Data

We used 12 waves of the Household, Income and Labour Dynamics in Australia (HILDA) survey data. HILDA is a nationally representative household-based panel study that began in 2001. Each wave has more than 7,000 households with more than 17,000 household members. Response rates have been consistently high (>90%; Summerfield, Reference Summerfield2011). The survey asks respondents about their employment, family circumstances, health and socio-economic characteristics through face-to-face interviews and self-completed questionnaires. We used data from Waves 5–16 as they contained measures of work conditions important to our modelling; some of the variables were not collected before Wave 5.

We limited the sample to employed adults aged 50–70, defined as older adults who worked in any wave (as long as they worked, they were included in our analysis sample) and who had data on mental health, vitality, wages, work hours and covariates. The maximum number of observations is 29,643 spanning 12 years (2005–2016). Due to missing data, attrition and recruitment of new respondents, the number of observations varies across waves, creating an unbalanced panel. As we used partner's employment and education as either control variables or instruments in models, the sample is further restricted to older workers with at least one partner employed. We tested our results by extending the sample to cover coupled and non-coupled observations in a sensitivity analysis.

Estimation method

Empirical estimation model

The 3SLS bootstrapping estimation as applied to a three-equation simultaneous model is as follows:

(1)$$H_{i, t} = \alpha _0 + \alpha _t + \alpha _1H_{i, t \hbox{-} 1} + \alpha _2\ln W_{i, t} + \alpha _3T_{i, t} + \alpha _4T_{i, t}^2 + \mathop \sum \limits_{\,j = 5}^n a_jXh_{\,j, i, t} + e_{i, t}$$

(2)$$T_{i, t} = \beta _0 + \beta _t + \beta _1T_{i, t \hbox{-} 1} + \beta _2\ln W_{i, t} + \beta _3H_{i, t} + \mathop \sum \limits_{\,j = 4}^k \beta _jXt_{\,j, i, t} + u_{i, t}$$

(3)$$\ln W_{i, t} = \gamma _0 + \gamma _t + \gamma _1H_{i, t} + \gamma _2T_{i, t} + \mathop \sum \limits_{\,j = 3}^m \gamma _jXw_{\,j, i, t} + v_{i, t}. $$

In this simultaneous equation system, the dependent variables are health (H_{i ,t}: mental health, vitality), weekly work hours (T_{i ,t}) and log of hourly wage rate (lnW_{i ,t}) of individual i in year t. We also controlled for a set of covariates (X_h, X_t, X_w) in the corresponding equations. The time-specific effects on health, work hours and wages were captured in α_t, β_t, γ_t. The lagged dependent variable is also entered as a predictor to adjust for large variation between observations due to its historical trend and auto-correlation (Arellano and Honore, Reference Arellano, Honore, James and Edward2001). The error terms (e, u, v) capture measurement errors and unobserved factors. Details of variables used in each equation are described below.

In health Equation 1, apart from work hours (T) and wage rate in logarithm (lnW), we added work hours squared (T ²) to test for non-linear relationships between work hours and health. We also included potential work experience, ethnicity, marital status, work flexibility, work intensity, employment type, occupation, gender, smoking, drinking, and lag of physical activity and lag of general health (all likely to be predictors of health). We added urbanity, state and year dummies to capture the differences in health due to geographic and time factors. In this paper, we focused on mental health and vitality rather than general or physical health because we used weekly work hours. The effects of work hours on mental health and vitality are likely to be observed in a relatively short period of time, whereas any impacts on general health and physical functioning may take several years before they become observable.

