Statistical methods

The statistical analyses were conducted with SAS V.9.4 and IBM SPSS Statistics V.24. Fixed effects time-dependent logistic regression analysis was used to test whether significant increase or decrease in unsocial working hour characteristics would show a parallel change in WLC (increase in working hour characteristic—increase in WLC or decrease in working hour characteristic—decrease in WLC). The fixed effects regression model has the advantage that each individual serves as his or her own control and time-invariant (fixed) confounding (non-observed individual and environmental factors) are eliminated.³⁰ First, a crude model was used and then model adjusting for various covariates. The covariates were marital status (married or cohabiting vs divorced, estranged, single or widow), number of small (0–6 years) and school-aged (7–18 years) children, stressfulness of the overall life situation during the past 12 months (a five-point Likert-type scale from not burdensome to extremely burdensome), and perceived health (a five-point Likert-type scale from good to poor).³¹ The number of small children was classified into zero children, one children, at least two children and missing information, whereas the number of school-aged children was classified into zero children, one children, two children, at least three children and missing information.

We tested interactions by age (≤39, 40–49 and ≥50 years) and sex and performed stratified analysis for interactions with P<0.10.³²