Effects Among the Affected

Area of Science:

• Causal inference
• Biostatistics
• Health services research

Background:

• Understanding sequential treatment effects is crucial in chronic disease management.
• Previous methods often struggle to capture how early treatment responses modify later treatment impacts.

Purpose of the Study:

• To propose a novel causal estimand that quantifies the modifying effect of an earlier treatment's response on a later treatment's effect.
• To develop a data-adaptive statistical method for estimating this causal parameter.

Main Methods:

• Defined a working marginal structural model based on the conditional average effect of an earlier treatment.
• Developed a targeted maximum likelihood estimator (TMLE) for the causal estimand.
• Utilized a sequentially randomized design for identification and influence curve-based inference.

Main Results:

• The proposed causal estimand allows for estimation of conditional average treatment effects using machine learning without strong assumptions.
• Simulation studies demonstrated the estimator's performance in finite-sample scenarios.
• Analysis of the HIV care retention trial showed that discontinuing cash transfers was most detrimental to those who initially gained the most benefit.

Conclusions:

• The novel causal estimand provides a robust framework for analyzing sequential treatment effects.
• The targeted maximum likelihood estimator offers reliable inference for complex causal questions.
• Findings highlight the importance of considering individual treatment response heterogeneity in HIV care interventions.