A cautionary tale: dealing with missing data in clinical trials for rheumatic diseases

OBJECTIVES: Missing data are found in nearly all clinical trials and it is important to use appropriate statistical techniques to analyse clinical trials with missing data. We discuss common statistical methods for tackling missing data and how to handle results when the analyses give different results.

METHODS: Using data from a placebo-controlled, randomised bovine Type I collagen (CI) study in diffuse cutaneous systemic sclerosis (dcSSc), we apply different statistical approaches to handling missing data. We also describe simple ways to ascertain the type of missing data in the data set, to the extent possible.

RESULTS: We examine eleven different methods to impute missing data. An analysis based on completers alone (complete case analysis and available case analysis) and the last observation carried forward (LOCF) methods require underlying assumptions which are rarely met in practice. Multiple imputation, mixed effects, and repeated measures try to account for the differences among patients and account for patient's specific response patterns, although the assumption that the missing data is directly related to the observed characteristics may well not be true. The joint likelihood based model combines the mixed effect model and logistic regression model to explicitly handle data not missing at random and so it is more realistic and potentially takes an additional step toward decreasing bias.

CONCLUSIONS: We discussed various ways of handling missing data and provide recommendations on how to arrive at a conclusion when different statistical approaches to analyse missing data analysis in clinical trials give conflicting answers.