How is MCMC implemented in ADMB?
AD Model Builder uses the hessian to produce an (almost) multivariate normal distribution for the Metropolis-Hastings algorithm. It is not exactly multivariate normal because the random vectors produced are modified to satisfy any bounds on the parameters. See the user manual for various modifications to the MCMC that are available, such as to deal with highly correlated parameters. Hybrid Monte Carlo sampling can also be used through the -hybrid command, but thorough benchmarks have not been conducted for this method.
