Strategy to reduce the number of parameters to be estimated in discrete choice models: An approach to large choice sets

Abstract

Most of the problems associated with discrete choices involve a small number of alternatives. In cases with many alternatives, however, in addition to the concerns regarding data collection, the calibration process may be a problem as it is usually computationally expensive. Moreover, the available tools may be restricted concerning a large number of parameters found in these models. Therefore, this article presents a procedure to reduce the number of parameters to be estimated in discrete-choice models using many alternatives without affecting the model's overall quality. The strategy uses the Classification and Regression Tree (CART) algorithm and can only be applied to variables related to individuals. To test the feasibility of the procedure, data from a household survey in the city of Santa Maria (RS, Brazil), prepared for the Municipal Urban Mobility Plan, was used. The model, a Multinomial Logit type, was then applied to predict the choice of urban destinations, and its results were compared to those of the calibration without the proposed procedure. The results obtained showed that the strategy, applied to the study case under predefined criteria, did not cause any losses to the overall quality of the model (fit measures). It was concluded that the procedure proves to be viable for large choice sets as long as it can be complemented by a researcher's knowledge, or information from the literature, regarding the influence of variables on the forecast of the phenomenon under study.