using SOEP data for 2002–10, this paper finds that job loss increases the probability of smoking initiation by 3 percentage points (55%) on average. However, there is little evidence that baseline smokers—i.e. individuals who smoked before the job loss—intensify their smoking or are less likely to stop smoking due to job loss. Job loss increases body weight slightly (by around 0.1 kg/m2), but significantly. These findings emerge regardless of who is considered: only individuals who lost their jobs due to plant closure, or all individuals experiencing job loss due either to plant closure or dismissal. The results are robust over various matching specifications and different choices of the conditioning variables. Further analyses indicate treatment-effect heterogeneity. In particular singles, younger individuals and individuals with lower health or socioeconomic status prior to job loss exhibit high rates of smoking initiation. The increase in body weight is larger for women and overweight individuals, but still well below 0.5 kg/m2. In general, the effects on smoking and body weight are smaller than comparable findings for the USA, which might be attributable to the more generous unemployment assistance in Germany.

1 Related Literature

  1. Top of page
  2. Abstract
  3. Introduction
  4. 1 Related Literature
  5. 2 Empirical Strategy
  6. 3 Data and Variables
  7. 4 Matching Quality
  8. 5 Results
  9. 6 Robustness
  10. 7 Heterogeneity Analysis
  11. 8 Conclusion
  12. Acknowledgments
  13. References
  14. Supporting Information

Initially, the economic literature on the impact of job loss focused on lost earnings (see, for example, Jacobson et al. 1993). Research in the field developed and started looking at the implications of job loss beyond mere financial consequences to acquire a more exhaustive picture of job loss. This is of particular importance for policymakers who are comparing the costs and benefits of labour market policies that prevent job losses—especially in the financial and economic crisis that enveloped the world starting in 2008. During this crisis, the rate of job losses increased or were expected to rise in many western countries; for example, in the USA, the rate of job losses is at a record high (Farber 2011).

Previous research finds that the loss of employment strongly reduces life satisfaction (Kassenboehmer and Haisken-DeNew 2009; Knabe and Rätzel 2011), increases the risk of divorce (Charles and Stephens 2004), and decreases fertility (Del Bono et al. 2012). Other studies analyse the health consequences of job loss. Unemployed individuals experience higher mortality (see Roelfs et al. 2011), and several studies provide evidence that this is (partially) due to the loss of employment and not merely due to poorer health status or other observable factors prior to job loss (Eliason and Storrie 2009; Sullivan and von Wachter 2009).1

At the macro level, however, increases in the unemployment rate are found to reduce mortality (Ruhm 2000). While infant health also improves at aggregated levels when unemployment is high (Dehejia and Lleras-Muney 2004), at the individual level the loss of employment is found to reduce the birth weight of own children (Lindo 2011). Furthermore, physical inactivity and body weight, as well as tobacco and alcohol consumption, decrease during economic downturns (Ruhm and Black 2002; Ruhm 2005). However, it is not clear whether the improvement in health behaviour is driven by the unemployed or by those individuals who are still in employment but working less.

The literature discusses several arguments as to why job loss might affect smoking and body weight. From a theoretical point of view, it is not clear whether job loss improves or worsens health behaviours.2 One of the most relevant mechanisms might be stress. Medical studies find that stress increases both eating (Adam and Epel 2007) and smoking (Kassel et al. 2003). Hence, to cope with the stress associated with job loss, individuals might increase calorie and nicotine intake as a form of self-medication. Job loss-related stress might result from reduced income, the fear of not finding a new job, and the loss of the non-financial benefits of work, such as respect of others—adhering to social norms—and therefore identity (Akerlof and Kranton 2000).

Job loss might also affect health behaviours through an income effect. The loss in income might reduce the number of cigarettes smoked. The income effect on body weight is ambiguous (consumption of less food versus cheaper, less healthy food). However, unemployment benefits are comparably generous in Germany. Recipients get 60% of their previous income (or 67% if they have a child) for up to 24 months, depending on the duration of their own contributions to unemployment insurance. Moreover, the unemployed can receive tax-financed unemployment assistance if unemployment insurance payments are below a certain threshold or if the payment period expires (see Caliendo et al. 2009). Other arguments as to why job loss might affect smoking and body weight include increased leisure time (with ambiguous effects on health behaviours) and a shift in the individual's time preference towards the present (Schunck and Rogge 2010).

Research on the impact of job loss on health behaviours contributes to the literature in several ways. First, it investigates one of the potential mechanisms as to why job loss increases mortality. Job loss might have negative health consequences if health behaviours deteriorate. Second, analysing the effect of job loss on smoking and body weight deepens our understanding of why aggregated health behaviours improve during economic downturns. Third, the effect of job loss on health behaviour is also relevant from a public health perspective. Smoking and excess body weight are first and third among the major causes of preventable deaths in industrialized countries, according to the World Health Organization (2009). Policies aimed at reducing the prevalence of smoking and being overweight may be more effective if they consider the vulnerability of specific groups, such as individuals who have lost their employment.

2 Empirical Strategy

  1. Top of page
  2. Abstract
  3. Introduction
  4. 1 Related Literature
  5. 2 Empirical Strategy
  6. 3 Data and Variables
  7. 4 Matching Quality
  8. 5 Results
  9. 6 Robustness
  10. 7 Heterogeneity Analysis
  11. 8 Conclusion
  12. Acknowledgments
  13. References
  14. Supporting Information

To investigate the causal effect of job loss on body weight, smoking status and the daily number of cigarettes smoked, this paper applies a regression-adjusted difference-in-difference (DiD) matching strategy, similar to the one proposed by Heckman et al. (1997). The general idea of this estimator is to compare individuals who have lost their jobs with (nearly identical) individuals who have not lost their jobs and see how the health behaviours of these two groups changed. This study focuses on the average treatment effect on the treated (ATT), i.e. changes in health behaviours brought about by the job loss of those who actually lost their employment. The identifying assumption for this ATT is that no unobserved variables exist that determine job loss and simultaneously influence a change in health behaviours; that is, in the absence of treatment (job loss), the health behaviours of treated (individuals experiencing job loss) and matched controls (similar individuals not experiencing job loss in the period) follow the same trend. In formal notation, the identifying assumption is given as

  • display math(1)

where inline image refers to the change in health behaviours from before (b) to after (a) the treatment in the absence of treatment, and D denotes the treatment group indicator. The propensity score P is the conditional probability of job loss, i.e. inline image, where Sb is a set of conditioning variables obtained in the pre-treatment period.

In general, DiD matching estimators of the ATT can be written as (Heckman et al. 1997; Smith and Todd 2005)

  • display math(2)

where n1 is the number of cases in the treatment group I1. The control group observations are indicated by I0, and ω(i,j) is a matching procedure specific weight. For instance, for k nearest neighbour matching, the weights ω(i,j) take on the value 1/k for the k nearest control neighbours of each treated i, and 0 for all other non-treated. The ATT can be also expressed in matrix notation as

  • display math(3)

where Δy is the vector of health behaviour changes, and W is a diagonal matrix with 1 as diagonal element for individuals of the treatment group and inline image for control group members. In the case of matching without regression adjustment, X is an n × 2 matrix, in which the first column consists of a vector of 1s, and the second column gives the values of D for each individual. With regression adjustment, X contains an additional column for each adjustment variable.3

I apply a three-step procedure to estimate the ATT. First, I estimate the propensity score, then I compute the weighting matrix of equation (3), and finally I perform a weighted regression to compute the ATT.

In the first step, I estimate the propensity score by probit regressions on the pooled sample. I do not revert to the propensity score directly, but instead use the linear index of the propensity score, which is particularly effective at improving the balance between treated and controls (Rosenbaum and Rubin 1985a). In order to prevent the comparison of treated and untreated that are not comparable, I restrict the analysis to the region of common support. The analysis excludes treated observations whose linear propensity score exceeds the maximum of the linear propensity score in the control group or falls below its minimum. For kernel matching, the Epanechnikov kernel works as an additional common support condition since it matches only control observations within a specific interval around each treated observation.

