MISSING DATA FOR REPEATED MEASURES: SINGLE IMPUTATION VS MULTIPLE IMPUTATION

Missing data is often a major issue in clinical trials, especially when the outcome variables come from repeated assessments. Single imputation methods are widely used. In particular, when data collection is interrupted at a certain time point, Last Observation Carried Forward (LOCF) is usually applied. Regulatory agencies advise to use the most conservative approach to impute missing data. As a drawback, single imputation methods do not take into account imputation variability. In this work we intend to compare single imputation versus multiple imputation methods in order to verify the effect on the successive inferential analysis, especially in terms of statistical significance of results. In particular we intend to verify if a more conservative single imputation method can be considered also a conservative method in term of statistical significance, respect to multiple imputation, where the higher variability can reduce the probability of having a significant result. We simulated a dataset representing a clinical trial testing the analgesic efficacy of a combination of drugs on moderate to severe pain after surgery. Pain is measured using a VAS scale. Analysis of covariance is applied to the primary efficacy variable, which is VAS change versus baseline. Both methods for handling missing data are applied. Multiple imputation in SAS uses PROC MI. We finally present statistical significant results. Analyzing results from several simulated dataset, we found out that multiple imputation consistently reduce the probability of finding statistical significance.