A0239
Title: Cumulative probability models and their utility in semi-parametric estimation of causal effects
Authors: Andrew Spieker - Vanderbilt University Medical Center (United States) [presenting]
Bryan Shepherd - Vanderbilt University Medical Center (United States)
Caroline Birdrow - Vanderbilt University Medical Center (United States)
Abstract: G-computation is a longitudinal generalization of standardization designed to accommodate time-varying confounding. While especially useful for estimating causal effects of time-dependent treatments, the parametric g-formula is sometimes criticized for its sensitivity to departures from assumptions. Cumulative probability models have recently been developed as a semi-parametric approach to modeling continuous outcome, through which one is able to model the CDF of an outcome conditional on covariates. We will show how cumulative probability models can be embedded within g-computation in order to bypass overly stringent parametric assumptions. We will then illustrate the utility of this methodology through Monte-Carlo illustrations and an application to a large cohort of women with endometrial cancer in order to compare cumulative medical costs associated with various adjuvant treatment strategies.
