Assessing the sensitivity of methods for estimating principal causal effects

Area of Science:

• Causal inference
• Statistical methodology
• Biostatistics

Background:

• Principal stratification framework allows analyzing treatment effects based on compliance.
• Complier Average Causal Effect (CACE) is a key estimand but its estimation is challenging.
• Existing CACE estimation methods have varied assumptions and limited guidance for researchers.

Purpose of the Study:

• To compare two distinct methods for estimating CACE: a maximum likelihood (joint) method assuming the exclusion restriction (ER) and a propensity score-based method assuming principal ignorability.
• To assess the sensitivity of each method to its own assumptions and the assumptions of the alternative method.
• To provide guidance on selecting appropriate CACE estimation procedures.

Main Methods:

• Detailed examination of the assumptions underlying the maximum likelihood (joint) method and the propensity score-based method.
• Sensitivity analysis using simulated data to evaluate method performance under assumption violations.
• Application to a motivating real-world example.

Main Results:

• The exclusion restriction (ER)-based joint method demonstrated greater robustness to its own assumptions compared to the propensity score method.
• Both methods showed improved performance when strong predictors of compliance were present.
• Each method performed well when the assumptions of the other method were violated.

Conclusions:

• The choice of CACE estimation method should be guided by the likelihood of satisfying its underlying assumptions in a specific research context.
• Strong predictors of principal stratum membership significantly enhance the performance of CACE estimation methods.
• Researchers should carefully consider method-specific assumptions and potential violations when estimating CACE.