Understanding how different alternatives for dealing with missing features may affect explanations is another interesting direction for future research. For example, if features are correlated, mean imputation and retraining the model without the removed feature may produce different results. For instance, a decision may change when imputing the mean for a removed feature, but if a new model without the feature is trained, the same decision may not change when removing the feature if the remaining features capture most of the information given by the removed feature. Therefore, while our proposed framework can work with either approach, future research should discuss the advantages of each approach in various settings.