Facets of Disparate Impact: Evaluating Legally Consistent Bias in Machine Learning