Occupational mortality studies have identified a large proportion of known human carcinogens and other exogenous toxins. However, such studies are subject to biases that arise from selective forces collectively known as the healthy worker effect. One of these forces is the healthy worker survivor effect, in which those who remain employed are healthier than those who leave employment. This differential survival will tend to attenuate estimates for the effect of an occupational exposure, so that weak carcinogen may be missed. In previous work, four methods to control for the healthy worker survivor effect were evaluated. Except for the G-Null test, all methods were found to be flawed because they make the unrealistic assumption of no relation between work status and prior exposures. In empirical data, exposure was found to strongly predict job-leaving. The control of confounding from variables such as job-leaving that are also intermediate in the causal pathway requires specialized methods. Our pilot project also showed, however, that the G-null test is limited in its statistical power by the rigidity with which exposure histories are matched. Most occupational cohorts are not large enough for a statistically powerful use of the G-null test. The G-null test has been generalized via structural nested failure time models, and these models, known as G-estimation methods, overcome the limitations. The proposed project will apply these models to occupational mortality data, and generalize the application to ordinal exposures. To address the problem of generalizability and robustness of the methods, a simulation study will generate occupational data under various assumptions regarding the relations among exposure, job-leaving, and mortality and will assess the bias when these data are analyzed using standard methods versus G-estimation. This project will: (a) evaluate feasibility and usefulness of the structural nested failure time models for control of the healthy worker survivor effect in occupational studies; (b) determine robustness of G-estimation to varying scenarios of differential job survivorship; and (c) determine in which situations ordinary methods are adequate. Uses of methods to control this bias is likely to improve the sensitivity of occupational studies enabling detection of carcinogens and other toxins that are less potent or are present at low exposure levels. Also, since occupational studies are often used to risk assessments to establish standards for environmental or occupational exposures, this project may have policy ramifications.
