A machine learning engineer needs to handle missing values in a dataset containing numerical features. The missingness is completely at random (MCAR). Which imputation strategy is most robust for downstream model performance?