Should we use posterior mean or posterior mode

[jpfeifer]

This looks like a problem with the convergence of the MCMC and the initial jumping distribution. Please download the unstable version and try the use_tarb option.

[Oriana]

Thank you. After I get a trial license for matlab statistical_toolbox I finally runned the model with use_tab option.
Nevertheless, the computation with Tailored randomized block MCMC algoritm has proved to be much more slower with a very low acceptance ratio. In order to avoid a lower acceptance ratio I tried mode_compute = 6 and after few hours I got the following error: Error using ==> chol
Matrix must be positive definite.

Note, I had also replaced the line [b]chol(hh) in the try-catch-statement of dynare_estimation_1.m by hh=1e-4*eye(size(hh)) on the unstable version.

What can possibly be done about this?

[jpfeifer]

Even if it takes longer, stay with the TaRB and lower mh_jscale to get a better acceptance rate. It will take longer, but you will need fewer draws. The reason is that when doing this, Dynare will recompute the Hessian so that the "Matrix must be positive definite" will not be an issue. You should use say 1000 draws in a first step and then use the trace_plot to check the movement of the MCMC. Check whether there is still a drift. If yes, restart with the most recent values.