Full Bayesian Hierarchical Logistic Regression in ACS NSQIP: Advancing Surgical Quality Benchmarking

ACS NSQIP benchmarks 30-day postoperative outcomes using hierarchical logistic regression with empirical Bayes smoothing, but this framework can be computationally burdensome at current program scale and may be less robust in sparse-data settings.

Using 2024 ACS NSQIP data, 14 "All Cases" outcome models were fit in 963,565 patients from 656 hospitals with matched generalized linear mixed model-empirical Bayes (GLMM-EB) and full Bayesian (FB) hierarchical logistic regression specifications using identical predictors and hospital random intercepts. Comparisons focused on hospital odds ratios, outlier classification, random-intercept variance, and computation time. Secondary analyses examined 4 lower-volume mortality models.

FB and GLMM-EB yielded highly concordant hospital estimates across 14 primary models, with mean correlation of log-odds 0.9958 and mean slope 0.9759. Concordance for outlier detection was high, with mean Jaccard indices of 0.8849 for high outliers and 0.8174 for low outliers. Mean random-intercept variance was slightly higher with FB than GLMM-EB (0.1813 vs 0.1763). Mean processing time was markedly lower with GPU-accelerated FB than GLMM-EB (5.32 vs 121.80 minutes). FB additionally provided posterior probabilities for high- and low-outlier status and supported Bayesian false discovery rate control. Differences between methods increased in the 4 smaller mortality models.

In ACS NSQIP benchmarking, modern FB hierarchical logistic regression produced hospital performance estimates similar to GLMM-EB while substantially reducing computation time and extending probabilistic interpretability. These findings support transition to an FB framework for scalable surgical quality benchmarking.