Feature engineering and preprocessing · insight-0115

Avoid single mean imputation when the analysis is inferential rather than predictive

Situation
A stakeholder wants coefficients, confidence intervals or group comparisons from data with substantial missingness.
Action
Diagnose the missingness mechanism, then use multiple imputation or a model that handles missingness explicitly, and report how results change across approaches.
Reason
A single imputed value is treated as observed, so the uncertainty about what was missing never enters the standard errors.
Boundary
Regression coefficients, causal estimates, survey analysis and KPI comparisons.

Symptom or risk

Mean imputation shrinks variance, weakens correlations and makes intervals too narrow; conclusions look more certain than they are.

For pure prediction, simple imputation plus indicators is often adequate if validated on held-out data.

Multiple imputation needs more modelling and computation, and its validity still depends on MAR-type assumptions.

Treating MNAR data (people skip questions because of the answer) as if any imputation fixes it; sensitivity analysis is needed.

Single imputation underestimates uncertainty; multiple imputation propagates it into estimates (FIMD, chapters 1-2).
Flexible Imputation of Missing Data (2nd edition, online)