The potential endogenous variables of Equation 1 were weekly work hours and wage rate. Instruments for these variables should be correlated with work hours and wage rate but not correlated with the error term of this equation (i.e. the excluded instruments have no direct effect on health outcome). The potential instruments were partner's employment status, partner's education and household non-wage income. We assume that these variables were likely to directly influence an individual's employment, work hours and wages, but only indirectly affect their health. We acknowledge that these instruments may violate the exclusion restriction, Cov(z, error terms) = 0. Under the 3SLS approach, we are unable to directly test this possibility, however, we provide first-stage F-statistics and instrumental variable (IV) test results for the excluded instruments we used (see Table S1 in the online supplementary material). These tests are equivalent to the weak instrument and relevant IV tests (for relevance conditions) – Cragg-Donald Wald F-statistics in 2SLS/IV models (Finlay et al., Reference Finlay, Magnusson and Schaffer2013). The very high F-values (ranging from 92.6 to 604) for the first-stage estimates (significantly greater than 10, a commonly used threshold for weak IV and relevance restriction test) and the test results for excluded instruments being jointly equal to zero in the first stage (all statistically significant at the 1% level) suggest that our instruments satisfy the relevance condition.

In work-hour Equation 2, we included both wage rate (log) and health on the right-hand side of the equation. We also controlled for marital status, gender, prior general health, occupation, location, state and year dummies. The potential instruments for the work-hour equation must have no direct effect on work hours, but be correlated with wages and health. We therefore used lag of log wage, lag of health outcome and socio-economic status such as financial distress as instruments.

In wage Equation 3, we followed the Mincerian earnings equation (Mincer, Reference Mincer1974) to build the wage model. This equation included education and potential experience (its squared term was not added as our sample includes older people (aged 50–70), the non-linear relationship between age or experience with income no longer exists). Gender, ethnicity, location, state and year dummies were also included to capture wage differences across groups, states and years. The prior health (or lagged health) variable was included to capture the influence of workers' health on their productivity and wages. The health variable also represents job-related hazards which may sometimes attract additional pay as a result (Haveman et al., Reference Haveman, Wolfe, Kreider and Stone1994). Work hours were also added in the wage equation as a predictor for wage rate, via wage premiums, overtime and penalty rates. Longer work hours lead to longer (equivalent) work experience that can alter wage rate. Commonly used instruments for the wage equation are those that predict education, but do not directly affect wages. These include father's education, mother's education, sibling's education, partner's education and parents' socio-economic status. In our case, partner's education level was a stronger predictor of other partner's education than father's education and mother's education, and we selected it as our excluded instrument.

Variables and measures

Covariates and variables

The covariates, variables and their definitions are presented in Table 1.

• • Stratification and interaction variables.

• In models for occupation (a ‘white’ dummy variable: white- versus pink/blue-collar workers), the white-collar group includes managers, professionals, clerical and administrative workers, and the pink/blue-collar group includes the remaining occupations. The white-collar jobs are high-status/well-paid jobs, while pink/blue-collar jobs are lower-status jobs which are often physically demanding (Schreuder et al., Reference Schreuder, Roelen, Koopmans and Groothoff2008). The ‘white’ dummy variable was interacted with work hours and its squared term to capture the shift of the coefficients between the occupations. We created four categories for age comparisons of the estimates: younger workers (aged 25–49), older workers (aged 50–70) and then two subgroups of older workers (50–59 and 60–70 years) (Table 2).

Table 1.

Covariate and variable definition and coding

Notes: OECD: Organisation for Economic Co-operation and Development. ANZSCO: Australian and New Zealand Standard Classification of Occupations.

We also model for health status using physical function distribution cut-offs at the 20th percentile (pc), 30th pc and median to see if the mental health and vitality tipping point varies by workers' physical health status. Physical health impacts of work hours are likely to be chronic, cumulative and longer term and so require different models to 3SLS. However, we expected that mental health and vitality tipping points would be at least partly a function of physical health status, and it is important to estimate this because of the ‘healthy worker’ bias implicit in older employed samples. For example, Medic et al. (Reference Medic, Wille and Hemels2017) showed that sleep disruption causes short-term effects on mental health-related health including increased stress responsivity, somatic pain, reduced quality of life, emotional distress and mood disorders, and cognitive, memory and performance deficits. However, it takes a much longer time to affect physical health outcomes. The long-term consequences of sleep disruption include dyslipidemia, cardiovascular disease and weight problems.

In our models, we were aware of potential reverse causality between general health and our health outcome variables (mental health and vitality). As such, we controlled for prior general health to avoid such possible reverse causality biases. We use physical functioning to stratify for health status.

