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 7,000 households [25]. A top-up sample of 2,000 responding households, totalling 4000 responding people, was added in 2011 to allow better representation of the Australian population using the same methodology as the original sample [26]. The response rates for the HILDA survey are above 90% for respondents who have continued in the survey and above 70% for new respondents being invited into the study [25]. The main variables examined in this study were available in all annual waves of HILDA (2001 to 2015). More information on the HILDA survey is available in S1 File.

Outcome variable

We considered employment (yes or no) as our primary outcome. A person was considered to have “failed” if they dropped out of work, either because they became unemployed (e.g., where they were still actively looking for work) or were “not in the labour force” (i.e., NILF), where a person was not actively searching for work. As mentioned below, we conducted a sensitivity analysis to examine differences in outcomes depending on whether a person was unemployed or NILF. A key inclusion criterion was that a person had been in employment for at least one wave before they could be considered at-risk of experiencing exit. A person was excluded from the study once exit occurred.

Exposure variable

Disability was determined from the following survey question “…do you have any long-term health condition, impairment or disability that restricts you in your everyday activities, and has lasted or is likely to last, for six months or more?” Specific examples of long-term conditions were shown, such as limited use of fingers or arms, or problems with eyesight that could not be corrected with glasses or contact lenses. These questions were asked at every wave. Data was available from 2001 to 2015. Information on impairments (available from 2003 onwards) suggests that most people reported a physical impairment (6.12% of the observations in HILDA) followed by unspecified impairments (5.88% of the observations in HILDA), sensory (2.51% of the observations in HILDA) and psychosocial problems (1.64% of the observations in HILDA).

We constructed a time-invariant measure of disability that identified if a person reported disability in any waves versus no waves (non-disability). Within the disability group, we conducted a sensitivity analysis to assess possible differences between the proportion of the time a person reported a disability (between 1 to 24% of the time they were in the cohort, 25% to 49% of their time in the cohort, 50% to 74% of their time in the cohort, and 75% to 100% of their time in the cohort).

Predictors of exit

Analytic approach

We constructed a survival analysis dataset with multiple observations per person. To be considered eligible for inclusion in the analysis a person had to be employed for one wave of employment before becoming observed as “at risk” of dropping out of employment. We considered time of origin as the first period a person was defined to be at risk (e.g., they were employed for at least once wave). People were censored if they had missing data on the outcome (e.g., once a person had missing data on employment), or did not observe the outcome at the end of their last contributed wave to HILDA. People “failed” once they reported being unemployed or “not in the labour force”. The study population was restricted to people between the ages of 17 and 64 years.

We estimated and plotted survival curves to assess differences in the outcome, over time, by disability status (two levels (yes/no)). We then used a log-rank test of equality of the survivorship functions to examine whether there were differences in exit from employment by disability.

Semi-parametric Cox regression models were used to examine time-varying predictors of exit from paid work those who did and did not experience any disability. A sensitivity test was performed to assess for possible differences depending on the proportion of time a person reported disability. We controlled for all identified time-varying and time-invariant confounders as discussed above. To determine whether disability was an effect modifier of employment-related factors and exit from employment, we assessed interaction terms of disability and employment variables. We checked if the interactions improved the fit of our model to the data using the likelihood ratio test (comparing a model with multiplicative effects of disability and employment to a model with main effects only).

To deal with the potential selection bias associated with who was left remaining in the sample at follow up (defined as not experiencing exit from employment), we derived weights representing the probability of remaining uncensored up to time t using the approach described in Fewell et al., [37]. We estimated the denominator using a logistic regression approach. The outcome was defined as remaining uncensored at end of study follow up. Specifically, participants whom remained in the sample were coded ‘1’; and censored participants were coded ‘0’. The predictors measured included time-varying and time-invariant covariates. In addition, a variable describing the number of years in the study and a spline variable representing the 5^th, 25^th, 50^th, 75^th, 95^th percentile of number of years in the study were included in the model. Finally, to control for each participant’s censoring history, we calculated the probability of complete censoring history to each year for each individual included in the study. The numerator was estimated in a similar way to the denominator, except that the initial logistic regression only included covariates measured at baseline. The stabilised weights were calculated by dividing the denominator by the numerator, and a mean value was then calculated for each individual included in the study. A further sensitivity analysis was conducted to assess differences based on whether an individual left employment due to NILF versus unemployment. This was calculated across the whole model, as well as for persons with and without disabilities. We also assessed whether removing the SF-36 MCS and PCS influenced the main results regarding disability, employment and exit from employment.

