Article Text
Abstract
Objectives This study investigated the extent that psychosocial job stressors had lasting effects on a scaled measure of mental health. We applied econometric approaches to a longitudinal cohort to: (1) control for unmeasured individual effects; (2) assess the role of prior (lagged) exposures of job stressors on mental health and (3) the persistence of mental health.
Methods We used a panel study with 13 annual waves and applied fixed-effects, first-difference and fixed-effects Arellano-Bond models. The Short Form 36 (SF-36) Mental Health Component Summary score was the outcome variable and the key exposures included: job control, job demands, job insecurity and fairness of pay.
Results Results from the Arellano-Bond models suggest that greater fairness of pay (β-coefficient 0.34, 95% CI 0.23 to 0.45), job control (β-coefficient 0.15, 95% CI 0.10 to 0.20) and job security (β-coefficient 0.37, 95% CI 0.32 to 0.42) were contemporaneously associated with better mental health. Similar results were found for the fixed-effects and first-difference models. The Arellano-Bond model also showed persistent effects of individual mental health, whereby individuals' previous reports of mental health were related to their reporting in subsequent waves. The estimated long-run impact of job demands on mental health increased after accounting for time-related dynamics, while there were more minimal impacts for the other job stressor variables.
Conclusions Our results showed that the majority of the effects of psychosocial job stressors on a scaled measure of mental health are contemporaneous except for job demands where accounting for the lagged dynamics was important.
- lagged effects
- psychosocial job stressors
Statistics from Altmetric.com
What this paper adds
There has been substantial debate about whether psychosocial job stressors have lasting effects on mental health.
Our study suggests that the relationship between mental health and psychosocial job stressors is mainly contemporaneous.
We found persistent effects of individual mental health, which suggests that an individual's reporting of mental health in one wave was related to their reports in subsequent waves.
These results indicate the importance of assessing the possible effects of prior exposures on outcomes.
Introduction
There is substantial evidence that exposure to psychosocial job stressors such as low job control, high demands, job insecurity and effort-reward imbalance are associated with worse mental health.1 ,2 However, there is a possibility that this relationship may be subject to reverse causation, that is, that poor mental health may predict employment in jobs with poor psychosocial job quality, or that people with worse mental health may be more likely to report job stressors.3 ,4 However, there have now been a number of studies (mostly published in the psychological literature using structural equation modelling) that have suggested that the job stress–mental health relationship is stronger than the mental health–job stress relationship.4–8
Aside from questions about the direction of effects, there has also been some interest in the dynamics of this relationship. Specifically, whether prior (ie, time lagged) exposure to job stressors has continuing long-term impacts on mental health or whether the relationship between job stressors and mental health is mainly contemporaneous (eg, occurring in the same time frame).9 ,10 For example, Amick et al9 used survival analysis to assess the role of previous exposures to psychosocial job stressors on mortality, finding that working in a low-control job was associated with a 43% increase in the chance of death. There is a need to estimate the dynamics of this relationship because not controlling for it is likely to bias the result about the ‘average’ effects of exposure to psychosocial job stressors on mental health.
In addition to dynamic effects, there is a need to control for persistent individual differences between persons (ie, unobserved permanent heterogeneity), such as some individuals being more likely to consistently report better mental health than others. An analytic approach for dealing with unobserved permanent heterogeneity (also called time invariant individual differences) are fixed-effects models, in which each individual acts as their own control, therefore providing estimates that are not confounded by stable individual factors.11 However, on their own, fixed-effects models do not deal with lagged effects, as they pool observations where a person was exposed to psychosocial job stressors (regardless of the timing and sequencing of observations) and compares them to observations when a person was unexposed.
