Variance-component-based nested logit specifications: Improved formulation, and practical microsimulation of random disturbance terms

Abstract

The initial motivation leading to the results in this paper was a problem most choice modeling researchers may have not considered: how to simulate random disturbance terms from nested logit (NL) models. We develop an approach using results from Cardell (1997), who proved the existence of a probability distribution (C(λ)) that could be used to formulate NL models based on statistically independent variance components. These components can be interpreted as unobserved preference heterogeneity for the choice ‘dimensions’ used to define NL tree structures. Simulation aside, we consider this formulation to have other practical advantages for empirical work, but it does not appear to have penetrated the literature (possibly due to notational obstacles). We use notation from Daly (2001) to implement an equivalent representation, which also establishes mathematical equivalence between Cardell (1997) and other important results in the NL literature.