Selecting Predictor Subsets: Considering Validity and Adverse Impact