Study protocol for examining job strain as a risk factor for severe unipolar depression in an individual participant meta-analysis of 14 European cohorts

Introduction

Unipolar depression is prevalent and incurs substantial costs for the individuals affected and society at large^1,2. The disorder is thought to develop in a complex interplay of biological, psychological and social factors^3–5. Following the diathesis-stress framework⁶, etiological determinants throughout the life-course may affect vulnerability to depression or act as triggering factors.

According to the job strain model⁷, a psychosocial work environment characterized by high psychological demands and low control may result in stress-reactions and lead to adverse health outcomes. Accordingly, job strain has been linked to several health conditions, including coronary heart disease⁸ and unipolar depression^9,10. There are, however, indications of publication bias in the field, suggesting that the published literature may be biased towards studies showing stronger associations between job strain and depression¹⁰. In addition, many previous studies on job strain and depression have applied outcome measures with uncertain diagnostic validity, for example self-reported symptoms⁹. Hence, the applicability of these findings to clinically diagnosed depression is uncertain⁹. In this study protocol, we set out our plan to obtain data from 14 European cohort studies on work environment and health, and to examine the association between job strain and subsequent first hospitalisation due to a diagnosis of unipolar depression. The purpose of this planned project is to examine whether stressful working conditions characterized by high psychological demands and low control, i.e. job strain⁷, are a risk factor for the development of unipolar depressive disorder. We hypothesize that individuals experiencing job strain are more likely to become hospitalized with a diagnosis of unipolar depression than individuals without job strain. Furthermore we aim to explore whether the association between job strain and depression is similar across strata of sex, age and socioeconomic status (SES).

Data

The IPD-Work Consortium combines data from a number of European work environment studies. For the analyses on job strain and depression we include data from 14 studies that are linked with hospital admission registers including psychiatric admissions. These studies encompass a total of approximately 120,000 individuals. Table 1 gives an overview of the included studies.

In each study, data on psychiatric hospital treatment are available through national registers^11–14. Most registers include both inpatient and outpatient treatments. Outpatient data are available since 1995 in the Danish register, since 2001 in the Swedish register, since 1998 in the Finnish register and since 2003 in the UK register. To maximize the number of cases, these data are included when available.

Study population & design

The study is designed as a prospective cohort study. Participants will be included if they are employed at baseline of the respective study and have provided data on job strain, sex, age, cohabitation and socioeconomic status (SES). To ensure a prospective design we exclude all individuals with a hospital-based diagnosis of unipolar depression before or at baseline. Data are analysed using a two-step individual participant data meta-analysis; i.e. we first obtain study-specific risk estimates using harmonised exposure and outcome data and then combine these estimates using meta analytic techniques¹⁵. The study-specific risk estimates are calculated using Cox regressions analysis and the pooled estimates by random effects meta-analysis (see section “main analyses” for details).

Main analyses

All study-specific analyses will be conducted using Cox proportional hazards regression with the occurrence of the first hospital record of unipolar depression as the failure-date, and censoring for migration (where available), death and end of follow up. All summary risk estimates will be calculated by pooling study-specific risk estimates and standard errors using inverse variance weighted random effects meta-analysis. Pooling will be conducted in R (www.rproject.org) using the meta package²⁵ and the degree of heterogeneity between the study-specific estimates will be assessed by I²²⁶. Table 3 gives a ghost table for the main results.

Statistical power calculations

The expected numbers of cases of hospital-treated unipolar depression in each of the cohorts are presented in Table 4. The estimates are based on observations in the Danish register data concerning the incidence of hospital treatment for unipolar depression in gainfully employed Danes. These numbers were applied to the studies to estimate the expected number of cases, and the observed numbers of cases in the databases may diverge from this estimation. If there are no observed cases amongst individuals exposed to job strain in a study, this study will not be included in the analyses, as a risk estimate cannot be obtained.

The expected statistical power as a function of the hazard ratio is shown in Figure 1. The planned analysis is expected to be powered to show an association of 1.23 with >90% power. This is under the assumption that the actual number of cases will match the expected number of cases in Table 4 and that all studies will provide cases and thus are included in the analysis.

Figure 1. Expected statistical power as a function of the hazard ratio.

Sensitivity analyses

The following section specifies the planned sensitivity analyses which will examine the robustness of the results. All statistical tests for the sensitivity analyses will be two-sided with a significance threshold of P<0.05. We will conduct a series of sensitivity analyses which may lead to concerns of mass significance due to multiple testing. To avoid inflating the type I error due to multiple testing, we will consider the sensitivity analyses nested within the main hypothesis test. Thus, their interpretation will depend on the results of the main analysis: if our main hypothesis is confirmed, we will consider the significance tests of the nested hypotheses valid and the tests which yield p-values <0.05 statistically significant. The sensitivity analyses may in this case be considered as an examination of the conditions under which the average population effect found in the main analysis holds. However, if the main hypothesis is not confirmed, we will not consider the tests of the sensitivity analyses (nested hypotheses) with p<0.05 confirmatory, i.e. the results of the sensitivity analyses will be considered exploratory and hypothesis generating. This strategy will retain the overall probability of a type I error under 0.05, whenever the main null hypothesis is true.

