B1704
Title: Binary exposure and longitudinal cognition outcomes in the presence of non-ingorable dropout and death
Authors: Maria Josefsson - Umea School of Business, Economics and Statistics (Sweden) [presenting]
Michael Daniels - University of Florida (United States)
Abstract: G-computation, g-estimation and inverse probability weighting, have been proposed as alternatives to regression for causal inference of time-varying exposures. Although missingness due to loss to follow-up easily can be incorporated in either method in the presence of ignorable missing data, standard methods are generally invalid when the missingness is non-ignorable or due to death. We propose a Bayesian non-parametric method to simultaneously address both issues. In particular our approach incorporates Bayesian additive regression trees for G-computation estimation of the survivors average causal effect (SACE), i.e. the causal effect on the subpopulation of those surviving irrespective of exposure. The method allows to perform sensitivity analyses for assumptions about missing data mechanisms, i.e. lost to follow-up, death and unmeasured confounders. We illustrate the methodology using longitudinal observational data for studying the effect of widow(er)hood on cognition, where study participants death or drop-out, up until or after the time of the event, are complicating factors.
