It is unclear if the overdispersion restriction is causing likelihood convergence issues in the joint model, where premature convergence is indicated via a more negative log likelihood.
This paper underscores the importance of seeking a methodological consensus for analyzing crash decomposition by type