Title: MCMC strategies for the analysis of discrete graphical models of marginal independence
Abstract: We present and compare different MCMC strategies for the analysis of marginal log-linear models with a bi-directed graph representation. We consider two alternative parameterisations : one based on cell-probabilities (pi-parameterisation), and the other based on the marginal log-linear parameters (lambda-parameterisation). Working with the pi-parameterisation it is possible to perform a conjugate and conditional conjugate analysis based on product of Dirichlet priors (Ntzoufras and Tarantola, 2013). On the other hand, with the lambda-parameterisation a conjugate analysis is not feasible and more sophisticated techniques are required. An application to real data will be presented.

Joint work with Ioannis Ntzoufras (AUEB, GR)