It is well-known that the multinomial logit (MNL) model is characterised by the independence of irrelevant alternatives (IIA) property. This property implies that MNL is unable to take account of patterns of similarity between alternatives. Nested logit (NL), which was proposed originally by Ben-Akiva (1973) and subsequently derived from random utility theory by Williams (1977) and McFadden (1978), is a restricted form of MNL that offers a (limited) account of IIA violations. This paper builds on a previous empirical approach (Batley, 2000) by presenting a theoretical comparison of NL and PRETREE. A particular focus of the paper is similarity between choice alternatives, and the analysis considers the ways in which the two models seek to capture IIA violations. The paper critiques work by Gensch and Ghose (1997) in marketing research, which demonstrated the apparent inability of PRETREE to capture all IIA violations. It is suggested that Gensch and Ghose's paper is based on a misunderstanding of the concept of IIA. Indeed it is shown that the phenomenon identified by Gensch and Ghose is common to both NL and PRETREE. The paper also challenges the suggested equivalence between NL and PRETREE, identifying flaws in McFadden's (1981) influential study.
Samenvatting