Highlights • Factors affecting the motorcyclist injury severity outcomes were explored. • Motorcycle crashes on rural and urban area were investigated separately. • Correlated random parameters and heterogeneity in means were employed. • Non-transferability was found between rural and urban models. • Policy-related recommendations were discussed base on the modeling results.

Abstract In Thailand in 2016, more than 70% of all deaths due to road accidents were motorcyclist deaths. This study uses a correlated random parameters ordered probit model with heterogeneity in means (CRPOPHM) to obtain insight into differences in the significant factors determining the severity of motorcyclist injury between motorcycle crashes in urban and rural roadways, using data on motorcycle crashes in Thailand from 2016 to 2019. Using a rating system for injury severity level from minor injury to severe injury and to fatal injury, a wide range of potential risk factors are considered, including rider characteristics and actions, roadway characteristics, environmental and temporal characteristics, and crash characteristics. The findings indicate that, although some factors are significant in both urban and rural models, factors such as male rider, illegally overtaking, drowsiness, four-lane or wider highway, flush or depressed median, road on slope, weekend, nighttime with light, crash with van or minibus, and rear-ending or sideswiping crash, are significant only in the rural model, whereas the factors barrier median, occurring between 18:00 and 23:59, and striking a passenger car are statistically significant in only the urban model. These findings further suggests that difference in effect of unobserved characteristics could be seen in different crash locations, and splitting the model estimation between both location types could be done to develop effective guidance for policies to mitigate the severity of motorcyclist injuries. In addition, practical policy-related recommendations drawn from the results of the analysis are provided. With respect to methodology, the proposed CRPOPHM method outperforms lower-ordered models in terms of statistical fit and captures unobserved heterogeneity to a greater extent.