Accounting for missing values in sample size calculations

Paolo Fina, Senior Statistician, CROS NT

Often in the sample size calculation, subjects are classified in ‘evaluable’ and ‘not evaluable’ and subjects with a missing value for the primary endpoint are included into the ‘not evaluable’ category . The sample size is then determined in terms of ‘evaluable’ subjects and an extra enrollment Is scheduled to take into account for the presence of missing values. Even if largely used, this approach is not consistent with the ITT principle, ICH guidelines and it appears to be related to an old way to deal with missing data based on the idea of the replacement. In ITT analyses where missing data are managed by means of imputation techniques like LOCF or others, all the subjects are ‘evaluable’ and it’s the effect of these techniques on the observed treatment effect which should be considered. When missing values are managed using ML methods (for instance MMRM) again all the subjects are ‘evaluable’ but completers and not completers provide a different level of information. In this case we should consider the ‘effective sample size’ . During the talk it will be discussed how to derive the ‘effective sample size’ of a longitudinal study analyzed using a MMRM approach.