An Early Stopping Bayesian Data Assimilation Approach for Mixed-Logit Estimation

Abstract

The mixed-logit model is a flexible tool in transportation choice analysis, which provides valuable insights into inter and intra-individual behavioural heterogeneity. However, applications of mixed-logit models in practice are limited by the high computational and data requirements for model estimation. When estimating mixed-logit models on small samples, the Bayesian estimation approach becomes vulnerable to over and under-fitting. This is problematic for investigating the behaviour of specific population sub-groups or market segments, where a modeller may wish to estimate separate models for multiple similar contexts, each with low data availability. Similar challenges arise when adapting an existing model to a new location or time period, e.g., when estimating post-pandemic travel behaviour.In order to address the data and transferability issues of the mixed-logit model, in this paper we propose a new Early Stopping Bayesian Data Assimilation (ESBDA) simulator for estimation of mixed-logit which combines a Bayesian statistical approach with Machine Learning (ML) model estimation methodologies. The ESBDA simulator is intended to improve the transferability of mixed-logit models and to enable the estimation of robust choice models with low data availability. This approach can therefore provide new insights into people's choice behaviour where the traditional estimation of full mixed-logit models was not previously possible due to low data availability, and open up new opportunities for investment and planning decisions support.To assess the performance of the new approach, we benchmark the ESBDA estimator against the Direct Application approach and two reference simulators: (i) a basic hierarchical Bayes (HB) simulator with random starting parameter values; (ii) a Bayesian Data Assimilation (BDA) simulator without early stopping. Two case-studies are used to investigate the relative performance of the simulators in varied contexts.