In the second step, I compute the weights within cells defined by baseline smoking status (yes/no) and survey year according to the distance in the linear propensity score.4 This is equivalent to exact matching on survey year and smoking status. As equations (2) and (3) show, matching procedures basically differ with respect to the weights.5 Asymptotically, all matching procedures produce the same results because they reduce to exact matching in infinite samples (Caliendo and Kopeinig 2008). For finite control groups, there is no one best matching procedure for all situations (Caliendo and Kopeinig 2008). Yet Heckman et al. (1997) and Smith and Todd (2005) argue for the use of kernel matching, which is the method primarily used in this study. To test the sensitivity of the results with respect to different matching procedures, I also apply 5-to-1 nearest neighbour caliper matching as a robustness check. For kernel matching, the weights take on the form

  • display math(4)

where P is the linear index of the propensity score, K[·] is a specific kernel function and bn is a bandwidth parameter. There is a general agreement that the choice of the kernel is less crucial than the choice of the bandwidth. I use the Epanechnikov kernel due to its slight superiority in terms of efficiency. The chosen bandwidth is 0.06, as applied by Heckman et al. (1997). Although Smith and Todd (2005) find their results to be robust against different bandwidth choices, in the section on robustness checks I apply a different bandwidth.

The third step is the regression step, where I compute the ATT according to equation (3). I include all conditioning variables S from the propensity score equation in X. While the calculation of the weights is performed within cells, the ATT is computed for the pooled sample.

The applied regression-adjusted DiD matching estimator improves conventional matching estimators in several ways. First, the DiD matching estimator eliminates the time-invariant outcome difference between individuals in the treatment and control group. The pure cross-sectional matching estimator assumes that all relevant variables that determine job loss and influence the level of health behaviours are included in the dataset and incorporated into the model. This is a much stronger assumption. Second, the regression-adjusted matching estimator remains consistent if either the propensity score equation or the regression equation is correctly specified. Therefore the researcher has two opportunities to deal correctly with the selection bias, and hence the model can be regarded as doubly robust (Bang and Robins 2005). The inclusion of all variables from the propensity score estimation reduces the effect of prevailing covariate imbalances after matching, and thus decreases the small-sample and asymptotic bias of the matching estimator (Abadie and Imbens 2006). Furthermore, including relevant variables in the regression step decreases unexplained variance in the outcome, and hence the standard errors of the treatment effect estimates. This is similar to including control variables in randomized experiments. Third, I combine propensity score matching with exact matching on survey year and baseline smoking status, which I regard as crucial variables for the purpose of this study. Smoking initiation and smoking cessation are different decisions, and the distinction between these two decisions is useful (DeCicca et al. 2008). Therefore I display smoking results separately for smokers and non-smokers, and match exactly on the smoking status to compare only like and like. Performing inexact matching on survey year would imply comparing individuals from different time periods. Additionally, matching on the survey year allows me to control for general macroeconomic trends (like cigarette taxes, which are the same everywhere in Germany at a given point in time). Fourth, the matching and analysis steps are clearly separated. Therefore the quality of the matching procedure is evaluated prior to the computation of the ATT.

To highlight some benefits of the applied matching procedure, the presentation of the results in Table 3 starts with a simple specification without pre-treatment health behaviour information and without exact matching. The other specifications gradually incorporate the more sophisticated procedures.

There is a debate in the literature as to how to estimate the variance of propensity score-based matching estimators (see Caliendo and Kopeinig 2008; Stuart 2010). The crux is that the variance estimation should take into account uncertainty in the propensity score estimation, although there is evidence that standard errors are overestimated if the estimated propensity score is used instead of the true propensity score (see Stuart 2010). Hence if the uncertainty in the propensity score model is not considered, the error might be on the conservative side. Usually, bootstrapping constitutes a popular way of estimating standard errors when they are difficult to compute analytically or when the theoretical distribution of the relevant statistic is unknown. However, there is no formal justification for the application of the bootstrap in the case of matching, and Abadie and Imbens (2008) show that bootstrapping fails in the case of nearest neighbour matching. Therefore this study does not rely primarily on bootstrapped standard errors.6 Instead, I use robust standard errors from the weighted regressions. Robustness tests in the full working paper (Marcus 2012) show that the applied standard errors are slightly more conservative than comparable bootstrapped standard errors and standard errors computed according to the formula suggested by Lechner (1999).

3 Data and Variables

  1. Top of page
  2. Abstract
  3. Introduction
  4. 1 Related Literature
  5. 2 Empirical Strategy
  6. 3 Data and Variables
  7. 4 Matching Quality
  8. 5 Results
  9. 6 Robustness
  10. 7 Heterogeneity Analysis
  11. 8 Conclusion
  12. Acknowledgments
  13. References
  14. Supporting Information

This study uses data from the 2002–10 waves of the German Socio-Economic Panel Study (SOEP), which is one of the largest and longest-running survey panels in the world. Annually, it conducts interviews with more than 20,000 individuals in over 10,000 households in Germany (Wagner et al. 2007). The SOEP provides a wide range of information at the individual and the household level, for example about working and living conditions. In even-numbered years, starting in 2002, the SOEP includes a detailed health module, which I use to construct the outcome measures.

Outcome variables

This study analyses two different types of health behaviour, smoking and body weight. I use information about smoking behaviour and body weight from the 2002, 2004, 2006, 2008 and 2010 waves. The SOEP does not elicit this information in the odd-numbered years. I look at two different smoking measures: the daily number of cigarettes smoked, and a dummy variable indicating whether or not an individual currently smokes. For the current smoking status, I distinguish between smoking initiation (the decision of non-smokers to start smoking),7 smoking continuation (the decision of smokers to remain smokers) and smoking participation (the smoking decisions of smokers and non-smokers combined), following DeCicca et al. (2008). I exclude observations of those who smoke solely pipes or cigars.

I use body mass index (BMI) as the weight outcome, where BMI is defined as the individual's body weight divided by the square of the individual's height, measured in kg/m2. The study does not consider overweight status (BMI > 25) or obesity (BMI > 30) directly because few individuals in the sample change between these (arbitrary) weight categories. A gain in body weight is not necessarily bad for health, especially for underweight people. Yet more than half of the treated individuals are overweight, and more than 15% of them are obese. Section 'Heterogeneity Analysis' looks at whether there are differences in job loss-induced body weight changes between overweight and normal weight individuals.

Treatment and control group

Treatment and control groups consist of individuals between 18 and 62 years of age8 who were neither self-employed nor civil servants, and who were working either full- or part-time during the pre-treatment period (denoted by t − 1). I include only individuals with non-missing smoking and BMI information before and after treatment. This drops the sample size by about 2%. The treatment group comprises individuals who lost their job due to plant closure or dismissal between two survey rounds, i.e. between t − 1 and t + 1, with information about the individuals' health behaviours. For the control group, I select only those individuals without job change between t − 1 and t + 1. Note that the SOEP includes the information on health behaviours only every two years. Hence between t − 1 and t + 1 a survey round without questions on health behaviours takes place (at time t). Since the question about job loss in the SOEP incorporates all job losses occurring during the survey year and the previous year, the information about job loss is derived from the interview in either t or t + 1. All other data originate from t − 1, the last wave with information about the individuals' health behaviours before the treatment. On average, this last interview took place eleven months before the job loss. The time between job loss and the following interview is, on average, thirteen months.

The study looks only at involuntary job loss, and therefore does not include individuals who voluntarily terminated their employment (including, for example, resignation or mutual agreement), because selection issues might arise. Further, it is not clear from a theoretical point of view why this should lead to a change in health behaviours. It might be argued that individuals are dismissed because of a change in smoking behaviour or body weight. Dismissals due to a high level of smoking or weight should not distort the results since the levels are taken into account in the matching procedure. To show that there is no reverse causality, I also perform analyses in which the treatment group consists of only individuals who lost their jobs due to plant closure, for which exogeneity is stricter. In the whole observation period there are 1768 incidences of involuntary job loss. Since only 520 occurred due to plant closure, I do not rely solely on these cases. The potential control group consists of more than 22,500 observations.

Conditioning variables

The identification strategy builds on the assumption that the model includes all variables that simultaneously influence the probability of job loss and changes in health behaviours. Therefore the choice of the conditioning variables S is crucial. I reviewed variables used in other studies analysing health effects of job loss,9 and include as many of these variables as possible. Since I rely on rich survey data, I can not only include more conditioning variables than any one of these other studies, but also use almost all conditioning variables from these studies. It is even possible to include perceived job security, which was found to be a good predictor of subsequent job loss also when controlling for other characteristics (Stephens 2004). With respect to the outcome variables, the data allow me to condition not only on smoking behaviour and BMI prior to job loss, but also on smoking history, i.e. whether an individual ever smoked 100 cigarettes in their life.

