Evaluating different strategies for estimating treatment effects in observational studies

Area of Science:

• Epidemiology
• Biostatistics
• Observational Data Analysis

Background:

• Estimating treatment effects from observational data is challenging due to confounding.
• Propensity score (PS) methods and other statistical approaches have advanced, but guidance on optimal strategies is lacking.

Purpose of the Study:

• To evaluate and compare various statistical methods for controlling confounding in observational studies.
• To provide guidance on selecting appropriate analytic strategies for different research scenarios.

Main Methods:

• Conducted extensive simulations using tree-based and smooth regression models.
• Evaluated established methods (regression, PS weighting, stratification, matching) and newer approaches (tree-based, local control, entropy balancing, genetic matching, prognostic scoring).
• Assessed numerous analysis strategies by combining different treatment choice and outcome models.

Main Results:

• No single analytic strategy consistently outperformed others across all simulated scenarios.
• Properly addressing interactions in treatment choice and/or outcome models significantly impacts results.
• A tree-structured treatment choice model combined with a polynomial outcome model (including second-order interactions) demonstrated strong performance.

Conclusions:

• The optimal method for estimating treatment effects in observational data depends on the specific scenario.
• Researchers must carefully consider and model interactions to improve the accuracy of confounding control.
• Further research is needed to explore heterogeneous treatment effect scenarios.