Identifiability of subgroup causal effects in randomized experiments with nonignorable missing covariates

Area of Science:

• Statistics
• Causal Inference
• Biostatistics

Background:

• Randomized experiments are ideal for causal effect estimation.
• Missing pretreatment covariate data complicates subgroup causal effect analysis.
• Nonignorable missing data mechanisms lead to non-identifiable parameters, yielding wide bounds.

Purpose of the Study:

• To address challenges in estimating subgroup causal effects with missing covariate data.
• To demonstrate the identifiability of causal effects and joint distributions under specific nonignorable missing data mechanisms.
• To evaluate statistical inference performance and apply methods to real-world data.

Main Methods:

• Investigated four interpretable nonignorable missing data mechanisms.
• Developed methods to identify causal effects and joint distributions.
• Conducted simulation studies to assess statistical inference performance.
• Applied methods to a randomized clinical trial dataset and a job-training program dataset.

Main Results:

• Identified causal effects and joint distributions for four nonignorable missing data mechanisms.
• Simulation studies demonstrated the performance of the statistical inference methods.
• Analysis of a randomized clinical trial indicated a nonignorable missing data mechanism fit better than an ignorable one.
• Methods showed potential applicability to observational studies.

Conclusions:

• The study provides a framework for causal effect estimation when pretreatment covariates are nonignorably missing.
• The proposed methods enhance the identifiability of causal effects in challenging data scenarios.
• Findings support the use of specific nonignorable missing data models in analyzing real-world data from randomized trials and observational studies.