What is the primary risk of using mean imputation on a feature that exhibits a non-random missingness mechanism?

Question

Data Science — Hard

What is the primary risk of using mean imputation on a feature that exhibits a non-random missingness mechanism?

Accepted Answer

Mean imputation assumes data is Missing Completely at Random (MCAR). If data is Missing Not at Random (MNAR), the missing values are systematically different from observed values. Imputing the mean ignores this dependency, leading to biased parameter estimates. While variance is reduced (not increased), the primary concern in predictive modeling is the systematic bias introduced by failing to account for the mechanism of missingness.