Imputation Methods: Mean, Median, KNN, MICE Compared

Imputation replaces missing values with estimates so you can keep the row instead of dropping it. The four methods you'll actually meet in practice — mean/median fill, K-nearest-neighbors, and MICE (multiple imputation by chained equations) — trade off bias, honesty about uncertainty, and effort very differently.

How do the main imputation methods compare?

The core trade-off: simple methods distort your data's distribution, while sophisticated methods cost setup time and compute. This table summarizes the practical differences for a numeric column with moderate missingness:

Method	Bias in the mean	Variance distortion	Preserves correlations?	Difficulty
Mean fill	None under MCAR; biased under MAR/MNAR	Severe shrinkage	No — pulls them toward 0	Trivial
Median fill	Same as mean, robust to outliers	Severe shrinkage	No	Trivial
KNN	Low under MAR if features predict the gap	Mild shrinkage	Mostly	Moderate (scaling matters, slow on big data)
MICE	Low under MAR	Honest — adds noise back in	Yes	Highest (iterative, needs pooling across imputations)

Why does mean imputation shrink variance?

Because every imputed cell gets the exact same value — the center of the distribution. Real missing values would have been spread out; replacing them all with one number piles probability mass at the mean, so the column's standard deviation drops and its histogram grows a spike. Correlations weaken too, since the imputed points form a flat horizontal line in any scatterplot.

Concretely: impute 20% of a column with its mean and the variance of that column falls by roughly 20% — which narrows confidence intervals and inflates significance artificially. Your analysis becomes more confident precisely because you invented data with zero spread. This is the main reason mean imputation is fine for a quick ML baseline but risky for statistical inference.

When should I use mean or median fill anyway?

When missingness is small (under ~5%), the column is a minor predictor, and you need something fast and deterministic — typically inside an ML pipeline where predictive accuracy, not inference, is the goal. Prefer median over mean for skewed columns (income, page views) so outliers don't drag the fill value. And pair the fill with a missingness indicator column so the model can still learn from the gap itself:

import pandas as pd

df['income_missing'] = df['income'].isna().astype(int)
df['income'] = df['income'].fillna(df['income'].median())

More fillna patterns — group-wise fills, forward-fill for time series — are covered in pandas fillna for missing data.

How do KNN and MICE work, without the math?

KNN imputation finds the k rows most similar to the incomplete row on its observed columns, then fills the gap with those neighbors' average. A 40-year-old engineer with a missing salary gets a salary typical of other 40-year-old engineers — not the company-wide mean. It respects relationships between columns, but you must scale features first (otherwise one big-range column dominates the distance), and it gets slow on wide or tall data.

MICE goes further: it fills every column's gaps with a regression model trained on all the other columns, then cycles through the columns repeatedly until the fills stabilize. Crucially, run as multiple imputation it produces several completed datasets with deliberately different random draws, you analyze each, and pool the results. That spread across datasets is what keeps your standard errors honest — single-shot imputation of any kind understates uncertainty. Under MAR missingness, pooled MICE estimates are unbiased, which is why it's the default in clinical and social-science research.

When should you NOT impute?

Three situations where imputation is the wrong move. First and most important: never impute the target variable in supervised learning or the outcome in a study — you'd be training the model on answers you made up, and errors in the imputation leak straight into your labels. Drop those rows instead. Second: when a value is missing because it doesn't exist (a "spouse_age" for single people) — that's structural missingness, and the fix is modeling or encoding, not filling. Third: when missingness is heavy (over ~40–50%) in a column of real importance — at that point the imputation model is mostly fiction, and dropping the column or collecting more data is more honest.

Also remember imputation cannot rescue MNAR data: if high values are missing because they're high, every method above will still underestimate them. And whatever you choose, fit the imputer on training data only, then apply it to test data — imputing before the split leaks information.

For the decision process across dropping vs filling, see deletion vs imputation and the broader missing data handling guide.