East Coast Ballroom
WITHDRAWN - Buttressing: A Multiple Imputation Framework for Joining Available Background Covariate Data with Observed Outcomes for Precision Gains (306665)
Thomas Belin, University of California, Los Angeles*Jay Jia Xu, University of California, Los Angeles
Keywords: Missing Data, Multiple Imputation, Randomized Controlled Trial, Sample Size, Treatment Effect
Researchers routinely face financial, time, and other resource constraints that limit the sample size of a study and the corresponding precision of estimation of quantities of interest. However, datasets containing background covariates on the total pool of eligible subjects identified for a study - subjects enrolled in the study as well as those subjects unable to be enrolled - are often readily available. If the background covariates are strongly predictive of the outcome, accurate predictions of outcomes for unenrolled subjects can be obtained from a model of the outcome on the background covariates. In such scenarios, it is possible that compared to just analyzing the observed outcomes, stronger inferences for the outcome can be achieved through including predicted values. Viewing this scenario as a missing data problem, we propose a multiple imputation procedure to predict plausible values of the outcome for unenrolled subjects in order to strengthen inferences concerning the outcome when estimating means and treatment effects, a procedure we refer to as ``buttressing." Using simulations and a real data example, we demonstrate the potential buttressing possesses.