A framework for causal estimand selection under positivity violations.

Area of Science:

• Observational studies
• Causal inference
• Health policy analysis

Background:

• Estimating causal effects using observational data presents challenges due to covariate distribution imbalances and lack of overlap between treated and control groups.
• Existing methods like inverse probability weighting (IPW) and overlap weighting (OW) involve tradeoffs between statistical bias and variance, and target different populations.

Purpose of the Study:

• To propose a framework for navigating the tradeoffs between bias and variance in causal effect estimation from observational data.
• To introduce a bias decomposition and metrics for selecting appropriate estimands based on research preferences.
• To aid researchers in balancing the preservation of the original research population with the reduction of statistical bias.

Main Methods:

• Developed a bias decomposition framework to differentiate between statistical bias and estimand mismatch.
• Proposed two design-based metrics to quantify tradeoffs.
• Introduced an estimand selection procedure incorporating domain-specific preferences.

Main Results:

• The proposed framework and procedure effectively illustrate the tradeoffs between bias and variance.
• The methodology allows for informed selection of estimands based on preferences for population preservation or bias reduction.
• Demonstrated the application of the framework using right heart catheterization data.

Conclusions:

• The framework provides a structured approach to address challenges in causal inference with observational data, particularly concerning limited overlap and covariate imbalance.
• Researchers can utilize the proposed metrics and selection procedure to make informed decisions about estimand targeting, optimizing for their specific research goals.
• This work enhances the reliability and interpretability of causal effect estimates derived from observational health studies.