Usage of load_mh_file

[yc1122]

Hi,

I am estimating a DSGE model but the MCMC draws do not mix well as shown from the Brooks&Gelman diagnostics and traceplots. So I'd like to run more replications to achieve convergence. Happy to learn that there is an option of "load_mh_file" that can do the work. But I want to make sure if it will start exactly from the last draw from the previous MH draws reloaded? And if the answer is "Yes", then I don't need to specify a burn-in period, ie. set "mh_drop = 0", right?

Furthermore, when I try to add more replications, the acceptance rate is 15% which is significantly different than 20% from the previous MH draws. I guess 14% is the acceptance rate ONLY for the added draws, then will the overall acceptance rate be recalculated based on the whole sample of simulations after posterior simulations are finished?

Thank you.

yc

[jpfeifer]

load_mh_file continues at the last draw of the previous file. The burnin is taken from all draws. That is, if your first run has 2000 draws and you specify an additional 2000 draws, and have mh_drop=0.5 the first 2000 draws will be used as a burnin. The burnin is not applied to any intermediate draws.

The acceptance rate is computed only for the current new draws. There currently is no way to compute the overall acceptance rate.

Note also that load_mh_file does not continue using the previous mode and Hessian unless it is tolds so by using the mode_file option and mode_compute=0. Otherwise, Dynare will recompute the mode given the starting values and use this mode and Hessian to continue the draws.

[ZBCPA]

load_mh_file continues at the last draw of the previous file. The burnin is taken from all draws. That is, if your first run has 2000 draws and you specify an additional 2000 draws, and have mh_drop=0.5 the first 2000 draws will be used as a burnin. The burnin is not applied to any intermediate draws.

The acceptance rate is computed only for the current new draws. There currently is no way to compute the overall acceptance rate.

Note also that load_mh_file does not continue using the previous mode and Hessian unless it is tolds so by using the mode_file option and mode_compute=0. Otherwise, Dynare will recompute the mode given the starting values and use this mode and Hessian to continue the draws.

Dear Johannes,

If initially I have run 200,000 draws already and got results, however the convergence is not perfect, so I plan to run more draws. To save time, I can use load_mh_file to let this 200,000 draws as burnin in the new estimation process and run another 300000 draws.Acording to your explanation, the mh_drop becomes 200000/(200000+300000)=0.4 in this case.
Should I use load_mh_file command like this and at same time hope to find a better mode:

Code: Select all

Estimation(load_mh_file,mode_compute=9,mode_file=hpdata_mh_mode, mh_replic=300000,mh_drop=0.4,......);

This just likes run 500000 draws, so it saves time. Am I right?

Many thanks,
Huan

[jpfeifer]

Yes, that is right. However, when loading mh_files you should not recompute the mode, but start with the mode-file used in the old estimation.