Pacific C
Bayesian Model Selection for Networks: Application to Patient-Sharing Networks (306604)
Victor DeGruttola, Harvard School of Public Health*Ravi Goyal, Mathematica Policy Research
Keywords: Networks, model selection, patient-sharing, graph enumeration
A Bayesian approach to conduct statistical network model selection is presented for a general class of network models referred to as the congruence class models (CCM). The CCMs generalize many common network models including the Erdos–Rényi-Gilbert model, stochastic block models and the exponential random graph models when nodal attributes are discrete. Estimating the posterior probability for each CCM is challenging as the computation of the evidence of the posterior is difficult, which is often the case for Bayesian model selection, as well as the likelihood. Evaluating the likelihood for a CCM requires calculating the number of labeled graphs with specific properties; we present a recursive formula to estimate this quantity. We utilize the presented Bayesian network model selection approach for CCMs to investigate the structure of patient-sharing networks, which are associated with patient care.