MIMICKING COUNTERFACTUAL OUTCOMES TO ESTIMATE CAUSAL EFFECTS

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Observational studies often involve time-dependent treatments that adapt to patient covariates.
• Traditional models like time-dependent Cox models struggle to estimate net treatment effects in such dynamic settings.
• Structural nested models offer a counterfactual framework but often assume deterministic relationships.

Purpose of the Study:

• To develop a method for estimating net treatment effects in continuous-time observational studies with dynamic treatment regimes.
• To relax the deterministic assumption in continuous-time structural nested models.
• To provide a robust approach for causal inference in complex longitudinal data.

Main Methods:

• Constructing random variables, as solutions to a differential equation, to mimic counterfactual outcomes.
• Utilizing these "mimicking" variables within the structural nested model framework.
• Validating the distribution of mimicking variables against observed data, even with past covariates.

Main Results:

• Demonstrated that counterfactuals in continuous-time structural nested models can be accurately mimicked by random variables.
• Proved that this mimicking approach allows parameter estimation without assuming deterministic counterfactuals.
• Established a method to estimate net treatment effects even when treatment is continuously adapted to covariates.

Conclusions:

• The proposed method of mimicking counterfactuals provides a flexible and powerful tool for causal inference in complex observational studies.
• This approach overcomes key limitations of existing structural nested models, particularly regarding deterministic assumptions.
• Enables more reliable estimation of treatment effects in dynamic, real-world health scenarios.