Analysing the estimation results

I just finished my estimation problem and I want to know if it was good estimation.

  1. How can I know if the parameters were well identificated?
  2. How can I know if the estimation converged? (multivariate MCMC diagnostic)
  3. How can I know about the acceptance rates?


  1. Use the identification command.
  2. If you used at least two chains, the convergence diagnostics are displayed.
  3. The acceptance rates are displayed by Dynare. They should also be stored in the MC_record field in the recent snapshot. Otherwise, they should be stored in the metropolis.log file in the metropolis subfolder.


I am also struggling with the assessment of the estimation quality, especially regarding the MCMC diagnostics.
Dynare spits out the MCMC/BrooksGelman statistics, and there are a few aspects to check to confirm convergence: The blue and red line have to get close and both lines have to settle down. As far as I see it right these check points are assessed by visual judgement.
My question:
Is there some sort of a “hard criterion” to judge properly whether the estimation has converged? Is there some sort of a test in dynare?

Thanks in advance!

We are working on implementing a hard critierion, but it will still take some time. In LeSage’s econometrics toolbox there is the coda-command that computes the Geweke and Raferty/Lewis diagnostics. You could use those functions.

thank you for the quick reply!
Well, as I am not a specialist in programming I am firstly looking for some rule of thumb. What do you think about the convergence statistics attached below? Are they o.k.?
MCMC7.pdf (8.91 KB)
MCMC5.pdf (8.08 KB)
MCMC1.pdf (8.53 KB)

Most of them look OK, but some of them are still borderline.

Thanks a lot for your reply!
What would you suggest? I use mode-compute=9 and I am already up to 200000 iterations (2 Markov chains). My acceptation rate is at about 28%. Does increasing the number of iterations help? Or adding another sequence? I am asking because estimation already lasts a long time.
Thanks in advance!

More iterations should help. Try using the results after your first 200000 draws (which you should have) as starting value for mode_compute=9. In the next sequence of convergence plots the drift is then hopefully gone.

Thanks jpfeifer!

I tried what you suggested, i.e. I used the posterior means of the previous simulation as starting values and extended the number of iterations to 300000. However, the results do not look much better. I attached some of thee MCMC diagnostics below.

Is it common to just judge theim visually?

Have a nice weekend!
MCMC5.pdf (7.31 KB)
MCMC4.pdf (7.35 KB)
MCMC3.pdf (6.93 KB)
MCMC2.pdf (7.96 KB)
MCMC1.pdf (7.69 KB)

Actually, it looks better now. If in doubt, try repeating with more MCMC draws.