Title: Estimating treatment effects in non-Markov multi-state models
Authors: Jon Michael Gran - University of Oslo (Norway) [presenting]
Abstract: Multi-state models, as a generalization of traditional time-to-event models, is a convenient framework for analysing transitions between a possible large number of states. Estimation and covariate adjustment can be based on any traditional hazard model, separately for each transition intensity, before overall outcome measures are derived, e.g. using the Aalen-Johansen estimator. We show that some of these outcome measures, such as state transition probabilities, can be very sensitive to violation of the Markov assumption. Others, like state occupation probabilities are not. We look at two general estimation procedures for non-Markov models based on landmark subsampling, and discuss the use of Cox proportional hazard models, Aalen additive models and inverse probability of treatment weighted Nelson-Aalen estimators for causal inference versus mere prediction. The motivating example is a study on the effects of national workplace initiatives on long-term sick leave and work participation, analysing a large scale dataset linked from numerous Norwegian population-wide registries.