There are two pathways through which lagged job stressors may impact on current mental health. One is the direct effect which can be controlled by simply inserting a lagged exposure variable (eg, psychosocial job stressor) into a fixed-effect model and the other is an indirect effect whereby lagged job stressors impact on lagged mental health and then mental health itself persists to some degree over time. However, introducing a lagged outcome variable (eg, lagged mental health) to capture persistence of the outcome into a fixed-effects analysis results in a correlation between the unobserved time invariant individual differences and the lagged outcome variables, which make standard fixed-effects estimators inconsistent.12 Hence, in order to fully assess the possibility of lagged effects (both direct and indirect), it is necessary to implement other analytical methods to correct these potential biases. In this study, we use a 13-year panel study of working Australians to answer the research question: what are the contemporaneous and lagged effects of psychosocial job stressors (fairness of pay, control, demands, insecurity) on mental health? The aim of this paper is twofold: first, to shed light on the potential contemporaneous and/or lagged effect of job stressors on mental health; and, second, to describe the application of econometric research methodologies to this subject area, with the intention that this will lead to further consideration of these approaches.
Method
Data source
The Household, Income and Labour Dynamics in Australia (HILDA) survey is a longitudinal, nationally representative study of Australian households established in 2001. It collects detailed information annually from over 13 000 individuals within over 7000 households.13 The survey covers a range of dimensions including social, demographic, health and economic conditions using a combination of face-to-face interviews with trained interviewers and a self-completion questionnaire. The initial wave of the survey began with a large national probability sample of Australian households occupying private dwellings.13 Interviews were sought in later waves with all persons in sample households who had attained 15 years of age. Additional persons have been added to the sample as a result of changes in household composition with a top-up sample of 2000 people added to the cohort in 2011 to allow better representation of the Australian population using the same methodology as the original sample.14 The response rates for new respondents who join the HILDA survey are above 70% and the (wave-to-wave) retention rate for respondents who continue in the survey is above 90%.13
Outcome variable
The Mental Component Summary (MCS) of the Short Form 36 (SF-36) measure was used as the primary outcome measure. The MCS represents a summary measure of mental health and well-being and comprises the eight subscales of the SF-36, most heavily weighted on the mental health, role limitations due to emotional problems, vitality and social functioning subscales.15 The SF-36 is a widely used self-completion measure of health status. It has been validated for use in the Australian population, and to detect within-person change over time.15 The SF-36 in the HILDA survey has been shown to be psychometrically sound, with good internal consistency, discriminant validity and high reliability.15 We used the standardised version of the SF-36. The mean score on the MCS in HILDA was ∼49.8, with an SD of 10.3, a minimum of 4.4 and a maximum of 73.9. Scores run from 1 to 100. Higher scores represent better mental health.
Exposures
We examined four psychosocial job stressors: job control (running from 1=low to 21=high), demands and complexity (from 1=low to 21=high), job security (from 1=low to 21=high) and fairness of pay (from 1=low to 7=high). Full details of the construction and validation of the job stressor measures are presented elsewhere.16–18 The individual scales are associated with more widely used measures of job demands and control, and other employment conditions such as casual status, hours worked and shift work.16–18
Confounders
We considered the following variables as independent time-varying causes of psychosocial job stressors and mental health: occupation skill level (low skill, medium skill and high skill), employment arrangement (permanent, casual/labour hire, fixed-term and self-employed) and long-term health condition, disability or impairment. We also conceptualised household structure as a proxy for living arrangement and relationship status. This variable included couple without children, couple with children, lone parent with children, lone person, and other.
Analytic approach
We used a number of analytic approaches to assess the contemporaneous and lagged effects of psychosocial job stressors on mental health. First, we modelled assuming contemporaneous effects only, using a fixed-effects regression model (Equation 1). This shows that , for the ith of N individuals measured at time t, is predicted by time-varying psychosocial job stressors and time-varying covariates . Defining εit as the random error term (representing ‘disturbances’ to the outcome, assumed to be homoskedastic, ie, have a constant variance across time), β0t as the intercept and as the impact of time-invariant covariates,
Fixed-effect model (contemporaneous) 1
Fixed-effects regression models describe the extent to which within-person average differences in the MCS score are associated with within-person average differences in psychosocial job stressors.11 Second, we conducted a first-difference model, in which the changes in mental health from one period to the next are modelled against changes in psychosocial job stressors (Equation 2). As can be seen, the fixed-effect (μi) and the constant (β0) are eliminated with the first-difference model (Equation 3).
