Accommodating for systematic and unobserved heterogeneity in panel data: Application to macro-level crash modeling

Abstract

The current research contributes to the burgeoning literature on multivariate models by proposing a hybrid model framework that (a) incorporates unobserved heterogeneity in a parsimonious framework and (b) allows for additional flexibility to accommodate for observed/systematic heterogeneity. Specifically, we estimate a Latent Segmentation Panel Mixed Negative Binomial (LPMNB) model to study the zonal level crash counts across different crash types. Further, we undertake a comparison exercise of the proposed hybrid LPMNB model with a Panel Mixed Negative Binomial model (PMNB) that accommodates for unobserved heterogeneity via a simulation setting. The analysis is conducted using the zonal level crash records by different crash types from Central Florida region for the year 2016 considering a comprehensive set of exogenous variables. The comparison exercise is further augmented by computing several goodness of fit measures along with elasticity analysis and the results offered by the LPMNB model highlight the value of the proposed model. Further, to offer insights on model selection incorporating computational complexity dimension along with other important attributes, we conduct a trade-off analysis considering four different attributes: (a) model fit, (b) prediction, (c) inference power and (d) computational complexity; across six different model strictures including traditional crash frequency models and our proposed LPMNB model.