The sufficiency criterion has long been used to evaluate the effectiveness of a ground motion intensity measure (IM) in capturing the link between ground shaking and structural response. However, a typical sufficiency-based evaluation of an IM only tests for the possibility of linear dependency and the interaction among the upstream parameters is not considered. To address these and other limitations, two new IM evaluation methodologies are proposed. The first methodology considers the loss of statistical information when an IM is used to predict the engineering demand parameters (EDPs) without including the upstream parameters (i.e., earthquake magnitude, source-to-site distance and epsilon). The best IM is the one that minimizes the loss of predictive performance when it is the only model input relative to when it is used as a predictor together with the upstream parameters. To consider the possible interactive effects, a machine learning model is used when both the IM and upstream parameters are used as inputs. The second methodology uses a causal inference approach where the effect of the IM on the EDP distribution is quantified while considering the earthquake magnitude, source-to-site distance and epsilon as control variables. The double machine learning approach is implemented for this purpose. The two methodologies are applied to a set of five steel specifical moment resisting frames. The results show that the statistical loss-based and causal inferencing approaches produce results that are more conclusive than the sufficiency-based approach and more consistent with the physical laws that govern the IM-EDP relationship.