Sensitivity analysis for principal ignorability violation in estimating complier and noncomplier average causal...

Area of Science:

• Causal Inference
• Statistical Sensitivity Analysis
• Econometrics

Background:

• Principal ignorability (PI) is crucial for identifying causal effects with noncompliance but is untestable.
• Assessing the sensitivity of effect estimates to PI violations is essential for robust causal inference.
• Noncompliance creates principal strata (compliers, noncompliers) that may differ in unobserved ways.

Purpose of the Study:

• To develop and apply sensitivity analysis methods for the principal ignorability assumption in one-sided noncompliance settings.
• To quantify the impact of potential PI violations on causal effect estimates.
• To provide practical tools for researchers to assess the robustness of their findings.

Main Methods:

• Sensitivity analysis is performed by allowing differences in outcome-mean-given-covariates functions between compliers and noncompliers under the control condition.
• Various sensitivity parameters are used, including odds ratio, generalized odds ratio, mean ratio, and standardized mean difference.
• Techniques are tailored to principal ignorability-based analysis methods like outcome regression, influence function (IF)-based, and weighting methods.

Main Results:

• The study demonstrates how to tailor sensitivity analysis techniques to different PI-based main analysis methods.
• It provides guidance on selecting appropriate ranges for sensitivity parameters.
• Illustrations using the JOBS II study showcase the application of these methods for various outcome types.

Conclusions:

• Sensitivity analysis is a vital tool for evaluating the robustness of causal effect estimates derived under the principal ignorability assumption.
• The proposed methods offer practical approaches for researchers to gauge the potential impact of PI violations.
• Understanding the sensitivity of results enhances the credibility and interpretability of causal findings in the presence of noncompliance.