Choosing a Method and Its Impact

Choosing the right strategy

Situation	Suggested approach
Under 5% missing, MCAR	Drop rows or use simple imputation
Numeric column with outliers	Median imputation
Categorical column	Mode or an "Unknown" category
Time series	Forward/backward fill or interpolation
Strong correlations between features	KNN or iterative imputation
Missingness carries meaning (MNAR)	Impute and add a missing-indicator
Over 60% missing in a column	Consider dropping the column
Missing target value	Drop the row

Comparing the methods

Method	Speed	Keeps sample size	Uses other columns	Risk
Drop rows	Fast	No	No	Bias, data loss
Mean / median / mode	Fast	Yes	No	Reduced variance
Forward / backward fill	Fast	Yes	No	Wrong if gaps are long
Interpolation	Fast	Yes	No	Assumes a pattern
KNN / Iterative	Slower	Yes	Yes	Complexity, overfitting

Impact of handling missing values well

• Improved data quality — a cleaner dataset is more reliable for analysis and training.
• Better model performance — models train on complete data and predict more accurately.
• Preserved data integrity — consistent, well-documented values for later analysis.
• Reduced bias — patterns reflect reality more closely.

Golden rules

1. First find out why the data is missing.
2. Convert placeholders like -999 to NaN before anything else.
3. Learn imputation values from the training set only to avoid leakage.
4. Put the imputer inside a Pipeline.
5. Document what you did and why.