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.