Online Program

An Evaluation of Multilevel Poisson Sample Selection Models

Theodore J. Iwashyna, VA Center for Clinical Management Research, Ann Arbor VA Health Services Research and Development 
*Kyle M. Kepreos, VA Center for Clinical Management Research, Ann Arbor VA Health Services Research and Development 
Anne E. Sales, VA Center for Clinical Management Research, Ann Arbor VA Health Services Research and Development 
Wyndy L. Wiitala, VA Center for Clinical Management Research, Ann Arbor VA Health Services Research and Development 

Keywords: Selection bias, Multilevel Poisson models, Multilevel zero-inflated Poisson models, Multilevel Poisson sample selection models, Empirical Bayes ranking, Monte Carlo methods

Hospital care focuses on improving patients’ long-term quality of life, yet hospital quality metrics typically focus on short-term outcomes. Patients must first survive the hospitalization in order for post-hospitalization outcomes to be observed. Although the bias from failing to consider such selection is known for single-level models, it is not generalized to the multilevel context. The current study’s objective is to evaluate sample selection bias on long-term count outcomes by comparing multilevel poisson models (MPM), multilevel zero-inflated poisson models (MZIP), and multilevel poisson sample selection models (MPSSM). We specifically sought to determine whether MPSSMs produced more accurate parameter estimates and empirical Bayes rankings. We simulated datasets in which our outcome was the post-discharge count of nursing home days, but a pre-specified selection process determined patient survival of hospitalization. Unobserved factors affecting in-hospital survival were correlated with unobserved factors affecting the long-term count outcome. Individual hospital effects were simulated and kept the same throughout all datasets. Preliminary results suggest that the MPSSM produces parameter estimates that are significantly closer to the true pre-specified parameters. Further, the MPSSM produces empirical Bayes ranks that recover the true ranking with greater accuracy. This suggests that when modeling multilevel effects on long-term count outcomes observed after a short-term selection process, both individual and hospital-level effects are more feasibly recovered using MPSSMs.

Acknowledgements: This work was supported by VA HSR&D IIR 11-109. The authors’ opinions do not necessarily represent the views of the US Department of Veterans Affairs or the US Government.