OBJECTIVES: There are conflicting results about whether using multilevel modeling (MLM) produces different rankings in hospital performance compared to using traditional statistical techniques. We would like to compare hospital rankings obtained from standard logistic regression (LR) relative to MLM modeling of risk-adjusted hospital mortality rates for stroke among hospitals in the south. METHODS: The 2012 U.S. Nationwide Inpatient Sample (NIS) was used to identify patients with a primary diagnosis for ischemic stroke, using ICD-9 diagnosis criteria. Stepwise backward selection technique with logistic regression and multilevel modeling was performed to examine the variation among hospitals in the south by adjusting for patient-level and hospital-level risk factors.  The analysis included hospitals with more than 30 stroke cases in the southern region of the U.S. The hospital IDs have been masked in order to comply with the NIS data user agreement. RESULTS: There were 19,071 stroke hospitalizations in 320 hospitals. The same rankings were observed for top performers in hospitals’ observed/expected ratios (O/E) by LR and MLM.  However, high O/E outlier status differed between the two statistical methods, yielding different rankings for worst performance. Hospital K, L, M were ranked as the worst 3 performers in LR whereas K, N, L were ranked the same in MLM.  O/Es of Hospital K were 4.49 (CI=1.62-9.84) and 3.942 (CI=1.42-8.64) in LR and MLM, respectively. O/E’s of Hospital L were 4.24 (CI=1.90-8.13) and 3.33 (CI=0.70-9.54) in LR and MLM, respectively.  LR identified Hospital M’s O/E ratio as 3.58 (1.47-7.26).  MLM identified Hospital N’s O/E as 3.33 (CI=0.70-9.54). CONCLUSIONS: Although both logistic regression and multilevel modeling produced similar rankings for top performers, the MLM approach was more conservative in its O/E ratio estimates which were shrunken towards the overall mean. The MLM method is recommended to better statistically adjust and avoid false positive identification of outliers.