A Comparison of Elasticities Derived from Multinomial Logit, Nested Logit and Heteroscedastic Extreme Value SP-RP Discrete Choice Models

Abstract

Developments in the estimation of discrete choice models which relax elements of the independence of irrelevant alternatives (IIA) property of the multinomial logit model (MNL) provide opportunities to explore the richer behavioural sensitivity of a choice model to changes in the levels of attributes influencing choice. Surprisingly, the literature offers limited evidence on the variations in sensitivity (ie elasticity) as we move from an MNL model based on revealed preference (RP) data, to MNL based on stated preference (SP) data, to combined RP-SP data estimated sequentially and jointly with partial relaxation of the differential variance in the unobserved effects by the ‘nested logit’ method, and then as free variance across all RP and SP alternatives by heteroscedastic extreme value (HEV) estimation. This paper draws on a data set collected in 6 Australian capital cities in 1994 to estimate a series of commuter mode choice models in the presence and absence of two ‘new’ alternatives (light rail and busway systems), to derive matrices of direct and cross point elasticities for travel cost and travel time. The evidence suggests that constraining the variance of the unobserved effects to varying degrees tends to over-estimate the elasticities sufficiently to distort the real behavioural sensitivity of specific attributes influencing choice. Furthermore, we seriously question the usefulness of studies which rely solely on SP data.