Bayes rules for optimally using Bayesian hierarchical regression models in provider profiling to identify...

Area of Science:

• Healthcare analytics
• Biostatistics
• Decision theory

Background:

• Growing use of "hospital report cards" to identify facilities with high mortality rates.
• Bayesian hierarchical models are proposed for provider profiling.
• Misclassification of hospital performance is a known issue, with quantifiable impacts via loss functions.

Purpose of the Study:

• To propose and develop Bayes rules for various loss function families to minimize misclassification costs in hospital report cards.
• To apply these decision rules to a real-world dataset of acute myocardial infarction patients.

Main Methods:

• Development of Bayes rules for generalized 1-0 loss, absolute error loss, and squared error loss functions.
• Application of these rules to a dataset of 19,757 acute myocardial infarction patients across 163 hospitals.

Main Results:

• The number of hospitals identified with high mortality is sensitive to the penalty ratio of false negatives versus false positives.
• The specific choice of loss function family had a limited effect on hospital classification.

Conclusions:

• Hospital report card design can be framed within a decision-theoretic approach to reduce misclassification costs.
• While the loss function choice can influence a small number of hospital classifications, the framework offers a method for cost minimization.