Inquiry on Comparing MCMC and SMC Estimation Results in Dynare

Estimation
1

Dear Dynare Community,

I am working on estimating a standard three-equation DSGE model and am comparing the results between MCMC and SMC methods. However, I am observing significant differences in the estimated parameters between the two approaches, which I did not anticipate.

Could you kindly advise if the estimation codes I am using are appropriate for a direct comparison between these sampling methods? I would appreciate any insights or recommendations on how to proceed with such a comparison.

The codes I am using are:
MCMC:

estimation(datafile=US_MacroData, mh_nblocks=2,mh_replic=5000, mode_compute=4,prior_trunc=0,
mh_jscale=.2) y pi i;

SMC:

estimation(
   datafile=US_MacroData,
    mh_nblocks=2,mh_replic=5000, mode_compute=4,prior_trunc=0,
    mh_jscale=.2,
    posterior_sampling_method='hssmc')  y pi i;

Thank you in advance.

Best regards,
Susie

2

Can you provide me with the codes to investigate the issue?

3

Thanks a lot for your reply. Sure, please see codes and data below:
DE_small.mod (2.2 KB)
US_MacroData.xlsx (19.0 KB)

4

I think there are two things at play:

  1. There is a documented bug in SMC that results in wrong results in version 6.1. It will be fixed in Dynare 6.2.
  2. The standard MCMC shows poor convergence if you start with rather poor starting values.
    The attached file shows that three posterior samplers create very similar posteriors in the current unstable version / 6.2.
    DE_small.mod (2.3 KB)
    US_MacroData.xlsx (19.0 KB)
5

Thank you for your helpful comments. I really appreciate it.