Collections > Electronic Theses and Dissertations > Healthy Worker Survivor Bias in a Cohort of Uranium Miners from the Colorado Plateau

Radon, a ubiquitous gas present in breathing air and concentrated in the indoor environment, is a well established risk factor for lung cancer. Primarily, evidence for this association originated in studies of miners occupationally exposed to high con- centrations of radon. Much work has been done to predict lung cancer risk due to lower dose exposures in residences using dose-response curves derived from long- term, high-dose miner studies and shorter-term, low-dose residential studies. While residential studies suffer from a high probability of exposure misclassification at low exposures, miner studies present an opportunity to apply more precise estimates of the lung cancer-radon association to risk assessments. However, potential bias due to the Healthy Worker survivor bias has not been addressed in previous studies of occupa- tional exposure to radon. The Healthy Worker survivor bias occurs when workers with poor prognosis leave work sooner than those with better prognosis, thus creating an apparent association between low cumulative occupational exposures and mortality. Healthy worker survivor bias has been shown to substantially bias dose-response esti- mates in other settings, but it has not been explored in occupational studies of radon exposure. We apply two g-methods designed for addressing healthy worker survivor bias that cannot be controlled using conventional statistical methods. We utilize data from the Colorado Plateau uranium miners cohort, which comprises 4,137 male ura- nium miners who agreed to participate in a health study between 1950 and 1960 and were followed up for mortality through 2005. Our results suggest that there may be healthy worker survivor bias of the association between cumulative radon exposure and both lung cancer and all cause mortality. This work highlights the need for non- standard approaches to controlling time-varying confounding in occupational data. We show that, under certain conditions, g-methods can control this confounding, but that careful consideration should be made in the choice of method.