Observability: Nonrandom Attrition

Chapter 12 — Observability: Nonrandom Attrition Missing outcome data is one of the most common and most damaging threats to experimental integrity. This chapter develops the implications of nonrandom attrition within the potential outcomes framework, covering tests for selective attrition and its determinants, and a toolkit of analytical responses: available case analysis, Horowitz-Manski bounds, inverse probability weighting, selection models, and Lee bounds. It also addresses missing covariates and shows how design choices made before data collection can minimize attrition and preserve the validity of inference.

1. Nonrandom attrition can frustrate causal inference, but under stronger assumptions an internally valid estimate for nonattritors and the study population can be recovered.
2. The implications of these new identifying assumptions are testable, and when they hold, statistical approaches to analyze the data are available.
3. In certain cases, covariates may be missing. While not a threat to internal validity, recovering CATEs or using controls to enhance experimental power are hindered.
4. Design considerations can attenuate missing data concerns and permit tests of attrition.