Negative bias in exposure-response trends in occupational studies: modeling the healthy workers survivor effect

Am J Epidemiol. 1996 Jan 15;143(2):202-10. doi: 10.1093/oxfordjournals.aje.a008730.


Many occupational studies analyze trends between cumulative exposure and mortality. The authors show that such trends are, in general, negatively confounded by employment status. Mortality rates for workers who leave work ("inactive" workers) are higher than for active workers because some workers leave because they are ill. The percentage of inactive relative to active person-time is higher in low categories of cumulative exposure, causing employment status to act as a negative confounder of exposure-response trends (the opposite occurs for time-since-hire). We illustrate these phenomena using 10 "negative" mortality studies, in which adjustment for employment status removes false trends. However, adjustment for employment status will lead to biased estimates when it acts as an intermediate variable between cumulative exposure and death, as occurs directly when exposure causes a disabling disease that, in turn, causes death or indirectly when exposure causes workers to leave work. The authors illustrate this problem using simulated follow-up data for leaving, disease incidence, and mortality. In the null case in which cumulative exposure affects neither disease incidence (or mortality) nor leaving rates, employment status indeed acts as a negative confounder of exposure-response trends, and traditional adjustment eliminates this confounding. However, when cumulative exposure affects disease incidence or rates of leaving, adjustment for employment status will not be adequate. Employment status falls under the general rubric of variables that are simultaneously confounders and intermediate variables.

MeSH terms

  • Bias*
  • Cohort Studies
  • Confounding Factors, Epidemiologic
  • Employment / statistics & numerical data*
  • Follow-Up Studies
  • Humans
  • Occupational Diseases / mortality*
  • Occupational Exposure / statistics & numerical data*
  • Poisson Distribution
  • Predictive Value of Tests
  • Proportional Hazards Models
  • Survival Analysis