The effect of reference price on brand choice decisions has been well documented in the literature. Researchers, however, have differed in their conceptualizations and, therefore, in their modeling of reference price. In this article, we evaluate five alternative models of reference price of which two are stimulus based (i.e., based on information available at the point-of-purchase) and three that are memory based (i.e., based on price history and/or other contextual factors). We calibrate the models using scanner panel data for peanut butter, liquid detergent, ground coffee, and tissue. To account for heterogeneity in model parameters, we employ a latent class approach and select the best segmentation scheme for each model. The best model of reference price is then selected on the basis of fit and prediction, as well as on the basis of parsimony in cases where the fits of the models are not very different. In all four categories, we find that the best reference price model is a memory-based model, namely, one that is based on the brand's own price history. In the liquid detergent category, however, we find that one of the stimulus-based models, namely, the current price of a previously chosen brand, also performs fairly well. We discuss the implications of these findings.