When I use Dynare to apply Bayesian Estimation, mode_compute = 6 is chosen to avoid non-positive definite matrix problem. However, this routine seems to automatically fixes the mh_jscale = 0.00015 to obtain a target acceptance rate around 1/3. The small jscale makes the posterior very narrow around the mode, compared to the priors. It is however desirable to make the MC travel a vaster region. Anyone could help me with this problem?
Thanks for answers,
If the mod file containing your model is titled ‘model’, then you can find the scale parameter which DYNARE computes using mode_compute=6 in the file ‘model_optimal_mh_scale_parameter’. Here you can modify the scale parameter (just keep the file name the same) to achieve an average acceptance ratio more to your liking.
Just to report on my experience with using mode_compute=6: I generally do not get convergence of the MH algorithm using mode=6. Using mode=4 or =7 (when it works) will produce good convergence diagnostics for the same model estimation in which mode=6 did not work. Mode=6 in my experience tends to produce spiky posteriors, and the convergence diagnostics indicate that no convergence has occurred (even after hundreds of thousands of replications in multiple chains.)