Hi all,
I got 300,000 draws for each block. Three blocks shown near 1/3 acceptance ratio while other two blocks didn’t. The code were as below:
Estimation::mcmc: Total number of iterations: 300000.
Estimation::mcmc: Current acceptance ratio per chain:
Chain 1: 32.5557%
Chain 2: 32.324%
Chain 3: 32.2273%
Chain 4: 0%
Chain 5: 0%
And I wonder if the terrible chain 4 and chain 5 has affected my convergence diagnostic. If so, how can I restart with chain 1 to 3 skipping the MH process?
Thanks!
That is indeed unusual. Have you tried to find out what is going on in those chains? Did you look at the trace_plots?
The 0% accepted chains were stuck at some constant.
This case took place again when I run the same .mod file specifying 400,000 draws or 500,000 draws of each block today.
Did you use Dynare parallel?
I didn’t. Thanks for letting me know about this.
I was curious as I had a similiar issue with one chain behaving strange when using Dynare parallel (unfortunately I don’t remember anymore which mod file caused the behavior. Anyways, this looks like a bug which we would need to investigate. Could you maybe post your mod file and data here. Alternatively, you could sent it to us via private message or email (willi@mutschler.eu).
I would also be interested whether the same happens with the current unstable version of dynare, i.e. if we already fixed this at some point in the development cycle.
It must have to do with starting the chains with overdispersed starting values. It seems that after the initial draw, no other valid draws were found. Did you mode_check-plots indicate that you were close to the boundaries of the determinacy region?
Can you try a lower mh_init_scale?
I made much customized changes to the code allowing the time-varying expectations, which might no be a typical case for debug.
As Johannes suggested I checked the mode_check plots, a parameter’s mode was chosen close to the lower bound. I made the prior bound broader and it worked. The mcmc generated 4 chains with accpetance ration near 1/3. The chain Initialization problems were solved, I think.
Thank you Johannes. I did get a boundary value from mode computation before mcmc.
