Adjusting for the healthy worker selection effect in cross-sectional studies.

BACKGROUND In a cross-sectional study of musculoskeletal disorders, women employed in highly repetitive manual work (garment assembly) were found to have approximately double the risk observed in a population with more varied tasks (hospital work). It was suspected that this estimate might be biased if garment workers with musculoskeletal pain were more likely than others to leave employment. METHODS Retrospective information on date of first onset of symptoms and years employed to date of pain onset, or to survey date (whichever was earlier), was used to calculate age and calendar period-specific rates of onset, conditional on remaining employed until the survey. RESULTS These rates, and the relative risk for garment work, increased over the 20-year period preceding the year of the survey. The trend was not explained by age or length of employment, or by any known changes in work demands that might have caused a true increase in incidence density. CONCLUSIONS In the absence of longitudinal cohort data, alternative explanations for these results cannot be excluded. However, with specified assumptions, the most plausible appears to be a healthy worker selection effect acting differentially between high and low exposure groups. This effect would have caused the smallest bias in the prevalence in the year immediately before the survey, and a better estimate of the true relative risk would be approximately five. Where date of onset has been obtained, this method may be used in other cross-sectional studies to estimate and reduce the magnitude of selection bias in a survivor population, if longitudinal data cannot be collected.