Estimating treatment effects from longitudinal clinical trial data with missing values: comparative analyses using different methods

Abstract

The selection of a method for estimating treatment effects in an intent-to-treat analysis from clinical trial data with missing values often depends on the field of practice. The last observation carried forward (LOCF) analysis assumes that the responses do not change after dropout. Such an assumption is often unrealistic. Analysis with completers only requires that missing values occur completely at random (MCAR). Ignorable maximum likelihood (IML) and multiple imputation (MI) methods require that data are missing at random (MAR). We applied these four methods to a randomized clinical trial comparing anti-depressant effects in an elderly depressed group of patients using a mixed model to describe the course of the treatment effects. Results from an explanatory approach showed a significant difference between the treatments using LOCF and IML methods. Statistical tests indicate violation of the MCAR assumption favoring the flexible IML and MI methods. IML and MI methods were repeated under the pragmatic approach, using data collected after termination of protocol treatment and compared with previously reported results using piecewise splines and rescue (treatment adjustment) pragmatic analysis. No significant treatment differences were found. We conclude that attention to the missing-data mechanism should be an integral part in analysis of clinical trial data.