Matching variables can be roughly divided into demographic, labour market-related, educational, regional and health-related (see Table 1). In order to not lose treatment observations with missing values on some conditioning variables, I set missing values to 0 and include binary variables indicating a missing value. The affected variables are marked in Table 1 (see the table notes). Hence I treat missing values as just another category of these variables. Matching is therefore not only on the observed values but also on the missing data pattern (Stuart 2010). The ATT estimates are very similar when observations with missing values are excluded; the standard errors, however, slightly increase (results are not shown, but are available on request). In total, the study conditions on 79 non-collinear variables.10

Table 1. Overview of the Conditioning Variables
  1. Notes

    aVariables with an additional category for missing values. bI group Bremen with Lower Saxony and Hamburg with Schleswig-Holstein due to few cases. cCoding according to Ziebarth and Grabka (2009). Since the data on drinking behaviour are available only since 2006, they are highly affected by missing values. Excluding these variables does not alter the results (see the online appendix).

Demographic
Female0 = male, 1 = female
Agein years, third-order polynomial
Migrant1 = individual or parents moved to Germany, 0 else
Non-German0 = German, 1 = foreign citizenship
Home owner0 = tenant, 1 = home owner
Spouse1 = married or unmarried spouse in the household, 0 else
Children1 = children under 18 in household, 0 else
Survey year4 categories (2002, 2004, 2006, 2008)
Labour market
Labour earningslogarithm of annual earnings in euros
Never unemployed0 = ever unemployed, 1 = never unemployed
Years unemployedyears with months in decimal form
Tenurein years
Blue collar0 = white-collar, 1 = blue-collar worker
Perceived job securitya3 categories (big worries, some worries, no worries)
Company sizea3 categories ( < 20, 20–200,  > 200 employees)
Industry sector10 categories
Educational
University0 = no university degree, 1 = university degree
Vocational training0 = no vocational training, 1 = vocational training
Secondary schoolinga4 categories (no degree/basic school, intermediate/other school academic school track (Abitur), technical school)
Regional
Residential district4 categories (<2000, 2000–20,000, 20,000–100,000, >100,000 inhabitants)
Federal state14 categoriesb
Regional unemploymentyearly information on the state level
Health
Self-rated health3 categories (very good/good, satisfactory, poor/bad)
Private health insurancea0 = public, 1 = private health insurance
Mental healthabased on SF12 questionnaire (see Andersen et al. 2007), cardinal
Physical healthabased on SF12 questionnaire, cardinal measure
Baseline smoker0 = baseline non-smoker, 1 = smoker before treatment
Number of cigarettesdaily number of cigarettes
Ever smokera0 = never smoked, 1 = ex-smoker (asked only in 2002 survey)
Heavy smoker0 = <20 cigarettes/day, 1 = ≥20 cigarettes a day
Partner smokesa0 = partner is non-smoker, 1 = partner smokes
Overweight0 = BMI ≤ 25, 1 = BMI > 25
Body weightin kg
BMIbody mass index, in kg/m2
Heightin centimetres
Drinking behavioura4 categories (regular, moderate, rare drinker, abstainer)c

4 Matching Quality

  1. Top of page
  2. Abstract
  3. Introduction
  4. 1 Related Literature
  5. 2 Empirical Strategy
  6. 3 Data and Variables
  7. 4 Matching Quality
  8. 5 Results
  9. 6 Robustness
  10. 7 Heterogeneity Analysis
  11. 8 Conclusion
  12. Acknowledgments
  13. References
  14. Supporting Information

The quality of matching can be assessed by comparing the means of the conditioning variables for treated and untreated after matching. If the means of treated and matched controls deviate very much, matching on the linear propensity score did not work well, and rather different observations are compared. To determine whether a mean difference is large, I look at the standardized bias. The standardized bias displays for each conditioning variable S the difference between the means of treated (s1) and matched (s0) controls as a percentage of the square root of the average of the variances (inline image) in the two groups (Rosenbaum and Rubin 1985b):

  • display math(5)

When the standardized bias exceeds 5%, the mean difference is regarded to be large (Caliendo and Kopeinig 2008), and balancing did not work very well.

Table 2 presents means and standardized biases for the pooled sample before and after matching.11 Before matching, those who lost their jobs differ. For instance, they are more likely to have a migration background, to have lower levels of income, to have more previous unemployment times, to have less tenure, to be blue collar workers, to work more often in small companies, to be more concerned about the security of their jobs and to have lower levels of secondary education. With respect to the health behaviours analysed, the treated are more likely to smoke (43% versus 34%), have a slightly higher BMI on average and are more likely to be overweight (53% versus 51%) before job loss. Matching strongly reduces these disparities in health behaviours as well as in other conditioning variables. Therefore matching seems to function properly. Small remaining differences between treated and matched controls are mitigated by the regression adjustment.

Table 2. Means of Treated, Controls and Matched Controls—Before Treatment
VariableTreatedControlsStandardized bias (%)
UnmatchedMatchedUnmatchedMatched
  1. Notes

    Summary statistics for treated, all controls and matched controls. The first three columns present the means of selected variables before treatment for treated, controls and matched controls, respectively. The last two columns display the standardized bias in % before and after matching. a indicates that the mean represents a percentage share.

Demographic
Femalea41.3347.1142.21−11.64−1.78
Age40.8742.7141.02−18.06−1.42
Migranta19.6315.3619.7411.27−0.28
Non-Germana15.3811.5915.6811.11−0.82
Home ownera40.7054.2740.79−27.41−0.19
Spousea73.5978.7674.43−12.15−1.92
Childrena39.9541.1239.55−2.380.83
Labour market
Labour earnings9.6610.139.66−38.07−0.36
Never unemployeda44.3768.3945.74−49.91−2.75
Years unemployed1.070.441.0540.040.87
Tenure6.5911.676.81−58.05−3.07
Blue collara49.0232.9048.3533.241.35
Small companya39.4420.7938.8441.511.22
Medium companya32.7829.7533.306.54−1.11
Large companya25.6046.9525.56−45.520.10
No company infoa2.182.522.30−2.23−0.78
Job worriesa33.9815.3234.6244.36−1.35
No job worriesa20.4438.8520.18−41.160.64
No job worries infoa1.661.441.991.83−2.46
Education
Universitya14.6422.0714.52−19.290.32
Vocational traininga76.1875.5475.911.480.63
Basic schoola32.7228.2832.599.660.28
Intermediate schoola47.3043.8547.416.93−0.22
Technical collegea3.966.524.08−11.49−0.58
Highest secondarya13.4319.6613.35−16.810.25
No schooling infoa2.581.702.586.130.04
Regional
Villagea11.318.7411.618.54−0.94
Small towna34.1035.4733.76−2.870.71
Small citya26.6927.1626.57−1.060.29
Big citya27.9028.6228.07−1.61−0.37
Regional unemployment11.0610.1111.0320.710.60
Health
Bad healtha11.779.6012.167.02−1.21
Good healtha53.9058.8453.48−9.960.84
Private health insurancea3.276.713.40−15.85−0.71
No insurance infoa0.230.200.420.73−3.36
Mental health48.1249.2848.01−9.970.89
Physical health50.7251.3850.61−6.111.01
No health infoa2.181.962.211.53−0.19
Baseline smokera42.7133.6042.7118.840.00
Number of cigarettes7.295.497.4418.27−1.37
Ever smokera63.6657.3463.7012.96−0.08
No ever smoker infoa5.175.775.12−2.670.20
Heavy smokera19.5214.7319.7012.74−0.46
Partner smokesa26.5223.1226.737.89−0.46
No partner smoking infoa29.7925.4528.849.732.10
Overweighta53.2151.1053.674.23−0.92
Body weight77.7277.0077.774.50−0.36
BMI25.8025.6725.862.96−1.25
Height173.19172.82173.094.111.15
Abstainera4.543.714.454.160.42
Rare drinkera10.9113.0111.09−6.47−0.58
Moderate drinkera16.7020.6716.71−10.18−0.01
Regular drinkera6.898.356.91−5.50−0.07
No alcohol infoa60.9654.2760.8513.580.24
Table 3. The Effect of Job Loss on Smoking Behaviour and Body Weight
OutcomeCross-sectionLongitudinal 
 Not in cellsIn cellsRegression-adjusted
(1)(2)(3)(4)
  1. Notes

    The table presents the effect of job loss on smoking and body weight. Each cell displays the ATT from a separate regression and its robust standard error in parentheses. Row names indicate the outcome. Specification (1) does not take into account baseline health behaviours, while specifications (2)–(4) use the change in health behaviour as outcome. Specification (3) performs matching within cells defined by survey year and smoking status, which specification (2) does not. Specification (4) additionally includes the conditioning variables in the outcome equation. The row ‘Median standardized bias’ displays the median of the standardized bias over all matching variables. ‘Off common support’ indicates the share of treated individuals (from a base of N = 1768) who are not considered in the estimation due to inappropriate matches. *, **, *** indicate p < 0.1, p < 0.05, p < 0.01, respectively.

