O15-4 Estimating absolute risk in the presence of confounders and competing risks: combining inverse probability weights and a cumulative incidence function in an occupational study of crystalline silica and lung cancer

Abstract

Estimates of absolute risk allow for the calculation of parameters such as attributable fractions, and the number of cases prevented under specific interventions and, thus, are of direct applicability to regulators. Furthermore, unlike contrasts based on the rate, hazard, or odds, risk ratios and risk differences, are collapsible parameters. However, most analyses of longitudinal data dealing with confounders, including occupational cohort data, focus on estimates of hazard ratios. Previous evidence links exposure to crystalline silica with increased mortality from lung cancer and non-malignant respiratory disease. In this study we estimate the covariate-adjusted absolute risk of lung cancer, treating non-malignant respiratory disease as a competing event, in a cohort of 2342 male California diatomaceous earth workers regularly exposed to crystalline silica, followed between 1942 and 2011. We applied inverse probability of exposure weights to control for age, race, asbestos exposure, and calendar time. We then estimated the risk of lung cancer mortality by applying a sub-distribution cumulative incidence function to the weighted pseudopopulation. For workers in the highest quintile of cumulative silica exposure, we estimated a cumulative lung cancer mortality risk of 10.1% at age 75, compared to a risk of 6.4% in workers in the lower exposure quintiles. These estimates were used to calculate the risk ratio and risk difference; bootstraps were used to calculate the confidence intervals. The corresponding risk ratio was 1.59 (95% confidence interval (CI): 0.59–4.30), and the risk difference was 3.8% (95% CI: −2.9, 10.5). Allowing for competing risks and adjusting for confounders in this study, we obtained estimates of absolute risk, risk ratios, and risk differences from longitudinal occupational cohort data, under specific assumptions. This method also allowed us to compare survival curves between exposure groups, which may be more informative than single time-point contrasts.