Optimal Dynamic Treatment Methods

Abstract

This paper reviews and develops methods for implementing in practice recent ideas in the field of optimal dynamic treatment allocation. Given longitudinal sequences of observational data on health status and treatment selection for a cohort of patients, the aim is to determine a regime, or decision rule, which can be used to select treatment in order to optimise some final response or outcome. The approach to this problem that has been taken in the causal inference literature is shown to be extendable to problems in the field of stochastic optimisation. New diagnostic techniques to aid in model assessment are developed, and an application in anticoagulation is presented.

Keywords

• anticoagulation
• bandit problems
• causal inference
• diagnostics
• regret functions
• wild bootstrap