The Bidirectional Relationship Between Obesity and Labor Market Status – Findings from a German Prospective Panel Study

Unlike many other studies, we applied a short observation period, analyzing the outcomes of obesity or unemployment three years after baseline. With regard to transitions into or out of obesity, we established three BMI cut-off values ​​according to the World Health Organization (WHO) classification [14]† In addition, with regard to transitions into and out of unemployment, we refer to self-reported unemployment versus other employment status. In addition to obesity or labor market status at baseline, we include other potential predictors suggested by the literature, such as subjective health measures and health behaviors, previous unemployment experiences over the life course, obesity of family members and educational levels.

Study population

The data comes from the panel survey “Labour Market and Social Security” (PASS). This is a representative household panel survey of the German residential population crossing households on social assistance benefits (UB-II subsample). The PASS panel was established in 2006 by the Institute for Labor Research (IAB) [15]† The PASS population consists of individuals nested in households.

About 50% of the initial sample included households receiving social assistance benefits, and ~50% of the initial sample did not. Households were only assigned to one of two subpopulations (population subsample or UB-II subsample) [16]† With regard to the turnover of the panel, the sample population was renewed in 2005 and 2011. In addition, the population receiving benefits was refreshed annually to respond to new entrants into benefit households. [15]† Every family member over the age of 15 is addressed annually with an individual questionnaire [16]†

The overall initial response rate in wave 1 (W1; 2007) at the family level was 30.1%, resulting in N= 12,794 households, including 18,972 persons in N= 6804 households (N= 9586 individuals) for the representative population subsample and N= 5990 households (N= 9386 persons) for the sub-sample of benefit claimants (UB-II) [17]† The panel continuity at the individual level was quite high for subsequent waves; for example, in wave 6, the response rate for wave 5 participants was 85.7% [18]†

Study design

To address the two-way analyzes of obesity and unemployment experiences, we only used PASS data from waves 3, 6, and 9, where the PASS survey individual questionnaire included a comprehensive health module. The height/weight and health-related behavior assessments are part of this comprehensive health module. We limited our sample to individuals who participated in at least two consecutive waves of PASS based on the comprehensive health module. Since three measurements over six years create a rather short longitudinal observation window, which severely limits the application of panel estimators and sample size models for less common forms of obesity (BMI ≥ 35 kg/m²).2 and BMI ≥ 40 kg/m2), we applied transition models where we defined the first measurement point of each three-year period as T and the second measurement point (three years later) as T + 3. Although all explanatory variables are observed in T, the corresponding outcome is reported in T + 3. So, and similar to Jusot et al. [10], we used wave 3 (2008/9) and wave six (2012) or wave six (2012) and wave nine (2015) tuples. Since individuals can be part of either tuples, we calculated cluster robust standard errors.

We narrowed the age range to participants aged 15 to 58 at T to exclude age-related transitions to regular retirement at T+3. Our dataset includes 5117 observations of individuals participating in waves 3 and 6 and 6244 observations of individuals participating in waves 6 and 9. In total, this results in 11,361 observations representing 8,226 individuals available for our analysis. A total of 3037, 2080 and 3207 persons participated in both tuples (w3/w6 and w6/w9), respectively, only in tuple w3/w6 and only in tuple w6/w9.

Categorization of obesity groups and labor market status

We calculated the BMI of individuals according to the WHO definition based on self-reported information on height and weight. Both variables were assessed in PASS with a comprehensive health module, applied every third wave/every three years. Based on the WHO classification of obesity, three BMI thresholds were assessed [14]†

  • BMI ≥ 30.0 kg/m2 define two groups, which are below the limit of BMI ≥ 30.0 kg/m2 and people with a BMI ≥ 30.0 kg/m2

  • BMI ≥ 35.0 kg/m2 defining two groups; and

  • BMI ≥ 40.0 kg/m2 define two groups.