All individuals
Smoking participation0.047***0.023***0.025***0.024***
(0.013)(0.008)(0.008)(0.008)
Number of cigarettes0.626**0.1760.3580.234
(0.266)(0.175)(0.267)(0.163)
Body mass index (BMI)0.1400.100*0.095*0.097*
(0.121)(0.053)(0.053)(0.052)
Baseline non-smokers
Smoking initiation 0.030***0.028***0.030***
 (0.009)(0.009)(0.009)
Baseline smokers
Smoking continuation 0.0170.0200.016
 (0.013)(0.014)(0.013)
Number of cigarettes 0.0460.4620.086
 (0.363)(0.586)(0.338)
Off common support (%)0.110.171.471.47
Median standardized bias1.000.930.690.69

After imposing the common support conditions, the sample comprises 1742 treated observations (513 with plant closure). This reduces the sample size by about 1.5%. The standardized bias is much smaller after matching, and never exceeds the critical value of 5%. The median standardized bias over all 79 conditioning variables is 10.07 before matching and 0.69 after matching; the mean standardized bias amounts to 0.86. In general, the median standardized bias is rather low compared to other studies (e.g. Lechner 2002; Kuhn et al. 2009). Also based on this criterion, the matching procedure succeeds.

5 Results

  1. Top of page
  2. Abstract
  3. Introduction
  4. 1 Related Literature
  5. 2 Empirical Strategy
  6. 3 Data and Variables
  7. 4 Matching Quality
  8. 5 Results
  9. 6 Robustness
  10. 7 Heterogeneity Analysis
  11. 8 Conclusion
  12. Acknowledgments
  13. References
  14. Supporting Information

Table 3 presents the estimates of the effect of job loss on body weight and smoking (ATT) for several specifications: (i) for all individuals, and separately for (ii) baseline smokers and (iii) baseline non-smokers. The table starts with the simplest specification in the first column. The other specifications gradually incorporate more sophisticated procedures. Specification (4) is the preferred specification. For each matching specification, Table 3 also presents the share of treated observations off the common support, i.e. treated individuals without adequate match, and the median of the standardized bias.

Specification (1) does not use any information about the pre-treatment health behaviours in the estimation of the linear propensity score or the outcome variable, or for regression adjustment.12 Hence I do not present results separately for baseline smokers and non-smokers. This specification most closely resembles a pure cross-sectional matching estimator. Matching is not exact on survey year and baseline smoking status in this specification. In this specification, job loss increases smoking participation by 4.7 percentage points, the daily number of cigarettes smoked by 0.6 cigarettes and BMI by 0.14 kg/m2—although the effect on BMI is not significant at common significance levels.

Specification (2) is the DiD matching estimator where the outcome is the change in health behaviour and where the propensity score estimation takes into account Yb, the pre-treatment health behaviours. For the pooled sample, job loss increases the probability of smoking by 2.3 percentage points and increases BMI by around 0.1 kg/m2. This is in addition to the usual weight gain between two observation periods, which is about 0.34 kg/m2 for the matched controls. Therefore the total body weight increase amounts to 0.44 kg/m2 for the treated. The effect on the current smoking status is principally driven by baseline non-smokers starting to smoke. There is no significant increase in the probability of baseline smokers remaining smokers. Smokers also do not smoke significantly more cigarettes, on average, after their job loss. Compared to specification (1), the effect of job loss on health behaviours is considerably smaller. If there were no selection on unobservable characteristics with time-invariant effects, the estimates of the ATT should be very similar. The large differences might be taken as an indication that the cross-sectional matching estimator of specification (1) is biased upwardly.

While for specification (2) matching is on the linear index of the propensity score of the pooled sample, specification (3) additionally performs exact matching on survey year and smoking status. The median standardized bias is slightly smaller, while the share of treated observations outside the common support increases. The results resemble the results of the previous specification.

In specification (4) there is additional regression adjustment for all conditioning variables used in the estimation of the propensity score. As expected, regression adjustment slightly decreases the standard errors compared to specification (3). Some estimated effects decrease somewhat in magnitude. Baseline non-smokers are around 3.0 percentage points more likely to smoke due to job loss. This implies an increase in the probability of starting smoking of more than 55%, as less than 5.5% of the baseline non-smokers start smoking. The 0.1 kg/m2 increase in BMI due to job loss denotes an increase of about 0.4% (0.3 kg) in total body weight and about 30% in body weight change.

The average increase in BMI differs from findings of Deb et al. (2011), who observe significant increases for only one latent group in their finite mixture models. This group consists of individuals who already engage in unhealthy behaviours before job loss. For this group, Deb et al. (2011) estimate a job loss-related increase in BMI by more than one unit. Section 'Heterogeneity Analysis' performs separate analyses for different at-risk groups and shows that for overweight individuals the effect is larger but still considerably below 0.5 BMI units.

The effect on smoking initiation is somewhat smaller than the effect in Falba et al. (2005), where it is estimated that job loss more than doubles the probability of smoking initiation. However, Falba et al. (2005) not only apply a different estimation strategy and analyse US individuals aged over 50, but also focus solely on ex-smokers and therefore on smoking relapse. In Section 'Heterogeneity Analysis', I perform separate analyses for ex-smokers. I find that for ex-smokers, job loss increases the probability of (re)starting smoking by about 50% (5.6 percentage points), which is again smaller than the effect in Falba et al. (2005).

To gauge whether the statistically significant effects in specification (4) are meaningful, the findings of this study are compared with other effects on smoking and BMI reported in the literature. Ruhm (2005) estimates that a one point drop in the employment rate in the USA reduces the estimated smoking participation rate by 0.13 percentage points on the aggregate level. In contrast, I find individual job loss to increase smoking participation by 2.4 percentage points. This effect is almost twenty times as large. If healthy living really improves during economic downturns, and smoking participation of individuals experiencing job loss increases, there must be some groups (e.g. employed individuals with reduced working hours) for which smoking rates decrease even more than the average estimated by Ruhm (2005). Exploiting large cigarette tax increases in the USA, DeCicca and McLeod (2008) find that a $1 increase in the cigarette excise tax reduces smoking participation by 1.0–1.5 percentage points. For BMI, job loss resulted in an estimated increase in BMI of about 0.1 units in Table 3. This is considerably less than the reduction in BMI of 0.64 kg/m2 resulting from a one point reduction in the employment rate in Ruhm (2005). In light of these other findings, the statistically significant increase in BMI seems to be of minor economic significance, while the increase in smoking participation seems to be rather high.

6 Robustness

  1. Top of page
  2. Abstract
  3. Introduction
  4. 1 Related Literature
  5. 2 Empirical Strategy
  6. 3 Data and Variables
  7. 4 Matching Quality
  8. 5 Results
  9. 6 Robustness
  10. 7 Heterogeneity Analysis
  11. 8 Conclusion
  12. Acknowledgments
  13. References
  14. Supporting Information

This section conducts robustness checks and analyses the plausibility of the identifying assumption. Further robustness tests can be found in the full working paper version of the paper (Marcus 2012).

For the treatment group in specification (5), I consider only individuals who lost their job due to plant closure, since dismissals might be endogenous. This reduces the number of treated to less than one-third of the original sample (520 observations). While plant closure is largely exogenous to the individual, frequently it does not come without warning. Those experiencing a plant closure might be a very selective group, as these individuals did not leave the plant earlier (Kletzer 1998). Table 4 shows that the two main effects—the effect on BMI and the effect on the current smoking status for non-smokers—are of similar magnitude to the base specification. However, the effect on BMI is no longer significant. For smokers, the effects on smoking status and the number of cigarettes remain insignificant but become negative. Therefore the effect on smoking participation turns insignificant in the pooled sample.

