Sampling of Alternatives in MEV Models: Evidence from Migration Aspiration Models with Large Choice Sets

Abstract

Certain case studies require the estimation of discrete choice models with a very large number of alternatives and attributes, which in turn increases the computational cost. This challenge becomes even more pronounced in the family of Multivariate Extreme Value (MEV) models such as the cross-nested logit (CNL) model due to the additional complexity in the model specification. In this study, we investigate the use of sampling of alternatives to make the estimation of MEV models more tractable. Using migration aspiration data, we assess how different sampling protocols and sample sizes approximate the parameters of the full choice set model. Our results show that importance sampling strategies informed by observed choice frequencies outperform random sampling, providing more stable estimates even with small subsets of alternatives. These findings highlight that carefully designed sampling can substantially reduce estimation time without compromising accuracy, thereby making MEV models a practical tool for studying migration destination choice.