Introduction
Hospital benchmarking is central to antibiotic stewardship, but meaningful comparisons of inpatient antibiotic use require patient case-mix adjustment. Reference Moehring, Yarrington and Ashley1 , Reference Goodman, Pineles and Magder 2 Current antibiotic-use benchmarks, including the Standardized Antimicrobial Administration Ratio (SAAR), do not account for patient-level factors. Reference O’Leary, Edwards and Srinivasan3
Moreover, the interpretation of benchmarking results depends on the metric being used. Counting by days of therapy (DOT) assigns equal weight to narrow- and broad-spectrum agents. Tracking days of antibiotic spectrum coverage (DASC) weights antibiotic exposure by each agent’s spectrum of activity, incorporating both duration and spectrum of therapy. Reference Kakiuchi, Livorsi and Perencevich4 However, the appropriate way to evaluate spectrum in a metric like DASC is still unclear.
We recently developed risk-standardized ratio (RSR) metrics for DOT and DASC, defined as each hospital’s predicted antibiotic use (ie, the actual risk-adjusted amount of antibiotic use) divided by its expected antibiotic use (ie, how the hospital would have prescribed antibiotics if, given its patient case-mix, it had prescribed in the same way as the reference population of Veterans administration hospitals). Reference Livorsi, Merchant and Cho5 By adjusting for patient-level factors, these RSR metrics may have greater face validity than the SAAR.
A key variable in these RSR models was discharge diagnostic tier, which classifies diagnoses as those that always, sometimes, or never require antibiotics. Reference King, Tsay, Hicks, Bizune, Hersh and Fleming-Dutra6 In the RSR models, diagnostic tier was treated as a fixed effect, implicitly assuming that the relationship between diagnostic tier and antibiotic use was uniform across hospitals. Reference Livorsi, Merchant and Cho5 But hospitals may differ in types of infections seen, prescribing thresholds, and stewardship practices; thus, tier-related changes in antibiotic utilization intensity may also vary by hospital. Random-slope models can evaluate such between-hospital heterogeneity by allowing the association between diagnostic tier and antibiotic use to vary across hospitals. Reference Snijders and Bosker RJ7
In this study, we examined whether the relationship between diagnostic tier and predicted DOT per days present (DOT/DP) and DASC per days present (DASC/DP) varied across hospitals by modeling tier as a hospital-specific random slope.
Methods
We conducted a retrospective cohort study using the Veterans Health Administration (VHA) administrative data from fiscal year 2021. The unit of analysis was the inpatient ward-stay, or a continuous period of time during which a patient resided on a specific hospital unit (eg, intensive care unit, medical ward); admissions could contribute multiple ward-stays. The analytic cohort included 737,384 ward-stays from 538,241 admissions across 122 VHA hospitals. The University of Iowa and Iowa City Veterans Health Care System Institutional Review Board approved this study and waived informed consent because the analysis used existing data.
Each ward-stay was assigned the discharge diagnostic tier of its overarching admission. Diagnostic tiers classified discharge diagnoses as Tier 1 (always warrant antibiotics), Tier 2 (sometimes warrant antibiotics), or Tier 3 (never require antibiotics). Reference King, Tsay, Hicks, Bizune, Hersh and Fleming-Dutra6 When multiple diagnoses were coded, the lowest-numbered tier was assigned. Outcomes were total DOT and DASC during each ward-stay. DP, defined as the number of days in each ward-stay, was used to account for duration of hospital exposure.
We fit separate zero-inflated negative binomial mixed-effects models for DOT and DASC because many ward-stays had zero antibiotic use. Diagnostic tier was modeled as a categorical variable. To evaluate between-hospital heterogeneity in tier-related antibiotic use, the primary models added hospital-specific random slopes for diagnostic tier to both the count and zero-inflated components, along with a hospital random intercept and an offset for log(DP). Otherwise identical random-intercept-only models that included diagnostic tier as a fixed effect served as comparators. For each hospital and diagnostic tier, model-predicted DOT and DASC were summed across ward-stays and divided by summed DP to estimate predicted DOT/DP and DASC/DP. Hospital-level estimates were summarized using means, standard deviations, and percentiles. Hospital-specific trajectories across diagnostic tiers were visualized for each metric. Predicted-to-expected (P/E) ratios were calculated as the sum of hospital-specific predicted DOT or DASC divided by the sum of expected values for the same ward-stays under a standardized hospital framework, analogous to the RSR metrics. Reference Livorsi, Merchant and Cho5 Overall and tier-specific P/E ratios and hospital quartile rankings were compared between the random-intercept-only comparator models and the primary models with hospital random intercepts and hospital-specific random slopes for diagnostic tier. Fixed-effect covariates used in both models are summarized in Supplementary Tables S1 and S2.
Results
Predicted DOT/DP and DASC/DP decreased stepwise across tiers in both models (Supplementary Table S3). In random-slope models (Figures 1A and 1C), mean predicted DOT/DP was 0.84, 0.48, and 0.15 for Tiers 1, 2, and 3, respectively; corresponding DASC/DP values were 11.20, 5.49, and 1.74. In random-intercept-only models (Figures 1B and 1D), mean predicted DOT/DP was 0.83, 0.48, and 0.15, and DASC/DP was 11.10, 5.43, and 1.75, respectively. Mean predicted values were therefore very similar between model structures. At the overall hospital level, adding hospital-specific random slopes did not change any hospital’s quartile classification by more than one quartile compared to the random-intercept-only models (Figure 2A and 2B). In contrast, tier-specific comparisons across hospitals showed larger changes in quartile classification, including shifts of two or more quartiles for some hospitals (Figures 2C and 2D).
