Abstract 316: Using Reference Effect Measures to Identify Sources of Variation in 30-day Readmissions for Percutaneous Coronary Intervention

Background: Multilevel models for non-normal outcomes are widely used in outcomes research to estimate effects of covariates on outcomes, e.g. hospital readmissions following percutaneous coronary intervention (PCI). We refine Reference Effect Measures (REM) to compare effects of individual covariates, sets of covariates and random effects on the same scale (e.g. odds ratio, OR) and present a novel approach for displaying these effects. We illustrate this method by studying these sources of variation in 30-day readmission rates for Veterans Administration (VA) patients undergoing PCI. Methods: We used mixed effects logistic regression with 13 patient and 3 hospital covariates to study 30-day readmission rates in a national cohort of 45,521 VA Clinical Assessment, Reporting and Tracking (CART) patients who received a PCI during 10/2007-9/2012 at 49 VA hospitals. OR was used as a REM to compare percentiles of hospital or patient risk with median hospitals or patients to assess levels of variation. Results: Overall 30-day readmission rate was 11.5% ranging from 6.8% to 17.3% across the 49 sites. The figure below shows effects of individual patient and hospital covariates, combined effects of patient and hospital characteristics (shaded bars), and hospital random variation (shaded curve) for 30-day readmissions. The OR for comparing a 97.5th percentile hospital with a median hospital, all covariates being equal, was 1.43, which was substantially less than the 2.88 OR found when comparing patients at the same percentile of the combined patient risk distribution. The largest patient covariate effect (congestive heart failure) had an OR of 1.58, equivalent to comparing a 99th percentile hospital with a median hospital, or an 81st percentile risk patient to a median risk patient. Combining all hospital characteristics, the OR for a 97.5th percentile hospital versus a median hospital was 1.16. Conclusions: REMs are simple to compute, interpret and graph, and provide direct comparison on the same scale across random effects and covariates. The methods apply generally to linear, logistic, log, count, and time to event outcomes. These methods showed that patient risk drives 30-day readmission rates for VA PCI patients and the figure provided a new way of visualizing variation across these effects. ![][1] [1]: /embed/graphic-1.gif