A0806
Title: A comparison of variable selection algorithms with an application to the ordinal Markov random field
Authors: Don van den Bergh - University of Amsterdam (Netherlands) [presenting]
Abstract: Bayesian variable selection plays a crucial role in network models, especially because the space of models is too large to enumerate. A range of Bayesian variable selection algorithms is compared in terms of their ability to accurately recover posterior inclusion probabilities and their speed of convergence. First, the performance of reversible jump MCMC, mixtures of mutually singular distributions, Rao-Blackwellized estimators, and sticky piecewise deterministic Markov processes is contrasted on a toy regression problem where enumeration is feasible. The methods are compared on accuracy and computational speed. Next, the comparison is extended to a more complex model where enumeration is infeasible: The ordinal Markov random field. In this network model, the number of parameters subject to selection grows quadratically with the number of variables, making it an excellent example to study how the methods scale.
