Researchers by default tend to choose complex models when analyzing nonindependent response variable data, this may be particularly applicable in the analysis of longitudinal trial data, possibly due to the ability of such models to easily address missing data by default. Both maximum-likelihood (ML) estimation and multiple imputation (MI) are well-known to be acceptable methods for handling missing data, but much of the recently published quantitative literature has addressed questions regarding the research designs and circumstances under which one should be chosen over the other. The purpose of this article is threefold. First, to clearly define the assumptions underlying three common longitudinal trial data analysis models for continuous dependent variable data: repeated measures analysis of covariance (RM-ANCOVA), generalized estimating equation (GEE), and a longitudinal linear mixed model (LLMM). Second, to clarify when ML or MI should be chosen, and to introduce researchers to an easy-to-use, empirically well-validated, and freely available missing data multiple imputation program: BLIMP. Third, to show how missing longitudinal trial data can be handled in the three data analysis models using three popular statistical analysis software packages (SPSS, Stata, and R) while keeping the published quantitative research in mind.
Read full abstract