Adjustment set selection for estimating optimal treatment rules under confounding

Area of Science:

• * Machine learning applications in causal inference and personalized medicine.
• * Statistical modeling for treatment effect estimation in complex health data.

Background:

• * Growing demand for individualized treatment strategies in medicine, especially for conditions like depression with varied patient responses.
• * Availability of multiple antidepressants with similar average efficacy but significant individual response heterogeneity.
• * Advancements in machine learning for causal inference and variable selection offer new approaches to treatment rule optimization.

Purpose of the Study:

• * To compare variable selection strategies within dynamic marginal structural modeling for estimating optimal individualized treatment rules.
• * To investigate the performance of various machine learning methods for propensity score variable selection.
• * To determine the optimal treatment rule for unipolar depression using real-world electronic health records.

Main Methods:

• * Comparison of outcome adaptive lasso, group lasso, doubly robust estimation, double-index propensity score, high-dimensional balancing propensity score, and causal ball lasso for variable selection.
• * Dynamic marginal structural modeling approach to estimate individualized treatment rules.
• * Analysis of electronic health records from 74,058 patients with unipolar depression, comparing selective serotonin reuptake inhibitors (SSRIs) and serotonin-norepinephrine reuptake inhibitors (SNRIs).

Main Results:

• * All investigated variable selection methods provided similar unbiased estimates for treatment effects.
• * Variable selection for the propensity score and inclusion of variables in the outcome model enhanced statistical efficiency.
• * Tailoring antidepressant treatment based on baseline symptom severity did not significantly impact symptom severity at 6 months post-treatment.

Conclusions:

• * While various machine learning methods can identify relevant variables for treatment rule estimation, their performance in excluding extraneous variables and computational efficiency varied.
• * The study found no significant benefit in tailoring antidepressant choice (SSRIs vs. SNRIs) based on baseline symptom severity for unipolar depression.
• * Further research is needed to refine individualized treatment rule development and identify patient characteristics that truly predict differential treatment response in depression.