First-difference model 2 3
First-difference models represent a difference in scores of an independent variable from year to year and thus model a ‘difference’ or change score. Just like the fixed-effects model, each individual acts as their own control by considering the change from one period to the next and estimates are not confounded by measured or unmeasured time-invariant factors. The difference between the fixed-effects model and first-differenced model is that the fixed-effects model uses all available periods for the individual (the mean) as the reference while the first difference only uses the preceding period as the reference. In the extreme case where we only have two periods worth of data, both will provide equivalent results. If all periods of data within an individual can be considered independent, then using all other years previous within an individual as controls (fixed-effects) is valid and provides more power than only using the preceding period as the control (first-difference model). However, if there are unmeasured time-varying covariates that are highly persistent over time, then using ‘close’ and ‘far away’ periods as controls (the fixed-effects model) may produce greater bias than only using a ‘close’ period as a control (the first-difference model).
Incorporating lagged effects and dynamics
Next, we also considered that there may be a persistent effect of mental health, whereby a person's mental health in the previous years would affect their mental health during the year of measurement. Further, that the effects of psychosocial job stressors on mental health may not only have impacts in the current year but also carry-on effects in the subsequent years (Equation 4). Not controlling for either of these effects is likely to bias our results above. To incorporate this, we estimated a dynamic fixed-effect model with two lags of mental health, current psychosocial job stressors and a 1-year lag of the psychosocial job stressors. Two lags of mental health were included in the models because the Sargan test, which tests whether the overidentifying assumptions in the model appear valid (ref. 19, p. 529), and the Arellano-Bond test for serial correlation in the first-difference residuals suggested a longer persistence than a single year lag.
Inclusion of a lagged outcome variable (mental health) in the model (Equation 1) means that the data within an individual cannot be considered independent and therefore using the data from other time periods as ‘controls’ is problematic. In particular, the lagged outcome variable from the latter time period is related to the outcome variable from the earlier time period. As the distance between data points (eg, years or waves) increases, they are likely to become ‘less’ dependent and therefore the bias reduces. For example, the relationship between responses in 2001 and 2006 is less dependent on one another than the responses between 2001 and 2002. However, this means that both the fixed-effect model and the first-difference model produce biased results on the estimated persistence parameter. Arellano and Bond20 popularised an approach to deal with this problem using the first-difference model (Equation 5) and applying a generalised method of moments (GMM) estimator where earlier lagged values of the explanatory and outcome variables are used as instrumental variables for the lagged change in the outcome variable and (MCSit-2 – MCSit-3). This works under the assumption that these instrumental variables are correlated with the first difference of the lagged outcome variable but uncorrelated with the first differenced outcome variable other than through the lagged difference itself.
Fixed-effect model with a lag 4
Fixed-effect with Arellano-Bond correction 5
We will compare these three approaches to understanding the role of contemporaneous and lagged exposure to psychosocial job stressors on mental health.
Results
The analytic models involved the selection of participants with data available on the main outcomes variable and exposures. There were 27 894 people (182 799 observations) in the HILDA data set over the period 2001 to 2013. Among these people, 20 658 (116 165 observations) were employed and were thus eligible for the study. Within this group, there were 19 529 participants (101 012 observations) with data available on mental health and job stressors. The sample characteristics in their first and last contributed wave for the Arellano-Bond model (the most restrictive model in terms of the sample) can be seen in table 1. This shows the overall change in covariates over the time people were included in the study (eg, changes in household structure, age, etc). The characteristics of the sample in other models can be seen in online supplementary tables S1 and S2.
