I was doing my MCMC for my model with mh-replics= 1000000. Initially the acceptance rate came out to be around 30% for more than half of the process. But later the rate started to decline, and by the time the iterations were completed it was close to 10 %. I was wondering what one might diagnose of this. Thanks.
The only thing that one may say without looking at the output and at the actual Metropolis run is that Metropolis must have moved to another part of the distribution. It may be that the computed mode wasn’t very good.
How big is the model? How long does it take to run 1000000 iterations?
There is a log file for the metropolis in dynare version 4. If your model is written in XXX.mod you just have to open the text file :
In this file you will find a lot of informations about what is going on in your mh. For instance, regularly the acceptation rate is saved (say every 2000 simulations) with the mean value of the parameters, so that you can find some explanations about the drop of your acceptation rate.
Many thanks, Michel and Stéphane. My model is of quite size. It took about a day to run a block of 1000000 iterations. But without this many iterations, problems like this won’t show up. Over half of the process, the acceptance rate is pretty stable at 0.3.
I would like to find the origin of the problem. Can you send me the model and the data? I will look at it in September.
I had the same type of problems with bvar-dsge estimation a couple of weeks ago. If you estimate this kind of model you should try with the last version of dynare v4 (the problem is not corrected in dynare v3). The problem was that for some values of the deep parameters the likelihood (and so the posterior kernel) appeared to be complex (I do not yet understand why). In this case the mh algorithm is trapped and the acceptation rate is zero.