Estimating the Impact of County Boundaries on State-wide Patient-Sharing Network Models

Background: In the field of public health, network models are useful for understanding the spread of both information and infectious diseases. Collecting network data requires determining network boundaries (ie, the entities selected for data collection). These decisions, if not made carefully, have potential outsized downstream effects on study findings. In practice, collaboration and coordination between healthcare organizations are often dictated by historical or geopolitical boundaries (eg, state or county boundaries), which may distort the underlying network under study, and thereby affect the reliability and/or accuracy of the network model. Objective: We compared natural communities in a patient-sharing network with those induced by geopolitical boundaries. Methods: Using data from the Healthcare Cost and Utilization Project (HCUP), we constructed a patient-sharing network among hospitals in California, splitting the data into a training set and a holdout set. We performed edge-betweenness clustering on the training set, and with the holdout set, we compared the resulting partition with the partition by counties using modularity. We also clustered contiguous counties that might function more cohesively together than individually. We performed spatially constrained hierarchical clustering on the network constructed from total patient flow between pairs of counties. The results were again compared via modularity on the holdout set to the county partition. Lastly, we built an individual-based model (IBM) using HCUP and American Hospital Association data to perform epidemic simulations. For each of several counties, we implemented this model to estimate the proportion of patients infected over time. We then reran the individual-based model using the entire state while dividing the results into corresponding counties. Results: In total, 680,485 patients transferred between 374 hospitals in 55 counties from 2003 to 2011. The out-of-sample modularity for the edge-betweenness clustering partition was 464% higher than that of the county partition. Aggregating the counties into half as many contiguous clusters was 319% higher, and aggregating them into 6 clusters was 489% higher (Fig. 1). The epicurves from the individual-based model ranged from small to significant deviations between state versus county boundaries (Fig. 2) . Conclusions: Collecting network data using externally imposed boundaries may lead to inaccurate network models. For example, counties serve as a poor proxy for their underlying communities, resulting in poor overall disease spread simulation results when county boundaries are allowed to drive network construction. These issues should be considered when building coordination partnerships such as the Accountable Communities for Health.Funding: NoneDisclosures: None