This paper proposes a reference-dependent (RD) model for closing the empirical loop for difference scores. The RD model suggests that testing equations should consider a difference score and its reference component simultaneously. In contrast, Edwards’ model, labeled as the no-discrepancy (ND) model in this paper, only uses its two components in testing equations without involving a difference score. Theoretical and empirical evidence are provided to show that the RD model offers many advantages over the ND model. It may support a difference score as a relevant predictor, while the ND model may reject this difference score. It may test theoretically relevant functional forms that the ND model may not represent. The ND model may recommend a relationship as the joint effects of the two components, while in fact the relationship is simply one of many possible, equivalent joint effects of the two components. Recommendations for future research are discussed.