Longitudinal Targeted Minimum Loss-based Estimation with Temporal-Difference Heterogeneous Transformer.

Area of Science:

• Causal inference
• Machine learning
• Longitudinal data analysis

Background:

• Estimating counterfactual outcomes under dynamic treatment policies is crucial in longitudinal studies.
• Existing methods often struggle with complex, long-term treatment strategies and potential biases from machine learning algorithms.

Purpose of the Study:

• To introduce Deep Longitudinal Targeted Minimum Loss-based Estimation (Deep LTMLE), a novel approach for estimating counterfactual means in longitudinal settings.
• To address biases in machine learning estimates and enable statistical inference with confidence intervals.

Main Methods:

• Utilized a transformer architecture with heterogeneous type embedding trained via temporal-difference learning.
• Applied the Targeted Minimum Loss-based Likelihood Estimation (TMLE) framework for statistical bias correction.
• Incorporated asymptotic statistical theory for generating 95% confidence intervals.

Main Results:

• Deep LTMLE demonstrated superior performance compared to existing methods in complex, long time-horizon scenarios.
• The method maintained effectiveness in small-sample, short-duration contexts, matching asymptotically efficient estimators.
• Successfully applied to estimate counterfactual mean outcomes for blood pressure management strategies in a cardiovascular epidemiology study.

Conclusions:

• Deep LTMLE provides a robust and statistically sound method for causal inference in longitudinal studies with dynamic treatments.
• The approach effectively handles complex scenarios and offers reliable statistical inference.
• Demonstrated practical utility in a real-world cardiovascular epidemiology application.