Table 4. The Effect of Job Loss on Smoking Behaviour and Body Weight—Robustness Checks
OutcomeMainPlant closurePlacebo
(4)(5)(6)
  1. Notes

    The table presents the effect of job loss on smoking and body weight. Each cell displays the ATT from a separate regression and its robust standard error in parentheses. Row names indicate the outcome. Specification (4) is the same as in Table 3 (N = 1768). The treatment group in specification (5) consists only of individuals who lost their job due to plant closure (base N = 520). Specification (6) is a placebo regression that tests whether the job loss influences changes in health behaviours before the job loss. The row ‘Median standardized bias’ displays the median of the standardized bias over all matching variables. ‘Off common support’ indicates the share of treated individuals who are not considered in the estimation due to inappropriate matches. *, **, *** indicate p < 0.1, p < 0.05, p < 0.01, respectively.

All individuals
Smoking participation0.024***0.0090.001
(0.008)(0.014)(0.010)
Number of cigarettes0.234−0.187−0.190
(0.163)(0.256)(0.204)
Body mass index0.097*0.099−0.075
(0.052)(0.077)(0.063)
Baseline non-smokers
Smoking initiation0.030***0.029**0.015
(0.009)(0.014)(0.011)
Baseline smokers
Smoking continuation0.016−0.021−0.021
(0.013)(0.024)(0.018)
Number of cigarettes0.086−0.732−0.591
(0.338)(0.556)(0.428)
Off common support (%)1.471.350.53
Median standardized bias0.690.690.51

To identify causal effects, all matching procedures assume that the conditioning variables S include all variables simultaneously influencing the probability of job loss and changes in health behaviours. There is no direct test for this assumption. However, I perform placebo regressions to add additional credibility to this assumption.

In the placebo regressions I check whether the job loss influences changes in health behaviours before the job loss. Applying a similar three-step procedure as before, specification (6) shows that the placebo job loss does not influence changes in smoking behaviour and BMI. All estimated effects are insignificant and close to zero. This specification adds plausibility to the assumption that there are no unobservables that influence both treatment and outcome in the preferred specification. The evolution of health behaviours of treated and matched controls is similar before the actual job loss.13

In the full working paper version (Marcus 2012), I show that the results are robust to applying different bandwidths in the construction of the kernel weights, to including only variables in the set of conditioning variables that are significant in stepwise probit regressions, to using different methods of standard error computation (formula suggested by Lechner (1999), bootstrapping) and to applying different matching procedures.14

Nevertheless, the ATT estimates might be biased. Several arguments suggest that the bias would be rather a downward bias. First, individuals who lost their job might misreport their true adverse health behaviours. This does not produce a bias in the DiD estimation if they misreport in the same way before job loss, or if they change the degree of misreporting in the same way as the matched controls. However, if job loss makes them a stronger underreporter, then the DiD estimation underestimates the true effect of job loss. It seems possible that the treated increase the degree of underreporting in order to avoid fulfilling prejudices against the unemployed. Though it is also possible that they overreport their adverse health behaviours. This would lead to an upward bias of the treatment effect. Second, those who experience a greater negative impact from their job loss (e.g. with respect to stress, finances or identity) might be more likely to drop out of the sample.15 These individuals might also be the most likely to cope with the job loss by increasing body weight and nicotine consumption. Third, often job losses do not come without warning. If the period before the actual job loss was already stressful and poor health choices were how individuals coped with this increased stress, then the pre-treatment outcome might incorporate the higher level of adverse health behaviours. There is some empirical evidence that the fear of unemployment harms mental health (Knabe and Rätzel 2011; Reichert and Tauchmann 2011). This argument also suggests a downward bias. However, the placebo regression in specification (6) does not indicate an impact of job loss on changes in health behaviour before the job loss.

7 Heterogeneity Analysis

  1. Top of page
  2. Abstract
  3. Introduction
  4. 1 Related Literature
  5. 2 Empirical Strategy
  6. 3 Data and Variables
  7. 4 Matching Quality
  8. 5 Results
  9. 6 Robustness
  10. 7 Heterogeneity Analysis
  11. 8 Conclusion
  12. Acknowledgments
  13. References
  14. Supporting Information

While the two previous sections focus on average treatment effects, this section analyses subgroup-specific treatment effects. Average effects might hide the fact that some subgroups show no reaction to job loss while others show particularly strong reactions. Therefore the analysis of treatment effect heterogeneity is crucial for acquiring a better understanding of the consequences of job loss on smoking behaviour and body weight.

To facilitate a comparison of only like and like, in addition to exact matching on survey year and smoking status, I perform exact matching on the particular grouping variable. I apply the same three-step procedure as before, but compute the weights in step 2 separately for each combination of year, smoking status and the grouping variable. This resembles specification (4) in Table 3, with the exception that matching is exact on the grouping variable. Since I impose the same common support conditions as before, sample sizes might vary. Furthermore, these analyses drop observations with missing information on the grouping variable. I do not perform exact matching on all grouping variables simultaneously; the procedure outlined is repeated for each grouping variable.16 Grouping variables in Table 5 include gender, age, partnership status, socioeconomic status, overweight status, health and unemployment status.

Table 5. Treatment Effects by Subgroups
 NAll individualsNon-smokersSmokers
 Smoking participationBMISmoking initiationSmoking continuationNo. of cigarettes
  1. Notes

    The table presents the effect of job loss on smoking and body weight for various subgroups defined by the row variables. Each cell stands for a separate ATT. Super columns indicate the smoking status before job loss; columns indicate the outcome. The first column displays the number of all treated observations in the region of common support for the various subgroups. Estimation resembles the procedure of specification (4) in Table 3, with the difference that matching is additionally exact on the grouping variable. *, **, *** indicate p < 0.1, p < 0.05, p < 0.01, respectively.

Over 504160.019−0.0670.0020.058**−0.499
Under 5013180.026***0.142**0.039***0.0100.155
Difference  0.0070.208*0.037**−0.0490.653
Male10180.027**−0.0140.041***0.015−0.390
Female7170.0170.278***0.0160.0140.671
Difference  −0.0100.292***−0.025−0.0011.061
Single4480.029*0.0950.067***−0.0110.372
Partner12770.022***0.111*0.021**0.023−0.178
Difference  −0.0060.016−0.046*0.034−0.550
Lower unemployment8890.037***0.1160.048***0.0300.182
Higher unemployment8470.0110.126*0.0160.0040.235
Difference  −0.026*0.010−0.032*−0.0260.053
Lower education6580.035***0.1160.052***0.025−0.789
Higher education10310.0150.0970.021**0.0010.269
Difference  −0.020−0.020−0.031−0.0231.058
Normal8030.033***0.0590.038***0.0220.165
Overweight9240.018*0.134*0.021*0.0120.043
Difference  −0.0140.075−0.017−0.010−0.122
Worse health8000.046***0.0520.048***0.048***0.468
Better health9230.0070.172**0.020*−0.016−0.625
Difference  −0.040***0.120−0.028−0.064**−1.093
Job found10540.024**0.0880.026**0.017−0.211
Unemployed6880.025**0.1120.036***0.0140.521
Difference  −0.001−0.024−0.0100.003−0.732
No unemployment5450.028**−0.0600.026*0.024−0.121
Some unemployment12230.022**0.169***0.032***0.0120.165
Difference  0.006−0.229***−0.0060.012−0.286

Table 5 shows the ATT for various subgroups and outcome variables.17 The first column displays the number of treated observations (smokers and non-smokers together) in the region of common support for the subgroup defined by the row name. Two similar studies (Falba et al. 2005; Deb et al. 2011) analyse only individuals over the age of 50. In order to make it easier to compare the results with these studies, I look at individuals who are aged 50 or older and contrast them with individuals younger than 50. The first panel of Table 5 shows that the effect on BMI differs significantly by age groups. Only the BMI of young individuals increases due to job loss. Young non-smokers are also significantly more likely to start smoking than their older counterparts. If this pattern is similar in the USA, the probability of smoking initiation might be even higher in the general US population than the results of Falba et al. (2005) indicate. Older baseline smokers are, however, more likely to continue smoking due to job loss (5.8 versus 1.0 percentage points).

