Ranking USRDS provider specific SMRs from 1998-2001.

Area of Science:

• Health services research
• Biostatistics
• Health informatics

Background:

• Provider profiling, including ranking and percentiling, is a common practice in health services research.
• Bayesian models offer a robust framework for complex inferences like provider rankings.
• Accurate provider assessment requires methods that account for uncertainty.

Purpose of the Study:

• To outline and demonstrate a Bayesian approach for provider profiling using Standardized Mortality Ratios (SMRs).
• To extend existing methods by minimizing classification errors and incorporating multi-year data.
• To evaluate the performance of different estimation and ranking strategies.

Main Methods:

• Utilized Bayesian models to compute ranks and percentiles based on Standardized Mortality Ratios (SMRs) from 1998-2001.
• Applied methods minimizing classification errors and combined evidence over multiple years using an autoregressive model.
• Incorporated a nonparametric prior and assessed performance against maximum likelihood estimates.

Main Results:

• Optimal Bayesian estimates significantly outperformed traditional methods (maximum likelihood, SMR=1 testing) in accuracy.
• Combining multi-year data via autoregressive modeling reduced uncertainty and improved percentile accuracy.
• Percentiles derived from posterior probabilities closely matched those from loss-minimizing methods.

Conclusions:

• Loss function-guided percentiles are crucial for accurate provider profiling.
• Bayesian approaches, particularly those combining evidence over time, enhance ranking reliability.
• Robust uncertainty assessments are essential for interpreting provider performance metrics.