Article summary
Datasets built from text, images, merged surveys, and voter files often lack key covariates because those features are latent (for example, sentiment in text) or simply not collected (for example, race in voter files).
Datasets built from text, images, merged surveys, and voter files often lack key covariates because those features are latent (for example, sentiment in text) or simply not collected (for example, race in voter files).
This note compares multiple imputation and listwise deletion using simulations that focus on data missing not at random (MNAR). Both approaches are known to produce biased estimates under MNAR; the simulations investigate how their relative performance changes across realistic scenarios.