Title: Using offline policy evaluation to advance methodological barriers in digital health
Authors: Jane Kim - Stanford University (United States) [presenting]
Abstract: Evaluating the efficacy of mHealth apps which deploy algorithms remains a critical priority for health care providers and patients who need this evidence in advance of offering and receiving such care. Contextual bandits are well-suited to address concerns regarding the adaptability of interventions and non-adherence in behavioral interventions; prior work has demonstrated the benefits of deploying bandits in real settings. A key consideration, however, is that it is not usually feasible nor ethical to deploy a new algorithm live in the context of testing health interventions. We will present offline policy evaluation (OPE) as a means to answer practical questions that investigators and intervention designers may have about design questions related to research. OPE can be used to evaluate the performance of algorithms using logged data that was generated under a different policy. Its core idea is to use importance sampling that allows the estimation of a function of a distribution when the data of interest were generated from an altogether different distribution. First, we introduce the method of offline policy evaluation and provide background information about sequential decision making. Then we illustrate its utility using a case example of a digital recommendation system that delivers messages to reduce stress.