B1335
Title: Causal inference and dynamic treatment rule estimation based on contrast-approximating linear models
Authors: David Whitney - London School of Hygiene and Tropical Medicine (United Kingdom) [presenting]
Stijn Vansteelandt - Ghent University and London School of Hygiene and Tropical Medicine (Belgium)
Karla DiazOrdaz - University College London (United Kingdom)
Abstract: Optimal treatment rules, which assign treatment based on subject characteristics in a way that optimizes expected outcomes, are of widespread interest in statistics, economics, engineering, and other fields. Approaches to estimating optimal treatment rules include both data-adaptive machine learning and model-based methods. Machine learning approaches, while powerful, have been criticized for being less accessible to non-specialist audiences. Additionally, many machine learning approaches are tailored to specific types of treatment (e.g. binary or continuous). Model-based strategies can mitigate these criticisms but do so at the expense of strong modelling assumptions that are unlikely to hold in practice. To address these concerns, we propose a flexible framework that treats modelling the data-generating process as distinct from defining a parsimonious and interpretable rule. Each treatment rule in our proposal corresponds to a working structural mean model. Our estimator of the model coefficients allows for machine learning of nuisance parameters and accommodates any type of treatment. If the working model is correctly specified, then the corresponding treatment rule is optimal. We illustrate the finite sample performance of our proposal relative to other methods for estimating optimal treatment regimes in simulation and real-world data applications.
