UC Berkeley

In many longitudinal studies researchers are interested in examining failure times associated with a particular risk in settings where subjects may fail due to only one of multiple competing risks. Using targeted learning, we present a method of estimating the cumulative incidence of the multinomial outcome constructed of all possible risks which does not require independence assumptions among the outcomes. We show analytically and with simulation that this method is suitable within the causal roadmap to provide causal contrasts with natural interpretations. We compare the minimal assumptions required for statistically and epidemiologically unbiased interpretations using this method to those required for cause-specific estimation. We apply this method to estimate the causal effect of average yearly cumulative exposure to metal-working fluids on mortality from chronic obstructive pulmonary disease and cancer among a cohort of autoworkers hired between 1938-1985.