Based on these cutoffs, we focused on upward or downward transitions. For this purpose, we estimated the relative risk for BMI transitions of individuals exceeding the respective cut-off values ​​between T and T + 3 (BMI gain or BMI loss above the threshold, e.g. transition to BMI ≥ 30 kg/m22BMI ≥ 35 kg/m2BMI ≥ 40 kg/m2 or vice versa). We decided to apply this cutoff design as opposed to a dimensional BMI model to accommodate the WHO categorization of obesity into three BMI classes. These lessons are often used to assess health risks [19,20,21,22,23,24]†

From a bidirectional perspective, labor market transitions are the second interesting result. The PASS information on labor market transitions is based on self-reported information about a person’s labor market status. Because we are mainly interested in transitions into and out of unemployment, we have collapsed a broader list of labor market activities into the following seven status groups: unemployed (independent of registration status with the German Federal Employment Offices), employed (including self-employed and dependent worker), student ( in secondary, post-secondary and higher education or training), domestic worker (including caring for children or relatives at home), sick leave, early retirement (under age 60) and a residual group of others.


To address the bidirectional relationship between obesity and labor market status, we fitted four sets of exploratory multivariate models by applying logistic regression models:

  • A first set of logit models was used to predict transition above a defined BMI cut-off value based on the three BMI cut-offs for obesity of Class I-III. The “transition to BMI ≥ 30 kg/m2” model included all participants with a BMI below 30 kg/m2 at T; the dependent variable indicated who became obese at T + 3. Analogously, we designed the models “transition to BMI ≥ 35 kg/m2” and “transition to BMI ≥ 40 kg/m2† The three models assess the effect of the independent variables on menopause risk based on the indicated cut-off values ​​for obesity.

  • A second set of logit models was used to predict transition from obesity based on our three obesity cutoffs. The baseline population consisted of individuals with BMI values ​​above the respective thresholds.

  • A third set of stepwise logit models predicts the transition to unemployment. Here, the baseline of the model included all participants employed at T as a reference group. At T + 3 they continued to work or reported being unemployed at T+3.

  • A fourth set of stepwise logit models sought to address transitions from unemployment: at baseline, all participants reporting unemployment at T were included, and the logit model estimated the individual risk of transitioning from unemployment to another labor market status at T+3.


In addition to obesity and labor market status, we included a range of additional risk factors as important explanatory variables based on the literature review. We used the measure of health-related quality of life (HrQoL) as a proxy for perceived health [25]† HrQoL was assessed with the SF-12, a reduced form of the 36-Item Short-Form Health Survey (SF-36) [26, 27]† A standardized scoring algorithm results in a score for the summaries of the physical and mental component [28]† Scores range from 0 to 100 points, with higher scores indicating a better self-assessment of health. A meta-analysis of eight cross-sectional studies revealed lower physical HrQoL scores in adults with a BMI 25 compared to those in normal-weight adults [29]†

In addition, we included smoking behavior and physical exertion as predictors of health behaviour. The smoking behavior was categorized as follows: never smoked, quit smoking and smoking.

Self-rated physical exercise activities (“How often do you participate in active sports, fitness training, or gymnastics?”) were measured on a five-point scale from “every day” to “never”. There is evidence regarding the effects of family members on obesity, especially in the case of younger and female subjects [30]† therefore, we included information on the presence of additional obese subjects in the participant’s household [30]†

As mentioned, the literature reports inconsistent findings on the effect of previous unemployment experiences on obesity, while there are consistent findings on the scars of repeated unemployment experiences. [31]† We have included the number of unemployment years over the life course. Since health and obesity-related behavior is related to social position or socioeconomic class, we controlled for respondents’ educational level as a measure of respondents’ socioeconomic status. Finally, the literature review revealed gender-specific effects on the risks for both obesity and unemployment.

As further control variables, we used past life-course unemployment experiences, obesity of family members, migration background and region (old versus new German Länder). To capture the oversampling of the German UBII household population, we included a subsample dummy (UBII sample versus population sample). The variable “wave” indicates the year of measuring point T. Due to the panel nature of our data, individuals may be included in both tuples. To control for repeated measures of individuals, we applied clustered standard errors [32]† Three levels of significance were defined: p0.05, p≤ 0.01, and p€0.001.

PASS adhered to the ethical standards of the Institute for Labor Research. Ethical approval was not required for this secondary analysis of anonymized data performed with Stata 14 (Statacorp LLC College Station, Texas, USA).

Leave a Reply

Your email address will not be published.