BACKGROUND: Comparing trauma centers in terms of patient survival is a key element of performance evaluation. The current standard in trauma center profiling is based on Ordinary Logistic Regression (OLR). However, OLR does not take account of the hierarchical structure of trauma systems. Hierarchical Logistic Regression (HLR) accounts for the clustering of patients within hospitals and is therefore more theoretically appropriate. The objective of this study was to evaluate whether HLR generates different profiling results than OLR. METHODS: The study was based on the Quebec Trauma Registry with mandatory participation of all 59 designated trauma centers in the province of Quebec, uniform inclusion criteria, and standardized data collection methods. Trauma profiling was based on adjusted odds ratios, which represent the odds that a patient will die in a specific hospital compared with an "average" hospital. Risk adjustment was performed with the Trauma Risk Adjustment Model score. Hospitals were ranked according to odds ratio, and outliers were identified by comparing each hospital with all other hospitals. Hospital ranks and statistical outliers generated by OLR and HLR were compared. RESULTS: The study population comprised 83,504 patients including 4,731 hospital deaths (5.7%). OLR identified 11 hospitals as statistical outliers whereas HLR flagged only four of these hospitals as outliers. In addition, 54 of 59 hospitals changed ranks and 24 hospitals changed by more than five ranks when HLR replaced OLR. CONCLUSIONS: This study shows that replacing OLR with HLR has an important impact on the results of hospital profiling. Along with the many theoretical advantages of HLR, these results support the adoption of hierarchical modeling as the standard method for trauma center profiling.
BACKGROUND: Comparing trauma centers in terms of patient survival is a key element of performance evaluation. The current standard in trauma center profiling is based on Ordinary Logistic Regression (OLR). However, OLR does not take account of the hierarchical structure of trauma systems. Hierarchical Logistic Regression (HLR) accounts for the clustering of patients within hospitals and is therefore more theoretically appropriate. The objective of this study was to evaluate whether HLR generates different profiling results than OLR. METHODS: The study was based on the Quebec Trauma Registry with mandatory participation of all 59 designated trauma centers in the province of Quebec, uniform inclusion criteria, and standardized data collection methods. Trauma profiling was based on adjusted odds ratios, which represent the odds that a patient will die in a specific hospital compared with an "average" hospital. Risk adjustment was performed with the Trauma Risk Adjustment Model score. Hospitals were ranked according to odds ratio, and outliers were identified by comparing each hospital with all other hospitals. Hospital ranks and statistical outliers generated by OLR and HLR were compared. RESULTS: The study population comprised 83,504 patients including 4,731 hospital deaths (5.7%). OLR identified 11 hospitals as statistical outliers whereas HLR flagged only four of these hospitals as outliers. In addition, 54 of 59 hospitals changed ranks and 24 hospitals changed by more than five ranks when HLR replaced OLR. CONCLUSIONS: This study shows that replacing OLR with HLR has an important impact on the results of hospital profiling. Along with the many theoretical advantages of HLR, these results support the adoption of hierarchical modeling as the standard method for trauma center profiling.
Authors: Mila H Ju; Mark E Cohen; Karl Y Bilimoria; Melissa S Latus; Lisa M Scholl; Bradley J Schwab; Claudia M Byrd; Clifford Y Ko; E Patchen Dellinger; Bruce L Hall Journal: J Am Coll Surg Date: 2014-05-10 Impact factor: 6.113
Authors: Ryan P Merkow; Karl Y Bilimoria; Karen L Sherman; Martin D McCarter; Howard S Gordon; David J Bentrem Journal: J Oncol Pract Date: 2013-02-26 Impact factor: 3.840