Our first set of sensitivity analyses examines whether the association between job strain and depression is modified by sex, age (≤35, 36–49, 50+ years) or SES (low, intermediate, high⁸) following indications from previous studies^28–30. If there are too few cases to obtain estimates for 3 categories of age and SES in more than half of the eligible studies, we will use the categories ≤49/50+ years and low SES/other instead. Following the STROBE recommendations³¹ we will present results on effect-modification so that both departure from additivity and multiplicativity may be assessed³². However, our conclusions on effect-modification will be based on departure from additivity, as such information is particularly important from clinical and public health perspectives^31,33,34. The statistical test will be based on the Central Limit Theorem, and Gauss propagation of error formulas.

If any statistically significant associations are found in the main analyses, we will conduct a second set of sensitivity analyses to examine how these associations are affected by accounting for SES, self-reported baseline mental health, and chronic physical disorders. Regarding mental health, we will a) adjust for mental health (continuous score, see Table 5 for measures) and b) exclude individuals with poor mental health (the poorest quintile, based on the measures presented in Table 5). Regarding chronic physical disorders we will exclude individuals with baseline coronary heart disease, stroke, cancer, chronic obstructive pulmonary disease, musculoskeletal disorders or diabetes, where data are available, as chronic physical disorders are associated with increased risk of mental disorder³⁵. In a third and final sensitivity analysis we will examine how unipolar depression is related to the separate dimensions of demands and control (standardized continuous scores and the job strain model quadrants, i.e. comparing the three other combinations of demands and control to individuals with low demands and high control (cf.⁸).

Discussion

A major strength of the planned analyses is the register-based outcome with a specific diagnosis. The diagnoses are based on clinical assessments, which are independent of job strain assessments, and have high validity⁴². The use of such outcomes has not been possible in most previous analyses on work environment and depression^9,10, as the relatively low incidence of hospital-treated unipolar depression necessitates an exceptionally large dataset. Some exclusively register-based analyses (e.g.⁴³) have been conducted previously using hospital discharge registers. Such analyses have, however, not examined work environment variables measured at the individual level but exposures approximated using job exposure matrices. Although such approximation is useful when exposure measures are unavailable, the lack of individual level measurement hampers the interpretation of the findings as they are open to the ecological fallacy⁴⁴.

Measuring depression incidence exclusively through hospital treatment registers also has limitations. It is likely that only the most severe cases of depression are treated in a hospital. Consequently, the results of the planned analyses will not be extendable to mild and moderate depression, and severe depression not treated in a hospital, if the aetiology of depression varies with severity and treatment. Furthermore, the exclusion of individuals with depression before or at baseline, to ensure a prospective study design, is also based on hospital-treated depressive episodes. Hence, it is possible that participants are suffering from, or have suffered from, untreated depression before baseline. However, in this consortium study, the only way to obtain accurate retrospective diagnoses is based on hospital treatment. Although data on treatment with antidepressants for example may also be obtained from registers, these medications are used to treat a range of conditions, other than depression, such as anxiety, pain and incontinence⁴⁵, and consequently do not provide any accurate diagnostic information.

Another potential limitation of the planned analyses is the self-reported exposure data which could be sensitive to reporting biases due to negative affect at baseline; a bias which may result in inflated risk estimates, if individuals with reduced mental wellbeing report their exposure more negatively and are at increased risk of developing depression⁴⁶. We address this limitation via the sensitivity analyses by adjusting for baseline mental health. Although this analysis may be overadjusted, at least if negative affect at baseline is a consequence of the working environment, any remaining association between job strain and depression will provide a strong argument that the results are not explained by reporting bias.

The planned analysis uses data from 14 studies conducted in 4 European countries. The studies differ in design, timing, and study population. Whereas some studies (e.g. DWECS, SLOSH) include the general working population, others are restricted to employees of specific organisations or occupations (e.g. FPS, Whitehall). Although this combination of different study populations means great gains in statistical power, the generalizability of the findings must be considered. Ideally, we may generalize our findings to the working population of (Northern) Europe. Such generalization would be supported by low degrees of heterogeneity in the pooled estimates, as we found for the association between job strain and coronary heart disease⁸. If there are greater levels of heterogeneity in the findings, however, the generalizability of the associations outside the examined cohorts is less clear. In that case we may need additional post hoc sensitivity analyses to examine the reasons for heterogeneity, for example the length of follow up, the period with hospital data available before baseline, and whether or not outpatient hospitalisation data are included throughout the follow up period.

All participating studies have been approved separately by the relevant national ethical committees (see appendices of^8,17 for details). The results of this planned study will be published in an article in a scientific peer-reviewed journal. This planned study will constitute the largest in the field to date and as such is likely to set the parameters of research in this field for some time to come.

Project organization

The project is organized as part of the IPD-Work (“Individual-participant-data meta-analysis of working populations”) Consortium⁴⁷. IPD-Work was funded by the NEW OSH ERA⁴⁸, and is led by Professors Mika Kivimäki (Finnish Institute of Occupational Health and University College London, UK), Töres Theorell (Stress Research Institute, University of Stockholm, Sweden), Reiner Rugulies (National Research Centre for the Working Environment, Denmark) and Nico Dragano (Department of Medical Sociology, University of Düsseldorf, Germany). The IPD-Work Consortium is coordinated by Prof. Mika Kivimäki. The principal investigator in the analyses described in the present protocol is Dr. Ida E. H. Madsen, researcher at the National Research Centre for the Working Environment, Denmark.