Design-Based Causal Inference with Missing Outcomes: Missingness Mechanisms, Imputation-Assisted Randomization Tests,...

Area of Science:

• Causal Inference
• Statistics
• Experimental Design

Background:

• Design-based causal inference offers strong validity through study design, avoiding distributional assumptions.
• Outcome missingness is a significant challenge in applying design-based causal inference.
• Existing methods may struggle with complex missingness mechanisms or model misspecification.

Purpose of the Study:

• To systematically address outcome missingness in design-based causal inference.
• To develop a flexible framework for randomization tests with missing outcomes.
• To ensure finite-population-exact type-I error control under various missingness conditions.

Main Methods:

• Proposed a general outcome missingness mechanism for finite-population-exact randomization tests.
• Introduced an "imputation and re-imputation" framework for handling missing outcomes.
• Extended the framework for covariate adjustment and confidence region construction.

Main Results:

• The proposed framework ensures finite-population-exact type-I error rate control.
• Robustness demonstrated even with misspecified imputation models, unobserved covariates, or interference.
• Successful application in simulations and a large-scale randomized experiment.

Conclusions:

• The "imputation and re-imputation" framework effectively handles missing outcomes in design-based causal inference.
• Achieves finite-population-exact type-I error control, enhancing statistical rigor.
• Provides a robust method for covariate adjustment and confidence intervals with missing data.