The Healthy Worker Survivor Effect Dissected

Abstract

Objectives: The healthy worker survivor effect is a well-recognized bias usually described as a form of selection bias or confounding. A more precise epidemiologic explanation, however, has been elusive. We illustrate how directed acyclic graphs (DAGs) can be used to distinguish several components of the healthy worker survivor effect and suggest methods for bias correction in occupational cohort studies. Method: Although generally referred to a single effect, we demonstrate using simulation studies that there are in fact four distinct aspects of the healthy worker survivor effect. Two aspects, (1) time-varying confounding by variables on the causal pathway and (2) heterogeneity in susceptibility, are functions of the underlying process of the exposure and disease under study. The other two, (3) left truncation and (4) right truncation, are functions of how the data are collected, ie the study design. We evaluate the bias induced by each aspect of HWSE on dose-response parameter estimates and apply methods designed to reduce the bias. Results: We find that causal techniques, eg, g-estimation and marginal structural models, can correct for time-varying confounding. These methods will yield valid estimates even in the presence of nonpositivity, which is characteristic of occupational studies. Left and right truncation can be corrected using inverse probability of censoring weights; methods to address heterogeneity in susceptibility will also be explored. Conclusion: The healthy worker survivor effect occurs due to the presence of four factors that may function separately or in concert to produce a downward bias if not accounted for. We provide guidance for methodologic approaches to reduce this pervasive bias.