It is often argued that work is more crucial for the identity of men, and hence their reaction can be expected to be stronger. Roelfs et al. (2011) cite some studies that find stronger effects of job loss and unemployment on various outcomes for men. The second panel of Table 5 shows that the effect of job loss on smoking initiation is stronger for men (4.1 versus 1.6 percentage points)—although the difference is not statistically significant from zero. The only significant gender difference is with respect to weight gain, which is considerably larger for women (0.28 versus −0.01 kg/m2). Overall, there appears to be no clear evidence for stronger effects of job loss on health behaviours of men.

The next panel of Table 5 compares individuals with and without a cohabiting spouse. Spouses can give financial and emotional support, thus mitigating stress and financial hardship associated with job loss. Furthermore, spouses can work as a control authority preventing individuals from increasing smoking and body weight after job loss. The data support these theoretical expectations for smoking behaviour. Among baseline non-smokers, single individuals are the group showing the highest increase in the probability of starting smoking (6.7 percentage points); as a result of the loss of employment, they are significantly more likely to start smoking than individuals with a spouse in the household.

Clark (2003) finds the effect of unemployment on mental health to be smaller in regions with higher unemployment. In order to test whether the loss of employment has similar heterogeneous effects on health behaviours, I compare the effects in regions with high and low unemployment.18 The fourth panel of Table 5 shows that the effects of job loss on smoking initiation and participation are significantly higher in regions with lower unemployment. This finding suggests that job loss is less harmful when unemployment is more common in the region, and hence less of a social stigma.

Individuals with less education might react particularly strongly to job loss as they might have less savings to compensate foregone earnings. They might also be less aware of the dangers of adverse health behaviours and have smaller social networks to provide emotional support (Cutler and Lleras-Muney 2010). The results in the fifth panel of Table 5 support these considerations. Less educated individuals19 are more prone to initiate smoking due to job loss. This difference is substantial (3.1 percentage points) but not significant. A similar picture emerges when the results are grouped according to other indicators of socioeconomic status, for example above versus below median labour income, blue collar versus white collar (results not shown). Socioeconomic gradients in coping with job loss are found by other researchers as well (see Roelfs et al. (2011) for a discussion). For changes in BMI, though, no substantial differences are found between socioeconomic groups.

Deb et al. (2011) observe job loss-related increases in adverse health behaviours for individuals who have poor health behaviours prior to job loss. I cannot confirm this finding for smoking, but the effects on BMI go in the same direction as the findings in Deb et al. (2011). Heavy smokers (smoking at least 20 cigarettes a day) and light smokers do not differ significantly in their reactions (results not shown). Yet overweight individuals (BMI > 25) are more prone to increase their body weight as a consequence of job loss (see the sixth panel of Table 5). However, this difference is not statistically significant, and also for overweight individuals the increase in BMI resulting from job loss amounts to clearly less than 0.5 BMI units. The next panel shows that individuals in the lowest self-rated health categories (satisfactory, poor or bad) are significantly more likely to smoke due to job loss. This might further increase existing health differences, since smoking increases morbidity and mortality. There are no significant differences in the effects on BMI.

For the last two grouping variables, there is no need to recalculate the weights as the groupings rather reflect differences in the treatment. The penultimate panel differentiates between whether the individual had found a job at the time of the following interview or was still unemployed. Although those in unemployment are slightly more likely to have started smoking (3.6 versus 2.6 percentage points), there are no statistically significant differences between the two groups. This underlines that the job loss event is inherently stressful. The last panel differentiates between individuals who directly found a new job without intervening unemployment and individuals with some unemployment after the job loss. There are no significant differences between the two groups with respect to effects on the smoking behaviour, again pointing to the stressfulness of the job loss itself. However, individuals moving directly to a new job do not show any increases in BMI due to the job loss. Only individuals with some unemployment after the job loss significantly gain weight. This might indicate that it is the lack of physical work on the job that is driving the BMI results.

I analysed differences in the effects of job loss for further subgroups.20 There is no indication of a time trend in the effects of job loss: individuals who lost their job in the first half of the observation period (i.e. before the 2006 survey) do not differ from individuals experiencing job loss in the second half (results not shown). However, among baseline non-smokers individuals with a smoking history (i.e. individuals who had ever smoked at least 100 cigarettes in their life) have a substantially and significantly higher chance of (re-)starting smoking—although also individuals who never smoked exhibit a positive (but insignificant) probability of starting smoking due to the loss of employment.

Looking at the effects for all groups of Table 5 together, it emerges that for each subgroup, job loss increases the probability of starting smoking. This effect is significant for all groups, except for individuals over the age of 50 and females. The increase in smoking initiation emerges as the most robust finding affecting almost all groups in society. While there is no evidence of an overall effect of job loss on the probability of continuing smoking, for some subgroups (over age 50, worse health) the probability of continuing smoking increases significantly. However, job loss does not cause a significant increase in smoking intensity among baseline smokers for a single subgroup. For most groups I estimate a job loss-related increase in BMI, which is, however, significant for only some subgroups.

None of the analysed groups experiences a strong increase in BMI due to job loss. However, it might be that job loss has no strong effect on BMI on average (even within the groups), but only increases the chance to experience more extreme weight changes.21 Therefore I look at whether job loss affects more extreme weight changes. Table 6 shows that job loss has no significant effect on whether the individual experiences any positive weight gain or a larger weight loss. Yet job loss raises the chance that the individual's weight increases by more than 1, 2 or 3 BMI units. This indicates that job loss might strongly increase BMI for some individuals, but only marginally on average.

Table 6. The Effect of Job Loss on Body Weight—Extreme Weight Changes
OutcomeMain specification
(4)
  1. a

    Notes

    See Table 3. Row names indicate the outcome. All outcomes are binary variables that take on the value 1 if the respective condition is met. **, *** indicate p < 0.05, p < 0.01, respectively.

BMI change  < −3−0.005 (0.004)
BMI change  < −20.003 (0.006)
BMI change  < −10.011 (0.010)
BMI change  > 00.006 (0.013)
BMI change  > 10.041*** (0.012)
BMI change  > 20.020** (0.009)
BMI change  > 30.016** (0.007)

8 Conclusion

  1. Top of page
  2. Abstract
  3. Introduction
  4. 1 Related Literature
  5. 2 Empirical Strategy
  6. 3 Data and Variables
  7. 4 Matching Quality
  8. 5 Results
  9. 6 Robustness
  10. 7 Heterogeneity Analysis
  11. 8 Conclusion
  12. Acknowledgments
  13. References
  14. Supporting Information

This paper investigates the causal effect of job loss on smoking behaviour and body weight. Using data from the German SOEP, the paper finds that job loss increases the probability of smoking initiation by 3 percentage points. Job loss increases the body mass index (BMI) slightly (by about 0.1 kg/m2 or 0.3 kg on average), but significantly. However, there is little evidence that baseline smokers intensify smoking or are less likely to stop smoking due to job loss. These findings emerge whether only those individuals who lost their jobs due to plant closure are considered, or all individuals experiencing job loss due either to plant closure or dismissal. The estimated causal effects of job loss can partly explain the positive association between unemployment and adverse health behaviours.

Further analyses indicate treatment effect heterogeneity. In particular, singles, younger individuals and individuals with lower health or socioeconomic status prior to job loss exhibit high rates of smoking initiation. The increase in body weight is larger for women and overweight individuals, but still well below 0.5 kg/m2. There is also some indication that job loss raises the chance of experiencing a larger BMI increase. Compared to other effects on smoking and body weight (see, for example, Ruhm 2005; DeCicca and McLeod 2008), the increase in smoking seems considerable, while the average increase in BMI is rather small. In general, the effects on smoking and BMI fall below comparable findings for the USA (Falba et al. 2005; Deb et al. 2011), which might be attributable to the more generous unemployment assistance in Germany.

This study facilitates a better understanding of the implications of job loss, which is not only of particular importance in times of financial and economic crises. While previous studies find detrimental health effects of job loss (Eliason and Storrie 2009; Sullivan and von Wachter 2009), this study emphasizes worsening health behaviours as a potential mechanism for these findings. While other economic studies find that health behaviours improve at aggregated levels when unemployment is high (Ruhm and Black 2002; Ruhm 2005), the present results indicate that simply transferring these findings to the individual level might lead to an ecological fallacy.22 The improvement of health behaviours during economic downturns appears not to be driven by newly unemployed (but possibly by employed individuals with reduced working hours). This micro–macro difference resembles other findings on the link between job loss/unemployment and health: while mortality (Ruhm 2000) and infant health (Dehejia and Lleras-Muney 2004) improve at aggregated levels when the unemployment rate is high, at the individual level job loss increases mortality (Eliason and Storrie 2009; Sullivan and von Wachter 2009) and decreases infant health (Lindo 2011). The findings of this paper also emerge to be important from a public health perspective by highlighting that job loss is a crucial life event with a rather strong impact on smoking, which is one major cause of preventable deaths.

