Doubly Robust Estimation for Time‑Varying Treatments
MLA Citation
Summary
This seminal paper by Bang and Robins introduces a doubly robust estimator for causal effects in longitudinal studies with time-varying treatments. The authors address a critical challenge in causal inference: when estimating the effect of a treatment that changes over time (like medication dosage or job training participation), researchers must account for time-varying confounders that are themselves affected by prior treatment.
The key innovation is an estimator that combines two approaches: inverse probability weighting (IPW) and g-computation. The estimator is "doubly robust" because it yields consistent estimates if either the model for treatment assignment (propensity score) or the model for the outcome is correctly specified. This property provides important protection against model misspecification.
The authors demonstrate their method through simulations and a real-world example: estimating the effect of highly active antiretroviral therapy (HAART) on CD4 cell count in HIV patients. They show that their doubly robust estimator performs well even when one of the two models is misspecified, while traditional methods fail under the same conditions.
This paper has become foundational for causal inference in longitudinal settings, influencing subsequent work in epidemiology, economics, and social sciences. It provides a practical solution to the complex problem of time-varying confounding, where standard methods like regression or matching can produce biased results.
Key Contributions
-
Introduces doubly robust estimation for causal effects in longitudinal studies with time-varying treatments
-
Combines inverse probability weighting (IPW) and g-computation into a single estimator
-
Provides protection against model misspecification - consistent if either treatment or outcome model is correct
-
Addresses time-varying confounding where confounders are affected by prior treatment
-
Demonstrates practical application with HIV treatment data showing robustness to misspecification
-
Establishes foundation for modern causal inference methods in longitudinal settings