Abstract
Mechanistic network models specify the rules whereby networks grow and evolve. Typically, such models consist of one or more mechanisms, each of which leaves its distinct fingerprint to the observed network. Despite the prominence of such models, there has been little work to date on how to uncover the specific mechanisms that drive network formation. The main challenge is that, unlike statistical network models, mechanistic models generally have intractable likelihood functions, and therefore standard frequentist and Bayesian approaches to inference and model selection are not applicable. In this article, we treat a mechanistic model as a mixture model consisting of both informative and uninformative mechanisms. We develop a framework based on approximate Bayesian computation and summary statistic selection to identify the weights of each component mechanism and its parameters. We apply the method to three simulation studies and an empirical network of households in a public health setting in India. We find that the method is generally able to uncover both mechanism weights and mechanism parameters accurately.
Make this part of your paper trail
Save this paper to a shelf, write a review, and keep your own notes.