A practical question that our generalization raises is what does it mean to “remove” features from a data instance that will be input to a model-based decision-making procedure? Multiple solutions have been proposed for dealing with features missing from an instance when applying predictive models (SaarTsechansky and Provost, 2007), 6 such as imputing default values for the missing features, retraining models with only the available features, etc. The generalized framework is agnostic to which method is used to deal with the removed features—taking the position that this decision is domain- and problemdependent. What matters is that the decision may change when some of the features are not present at the time of decision making, and thus the change in the decision can be attributed to these missing features. In the empirical example presented below, we use mean imputation for continuous variables and mode imputation for categorical variables. We chose this approach because it is usually far cheaper than retraining models, and it is one of the most commonly used and recommended techniques to deal with missing values. Saar-Tsechansky and Provost (2007) discuss other alternatives and their implications for model predictions; any of them could be used in conjunction with this framework.