Descriptive statistics

Table 2 displays differences in some key variables across age, occupation and health status groups. In relation to younger workers (25–49), older workers (50–70) have slightly better mental health and vitality, but worse general health, they worked about 2 hours less each week, and spent 1.4 hour less on unpaid work. They were better off in terms of non-wage income and financial distress, and they smoked and drank less.

Table 2.

Sample characteristics by age, occupation and health strata (employed people sample)

Notes: Means were adjusted for sample weights. Ethnicity, state and year dummies were used in the models but are not shown. <30th pc PF: a group whose general health is below the 30th percentile of physical functioning distribution for employed people aged 50–70. >50th pc PF: a group whose physical functioning is on or above the 50th percentile of physical functioning distribution.

White-collar workers worked longer hours than their pink- and blue-collar counterparts and received a significantly higher weekly wage (Aus $1,276 versus $917). They had higher non-wage income, were more educated and faced less financial distress than pink- or blue-collar counterparts. White-collar workers also enjoyed more favourable working conditions including flexible work, and job security (fixed-term and permanent jobs) and job control, and they smoked and drank less than their pink/blue-collar counterparts. Workers with better health (physical functioning) were more advantaged than the poorer health status group in almost all aspects, even though they had quite similar paid hours and work experience.

Work-hour heath tipping points by age

In Table 3, only the summarised estimates for work hours and its squared term were reported to remain succinct (for full set of estimates, see Table S2 in the online supplementary material). A non-linear relationship between work hours and health (mental health and vitality) was found. The estimates of work hours and its squared terms are highly statistically significant (p < 0.001). An inverted U-shaped relationship was observed due to the positive coefficient of work hours and negative coefficient of its squared terms (the tipping point is where the first derivative of the health function equals zero, solution to this problem is the value of work hours or tipping point). Accordingly, the tipping point for mental health was 39.4 and 39.8 hours for the groups aged 25–49 and 50–70, respectively (Table 3; Figure 1). The vitality tipping point was 38.8 and 39 hours for the groups aged 25–49 and 50–70, respectively (Table 3; Figure 1, bottom panel). The thresholds for both groups were close to the overall workforce (25–64 years) estimations (see Dinh et al., Reference Dinh, Strazdins and Welsh2017). Overall, the estimated tipping points for both age groups were almost identical with no significant difference for both mental health and vitality. Table S1 in the online supplementary material provides the first-stage estimates and IV test results.

Figure 1.

Work hour–health tipping points by age.

Table 3.

Summary of work hour–health tipping points, by age

Notes: Bootstrapped standard errors (SE) in parentheses with 1,000 replications. Apart from the model specification in Equations 1–3 and variables mentioned in the ‘Variables and measures’ section, the health equation was further adjusted for work flexibility, work intensity, ethnicity, marital status, employment contract type, work experience, gender, unpaid time, occupation, lag of general health status, smoking and drinking status, lag of physical activity, and state, urbanity and year dummies. The work-hour equation was controlled further for marital status, gender, lag of general health, occupation, and state, urbanity and year dummies. The wage equation was controlled further for education level, work experience, ethnicity, gender, and state, urbanity and year dummies. Tables 4, 5 and 8 also have the same set of covariates.

Significance levels: * p < 0.01, ** p < 0.05, *** p < 0.01.

Sub-age group analysis was undertaken for the 50–59 and 60–70 age groups. All estimates of work hours and its squared terms were statistically significant except for squared work hours for the 60–70 age group. Significance declined by age group due to reduced sample size. The tipping point for the 60–70 age group was slightly lower (3–4 hours per week) than for the 50–59 age group. This was somewhat less than we anticipated as we expected deteriorating physical health amongst those age 60–70 would significantly reduce their work hour–health limit.

The predicted health curves for the older subgroups were above younger group curves for both mental health and vitality, suggesting that they have better mental health and vitality compared to the youngest group (25–49 years) across all work hours (Figure 1).

