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