Working hours and self-rated health over 7 years: gender differences in a Korean longitudinal study

The Korean Labour and Income Panel Study (KLIPS) is an ongoing longitudinal survey of a nationally representative sample. The same individuals were re-interviewed each year about various social and economic activities, including employment, income, education, and training. The study subjects used here were drawn from seven waves of the KLIPS between 2005 and 2011 (from the 8th to the 14th wave), where relevant information was available. Initially, 4086 workers who had no missing values on SRH and working hours were included. After exclusion of individuals who did not participate in all of the seven consecutive years, a total of 1578 workers remained for analysis of balanced panel data to make equal contributions from all of the participants. Ethical approval was obtained from the Institutional Research Board at Konkuk University Hospital, and secondary data analysis was performed without identifying personal information.

As a health measure, SRH over the last 12 months was used. A 5-point scale of SRH was collapsed into two categories: a rating of “very good” or “good” health status was identified as “good health” and a rating of “fair”, “poor”, or “very poor” health status was regarded as “poor health”. For participants who worked on a regular basis, working hours were calculated by asking a set of two questions: “What are your current regular working hours for a week excluding meal times?” and “On average, how many hours do you work overtime per week?” For workers who did not work on a regular basis, working hours were assessed by asking: “How many hours do you work in a week on average?” Working hours with a minimum of at least 20 h were categorized into five groups to reflect the current status of working hours authorized by Korean law, and the significance for each category is indicated in parenthesis; 20–35 h (less than standard working hours), 36–40 h (standard working hours), 41–52 h (overtime within legally permitted working hours per week excluding weekend work), 53–68 h (overtime within legally permitted working hours per week including weekend work) and 69 h or more (legally prohibited working hours in any circumstances). Binary variables were generated for working days per week (i.e. ≤5 days or >5 days), shift work and job satisfaction (i.e. yes or no). Based on the Korean Standard Industrial Classification, industrial sectors were grouped into five categories as follow; agriculture, fishing, mining and construction; manufacturing; retail, restaurant and hotel; financing, estate, transportation and information; health, social, educational and household services. Initially, occupations were classified into 10 major groups according to the Korean Standard Classification of Occupation and then collapsed into three categories; manager and professional (managers, professionals, technicians and associate professionals), white collar workers (clerks, service and sales workers) and manual workers (skilled agricultural and fishery workers, craftsmen and related workers, plant and machine operators and elementary occupations).

Relevant socio-demographic and behavioural covariates were identified from previous studies. Three categories of educational attainment were distinguished: middle school or less, high school, undergraduate and postgraduate. Monthly income was based on a single question about the amount of wages earned in the previous month; data were then grouped into quintiles. Employment status was grouped into three categories: regular, temporary and day labour, on the grounds of the contract period. Smoking habits were classified into two categories: current smoker vs ex- and non-smoker. Similarly, alcohol consumption was graded into two categories: current alcohol drinkers vs ex- and non-drinkers.

To explore the basic characteristics of the study population and basic associations, simple descriptive statistics were reported from the data pooled over seven years. For the analyses of association between working hours and SRH, the generalized estimating equation was applied to adjust for confounders and to consider the serial correlation between repeated measures within an individual. To clarify the temporal intervals, SRH in t + 1 year was secured to be ahead of predictors in the preceding t year (a lag time of 1 year). Application of a set of covariates for adjustments was identified using the Directed Acyclic Graph (DAG) approach [24], which is able to achieve a model selection accounting for a conceptual causal relationship. When a list of variables becomes sufficient enough to minimize confounding and selection bias under a causal assumption [25], DAG presents a model described as ‘minimal sufficient adjustment set’ of covariates. Models were fitted with different adjustments for covariates and stratified by gender. Model 1 was fitted only for working hours. Model 2 identifying a minimally sufficient adjustment set included age, working days, monthly wage, educational level, shift work, job status, along with long working hours. Then, additional to the covariates defined in the Model 2, further adjustment was made for behavioural factors (smoking and alcohol) and job satisfaction (Model 3). We tested this model with DAG and found that it is sufficient to build a causal structure between long working hours and SRH (Additional file 1). Statistical inference applied to panel data depends on time period in use, as Individual effects change over time; that is, for example, SRH declines with ages. It is therefore argued that the time-varying nature of SRH can be better modeled by introducing a longer panel [26, 27]. Panels have lasted for one to seven waves in the current study and, to minimize the period-dependent heterogeneity, we included only individuals who provided entire information on the variables specified in the Model 3 over all seven years (six observations per each individual, as the associations of predictors with SRH in the current study was defined to occur over a period of year t to t + 1). We also conducted analyses involving a different timescale with cross-sectional association (no time lag) and a similar magnitude and pattern of associations was observed (data not presented). To assess potential bias associated with the categorization, we repeated the same analysis with a slightly different definition of SRH; good (“very good”, “good” or “fair”) versus poor (“poor” or “very poor”) health status and associations were similar (data not presented) to the results presented here. Statistical analyses were conducted with STATA ver13 (StataCorp LP, Texas, USA) and an online program “dagitty” [28] for DAG analysis.