Work hour–health tipping points by occupation

Table 4 and Figure 2 display the work hour–health threshold by occupation (white-collar versus pink/blue-collar workers) (for full set of estimates, see Table S3 in the online supplementary material). The work hour–health tipping point for white-collar workers aged 50–70 was 44 and 46 hours for vitality and mental health, respectively, compared to 37 and 37.3 hours for pink/blue-collar workers. The large difference (around 8 hours) in tipping points between white- and blue/pink-collar occupations was considerably greater than differences between age subgroups.

Figure 2.

Work hour–health tipping points by occupation.

Table 4.

Summary of work hour–health tipping points among workers aged 50–70, by occupation

Note: See Notes of Table 3.

Significance level: *** p < 0.01.

More interestingly, the curvilinear relationship is stronger for pink/blue-collar workers than white-collar workers, suggesting that longer work hours have a greater effect (in both positive and negative directions) on their mental health and vitality (Figure 2).

Work hour–health tipping points by health status

Work hour–health tipping points for health subgroups – relatively poor health (under the 20th pc, under the 30th pc) and better health (over the 50th pc or above the median health) – for physical functioning are displayed in Table 5 and Figure 3 (for full set of estimates, see Table S4 in the online supplementary material). Those in the groups of <20th pc and <30th pc health distribution (physical functioning) were considered to have relatively poor health, with those in the >50th pc group health were considered to have relatively good health or ‘better health’, compared to the 50–70 group average. Our estimates show a large difference in tipping points between those with relatively poor health and those with better health. The gap was 7 hours for mental health and 11 hours for vitality when stratified by physical functioning (Table 5). Further, workers with relatively poor health (physical functioning) had significantly lower predicted mental health and vitality scores (lower curves) (Figure 3).

Figure 3.

Work hour–health tipping points by physical functioning status.

Note: pc: percentile.

Table 5.

Estimated tipping points for mental health and vitality by physical functioning status (ages 50–70)

Notes: See Notes of Table 3. pc: percentile.

Significance levels: ** p < 0.05, *** p < 0.01.

Health selection

Our sample likely represents a particularly healthy group due to health selection (which increases as people age) and is not representative of the population aged 50–70. Consequently, our work-hour–health limit estimates may be inflated and not generalisable to older adults who have already exited the labour market. Therefore, we performed some exploratory analysis, comparing our (relatively healthy) employed sample to counterparts not in the labour market after they turned 50.

Simple estimates in Table 6 show clear health gaps between our sample and non-working counterparts, with employed people aged 50–70 reporting better health for all health measures. Furthermore, the standard deviations for the non-working group were larger, indicative of health variation within this group.

Table 6.

Health differences between working and non-working, aged 50–70

Source: Authors’ estimation from the Household, Income and Labour Dynamics in Australia (HILDA) survey, 2005–2016, adjusted for sample weights.

Using a longitudinal logistic model, all prior health measures strongly affected the probability of labour market participation for those aged 50–70, thereby supporting the occurrence of health selection (Table 7).

Table 7.

Health conditions and working status, longitudinal (Random Effect) logistic model, aged 50–70

Notes: Standard errors in parentheses. Models also controlled for lags of net wealth, financial distress, non-wage income, age, unpaid time, marital status, gender, education level, ethnicity, state and year dummies.

Significance level: *** p < 0.01.

Sensitivity analyses

To test the robustness of our estimates, some model specifications were altered. First, we excluded work hours from the wage equation (Equation 3), to be in line with the labour supply and wage equation in labour economics literature, resulting in almost no change to tipping points. Second, as contemporary weekly work hours may not capture the regularity of working long (or short) hours, we used ‘percentage of time spent in jobs in last 12 months’ to approximate actual annual work hours, which yielded similar results (available upon request). Third, the use of partner's employment and individual characteristics, such as education, as either covariates or instruments in the main analysis restricted the sample to coupled people. Thus, to improve generalisability, partner's variables were excluded, and our sample expanded to include both coupled and non-coupled people aged 50–70. We used lags of log wage, non-wage income, financial distress index, lags of health outcomes and lags of weekly work hours as instruments in the first stage of estimation. Tipping points remained relatively unchanged (Table 8 versus Table 2).

Table 8.

Summary of estimated tipping points for mental health and vitality using the whole working sample

Note: See Notes of Table 3.

Significance level: *** p < 0.01.

