Missing Value Imputation & MICE

Missing Value Imputation Strategies & MICE (Multiple Imputation by Chained Equations) establish the mathematical decision hierarchy treating missing data based on Rubin's three foundational mechanisms: 1) MCAR (Missing Completely at Random); 2) MAR (Missing at Random, conditionally dependent on observed features); 3) MNAR (Missing Not at Random, dependent on the missing value itself, requiring Missingness Indicator binary masks); algorithmic spectrum spans: Univariate baselines (mean, median, mode, sentinel constants like -999) $\to$ Native Tree Branching (LightGBM/XGBoost learning default split directions for NaN) $\to$ MICE (Iterative Imputation running chained round-robin regressions across feature columns until convergence).

Financial credit score imputation, healthcare clinical diagnostic records, and dense tabular ML modeling.

Dropping rows with missing entries discards 50%+ of training samples; naive constant imputation distorts feature covariance; MICE preserves multi-variable dependencies and statistical variance.

Contrast MCAR, MAR, and MNAR missing mechanisms, explaining why Missingness Indicators are mandatory under MNAR?

Detail the step-by-step round-robin algorithm of MICE and how it handles mixed continuous and categorical feature types?

Explain how gradient boosting trees (XGBoost/LightGBM) compute optimal default split directions for missing values at each node?

Why does time-series imputation prohibit Backward Fill and mandate Forward Fill (LOCF) to eliminate future lookahead leakage?

How to bundle trained imputation transformers into serialization pipelines ensuring online features inherit train-set parameters?