On the use of mixed RP/SP models in prediction: accounting for random taste heterogeneity