Implementing random scan Gibbs samplers (Q2488391)

A Gibbs sampler is considered for the simulation of a multivariate distribution \(X\). It is proposed to change the probabilities of coordinate selection at each step taking into account the information from previously generated data. If the sampler is used to estimate a functional moment \({\mathbf E} h(X)\), then the variance of the estimator can be used as a criterion of optimality. The authors use a Gaussian approximation to derive a feasible criterion. Properties of the obtained adaptive sampler are studied via simulations. It is applied to the analysis of real medical data.