American Statistical Association New York City Metropolitan Area Chapter Mailman School of Public Health Columbia University Department of Biostatistics Colloquium

In longitudinal studies, dropouts are known to cause analytic problems. Missing outcomes require making additional assumptions on auxiliary models that are not of primary interest, such as the missing mechanism or the conditional distribution of missing data. Since inferences rely on the correct specification of auxiliary models, a mis-specified auxiliary model may introduce bias. In this talk, we propose an estimating equation approach that does not require the auxiliary model assumptions in the generalized linear mixed models under some subclass of nonignorable missingness mechanisms. Our method is based on the pairwise conditioning of ordered outcomes. We also present a semiparametric inferential procedure without specifying the distributional assumption of outcomes.