Targeted minimum loss based estimation of causal effects of multiple time point interventions

Area of Science:

• Causal inference
• Statistical modeling
• Longitudinal data analysis

Background:

• Time-dependent confounding complicates intervention effect estimation in longitudinal studies.
• Previous targeted maximum likelihood estimators (TMLE) required estimating the full data distribution.
• Enhancing finite sample performance necessitates focusing on the most relevant data components.

Purpose of the Study:

• Develop a new closed-form TMLE for intervention-specific mean outcomes in general longitudinal data structures.
• Improve the efficiency and finite sample performance of TMLE by targeting only essential parts of the data-generating distribution.
• Extend the TMLE framework to accommodate other causal parameters, including those from marginal structural models.

Main Methods:

• The new TMLE represents the target parameter as an iterative sequence of conditional outcome expectations.
• It estimates and updates this sequence using the general TMLE algorithm, focusing on relevant data parts.
• The approach integrates innovative ideas from Bang and Robins (2005) into the TMLE framework.

Main Results:

• A new closed-form TMLE is developed for intervention-specific mean outcomes in longitudinal data.
• The proposed TMLE effectively targets the relevant parts of the data-generating distribution for improved estimation.
• Theoretical properties are supported by a small-scale simulation study, demonstrating practical utility.

Conclusions:

• The novel TMLE offers a more efficient approach to estimating intervention effects in the presence of time-dependent confounding.
• By targeting specific conditional expectations, it enhances the finite sample performance of causal effect estimation.
• This work provides a valuable extension to the TMLE methodology for complex longitudinal data analysis.