Fourth, one may argue that individual mental health/vitality may also affect choice of occupation and job characteristics (flexibility, intensity). That is, these job characteristics are not exogenous to mental health and vitality, particularly in the long term. For example, a person may move from their main occupation/job in the long term due to economic structure change or a significant change to their health status, but may not in the short term, which is our current paper's focus. To consolidate our findings, we have tried models with lags of the occupation and job characteristics to make sure they are exogenous in the health equation; the estimated tipping point did not change much (a small change in tipping point may be due to a shrunk sample size). Financial distress affects mental health in the very short term, and may also affect work hours, likely over a longer timeframe as people need time to adjust their work hours. We also tried models with lags of financial distress as an instrument in the work-hour equation, and the results did not change much (results available upon request). Finally, we explored Tobit models and propensity score matching to estimate tipping points on the non-working 50–70 years group. These models required extensive imputation of all employment-related data, problematic for both approaches. Under the matching or propensity score matching approach, however, we were able to estimate tipping points that covered the whole sample. These were between 1 and 2 hours per week lower, likely due to the poorer health status of the non-working group. These results supply suggestive evidence that our estimates, because they are based on a selected, healthy 50–70 years sample, are over-estimating the population work hour–health tipping points. They confirm the stratified models by health status which showed that among the less than optimally healthy older adults, tipping points are considerably lower (see Table 5; Figure 3).

Findings summary and discussion

Using Australian population-representative data to estimate work hour–health thresholds for older workers (aged 50–70), we found that their estimated tipping points were almost identical to those of younger workers (aged 25–49) for both mental health and vitality. We find, on average, working a full-time week of 39–40 hours does not compromise older workers' mental health or vitality. Thus, many older-age adults can work as long as younger adults. Our study supplies further evidence that an ageing workforce has potential to be equally productive, and employment needs to be re-engineered to enable this (Riley and Riley, Reference Riley and Riley1994). However, there are important caveats and contexts to these findings; we further show that the work hour and health relationship depends heavily on the type of job, and on the physical health status of an older worker.

Our estimates show that among the optimally healthy, a full-time working week appears to be both feasible and health supporting, and good health enables older adults to maintain similar levels of engagement as to when they were younger. For these older adults, extended employment can be a potential pathway for financial security as well as better engagement socially, and underscores the significance of labour markets as key institution for successful ageing (Zaidi et al., Reference Zaidi, Gasior, Hofmarcher, Lelkes, Marin, Rodrigues, Schmidt, Vanhuysse and Zolyomi2013; Rowe and Kahn, Reference Rowe and Kahn2015). However, optimal health is found among only a proportion of all older adults and is closely related to occupational history, status and earnings. Among the less-healthy, lower-status older workforce, large gaps in work-hour capabilities were observed. When we stratified by physical functioning we found that, for these older adults, their working week was shortened by between 7 and 11 hours and they were also earning, hour for hour, lower wages. Less-healthy workers therefore face a dilemma as they age: working longer hours to increase financial security confronts them with further health trade-offs. For the growing group of less than optimally healthy older adults, employment offers a mixed bag of supports for successful ageing.

Similarly, occupation had a significant influence on work hour–health thresholds, due to variation in work environments and health demands that connect jobs with earnings and status. Blue- and pink-collar occupations are usually associated with more physically demanding and hazardous environments (Ravesteijn et al., Reference Ravesteijn, van Kippersluis and van Doorslaer2018), offering less pay, control, flexibility and security (Costa and Sartori, Reference Costa and Sartori2007; Eurofound, 2017), all of which negatively impact health. By contrast, higher-status white-collar jobs generally offer greater job security, control and higher wage rates (Matthews and Weaver, Reference Matthews and Weaver1996; Schreuder et al., Reference Schreuder, Roelen, Koopmans and Groothoff2008). Our analysis demonstrates the protective health effects of white-collar working conditions, as these workers were able to work between 7 and 9 hours longer than their blue- and pink-collar counterparts without compromising health. As is the case for older workers with poorer health, the value of extending employment to support successful ageing is qualified by the job conditions and status of workers. There is a major challenge emerging for successful ageing if it is to be widespread and fair, as contemporary labour markets are increasingly polarising into ‘good’ and ‘bad’ jobs (with a corresponding polarisation in income and health). The structural lag in this instance is selective, unequal and increasingly pernicious to the already vulnerable. Our findings indicate that age-integrated reform needs to target employment inequality if it is to achieve a structural change that addresses population ageing (Riley and Riley, Reference Riley and Riley1994).

