OBJECTIVES: Clinical risk-adjustment is important in evaluating hospital performance. However, the choice of risk-adjustment models may impact evaluation results. This analysis of the Hospital Quality and Service Recognition (HQSR) program evaluated the performance of 15 hospitals in a regional managed care organization. Two common estimating procedures were used to determine differences between hospitals in risk-adjusted complication rates and length of stay. METHODS: The HQSR program employs a multi-dimensional scoring algorithm that includes, among other measures, clinical complications and length of stay (LOS) of common maternity and surgery inpatient episodes. Both a fixed-effect model and a random-effect model were applied to calculate risk-adjusted complication rates and risk-adjusted LOS, by making the patient case-mix constant for all hospitals. The fixed-effect model treated patients from different hospitals as distinct groups, and estimated the hospital effect in a traditional regression framework. The random effect (mixed) model, on the other hand, assumed hospitals were sampled from a normally distributed population, and estimated effects based on an empirical Bayesian method. RESULTS: For maternity, mean risk-adjusted complication rates by fixed-effect and random-effect models were 9.02% (S.D.=4.6%) and 8.60% (S.D.=3.60%), and the mean risk-adjusted LOS by fixed and random-effect models were 2.47 days (S.D.=0.22) and 2.42 days (S.D.=0.14) respectively. The fixed-effect model estimates were closer to the true mean rates (10.04% and 2.47 days), with higher standard deviations. The two models resulted in similar hospital ranks for both measures, but the fixed-effect model showed greater variability in hospital scores. CONCLUSIONS: Risk-adjustment methods with different underlying assumptions give different results and scores. Although no gold standard exists for empirical model selection, the normality assumption underlying the random effect model may underestimate the difference among hospitals.