Personalizing Path-Specific Effects

TL;DR

This paper develops a method to personalize treatment policies by maximizing specific causal pathway effects using counterfactuals, identification algorithms, and machine learning, with validation through simulations.

Contribution

It introduces a novel framework for tailoring treatment strategies to maximize path-specific effects, combining causal inference with machine learning techniques.

Findings

General identification algorithm for path-specific counterfactuals

Proof of completeness for the identification method

Simulation results demonstrating effectiveness

Abstract

Unlike classical causal inference, which often has an average causal effect of a treatment within a population as a target, in settings such as personalized medicine, the goal is to map a given unit's characteristics to a treatment tailored to maximize the expected outcome for that unit. Obtaining high-quality mappings of this type is the goal of the dynamic regime literature (Chakraborty and Moodie 2013), with connections to reinforcement learning and experimental design. Aside from the average treatment effects, mechanisms behind causal relationships are also of interest. A well-studied approach to mechanism analysis is establishing average effects along with a particular set of causal pathways, in the simplest case the direct and indirect effects. Estimating such effects is the subject of the mediation analysis literature (Robins and Greenland 1992; Pearl 2001). In this paper, we consider how unit characteristics may be used to tailor a treatment assignment strategy that maximizes a particular path-specific effect. In healthcare applications, finding such a policy is of interest if, for instance, we are interested in maximizing the chemical effect of a drug on an outcome (corresponding to the direct effect), while assuming drug adherence (corresponding to the indirect effect) is set to some reference level. To solve our problem, we define counterfactuals associated with path-specific effects of a policy, give a general identification algorithm for these counterfactuals, give a proof of completeness, and show how classification algorithms in machine learning (Chen, Zeng, and Kosorok 2016) may be used to find a high-quality policy. We validate our approach via a simulation study.