Bounding causal effects with an unknown mixture of informative and non-informative missingness

Area of Science:

• Causal inference
• Biostatistics
• Epidemiology

Background:

• Missing outcomes are common in observational and experimental data.
• The reasons for missing outcomes are often unknown, complicating causal effect estimation.
• Existing methods may not adequately handle informative missingness, where missingness depends on the outcome.

Purpose of the Study:

• To propose bounds on causal effects that account for mixed missingness (informative and non-informative components).
• To develop flexible and statistically robust estimators for these bounds.
• To extend the methodology to causal quantities meaningful in the presence of competing risks.

Main Methods:

• Developed a mixed missingness framework to derive bounds on causal effects.
• Incorporated user-specified sensitivity parameters for bound estimation.
• Utilized influence-function based estimators for non-parametric and machine learning approaches.
• Achieved root-n convergence rates and asymptotic normality.

Main Results:

• Derived bounds on causal effects under various assumptions within the mixed missingness framework.
• Demonstrated the ability of influence-function based estimators to achieve desirable statistical properties.
• Showcased the applicability of the methods to identify and estimate bounds for causal quantities in competing risk scenarios.

Conclusions:

• The proposed bounds and estimation methods offer a robust approach to causal inference with informative missing outcomes.
• The methodology is flexible, accommodating non-parametric and machine learning techniques.
• The study provides a valuable tool for analyzing complex health data, such as the impact of antipsychotics on diabetes risk.