Delayed Rejection Algorithm to Estimate Bayesian Social Networks

Statistical social network analysis has become a very active and fertile area of research in the recent past. Recent developments in Bayesian computational methods have been successfully applied to estimate social network models. The Delayed rejection (DR) strategy is a modification of the Metropolis-Hastings (MH) algorithms that reduces the variance of the resulting Markov chain Monte Carlo estimators and allows partial adaptation of the proposal distribution. In this paper we show how the DR strategy can be exploited to estimate dyadic independence social network models leading to an average 40% variance reduction relative to the competing MH algorithm, confirming that DR dominates, in terms of Peskun ordering, the MH algorithm.