Policies aimed at preventing smoking initiation might be more effective if they consider the vulnerability of specific groups, such as individuals who lost their employment. In general, the findings emphasize that labour market policies to prevent job losses are more beneficial than previously thought. Policymakers should take the worsening of health behaviours into account when comparing the costs and benefits of such labour market policies.

Acknowledgments

  1. Top of page
  2. Abstract
  3. Introduction
  4. 1 Related Literature
  5. 2 Empirical Strategy
  6. 3 Data and Variables
  7. 4 Matching Quality
  8. 5 Results
  9. 6 Robustness
  10. 7 Heterogeneity Analysis
  11. 8 Conclusion
  12. Acknowledgments
  13. References
  14. Supporting Information

Valuable comments by Lex Borghans, Adam Lederer, Mathis Schröder, Thomas Siedler, Martin Simmler, Carla Welch, Nicolas Ziebarth and the two anonymous referees are gratefully acknowledged. I would also like to thank seminar participants at the 24th Annual Conference of the European Association of Labour Economists (EALE), the 15th IZA European Summer School in Labor Economics, the Warsaw International Economic Meeting, the 4th Annual Meeting of the German Association of Health Economists (dggö), and DIW Berlin.

Notes
  1. 1

    However, there are other studies with sophisticated identification strategies that do not find significant adverse health effects of job loss (Browning et al. 2006; Böckerman and Ilmakunnas 2009; Salm 2009; Schmitz 2011). These studies look at morbidity and health status indicators rather than at mortality. Kuhn et al. (2009) find increases in public health costs, resulting from job loss, predominantly for the treatment of mental health problems.

  2. 2

    In a strict sense, body weight is a state rather than a behaviour. But for the purpose of this study, body weight is considered to be a health behaviour, as it actually combines two health behaviours: eating (calories in) and exercising (calories out).

  3. 3

    Morgan and Harding (2006) use a similar notation for introducing propensity score weighting (PSW). PSW differs notationally only with respect to the construction of W. For each control group member j, the diagonal element of W equals inline image. Hence ω(j) is purely a function of the estimated propensity score inline image, and not of the distance to any treatment observation. PSW asymptotically yields the same estimate as the mean difference between treated and matched controls. The matching estimator presented here, however, yields (in the absence of regression adjustment) numerically identical ATT estimates as the mean difference between treated and matched controls because this weighting method is basically a rewriting exercise. PSW is sensitive to misspecification of the propensity score equation because the propensity score enters the weighting function directly (Caliendo and Kopeinig 2008; Stuart 2010). The present estimator does not suffer from this shortcoming as it is just a rewriting of matching estimators and the propensity score does not enter the weighting directly.

  4. 4

    I use the program ‘psmatch2’ (Leuven and Sianesi 2003) in Stata 11.2 to compute ω(j).

  5. 5

    See Smith and Todd (2005), Caliendo and Kopeinig (2008) and Stuart (2010) for excellent descriptions of the various matching procedures.

  6. 6

    Some people argue that bootstrapping might be valid in the case of kernel matching, which does not rely on a fixed number of matches and is thus smoother (Todd 2008).

  7. 7

    This includes both individuals who smoke for the first time and individuals who restart smoking. Section 'Heterogeneity Analysis' differentiates between these two groups.

  8. 8

    The estimated effects are very similar when the sample is restricted to smaller age ranges (e.g. individuals between 25 and 55 or between 30 and 50; see the online appendix).

  9. 9

    These are Browning et al. (2006), Böckerman and Ilmakunnas (2009), Eliason and Storrie (2009), Falba et al. (2005), Salm (2009), Schmitz (2011), and Sullivan and von Wachter (2009). Additionally, I looked at the variables highlighted by Wichert and Wilke (2012) as determinants of job loss in Germany.

  10. 10

    The online appendix provides marginal effects for the conditioning variables in the propensity score equation.

  11. 11

    Matching is also on the square and cube of age as well as on sets of binary variables for federal state, survey year and industry (see Section 'Data and Variables'). Table 2 excludes these variables due to space limitations. Also, for these variables the standardized bias falls below the critical value of 5% after matching. The calculation of the median standardized bias in Tables 3 and 4 considers these variables.

  12. 12

    Variables concerning smoking status, smoking intensity and smoking history as well as the BMI are excluded in this specification.

  13. 13

    The working paper version (Marcus 2012) shows that a placebo regression that does not take into account any information about health behaviours before the (placebo) treatment violates the unconfoundedness assumption. This indicates that the matching DiD estimator clearly outperforms the cross-sectional estimator.

  14. 14

    Also, fixed effects regressions and OLS regressions, where I regress changes in health behaviours on the treatment indicator and the same set of covariates, yield very similar results (see the online appendix).

  15. 15

    For instance, Dorsett (2010) finds unemployment to be related to panel attrition when comparing survey and register data.

  16. 16

    The online appendix shows that OLS regressions also provide similar results for the heterogeneity analysis.

  17. 17

    For reasons of brevity, Table 5 displays the effects on the number of cigarettes not for baseline smokers and non-smokers combined, but only for baseline smokers.

  18. 18

    I distinguish regions with high and low unemployment according to whether the unemployment in the federal state is below or above the annual median.

  19. 19

    Less educated means being in the lowest category of the CASMIN classification (inadequately completed, general elementary school or basic vocational qualification).

  20. 20

    These results are not displayed in Table 5 for the sake of clarity, but they are available on request.

  21. 21

    I would like to thank one of the anonymous referees for pointing this out.

  22. 22

    Although this study basically focuses on job loss and not directly on unemployment, the results in Section 'Heterogeneity Analysis' demonstrate that the effects on health behaviours are similar when analysing the entry into unemployment.

