COUNTERFACTUAL INFERENCE IN SEQUENTIAL EXPERIMENTS.

Area of Science:

• Statistics
• Machine Learning
• Biostatistics

Background:

• Sequential experiments with adaptive treatment policies are common in healthcare.
• Standard inference methods struggle with individualized, time-varying treatment effects.
• Existing models lack flexibility for complex counterfactual mean structures.

Purpose of the Study:

• To develop statistical inference for counterfactual means in adaptive sequential experiments.
• To provide guarantees for individualized treatment effects at each time point.
• To minimize assumptions on the adaptive treatment policy.

Main Methods:

• Introduced a latent factor model for counterfactual means, generalizing prior work.
• Employed a non-parametric nearest neighbors approach for estimation.
• Established non-asymptotic high-probability error bounds for counterfactual means.

Main Results:

• The proposed method provides accurate estimation of individualized treatment effects.
• Achieved non-asymptotic error bounds for counterfactual means.
• Derived asymptotically valid confidence intervals under regularity conditions.

Conclusions:

• The latent factor model and nearest neighbors method offer a robust approach for statistical inference in adaptive sequential experiments.
• The findings are applicable to complex clinical trial designs, including mobile health.
• The study provides a theoretical foundation for reliable inference in dynamic treatment regimes.