Direct likelihood analysis versus simple forms of imputation for missing data in randomized clinical trials

Abstract

In many clinical trials, data are collected longitudinally over time. In such studies, missingness, in particular dropout, is an often encountered phenomenon. We discuss commonly used but often problematic methods such as complete case analysis and last observation carried forward and contrast them with broadly valid and easy to implement direct-likelihood methods. We comment on alternatives such as multiple imputation and the expectation-maximization algorithm. We apply these methods in particular to data from a study with continuous outcomes. The outcomes are modelled using a general linear mixed-effects model. The bias with CC and LOCF is established in the case study and the advantages of the direct-likelihood approach shown. We have established formal but easy to understand arguments for a shift towards a direct-likelihood paradigm when analysing incomplete data from longitudinal clinical trials, necessitating neither imputation nor deletion.