References

  1. Top of page
  2. Abstract
  3. Introduction
  4. 1 Related Literature
  5. 2 Empirical Strategy
  6. 3 Data and Variables
  7. 4 Matching Quality
  8. 5 Results
  9. 6 Robustness
  10. 7 Heterogeneity Analysis
  11. 8 Conclusion
  12. Acknowledgments
  13. References
  14. Supporting Information
  • Abadie, A. and Imbens, G. W. (2006). Large sample properties of matching estimators for average treatment effects. Econometrica, 74(1), 23567.
  • Abadie, A. and Imbens, G. W. (2008). On the failure of the bootstrap for matching estimators. Econometrica, 76(6), 153757.
  • Adam, T. C. and Epel, E. S. (2007). Stress, eating and the reward system. Physiology and Behavior, 91(4), 44958.
  • Akerlof, G. and Kranton, R. (2000). Economics and identity. Quarterly Journal of Economics, 115(3), 71553.
  • Andersen, H. H., Mühlbacher, A., Nübling, M., Schupp, J. and 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), 17182.
  • Bang, H. and Robins, J. (2005). Doubly robust estimation in missing data and causal inference models. Biometrics, 61(4), 96272.
  • Böckerman, P. and Ilmakunnas, P. (2009). Unemployment and self-assessed health: evidence from panel data. Health Economics, 18(2), 16179.
  • Browning, M., Dano, A. M., and Heinesen, E. (2006). Job displacement and stress-related health outcomes. Health Economics, 15(10), 106175.
  • Caliendo, M. and Kopeinig, S. (2008). Some practical guidance for the implementation of propensity score matching. Journal of Economic Surveys, 22(1), 3172.
  • Caliendo, M., Tatsiramos, K. and Uhlendorff, A. (2009). Benefit duration, unemployment duration and job match quality: a regression-discontinuity approach. IZA Discussion Paper no. 4670.
  • Charles, K. K. and Stephens, M. (2004). Job displacement, disability, and divorce. Journal of Labor Economics, 22(2), 489522.
  • Clark, A. E. (2003). Unemployment as a social norm: psychological evidence from panel data. Journal of Labor Economics, 21(2), 32351.
  • Cutler, D. M. and Lleras-Muney, A. (2010). Understanding differences in health behaviors by education. Journal of Health Economics, 29(1), 128.
  • Deb, P., Gallo, W. T., Ayyagari, P., Fletcher, J. M., and Sindelar, J. L. (2011). The effect of job loss on overweight and drinking. Journal of Health Economics, 30(2), 31727.
  • DeCicca, P. and McLeod, L. (2008). Cigarette taxes and older adult smoking: evidence from recent large tax increases. Journal of Health Economics, 27(4), 91829.
  • DeCicca, P., Kenkel, D. and Mathios, A. (2008). Cigarette taxes and the transition from youth to adult smoking: smoking initiation, cessation, and participation. Journal of Health Economics, 27(4), 90417.
  • Dehejia, R. and Lleras-Muney, A. (2004). Booms, busts, and babies' health. Quarterly Journal of Economics, 119(3), 1091130.
  • Del Bono, E., Weber, A. and Winter-Ebmer, R. (2012). Clash of career and family: fertility decisions after job displacement. Journal of the European Economic Association, 10(4), 65983.
  • Dorsett, R. (2010). Adjusting for nonignorable sample attrition using survey substitutes identified by propensity score matching: an empirical investigation using labour market data. Journal of Official Statistics, 26(1), 10525.
  • Eliason, M. and Storrie, D. (2009). Does job loss shorten life? Journal of Human Resources, 44(2), 277302.
  • Falba, T., Teng, H., Sindelar, J. L., and Gallo, W. T. (2005). The effect of involuntary job loss on smoking intensity and relapse. Addiction, 100(9), 13309.
  • Farber, H. S. (2011). Job loss in the Great Recession: historical perspective from the Displaced Workers Survey, 1984–2010. IZA Discussion Paper no. 5696.
  • Heckman, J., Ichimura, H. and Todd, P. E. (1997). Matching as an econometric evaluation estimator: evidence from evaluating a job training programme. Review of Economic Studies, 64(4), 60554.
  • Henkel, D. (2011). Unemployment and substance use: a review of the literature (1990–2010). Current Drug Abuse Reviews, 4(1), 427.
  • Jacobson, L. S., LaLonde, R. J. and Sullivan, D. G. (1993). Earnings losses of displaced workers. American Economic Review, 83, 685709.
  • Kassel, J. D., Stroud, L. R. and Paronis, C. A. (2003). Smoking, stress, and negative affect: correlation, causation, and context across stages of smoking. Psychological Bulletin, 129(2), 270304.
  • Kassenboehmer, S. C. and Haisken-DeNew, J. P. (2009). You're fired! The causal negative effect of entry unemployment on life satisfaction. Economic Journal, 119(536), 44862.
  • Kletzer, L. G. (1998). Job displacement. Journal of Economic Perspectives, 12(1), 11536.
  • Knabe, A. and Rätzel, S. (2011). Scarring or scaring? The psychological impact of past unemployment and future unemployment risk. Economica, 78(310), 28393.
  • Kuhn, A., Lalive, R. and Zweimüller, J. (2009). The public health costs of job loss. Journal of Health Economics, 28(6), 1099115.
  • Lechner, M. (1999). Identification and estimation of causal effects of multiple treatments under the conditional independence assumption. IZA Discussion Paper no. 91.
  • Lechner, M. (2002). Program heterogeneity and propensity score matching: an application to the evaluation of active labor market policies. Review of Economics and Statistics, 84(2), 20520.
  • Leuven, E. and Sianesi, B. (2003). psmatch2. STATA module to perform full Mahalanobis and propensity score matching, common support graphing, and covariate imbalance testing, version 4.0.5; available online at http://ideas.repec.org/c/boc/bocode/s432001.html (accessed 3 March 2014).
  • Lindo, J. (2011). Parental job loss and infant health. Journal of Health Economics, 20(5), 86979.
  • Marcus, J. (2012). Does job loss make you smoke and gain weight? SOEP papers on Multidisciplinary Panel Data Research no. 432.
  • Montgomery, S., Cook, D., Bartley, M. and Wadsworth, M. (1998). Unemployment, cigarette smoking, alcohol consumption and body weight in young British men. European Journal of Public Health, 8(1), 217.
  • Morgan, S. L. and Harding, D. J. (2006). Matching estimators of causal effects: prospects and pitfalls in theory and practice. Sociological Methods and Research, 35(1), 360.
  • Morris, J. K., Cook, D. and Shaper, A. G. (1992). Non-employment and changes in smoking, drinking, and body weight. British Medical Journal, 304, 53641.
  • Reichert, A. and Tauchmann, H. (2011). The causal impact of fear of unemployment on psychological health. Ruhr Economic Papers no. 266.
  • Roelfs, D. J., Shor, E., Davidson, K. W. and Schwartz, J. E. (2011). Losing life and livelihood: a systematic review and meta-analysis of unemployment and all-cause mortality. Social Science and Medicine, 72(6), 84054.
  • Rosenbaum, P. R. and Rubin, D. B. (1985a). Constructing a control group using multivariate matched sampling methods that incorporate the propensity score. American Statistician, 39, 338.
  • Rosenbaum, P. R. and Rubin, D. B. (1985b). The bias due to incomplete matching. Biometrics, 41(1), 10316.
  • Ruhm, C. J. (2000). Are recessions good for your health? Quarterly Journal of Economics, 115(2), 61750.
  • Ruhm, C. J. (2005). Healthy living in hard times. Journal of Health Economics, 24(2), 34163.
  • Ruhm, C. J. and Black, W. E. (2002). Does drinking really decrease in bad times? Journal of Health Economics, 21(4), 65978.
  • Salm, M. (2009). Does job loss cause ill health? Health Economics, 18(9), 107589.
  • Schmitz, H. (2011). Why are the unemployed in worse health? The causal effect of unemployment on health. Labour Economics, 18(1), 718.
  • Schunck, R. and Rogge, B. G. (2010). Unemployment and its association with health-relevant actions: investigating the role of time perspective with German census data. International Journal of Public Health, 55(4), 2718.
  • Smith, J. A. and Todd, P. E. (2005). Does matching overcome LaLonde's critique of nonexperimental estimators? Journal of Econometrics, 125(1–2), 30553.
  • Stephens, M. (2004). Job loss expectations, realizations, and household consumption behavior. Review of Economics and Statistics, 86(1), 25369.
  • Stevens, A. H. (1997). Persistent effects of job displacement: the importance of multiple job losses. Journal of Labor Economics, 15(1), 16588.
  • Stuart, E. A. (2010). Matching methods for causal inference: a review and a look forward. Statistical Science, 25(1), 121.
  • Sullivan, D. G. and von Wachter, T. (2009). Job displacement and mortality: an analysis using administrative data. Quarterly Journal of Economics, 124(3), 1265306.
  • Todd, P. E. (2008). Matching estimators. In S. Durlauf and L. Blume (eds), The New Palgrave Dictionary of Economics, 2nd edn. New York: Palgrave Macmillan.
  • Wagner, G., Frick, J. and Schupp, J. (2007). The German Socio-Economic Panel Study (SOEP)—scope, evolution and enhancements. Schmollers Jahrbuch, 127(1), 13969.
  • Wichert, L. and Wilke, R. A. (2012). Which factors safeguard employment? An analysis with misclassified German register data. Journal of the Royal Statistical Society: Series A (Statistics in Society), 175(1), 13551.
  • World Health Organization (2009). Global Health Risks: Mortality and Burden of Disease Attributable to Selected Major Risks. Geneva: WHO Press.
  • Ziebarth, N. R. and Grabka, M. M. (2009). In vino pecunia? The association between beverage-specific drinking behavior and wages. Journal of Labor Research, 30(3), 21944.

Supporting Information

  1. Top of page
  2. Abstract
  3. Introduction
  4. 1 Related Literature
  5. 2 Empirical Strategy
  6. 3 Data and Variables
  7. 4 Matching Quality
  8. 5 Results
  9. 6 Robustness
  10. 7 Heterogeneity Analysis
  11. 8 Conclusion
  12. Acknowledgments
  13. References
  14. Supporting Information
FilenameFormatSizeDescription
ecca12095-sup-0001-OnlineAppendix.pdfapplication/PDF161K

Table A. The effect of job loss on health behaviors—OLS and fixed effects.

Table B. Determinants of job loss—propensity score equation.

Table C. Treatment effects by subgroups—OLS regressions.

Table D. The effect of job loss on health behaviors—for different age groups.

Table E. The effect of job loss on health behaviors—with and w/o control for alcohol consumption.

Please note: Wiley Blackwell is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.