Alternative tree structures for estimating nested logit models with mixed preference data

Abstract

The methods used with data from a single source are inadequate to the challenge of modelling choice behaviour with data from multiple sources. Two distinct formulations, namely the non-normalised nested logit and utility-maximising nested logit models, have been proposed to estimate discrete choice models with mixed revealed preference and stated preference data, in which each data type has the multinomial logit or nested logit form. The article uses two alternative nested logit model formulations to demonstrate how to correctly set up tree structures for estimating nested logit models with mixed preference data. This article provides formulae for recovering correct utility function, dissimilarity and scale parameter estimates. Estimations and correction procedures are empirically illustrated and can be applied to other nested logit models with multiple data sources.