CFE 2020: Start Registration
View Submission - CMStatistics
B0725
Title: Combining model-based and model-free methods for estimating interventions with indefinite horizons Authors:  Owen Leete - NC State University (United States)
Eric Laber - North Carolina State University (United States) [presenting]
Abstract: Mobile-health (mHealth) holds tremendous potential as a means of delivering precision medicine at scale. mHealth-based precision medicine strategies seek to tailor intervention decisions to the unique health trajectory of each patient. An optimal intervention strategy maximizes some cumulative measure of patient health over the intervention period. Most commonly used methods for estimation of an optimal intervention strategy can be broadly characterized as either model-based, in which the underlying generative model is estimated and subsequently used to identify an optimal strategy using simulation (g-computation), or model-free, in which semi-parametric estimating equations are used to identify an optimal regime. Model-based methods impose a structure that reduces variance but is subject to misspecification, which may induce bias. In contrast, model-free methods impose less structure which reduces the risk of bias but may increase variance. We propose a method that combines model-free and model-based estimators which is consistent if one (but not necessarily both) is correctly specified and efficient if both are correctly specified. Empirical results show the proposed approach mitigates the risk of misspecification and yields better patient outcomes than competing methods.