Predicted DOT/DP and DASC/DP across diagnostic tiers by model structure. Panel A shows predicted DOT/DP from the model with hospital random intercepts and hospital-specific random slopes for diagnostic tier; panel B shows predicted DOT/DP from the random-intercept-only model. Panel C shows predicted DASC/DP from the random-slope model; panel D shows predicted DASC/DP from the random-intercept-only model. Each line represents one VHA medical center. DASC, days of antibiotic spectrum coverage; DOT, days of therapy; DP, days present; VHA, Veterans Health Administration.
Figure 1. Long description
Panel A: A line graph shows predicted DOT/DP at Medical Center from the model with hospital random intercepts and hospital-specific random slopes for diagnostic tier. The x-axis is labeled ‘Final Tier’ with values 1, 2, and 3. The y-axis is labeled ‘Predicted DOT/DP at Medical Center’ with values ranging from 0.0 to 1.0. Each line represents one VHA medical center. Panel B: A line graph shows predicted DOT/DP from the random-intercept-only model. The x-axis is labeled ‘Final Tier’ with values 1, 2, and 3. The y-axis is labeled ‘Predicted DOT/DP at Medical Center’ with values ranging from 0.0 to 1.0. Each line represents one VHA medical center. Panel C: A line graph shows predicted DASC/DP at Medical Center from the random-slope model. The x-axis is labeled ‘Final Tier’ with values 1, 2, and 3. The y-axis is labeled ‘Predicted DASC/DP at Medical Center’ with values ranging from 0 to 15. Each line represents one VHA medical center. Panel D: A line graph shows predicted DASC/DP from the random-intercept-only model. The x-axis is labeled ‘Final Tier’ with values 1, 2, and 3. The y-axis is labeled ‘Predicted DASC/DP at Medical Center’ with values ranging from 0 to 15. Each line represents one VHA medical center.
Overall and tier-specific P/E ratios from random-intercept-only and random-slope models for DOT and DASC. Panels A and B show overall P/E ratio comparisons for DOT and DASC, respectively. Panels C and D show tier-specific P/E ratio comparisons for DOT and DASC, respectively. Each point represents one VHA medical center; in tier-specific panels, each point represents one medical center within a diagnostic tier. Point color indicates the absolute change in hospital quartile ranking between models, calculated as the difference between the quartile assigned using the random-intercept-only model and the quartile assigned using the random-slope model. P/E, predicted-to-expected; DASC, days of antibiotic spectrum coverage; DOT, days of therapy; VHA, Veterans Health Administration.
Discussion
In this national VHA cohort, predicted DOT/DP and DASC/DP decreased from Tier 1 to Tier 3, supporting the expected clinical ordering of diagnostic tiers. Reference Livorsi, Merchant and Cho5,Reference King, Tsay, Hicks, Bizune, Hersh and Fleming-Dutra6 Compared with models with only a random intercept, models with hospital-specific random slopes allowed the magnitude of these tier-related decreases to vary modestly across hospitals. However, adding random slopes had little effect on overall hospital-level performance, suggesting that the prior random-intercept framework remains broadly similar for overall benchmarking. The main added value of the random-slope approach appears to be in revealing heterogeneity across hospitals in their intensity of antibiotic use within each tier.
The observed between-hospital variation in tier-related antibiotic use should not be interpreted solely as modifiable differences in prescribing practice. It may also reflect residual heterogeneity within the broad 3-tier diagnostic classification, differences in coding practices, or other unmeasured case-mix differences. There may therefore be a need to use data more granular than tiers 1–3 to accurately adjust for discharge diagnoses. For example, the diagnoses included in tier 1 run the gamut of bacterial infections seen in hospitalized patients, some of which are treated with short courses and/or narrow-spectrum antibiotics while others require prolonged courses and/or broad-spectrum antibiotics. Further work is needed to better understand the optimal approach to risk-adjustment.
Our findings suggest that hospital-specific tier effects may be useful not only as model components but also as descriptive outputs when reported alongside overall RSR metrics. The RSR-DOT and RSR-DASC models summarize each hospital’s predicted antibiotic use relative to expected use under a standardized hospital framework. Reference Livorsi, Merchant and Cho5 If a random-slope model were summarized only as a single hospital-level RSR, heterogeneity across tiers 1–3 would be incorporated into the overall estimate. Reference Livorsi, Merchant and Cho5 Displaying hospital-by-tier predicted DOT/DP and DASC/DP can make this heterogeneity more explicit.
This study has limitations. VHA hospitals serve an older, predominantly male population, which limits generalizability. Diagnostic tiers were based on administrative discharge diagnoses and may be affected by coding practices.
Conclusions
Incorporating hospital-specific random slopes for discharge diagnostic tier into RSR metrics showed that predicted antibiotic metrics decreased from Tier 1 to Tier 3 while also capturing between-hospital differences in antibiotic use intensity within each tier. However, adding random slopes had little effect on overall hospital-level performance and provided similar results for benchmarking compared to our prior approach.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/ice.2026.10561
Data availability statement
Study protocol and analytical code can be made available upon request to the authors. Study data cannot be shared per the data usage agreement and data security protocol of the US Department of Veterans Affairs.
Acknowledgments
DL had full access to all data in the study and takes responsibility for the integrity and accuracy of the data analysis. SI used ChatGPT 5.5 (OpenAI, San Francisco, CA, USA) for grammar and typographical error checking of this manuscript.
Author contribution
The authors reviewed and edited the output as needed and take full responsibility for the content of the manuscript.
Financial support
This work was supported by an investigator-initiated research grant (I01 HX003595) from the US Department of Veterans Affairs Health Services Research and Development Service (PI: D. Livorsi and M. Goto).
Competing interests
All authors report no relevant conflicts of interest.
Open access
