Keywords: Two-part models, hospital benchmarking, antibiotic days
Despite attention to antibiotic stewardship, few measures exist to equitably compare antibiotic use between hospitals. Variation exists due to treatment indication: not every patient should receive antibiotic therapy, and if given, antibiotics are administered for different amounts of time. To compare hospital antibiotic use after adjusting for this known variation, we used a two-part modeling approach. Using registry data from 29 Michigan trauma centers, we created a two-part model to account for both the probability of any antibiotic use (first part: logistic regression) as well as duration of usage (second part: negative binomial distribution), adjusting for injury severity, presence/type of infection, comorbidity, etc. We performed observed-to-expected adjustments to calculate hospitals' risk-adjusted antibiotic days, bootstrapped O/E ratios to create confidence intervals, and flagged potential high outliers as hospitals whose CIs lay above the overall mean. Using two-part models, we were able to objectively benchmark hospital antibiotic day use, and conclude that two-part models can be a useful tool for analysts comparing count outcomes with large zero mass across hospitals.