Table 2 shows the results for the fixed-effects model (1), first-difference model (2) and Arellano-Bond model (3). The number of persons and observations included in each model can be seen at the top of the columns. The average number of waves included in each model was 5.2 (1), 5 (2) and 4.9 (3). The psychosocial job stressors can be seen in the first column of the table: including perceived fairness of pay (1=low to 7=high), job control (1=low to 21=high), job demands (1=low to 21=high) and job security (1=low to 21=high).
supplementary tables
As can be seen, the contemporaneous effects for the Arellano-Bond model (3) are similar to those in models 1 and 2. In the fixed-effect model, greater perceived fairness of pay was associated with an improvement in mental health (0.38 on the MCS 100 point scale for every 1 unit increase). There were similar improvements in mental health associated with this predictor in the other two models, but coefficients were of a lesser magnitude (0.32 and 0.34 for the first-difference and Arellano-Bond models, respectively). The Arellano-Bond model produced a slightly higher effect of job control (0.15) than the other two models. Job security had the greatest impact on MCS, given that it was on a 1–21 scale and had greater variation in the sample. Job security was higher in the Arellano-Bond and fixed-effect models (0.37). Higher job demands were associated with a reduction in mental health. The coefficient was the same across all three models.
Looking at the direct lagged effects (model 3), it is also apparent that most of the relationship between job stressors and mental health appears to be contemporaneous; all of the lagged coefficients are small relative to contemporaneous effects (eg, table 2, fourth row, fairness of pay (t−1)), and none are statistically significant (only security (t−1) shows a trend towards statistical significance). The Arellano-Bond models also demonstrate a lagged persistent effect of prior mental health on current mental health, both from the previous year and from the level reported 2 years previously (t−1 0.12, 95% CI 0.10 to 0.15, p<0.001; t−2 0.03, 95% CI 0.02 to 0.05, p<0.001).
As a means of visually summarising and comparing our results, we then estimated the hypothetical total impact over time of a 1 SD permanent change (eg, a change that is maintained across all HILDA waves) for each of the job stressor measures on mental health, based on the Arellano-Bond model coefficients and the fixed-effect model. Figure 1 shows that most of the impact is contemporaneous for each of the stressors apart from job demands. As can be seen, fairness and demand impacts grow over time with a full effect reached after about 2 years, which is a combined impact of the direct effect of these lagged job stressors and the indirect effect through the persistence of mental health itself. We find that by year 4 accounting for the dynamic relationship has more than doubled the estimated impact of job demands (albeit with a small increase on the MCS). Model estimates with bootstrapped CIs for the Arellano-Bond model can be seen in online supplementary table S3.
Discussion
Our study finds that the majority of the effects of psychosocial working stressors on a scaled measure of mental health are contemporaneous. We also observed a continued persistence effect of mental health, whereby a person's prior reported mental health continued to influence their mental health in the subsequent 2 years.
This study contributes to knowledge by providing evidence about the extent to which the relationship between jobs stressors and mental health is contemporaneous or related to prior exposures. As mentioned in the introduction, there has been some research suggesting that exposure to job stressors may have continuing long-term impacts on health.9 ,21 Similar to our results, Fletcher et al,21 found small effects for lagged job stressors on health over a 5-year period, although this study measured physical and environmental work stressors, rather than psychosocial stressors. Amick et al9 examined similar psychosocial job stressors in a longitudinal cohort; however, this was assessed in relationship to all-cause mortality, while the outcome in our paper was a scaled measure of mental health. However, comparing our results to this previous paper is problematic both because this used a discrete outcome (rather than scaled) and different analytic strategy.
