A1714
Title: Estimation of treatment effects with data coarsened over informative observation intervals
Authors: Elizabeth Juarez-Colunga - University of Colorado Anschutz Medical Campus (United States)
Paula Langner - Veterans Health Administration (United States) [presenting]
Abstract: When timing of observations in a longitudinal study is informative or related to the outcome, estimation may be affected if the observation process is ignored. When an outcome is coarsened over each observation interval, the subsequent information loss further impacts estimation. The effect of informative variation in observation intervals on treatment effect estimation is explored when longitudinal count data are coarsened. Whether accounting for the observation process within the framework of a joint model can retain efficiency is investigated. A joint model of a recurrent event outcome and observation process, linked through a shared random effect, is proposed. Two forms of a coarsened outcome are considered, where either full counts for each interval or a dichotomized outcome indicating the presence or absence of any event in the interval are observed, with focus on the relative efficiencies of the treatment effect estimates. Simulation studies demonstrate that when the baseline event rate is low, the treatment effect estimate based on binary data is relatively efficient compared to count data, even while considering a range of other population parameters. In these cases, the joint model does not substantially improve over the outcome-only model, which ignores the informative observation times. However, with a higher baseline event rate, bias and variance of treatment effect estimates based on binary data are significant.
