Data Augmentation in Limited-Dependent Variable Models.

This paper proposes a scheme that speeds up the convergence of Markov Chain Monte Carlo (MCMC) algorithms in the context of limited-dependent variable models. The algorithm reduces autocorrelations more than the recently proposed Parameter Expansion Data Augumentation (PX-DA) algorithm. In addition, the paper provides an algorithm to sample a variance-covariance matrix with restrictions directly from the conditional posterior distribution. Finally, it is shown that the PX-DA algorithm, as applied to the multivariate probit model, can be seen as sampling from a different parameterization of the model. However, in some cases the PX-DA algorithm is not invariant to reparameterizations, and a slightly different algorithm is proposed.