Targeted estimation of nuisance parameters to obtain valid statistical inference.

Area of Science:

• Statistics
• Causal Inference
• Machine Learning

Background:

• Estimating treatment-specific means requires controlling for baseline covariates.
• Propensity scores and outcome models are key statistical components.
• Data-adaptive methods like super-learning are crucial for nuisance parameter estimation.

Purpose of the Study:

• To develop improved estimators for treatment-specific means.
• To address sub-optimal bias/variance trade-offs in existing methods.
• To enhance statistical inference for causal effects.

Main Methods:

• Utilizing ensemble learning, specifically super-learning, for nuisance parameter estimation.
• Implementing targeted bias reduction for improved estimator performance.
• Developing novel targeted minimum loss-based estimators (TMLEs), including collaborative TMLEs (C-TMLEs).

Main Results:

• Demonstrating that targeted nuisance parameter estimation leads to second-order bias reduction for the estimand.
• Proving theorems that establish asymptotic linearity of the treatment-specific mean estimator.
• Constructing C-TMLEs with known influence curves for statistical inference, even with variable selection.

Conclusions:

• Targeted TMLEs offer a robust approach to estimating treatment effects.
• The proposed methods enhance the reliability of statistical inference in causal studies.
• These advancements are applicable to various causal inference scenarios, including inverse probability of treatment weighting.