Handling Missing Data in Clinical Trials

Mike Kenward, Professor of Biostatistics, Department of Medical Statistics, London School of Hygiene and Tropical Medicine

The problem of handling missing data in clinical trials is explored, particularly in the light of two recent publications associated with the US and European regulators. The importance of assessing methods in the light of the target of the analysis, the estimand, is emphasised, and two types of estimand, de-jure and de-facto, are introduced. A key distinction is made between analyses in which missing data are 'defined away' and in which they represent a nuisance to be accommodated. The relationship between different types of estimand and Rubin's framework of missing data mechanisms (MCAR, MAR, MNAR), and the implications for subsequent analysis, is discussed. The advantages and disadvantages of alternative statistical analyses will then be considered in the light of the points raised. The role of sensitivity analysis will be explored, and one particularly promising route, based on the so-called pattern-mixture framework, will be developed and illustrated.