|
Abstract:
|
In network analysis, one frequently must perform inference based upon only one sampled network. This poses a challenge for bootstrap-based approaches, which typically require an iid sample. A class of network models called latent space models overcome these difficulties by generating a network based on unobserved geometric structure, but this raises the question of whether inference in such models should be conducted by conditioning on this latent structure or by marginalizing over it. We develop bootstrap schemes for both cases, i.e., conditional and marginal bootstrap methods for network data. We establish bootstrap validity for both schemes for a broad class of network statistics, including modularity, which has not previously been addressed within the network bootstrap literature. Our experiments include simulated data as well as an application to Microsoft Bing search data.
|
