Discuss the potential drawbacks of using median imputation for a numerical feature in a dataset where the missing values are not missing at random. Explain how this could affect the model's ability to generalize to new data and suggest an alternative approach to handle such missing data.