Setting the context for behaviourally more plausible models

Many applications in marketing, transport, and the environment use the simple multinomial (MNL) logit model presented in chapter 3. This approach is common to studies using stand-alone stated preference (SP) or revealed preference (RP) data, as well as cases with multiple data sets, such as combined SP and RP data (see chapter 8). A great majority of empirical studies go no further than this. Some studies progress to accommodating some amount of difference in the structure of the random component of utility, through a nested logit (NL) model. The NL model partitions the choice set to allow alternatives to share common unobserved components among one another compared with a non-nested alternative.

Despite the practitioner's support for the MNL model and occasionally for the NL model (the latter being the main focus of this chapter), much research effort is being devoted to increasing the behavioural realism of discrete-choice models. This effort is concentrated on relaxing the strong assumptions associated with IID (independent and identically distributed) error terms in ways that are behaviourally enriching, computationally tractable and practical. Choice models are now available in which the identically distributed structure of the random components is relaxed (e.g., Bhat 1995, 1997b, Hensher 1997b). Extensions that permit non-independence between alternatives, such as mixed logit (ML) and multinomial probit (MNP) models, have also been developed, adding further behavioural realism but at the expense of additional computational complexity (see Greene 1997; Geweke, Keane and Runkle 1994; Bolduc 1992; Daganzo 1980, McFadden and Train 1996; Brownstone, Bunch and Train 1998).