An application of the propensity score method in comparisons of health status between 12-h rotating shift and day schedules

Background: In observational studies, imbalanced background characteristics hamper addressing causal questions, because a randomisation is usually not feasible for ethical and practical reasons. The Propensity score (PS) method reduces the measured background characteristics to a single composite characteristic and therefore enables a straightforward assessment. The present analysis aimed to compare the estimation by means of PS with a previous assessment of shift systems and their effects on the employee's health using conventional regression models. Materials and Methods: Two different shift schedules (3×12 and 4×12) were compared with a daytime working system to investigate potential differenzial effects on the employee's health status assessed with the Work Ability Index (WAI). A total of 924 participants (278 3×12, 321 4×12 shift workers, and 325 day workers) were recruited for the study. The outcomes of interest were the WAI sum score and its dimensions, dichotomized at the medians. PS was defined as the probability of being exposed in any of the two working time schedules, as a function of the individuals' background characteristics. A logistic regression model was used to estimate PS for each subject. Stratified comparisons were then made using Mantel-Haenszel test. Results: Age, BMI, living alone, smoking status, and the number of children represented major imbalances between the groups. Based on the PS, no significant differences in terms of WAI and its individual dimensions were found in either of the pair comparisons. The results were consistent with those found with the model-based analyses. Discussion: The PS method is considered especially advantageous in comparisons of more than two groups, in case of different underlying heterogeneities between the pairs. Although such heterogeneity exists in the present dataset, the outcomes across both analyses were comparable. Nevertheless, the PS method can only adjust for measured confounders and will never overcome its a posteriori nature in terms of assignment of exposure.