A note on specification tests for the multinomial logit model (Q1108729)

Choice probabilities generated by \textit{D. McFadden}'s random utility model [Conditional logit analysis of quantitative choice behaviour, in P. Zarembka (ed.), Frontiers in econometrics (1973)] are both multinomial logit and have the independence of irrelevant alternatives property. Various authors have suggested specification tests of this property. This paper identifies the alternative against which these tests are constructed and obtains classical specification tests. It is shown that both the null hypothesis and local asymptotic power of the classical tests is the same as for a proposed Hausman test. Finally, the discussion throws some light on the nature of the IIA property and on the performance of the logit model in applications where one would expect the IIA property to be violated but where the alternative set is fixed.