Application of multistate modeling to clinical data analysis

Abstract

Multistate models have been used for decades to analyze the economics of expensive and long-lasting treatments. More recently they also served to address questions in clinical drug development. It seems timely to introduce the broader pharmacometrics community to the technical aspects and the problem-solving capabilities of these models. A minimal model is introduced that can answer questions of interest to drug developers, regulatory agencies, and patients (with their carers and payers). A clinical study is simulated where 1000 patients are randomly allocated (1:1) to placebo and active treatment. After a recruitment phase, deaths are counted, and an administrative data cutoff occurs 858 days after the first patient is randomized. The minimal model has one initial state, two transient states, and two absorbing states. Fully parameterized semi-Markov processes govern the unidirectional transitions between states. Simulations explore the influence of parameter uncertainty and sample size on the validity of statistical inferences. The questions of interest to stakeholders are addressed predominantly with graphic displays. All programming codes are made available. Both drug developers and regulators are invited to re-evaluate the methods currently in use to assess the benefits and risks of new treatments.