Which imputation strategy is most appropriate for a feature exhibiting Missing Not At Random (MNAR) behavior?

Question

Select the correct answer: Single Choice

Accepted Answer

When data is MNAR, the missingness is related to the unobserved value itself. Simple imputation (mean/median) or even complex methods like MICE assume the data is Missing At Random (MAR). Creating an indicator variable explicitly captures the 'missingness' as a feature, allowing the model to learn the systematic bias inherent in the data collection process, which is the only way to mitigate MNAR bias without external data.