Unemployment and Subsequent Depression: A Parametric G-Formula Approach Combined with Individual-Specific Fixed-Effect Intercepts

Mikko Myrskylä, London School of Economics and Political Science
Ben Wilson, Stockholm University
Maarten Jacob Bijlsma, Max Planck Institute for Demographic Research
Pekka Martikainen, Centre for Health Equity Studies
Lasse Tarkiainen, University of Helsinki

The effects of unemployment on depression are difficult to establish because of confounding and limited understanding of mechanisms at the population level. One the one hand, due to longitudinal interdependencies between exposures, mediators and outcomes, intermediate confounding is an obstacle for mediation analyses. On the other hand, unmeasured baseline confounding such as selection into unemployment and depression (e.g. due to behavioral problems) can bias even the most advanced statistical method for longitudinal analysis. Using longitudinal Finnish register data on socio-economic characteristics and medication purchases, we extracted individuals who entered the labor market between ages 30 and 55 in the period 1996 to 2013 (n = 285,512). To this data we apply the parametric G-formula to account for longitudinal interdependencies. We contrast two G-formula approaches; one additionally controlling for individual-level fixed effects to account for baseline confounding, and the other not. With the parametric G-formula we estimated the population-averaged effect on person-years spent with antidepressant purchases of a simulated intervention which set all unemployed person-years to employed. In the data, 85% of person-years were employed and 6.7% unemployed, the rest belonging to other status (e.g. early retirement, studying or conscription). In the intervention scenario of the non-fixed effects G-formula, employment rose to 93% and the hazard of first antidepressant purchase decreased by 4%. These effects were especially prominent for less educated men and women. These effects seemed to operate through changes in household formation. However, the G-formula controlling for individual-level fixed effects indicates that these results should be taken with a grain of salt: its estimate shows an effect estimate of close to 0%, indicating no effect of unemployment on antidepressant purchases. This gives an early indication that baseline confounding plays a strong role in the study of unemployment and depression.


Depressionis a leading contributor to the global burden of disease, having a lifetimeprevalence of 10 to 15% worldwide. For individuals suffering from it,depression extends far beyond its direct symptoms, to large increased risks ofsuicide and possibly to cardiac. Unemployment is an important determinant of depression.Unemployment results in economic uncertainty, loss of work-place socialcontacts, time structures and purposeful activity. However, vice versa,depression may also lead to unemployment. Individuals with poor (mental) healthhave more difficulty in both finding and retaining employment, also known ashealthy hire and healthy worker survivor effects, respectively.


Ourdata are an 11% random sample of the population permanently residing in Finlandat the end of any of the years in the period 1996-2013. The data wereconstructed from register data by Statistics Finland (permission TK-53-339-13),and contains individual-level linked information on labor market records,census records and death records, with a further linkage to social carerecords, medication records and sickness absence allowance records maintainedby the Social Insurance Institution of Finland.

Settingand study population

Weextracted individuals who were 30 to 55 years old in the period 1996 to 2013 inFinland. Individuals that purchased an antidepressant prior to entering thelabor market were excluded from the study. We identified 285512 individualsthat met this criterion. Ca. 55% of this sample were male, and 45% female. Dueto the nature of the data source, non-administrative right censoring occurredto ca. 3% of the total n. Intermediate censoring occurred to less than0.5% of total person-years, which could occur if the individual out-migrated,died or was institutionalized.


Theoutcome variable of interest is antidepressant purchase (WHOanatomical-therapeutical-chemical (ATC) code N06A and N06CA; the categories‘Antidepressants’ and ‘Antidepressants in combination with psycholeptics’respectively). This is currently measured as a binary variable, which is 1 whenan individual purchases an antidepressant in a respective year, and 0otherwise.

Exposuresand mediators

Theprimary (time-varying) exposure is employment status. Employment status is acategorical (multinomial) variable which indicates if an individual isemployed, unemployed, student, or other (includes pensioners and conscripts).Employment status was measured once per calendar year. Variables that maymediate the effect of unemployment on antidepressant consumption are income,household status, and health conditions other than depression.


Theissue of longitudinal interdependencies requires a longitudinal approach. TheG-formula is a very flexible statistical approach that can account for issuesof reserve causality while allowing for the modelling of outcomesand mediators of any type, distribution and functional form. Importantly, theG-formula can be used to account for measured time-varying confounders that arealso affected by prior exposures (i.e. intermediate confounders). However, evenwith this method, it is likely that certain forms of heterogeneity remain that predisposeindividuals both to unemployment and antidepressant consumption. Such factors,such as behavioral problems, are likely to affect unemployment and depression positively.When these factors are of a time-constant nature, they could be accounted forby applying an individual-level fixed effects approach. Therefore, weadditionally apply a G-formula with an individual-level fixed effects approachto compare estimates. The G-formula operates by first estimating multivariaterelationships (Figure 1), and then performing a microsimulation in which somevariable has been intervened on (here: anyone predicted to be unemployed isinstead set to be employed).



Thisstudy is ongoing. Results are preliminary. In the no-unemployment scenario,actual employment increased on average by 7 percentage points compared to thenatural course data, rising from 86% in the natural course (close to theempirical employment rate) to 94% in the intervention scenario. An increase inemployment also appeared to lead to slight increases in family formation, with1 percentage point increases in cohabitation, and in marriage, with reductionsin single status. The average effect of reducing unemployment to 0, and havinga corresponding rise in employment, was a 4% reduction in person-years withantidepressant purchases. Overall, the effect appeared to increase over time:at the end of follow-up, the reduction was 10%. These results are populationaveraged. The effect among the individuals intervened on, namely thoseindividuals who were unemployed, is larger, namely a ca. 60% reduction inperson-years spent with at least one yearly antidepressant-purchase. However,when we applied G-formula in which we controlled for baseline confounding usingperson-specific fixed-effect intercepts, we found virtually no change inantidepressant purchasing when individuals who would otherwise have beenunemployed become employed. This gives an early indication that baselineconfounding plays a strong role in the study of unemployment and depression.


Figure 1.Directed Acyclic Graph (DAG) showing the relations between Antidepressantpurchasing (A), Employment (E), Income (I), Household Status (H) and otherHealth Conditions (C). Baseline variables not shown (but are assumed to affectall time-varying covariates at all timepoints t).

Presented in Session 1176: Health, Wellbeing, and Morbidity