Analysing longitudinal continuous quality of life data with dropout

Quality of Life (QL) is becoming an increasingly popular endpoint in phase III cancer clinical trials. However, there is still no agreement as to what is the optimal approach to analysis. In this paper we review some concepts which should be considered during a QL analysis. We present two modelling approaches that have been substantively developed in other research fields: selection models and pattern-mixture models. These models are compared using data from an EORTC clinical trial in poor-prognosis prostate cancer patients. It is illustrated that, although selection models and pattern mixture are probabilistically equivalent, they may shed completely different light on data from a modeller's point of view.