Semiparametric Regression Modeling of Incomplete Longitudinal Outcomes: Stratifying on Informative Missingness Predictors

We introduce a new method for analyzing continuous or discrete longitudinal outcomes that are incomplete. Our two-step method, informative missingness predictor stratified generalized estimating equations (IMPS-GEE), incorporates into the analysis of the outcomes certain extraneous covariates that are potentially predictive of missingness and also related to the covariates of interest and the outcomes. In the first step, we use generalized estimating equations to fit a model for the outcomes and covariates of interest, stratified on the levels of the extraneous covariates. In the second step, we combine the strata-specific estimates of how changes in the covariates of interest affect mean outcomes, using weights equal to the estimated probabilities of belonging to the strata. In this article, we describe in detail how to implement IMPS-GEE for a single extraneous covariate measured at baseline and also for multiple, time-varying extraneous covariates. We show analytically that there are settings where IMP...