Imputation strategies for missing longitudinal data.

Abstract

Longitudinal studies are commonly used to study processes of change. Because data are collected over time, missing data are pervasive in longitudinal studies, and complete ascertainment of all variables is rare. An imputation strategy for completing longitudinal data sets with missing covariate information is proposed which uses Bayesian methodology to improve upon imputation techniques commonly used in health services research. The proposed imputation strategy is applied to a longitudinal study of rural hospital closure and conversion with missing covariate data. It is shown to predict unobserved values well in this example. In addition, multiple imputation techniques are used to better reflect uncertainty in the imputation process. This imputation strategy is easily extended to other longitudinal studies with missing covariates as long as the missing data mechanism can be assumed to be ignorable.