Direct effect models.

Area of Science:

• Causal inference
• Epidemiology
• Biostatistics

Background:

• Causal effects are often mediated by intermediate variables.
• Understanding direct effects is crucial for mechanistic insights and interventions.
• Existing definitions of natural direct effects require assumptions beyond sequential randomization.

Purpose of the Study:

• To provide an alternative counterfactual definition of the natural direct effect.
• To develop an estimation approach for the natural direct effect based solely on the sequential randomization assumption.
• To propose novel estimators for direct effect modeling.

Main Methods:

• Developed an alternative counterfactual definition for natural direct effect.
• Proposed a novel modeling approach for direct effect estimation.
• Introduced inverse probability of censoring weighted (IPCW), double robust IPCW, likelihood-based, and targeted maximum likelihood-based estimators.

Main Results:

• The proposed definition of natural direct effect is identifiable under the sequential randomization assumption.
• The novel estimation approach allows direct modeling of the natural direct effect.
• Various robust and efficient estimators are presented for the causal model parameters.

Conclusions:

• This work offers a new definition and estimation strategy for natural direct effects in causal inference.
• The methods presented relax stringent identifiability assumptions, enhancing applicability.
• The proposed estimators provide robust tools for analyzing direct causal pathways.