HIGHLY ADAPTIVE LASSO: MACHINE LEARNING THAT PROVIDES VALID NONPARAMETRIC INFERENCE IN REALISTIC MODELS.

Area of Science:

• Causal inference
• Real-world data analysis
• Statistical methodology

Background:

• Estimating treatment effects on health outcomes from real-world data (RWD) is complex.
• Semiparametric methods like targeted maximum likelihood estimator (TMLE) provide asymptotically linear estimators.
• Asymptotic efficiency requires Donsker class likelihoods and faster nuisance parameter convergence rates.

Purpose of the Study:

• To introduce the Highly Adaptive LASSO (HAL) as a method to achieve asymptotic efficiency in causal inference.
• To demonstrate HAL's capability in enabling robust statistical inference for complex causal parameters.
• To enhance decision-making in health research through reliable uncertainty quantification.

Main Methods:

• Utilizing the Highly Adaptive LASSO (HAL) as an empirical risk minimizer.
• Employing a function class of càdlàg functions with a bounded sectional variation norm, known to be Donsker.
• Ensuring nuisance parameter estimates converge at a rate faster than $$.

Main Results:

• HAL satisfies the conditions for asymptotic efficiency of causal effect estimators.
• HAL's flexible function class captures realistic data patterns.
• HAL enables robust inference for non-pathwise differentiable parameters like CATE and causal dose-response curves.

Conclusions:

• The Highly Adaptive LASSO (HAL) method guarantees asymptotic efficiency for causal effect estimation in RWD.
• HAL provides essential statistical uncertainty quantification for precision health parameters.
• HAL bridges the gap in statistical inference for machine learning applications in health research.