Individual level models vs. sample level models: contrasts and mutual benefits

• Published in: Transportmetrica (Abingdon, Oxfordshire, UK) Vol. 11; no. 6; pp. 465 - 483
• Main Authors: Dumont, Jeffrey, Giergiczny, Marek, Hess, Stephane
• Format: Journal Article
• Language: English
• Published: Abingdon Taylor & Francis 03.07.2015; Taylor & Francis Ltd
• Subjects: Bayesian analysis; Decision making; Estimates; Estimating; Heterogeneity; hierarchical Bayes; individual level models; Mathematical models; prediction; Transport
• ISSN: 2324-9935; 2324-9943
• DOI: 10.1080/23249935.2015.1018681

With a view to better capturing heterogeneity across decision makers and improving prediction of choices, there is increasing interest in estimating separate models for each person. Almost exclusively, this work has however taken place outside the field of transport research. The aim of the present paper is twofold. We first wish to give an account of the potential benefits of a greater focus on individual level estimates in transport applications. Secondly, we wish to offer further insights into the relative benefits of sample level and individual level models (ILMs) by drawing on a data set containing an unusually large number (144) of decisions on holiday travel per individual. In addition to comparing existing approaches, we also put forward the use of a novel technique which draws on the relative benefits of both sample level and ILMs by estimating ILMs in a Bayesian fashion with priors drawn from a sample level model. Our results show only limited differences between ILMs and conditionals from sample level models when working with the full set of choices. When working with more realistic sample sizes at the person level, our results suggest that ILMs can offer better performance on the estimation data but that this is a result of overfitting which can lead to inferior prediction performance. Our proposed Bayesian ILM model offers good intermediary performance. The use of best-worst data rather than simple stated choice, as is done commonly in published ILM work, does not lead to major changes to these findings.