Abstract
The European Medicines Agency issued a guideline on missing data in 2009. In 2012, the U.S. National Research Council published a book on the prevention and treatment of missing data in clinical trials, which has a great impact on medical fields. The most striking fact is that last observation carried forward (LOCF) is not recommended as a primary approach unless the assumption is confirmed. This article overviews some issues on missing data from statistical viewpoints. Prevention of missing data is very important in designing and conducting clinical trials. Once missing data are generated, an appropriate statistical method to handle the missing data is needed after ascertaining the missing mechanism as proposed by Rubin. Most statistical methods assume a mechanism of missing at random (MAR), where a systematic difference between complete and incomplete cases can be explained by a set of observed data. When there are confounding variables influencing missing data, mixed-effects models for repeated measures (MMRM) or other methods can be applied to adjust for confounding factors. Excluding all cases with missing data would usually induce a significant bias. Simple imputation such as LOCF will not be valid unless the underlying assumption is verified. Multiple imputation strategy may be valid under the assumption of MAR. Sensitivity analysis is necessary to confirm the consistency among results obtained by different approaches; for instance comparing intention-to-treat with per protocol set. More accurate reporting on missing data should be strengthened. Pre-specification of handling of missing data is also important in developing a study protocol.
