SIMULACIÓN PARA EVALUAR EL TRATAMIENTO DE DATOS FALTANTES DE ESTRUCTURA LONGITUDINAL EN EL CONTEXTO DE ENSAYOS CLÍNICOS

Abstract

Longitudinal studies often suffer from missing data. In this work different methods are compared, by means of a simulation study, to address the estimation task under missing data in longitudinal settings. Completely random, random and non-random missingness mechanisms are studied. Methods used include complete cases, zero imputation, imputation by the mean and last observation carried forward. These methods are refined by the conditional mean imputation, regression and stochastic regression. Such methods are contrasted with maximum likelihood. The goodness of each one in terms of the proportion of missing values is assessed. A qualitative approach that allows a quick and clear assessment of each method is introduced, by means of their classification as Excellent, Good, Regular, or Bad. In this way a simulator written in the IML (Interactive Matrix Language) language of SAS (Statistical Analysis System) is obtained, which allows the evaluation. We conclude that naive imputation methods produce biased estimators. Simple refined methods such as stochastic regression imputation show an acceptable behavior for low proportions of missing values but underestimate variability parameters. The complete case method shows good behavior. Maximum likelihood behaves the best