Sampling Alternatives in Multivariate Extreme Value Models

This report investigates the use of sampling alternatives in the estimation of Multivariate Extreme Value (MEV) models, particularly focusing on the cross-nested logit (CNL) model. The study addresses the computational challenges associated with discrete choice models that involve a large number of alternatives and attributes, which can lead to high estimation times. By utilizing migration aspiration data, the report assesses various sampling protocols and sample sizes to determine their effectiveness in approximating the parameters of the full choice set model. The findings indicate that importance sampling strategies, which are informed by observed choice frequencies, yield more stable estimates compared to random sampling, even with smaller subsets of alternatives. This research highlights the potential of well-designed sampling methods to significantly reduce estimation time while maintaining accuracy, thereby enhancing the practicality of MEV models for analyzing migration destination choices. The report also outlines the methodology, experimental design, and results of the analysis conducted.