When modeling opioid-related mortality as a crude rate in a linear model, including an AR term significantly improves estimation performance with regard to RMSE. | |
When modeling counts of opioid-related mortality, a negative binomial model performs better than a Poisson model. | |
Linear AR models performed optimally with respect to bias, RMSE, Type I error, and correct rejection rates in the context of estimating state-level policy effects of opioid-related mortality | |
Sample size matters for SE estimation. For linear and log-linear models, clustered SEs significantly improved estimation when the treated group comprised 15+ states, yet they had worse performance than unadjusted SEs in the case of only a single treated state. |