查看文章

Methodological advances in discrete outcome modelling have largely focused on enabling representation of complex behaviors and improving forecasting accuracies. An important behavioral aspect often investigated is the presence of heterogeneity in the effects of contributory variables and how they vary in between alternative outcomes, across observations, and over discrete outcome events (Yuan et al., 2015). While multinomial-Logit models provide a fundamental basis for discrete outcome analysis under the random utility maximization framework, the basic structure comprises of shortcomings, including the ability to only capture effects that vary systematically with observed variables.

Over the years, research has focused on proposing improvements and flexibilities to the standard multinomial-Logit models to address some of the limitations. Among them, the Mixed-Logit models and Latent class models have by far been the most used to analyze heterogeneity in the effects. Mixed-Logit models use parametric distributions to accommodate effects that are heterogeneous (Vij and Krueger, 2017), offering additional insights regarding the disaggregate discrete processes. However, due to availability of several parametric distributions with different properties, the selection of an appropriate distribution has been recognized as an analyst-intensive task (Keane and Wasi, 2013, Vij and Krueger, 2017, Beeramoole et al., 2023, Paz et al., 2019). In contrasts, latent class models employ non-and semi-parametric distributions to relax some of the limitations of using parametric distributions, such as alleviating the need to prespecify the shape or functional …