A major strength of this study is the use of a full range of working hours including short working hours, instead of narrowly focusing on overtime. The availability of a relatively wide range of covariates, including working conditions, socioeconomic and behavioural factors, was important to allow examination of certain details. By providing different sets of adjustment covariates, we were partly able to test the separate effects of the factors. Further, compared to cross-sectional data, panel data can provide better estimates by controlling for unobserved heterogeneity and offers gains in establishing causality. Study strengths include the application of DAG, which was used for model selection to identify the minimally sufficient adjustment sets. Although DAG approach is based on the deduction of causal structure among variables under investigation, it is supposed to reduce the degree of confounding and selection bias by better representing the underlying causal relationship [25]. There were some potential limitations to our study. SRH may be too broad to detect specific influences of working hours on health. However, SRH is strongly correlated with objective health status [29] and is used as a relevant single measure, particularly when various health influences are expected to include both physical and mental aspects [30]. Though sample restriction to workers who stayed in employment over all of the 6 study years was motivated by the preference to balanced data over unbalanced data, this might result in ‘the healthy worker survival effect’. In a sensitivity analysis, comparing between those who stayed in the sample and those who did not, a lower proportion of poor self-rated health, longer working hours, more frequent regular job, and higher amount of wage were favourable among the former (data not presented). Selective loss of unhealthy workers consequently may have contributed to underestimates of the adverse impact of long working hours on SRH. Another limitation concerns unmeasured confounding factors such as sleep hours and occupational stress, which could not be assessed due to a lack of available data.

Consistent with some previous studies but not all, we found a negative association between long working hours and SRH. The ways through which long working hours affect workers’ health status are yet to be explored. Most obviously, long working hours were closely linked with work demands, thereby leading to an increase of occupational stress and bringing on poor general health. Other explanations include extended exposure to hazardous physical, chemical and psychological factors, and insufficient recovery due to short sleep duration [31]. Whether socioeconomic and behavioural factors modify the association between working hours and health is an important issue, and this study, with serial adjustments across models, showed substantial effects of socioeconomic factors but negligible effects of behavioural factors. This may seem to be inconsistent with the argument that long working hours disrupt health behaviours, which in turn partly mediate the relationship between working hours and health [31, 32]. However, this finding echoes previous studies [33], in which work environment explained a larger portion of socioeconomic differences in SRH than did behavioural factors. This implies that those exposed to both long and short working hours may co-exist with other adverse risks, which are more common to those of lower socioeconomic status. Another explanation may be provided from a data construction point of view; among workers who maintained paid employment for long years as seen in our sample, the adverse effects of long working hours through the behavioural path may be minimized for adaptation. This suggests that enabling people to stay in employment may help to promote healthy behaviours, which may be particularly prominent among women.

The worsening of SRH with increasing working hours was found only among women. This finding agrees with previous studies [2, 3, 34], including a study from the same population [35], which was limited to a cross-sectional analysis from one wave of the KLIPS. Two hypotheses can be addressed for explanation. First, female workers may be under the ‘double burden’: family and work responsibilities [36]. Though the hypothesis was not tested in the current study, in a study from Korea, female workers took most of the household tasks after work time [37]; women thereby became more susceptible to experiencing health problems. Second, this observation may be rooted in the unequal power relation between men and women, which consequently is reflected in the labour market position. Women often worked in more precarious and subordinate positions than men, as broadly shown in the current study (Tables 1 and 2) and other studies [38]. This subsequently leads to more exposure to hazardous and unfavourable working conditions towards women’s health [39]. However, the gender difference may also be because of the health measure used in this study; that is SRH. Women tend to report poorer SRH than men across countries and age groups [40], though the evidence on whether women’s assessment of their own health relative to their actual health status is more negative than men is unclear [41, 42]. Putting things together, long working hours may have a greater negative impact on women than men under the same conditions.

Interestingly, our study suggests that both the shortest as well as the longest working hours are associated with poorer SRH, resulting in a U-shaped association. This pattern, though not conclusive, was similarly observed in a meta-analysis for coronary heart disease [43], while other studies reported a gradual increase of poor health in response to an increase of working hours [44, 45], where working hours shorter than a standard category was mostly neglected. Reverse causation may be plausible. Despite the longitudinal nature of this study, poor health might affect the ability to work [23], as participants were not free from poor health at baseline. Thus, some of those who work short hours do physically light work because of health reasons, though such findings remain to be explored. This, together with another finding that the association was markedly attenuated by further adjustment for socioeconomic factors, implies that there has been bi-directional accumulation between poor health and lower socioeconomic status; those who work short hours do so as more of their socio-economical vulnerability than as of voluntary choice [46]. More attention in future studies needs to be paid to cover a wider range of working hours as the atypical nature of working hours is increasing at both ends of the range. A potential independent association of job satisfaction is also important. Job satisfaction represents the subjective combination of job security, control over workload, and levels of enjoyment [47]. Since the concept relates to a general condition rather than domain specific, job satisfaction is expected to be of potential importance for workplace policies aimed at improving employee’s global health status.