Combining RP and SP data: biases in using the nested logit ‘trick’ – contrasts with flexible mixed logit incorporating panel and scale effects

It has become popular practice that joint estimation of choice models that use stated preference (SP) and revealed preference (RP) data requires a way of adjusting for scale to ensure that parameter estimates across data sets are not confounded by differences in scale. The nested logit ‘trick’ presented by Hensher and Bradley in 1993 [Hensher, D.A., Bradley, M., 1993. Using stated response data to enrich revealed preference discrete choice models. Marketing Letters 4 (2), 139–152] continues to be widely used, especially by practitioners, to accommodate scale differences. This modelling strategy has always assumed that the observations are independent, a condition of all GEV models, which is not strictly valid within a stated preference experiment with repeated choice sets and between each SP observation and the single RP data point. This paper promotes the replacement of the NL ‘trick’ method with an error components model that can accommodate correlated observations as well as reveal the relevant scale parameter for subsets of alternatives. Such a model can also incorporate “state” or reference dependence between data types and preference heterogeneity on observed attributes. An example illustrates the difference in empirical evidence.