Methodologically, the main contribution of this paper is its application of econometric techniques to a research question thus far mainly studied via psychological or epidemiological analytic techniques. We would argue that econometric approaches have better considered how to effectively measure yearly change in exposures related to yearly change in outcomes (first-difference models), prior exposures on outcomes, unobserved permanent heterogeneity (eg, individual effects), and the potential bias related to inclusion of lagged effects in analytic models (Arellano-Bond models) than traditional epidemiological approaches. In this paper, we have demonstrated the likely problems associated with considering temporal effects and individual heterogeneity, particularly relating to the lagged outcome variable from latter time periods being related to the outcome variable from earlier time periods. A solution to this is to use even earlier lagged values of the explanatory and outcome variables as instrumental variables for the lagged change in the outcome variable (eg, 2-year lags).20
The presented analysis also demonstrates the persistent effects of mental health. Such problems have also been demonstrated in previous cohorts including the British Household Panel Survey, in which health persistent effects were estimated to account for about 30% of previously unexplained variation in health.22 In our analysis, these individual effects were controlled for using fixed-effect models, which, as explained above, use individuals as their own control in analyses, thus removing this source of bias from models.11 We23–28 and others29 have adopted fixed-effects to studying job stressors and mental health outcomes. However, thus far, there has been limited consideration of the time dependent nature of associations in fixed-effects models, with most previous researchers assuming a contemporaneous relationship.
The limitations of this study include the restrictions in terms of eligible sample across the different analytic models. This affects external validity as these requirements restricted the sample of people eligible for the study. To some extent, this also affects interval validity, as there may be differences in who is selected into and out of the models. This is particularly the case for the Arellano-Bond model. However, the similarity of the results between the three models reassures us that this selection is not a large issue. Nonetheless, this is a limitation. Both our outcome and exposure variables are self-reported; thus, there is a possibility for dependent misclassification bias, whereby the outcome may induce errors in the reporting of exposure. However, there is evidence that the association between working conditions and mental health is maintained even when using objective measures of the exposure,30–32 which supports assertions of this paper. In addition to the stressors contributing to the job quality measure used in this study, there are many other important psychosocial aspects of the work environment that were not included in this panel study that could influence our results (eg, social support and bullying at work), suggesting that this study provides a conservative estimate of the influence of workplace psychosocial stressors on mental health. There is also a need to consider factors connected to the HILDA study, such as, it generally has a lower proportion of persons born from overseas and lower socioeconomic groups.14
In conclusion, there is some evidence that previous exposures to job stressors have a continuing effect on mental health mainly through the persistence of mental health itself, although this is small compared with the contemporaneous effect. We suggest that future studies consider applying econometric methods in the analysis of cohort data given their ability to account for classic problems related to confounding and measurement error, albeit potentially at the cost of external validity.
Acknowledgments
This paper uses unit record data from the Household, Income and Labour Dynamics in Australia (HILDA) Survey. The data used in this paper were extracted using the Add-On Package PanelWhiz for Stata. PanelWhiz (http://www.PanelWhiz.eu) was written by Dr John P Haisken-DeNew (john@PanelWhiz.eu).
References
Footnotes
Contributors The article was conceived by AM, ZA and DP. AM, DP and ZA conducted the analysis. All the authors contributed to interpretation of results. AM drafted the manuscript with feedback from all the authors. All the authors contributed to the final draft of the manuscript.
Funding The study is funded by a National Health and Medical Research Council (NHMRC) Partnership grant (APP1055333), including contributions from the Victorian Health Promotion Foundation (VicHealth), WorkSafe Victoria, and Victoria Police. Additional support was also provided by a Victorian Health Promotion Foundation Centre (grant number 15732). DP is supported under the Australian Research Council's Discovery Early Career Awards funding scheme (project DE150100309). The Household, Income and Labour Dynamics in Australia (HILDA) Project was initiated and is funded by the Australian Government Department of Social Services (DSS) and is managed by the Melbourne Institute of Applied Economic and Social Research.
Disclaimer The findings and views reported in this paper, however, are those of the author and should not be attributed to the Australian Research Council, the Australian Government Department of Social Services or the Melbourne Institute of Applied Economic and Social Research.
Competing interests None declared.
Ethics approval The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1075, as revised in 2008.
Provenance and peer review Not commissioned; externally peer reviewed.
Data sharing statement Data are available on request from the Australian Government Department of Social Services.