Comparing methods of grouping hospitals

Objective To compare the performance of widely used approaches for defining groups of hospitals and a new approach based on network analysis of shared patient volume. Study Setting Non‐federal acute care hospitals in the United States. Study Design We assessed the measurement properties of four methods of grouping hospitals: hospital referral regions ( HRR s), metropolitan statistical areas ( MSA s), core‐based statistical areas ( CBSA s), and community detection algorithms ( CDA s). Data Extraction Methods We combined data from the 2014 American Hospital Association Annual Survey, the Census Bureau, the Dartmouth Atlas , and Medicare data on interhospital patient travel patterns. We then evaluated the distinctiveness of each grouping, reliability over time, and generalizability across populations. Principle Findings Hospital groups defined by CDA s were the most distinctive (modularity = 0.86 compared to 0.75 for HRR s and 0.83 for MSA s; 0.72 for CBSA ), were reliable to alternative specifications, and had greater generalizability than HRR s, MSA s, or CBSA s. CDA s had lower reliability over time than MSA s or CBSA s (normalized mutual information between 2012 and 2014 CDA s = 0.93). Conclusions Community detection algorithm‐defined hospital groups offer high validity, reliability to different specifications, and generalizability to many uses when compared to approaches in widespread use today. They may, therefore, offer a better choice for efforts seeking to analyze the behaviors and dynamics of groups of hospitals. Measures of modularity, shared information, inclusivity, and shared behavior can be used to evaluate different approaches to grouping providers.