Collaborative double robust targeted maximum likelihood estimation.

Area of Science:

• Statistics
• Semiparametric Models
• Causal Inference

Background:

• Targeted Maximum Likelihood Estimation (TMLE) is a powerful statistical method for estimating parameters in semiparametric models.
• Standard TMLE relies on correct specification of either the outcome model (Q) or the nuisance model (g) for consistency.
• Existing methods face limitations when both models are misspecified.

Purpose of the Study:

• To introduce and provide a template for Collaborative Targeted Maximum Likelihood Estimation (C-TMLE).
• To develop a robust estimation procedure that remains consistent and efficient even with misspecified models.
• To enhance statistical inference capabilities within complex semiparametric settings.

Main Methods:

• C-TMLE iteratively refines estimates of the nuisance parameter (g) using a loss function focused on the target parameter (Q).
• Likelihood-based cross-validation is employed to select the optimal TMLE candidate.
• Theoretical framework for "collaborative double robustness" is established, ensuring consistency under broader conditions.

Main Results:

• C-TMLE demonstrates "collaborative double robustness," achieving consistency when both Q and g are misspecified, provided g solves a specific score equation.
• The method exhibits improved adaptiveness and potential for super-efficiency compared to standard TMLE.
• Statistical inference, including confidence intervals and p-values, is supported.

Conclusions:

• C-TMLE represents a significant advancement in targeted learning for semiparametric models.
• The approach offers enhanced robustness and efficiency, expanding the applicability of TMLE.
• This work provides a flexible template for robust statistical inference in complex data settings.