Métodos de imputação múltipla para GEE em estudos longitudinais

Abstract

Missing data is a major challenge for longitudinal data analysis.This dissertation shows how missing data may have a great impact on the estimation of quantities of interest when one chooses to use the GEE model. This approach - flexible in the sense that the joint distribution of a subjects response vector does not need to be specified - yield valid estimates of the regression coefficients only with data missingcompletely at random (MCAR). Because this assumption is rarely true in practice, we explored another missing data mechanism. In order to correct the bias in regression coefficient estimates, we focus on multiple imputation, a technique proposed by Little & Rubin (1987) that has received great attention in the literature. It consists of predicting missing values in order to obtain complete data sets that can be analyzed using standard methods. We discuss five methods for imputing missing data, three of which consider a regression model and two use some form of matching. Besides the simulation results, in which we compared the performance of these imputation methods, among them one proposed, we present an application with real data. The results show that multiple imputation is an appropriate tool to remove the bias of the estimates in the GEE model, the largest gain obtained with regression-based models.