Understanding how imputation choices distort distributions and leak information across data splits.
You have a right-skewed feature with ~20% values missing not at random (MNAR), and you plan cross-validated model training. Which approach best avoids both distributional distortion AND data leakage?