Stationarity with Bayesian estimation

Hello. Apologies if this question has already been asked. I am wondering if Dynare enforces stationarity. For instance, if I run Bayesian estimation for a basic AR(1) process with N(0, 1) prior for the auto-regressive coefficient, will Dynare constrain the coefficient to be below 1 in absolute value?

The depends on the command you are running and the associated value of qz_criterium. For stoch_simul, unit eigenvalues are counted towards the stable roots, so 1 would be allowed. For estimation, that would not be the case unless you use diffuse_filter.

It is estimation that I am interested in. Thanks for the quick response Johannes.

Could I ask another question? Bayesian estimation produces _mh_mode file. Is this based on the full set of draws or the set of draws after burning a fraction defined by mh_drop?

It’s the mode over the draws after discarding the burn-in.


Dear Johannes

I was wandering if you could help with this recurring issue I have (unfortunately I cannot share the mode file). During Bayesian estimation on Mac (in parallel mode) I get this warning (screenshot attached). This particular screenshot was generated when estimating with 2mln draws (so the warning occurred around 100000th draw). When 2mln draws were complete, a total of 18 such warnings were issued. This does not happen on windows machine or when running chains sequentially. I am using Matlab Version 2022a, Dynare 5.1, Mac OS version is Monterey. File vg_run is simply instructing Matlab to clear workspace and run Dynare.

Can you try the suggestions in Parallel estimation MCMC diagnostics issue?

Thank Johannes. I have found this before and studied it carefully. Did all SSH requirements, still the issue I have persists. The difference, though, is it does not freeze on me. Estimation is finished, however, I am concerned that it might not be “correct”, given it fails to load the file on a few occasions (on line 702 in master Parallel, load([‘comp_status_’,fname,int2str(j),‘.mat’]);).

That files stores information on the progress of the MCMC. It’s saved in fMessageStatus.m. I don’t see a potential for corrupting the estimation itself.

Thanks Johannes. I added ‘-v7.3’,‘-nocompression’ in save to the fMessageStatus. Seems the error message is gone.