Handling with missing data in clinical trials for time-to-event variables

Missing data is often a major issue in clinical trials, especially when the outcome variables come from repeated assessments. In particular, time-to-event endpoints can be substantially affected by a too conservative treatment of missing data along the observation period. When neglected or not properly treated, missing data may bias the results, reduce power and lead to wrong study conclusions. The advantage of a more sophisticated statistical method versus the traditional clinical method, such as last observation carried forward (LOCF), is still under debate. We compare the two methods in a clinical study testing the efficacy of an anti-arrhythmic agent versus placebo on the time to Atrial Fibrillation (AF) recurrence, where the maintenance of normal heart rhythm or the occurrence of the AF event was to be daily evaluated by trans-telephonic ECG recorded by the patients. A Cox model is applied for the comparison between treatments. The dataset presents missing observations due to the fact that recording is missing or ECG is not assessable. Moreover a simulation is performed to provide an additional example. Both methods for handling missing data are applied. Multiple imputation in SAS uses PROC MI. We examined results and possible problems arising from the fact that PROC MI implements methods which are not suitable for this kind of data.