A key context and caveat to our findings is that all models were estimated on relatively healthy older workers who are not representative of the population. When we stratified by health status among those who were working (an imperfect solution to deal with health selection), we showed how important health status is for work hour–health tipping points, as there was a large disparity in work hour–health limits between those with poor health and those with good health. We further show how important poor health is for workforce exit in a series of regressions, and our exploratory work reported in the sensitivity analyses indicate that tipping points would be considerably lower if they were not estimated on optimally healthy mature-age adults. Together, our findings indicate it is health and job status, not age, that are the most important factors for determining retention among older-age adults, as well as their productivity and capacity to work. It is vital to recognise that good health and good jobs are requirements for remaining in the labour market and enabling successful ageing.

Contribution

This study represents the first demonstration of work hour–health tipping points for older Australian workers. It consists of an in-depth analysis by age, occupation and health status to illustrate the complex and dynamic relationship between work time, wages and health as they relate to extended working lives and population ageing. We used 3SLS to deal with reverse and reciprocal relationships and applied the bootstrapping 3SLS estimation technique to address heteroscedasticity in the equation errors. This enabled us to estimate a system of three simultaneous equations of health, work hours and wages, in what we believe to be the most robust way possible. Although the HILDA data are insufficient to control for sample selection bias, we were able to demonstrate the health selection bias in our estimates by conducting a series of extra analyses that found large differences in health outcomes between our analysis sample and the excluded sample. These findings provide further evidence of the impact of poor health on labour market participation amongst older workers, and the need for reforms of both the labour markets and the health system to redress it.

Our approach combines political economy and sociological insights into structural processes in labour markets (work hours and job status) with epidemiological and econometric methods and theory on work and health relationships. This inter-disciplinary analysis sheds light on social inequality processes as they relate to a key institution – the labour market – in the context of ageing. Our findings support current theory on the critical role of employment as a mechanism for successful ageing, but challenge assumptions that it offers a pathway for all. We also show that two key elements of successful ageing – mental and physical health and continued engagement in social and productive activities – are neither simply nor equally related. They can be in conflict, for example, when mature-age adults are working hours that compromise their health, if their health is already deteriorating or if they work in poor-quality jobs. Re-engineering labour markets to be age-integrated requires making them more equal and more health promoting in the jobs and conditions they offer (Rowe and Kahn, Reference Rowe and Kahn2015).

Study limitations and future research

The 3SLS approach with the bootstrapping estimation technique was used to address simultaneity, reciprocal relationships and heteroscedasticity in the error terms. This approach enables us to estimate robustly tipping points for the work hour–health relationship because it can address reciprocal relationships. However, inferences on causality cannot be made from this model because the 3SLS model does not allow us to test the exclusion restriction. Furthermore, the estimated tipping points are likely upward biased due to health selection (as suggested by sensitivity tests using OLS, fixed-effects approaches), which may have an increasingly strong effect on an older workforce, thus making the employed sample non-representative of the overall population aged 50–70.

Unlike previous studies, we adjusted for time spent in care and domestic work in our modelling, which is likely to affect overall work hour–health tipping points (Dinh et al., Reference Dinh, Strazdins and Welsh2017). However, time in unpaid work (e.g. caring) remains highly gendered even among older workers, and likely contributes to differences in tipping points by gender. These complex connections between time on and off the job, age, gender and care-giving are therefore worthy of future research. Furthermore, the impact of accumulated work time over long periods (annual hours rather than weekly hours as used in this study) could provide greater insight into health impacts and associations with chronic, physical health conditions including overweight and obesity which may take several years to become evident. Finally, our study does not address the role played by motivations and incentives, from employers in particular, to reward or limit older workers' work hours.