Employment related predictors of exist from employment

These factors predicting exit included: being employed in a low skilled occupation, being employed casually or on a fixed term basis, and being employed in a low psychosocial quality job (noting that this just crossed the bounds of significance for those with disabilities). There is considerable research demonstrating higher unemployment rates among low-skilled and “casual” (a measure of employment precarity) employed workers across OECD countries [38, 39], which may explain the higher rates of employment exit in these groups. The explanations for this are complex, but likely are related to wider economic structural factors, including a poor labour market, overall levels of country debt and the absence of labour market programs [38, 39]. It might be the case that, regardless of the presence of disability, low skilled workers are more likely to be in casual work, and therefore are more likely to lose work as employers respond to wider economic conditions [38]. We also found that workers in jobs with poor psychosocial job quality were more likely to fall out of the labour market. This makes sense, considering these jobs are characterised by high levels of job insecurity (as well as low control, high demands and pay inequity). Our results then suggest that workers perceptions that their jobs are tenuous are actually borne out in reality [40]. These findings highlight that people in certain jobs–particularly low status and lower quality work–are more likely to experience job loss. This is problematic in terms of health inequalities, considering research demonstrating the effects of unemployment on health [3].

The effect of disability on exit from employment

It is worth noting that the presence of a disability influenced exit from employment above health, as measured by the SF-36. Considering this, what else about disability (other than health) could explain the greater exit from the labour market? We would speculate that these factors could include discrimination and bullying, as identified in past research [41]. There is increasing recognition of the mental and physical health impacts of being exposed to workplace bullying [42] (defined as a process in which an employee is subjected to frequent negative acts (e.g. at least once a week) for a relatively long period of time (e.g. 6 months) by peers or superiors, against which defence or retaliation is hindered by the recognition of a formal or informal power imbalance [43]). Research also suggests that greater exposure to workplace bullying is associated with higher sickness absence [44] and levels of job insecurity and intention to leave a job [45]. Hence, exposure to these negative experiences while in the workplace may contribute to exit from the labour market among workers with disability. Aside from explicit bullying, past research has identified lack of support from colleagues and supervisors as being a factor contributing to the wellbeing of workers with disabilities [46]. This support is being particularly necessary in easing workplace accommodations, which may also be associated with a considerable source of stress for workers with disabilities [46]. There is also a strong likelihood that other negative working conditions, such as high job demands and low job control, also contribute to the attrition of people with disabilities from the labour market. Indeed research suggests that people with disabilities are more likely to experience these stressors [8].

Limitations

There has been a considerable amount of research demonstrating complexities in the measurement of disability [19, 47, 48], and we found this was also a challenge in the current study. The majority of our sample reported disability as a fluctuating condition. To take this into account, in addition to looking at individuals who report disability at all or no waves we conducted a detailed sensitivity analysis exploring the effects of altering disability measurement (e.g., between 1% to 24% of their contributed observations, 25% to 49% of their contributed observations, 50% to 74% of their contributed observations, and 75% and 100% of their contributed observations). The findings of these analysis were consistent with the main results reported in the main body of the paper, although lacked precision and were not statistically significant. Other limitations concern issues related to generalizability. A potential explanation for the differences in the study sample and the general population of people with disabilities is that we included only those persons who were in employment at the commencement of the study. Hence, the “healthy worker effect”, which sees the healthiest people being retained in work, may be having an effect here. Other issues include possible unmeasured confounding and dependent misclassification. In saying this, we would note that the outcomes of labour force status may be less prone to misreporting that more subjective factors, such as perceived psychosocial job quality. The strengths of this study are its large size and methodological approach, which allowed us to adjust for survivor bias introduced by censoring.