Nested Logit Modelling: Some Hard Facts
MUNIZAGA M A, Universidad de Chile and ORTUZAR J de D, Pontificia Universidad Catblica de Chile, Chile
During the last 25 years the multinomial logit (MNL) and nested logit (NL) models have set the standard for discrete choice modelling and travel demand forecasting. Recently, however, the NL has been the subject of almost passionate discussions at three l
During the last 25 years the multinomial logit (MNL) and nested logit (NL) models have set the standard for discrete choice modelling and travel demand forecasting. Recently, however, the NL has been the subject of almost passionate discussions at three levels: a historical perspective concerning the genesis of the model, a theoretical discussion where correct scaling, proper statistics and even the role of the model constants have been brought into question, and, finally, a general concern about its future in practice with the advent of more powerful forms such as the multinomial probit and random coefficients (or error component) models, which even practitioners (normally years behind the state of the art) are starting to perceive as real alternatives.
We have used MNL and NL models extensively in practice and have recently explored their performance in depth using a series of analytical and empirical tests. Therefore, we consider ourselves to be in good position to shed some light on the above issues. The rest of the paper is organised as follows. In section 2 we present a brief synthesis of the model history of development in order to give appropriate credit to its actual developers. In section 3 we present the NL model formally and then discuss a number of arguments that have been aired in the recent literature. In section 4 we discuss the results of a simulation exercise designed to test the modelÕs performance in comparison to both simpler and more complex model functions, and finally in section 4 we briefly summarise our main conclusions.
Association for European Transport