Hi everyone,
I have regularly worked with RBC models where the calibration consisted of matching the theoretical moments with the moments of the data. Now that I have had to resort to estimating parameters by Bayesian methods, given the complexity of the models.
In this case, I was able to obtain theoretical moments very similar to the real ones by means of a Bayesian estimation, even the parameters have reasonable values. However, looking at the forum posts I can notice that there are some criteria about the estimates for example, in my estimate I get:
Estimation::marginal density: There’s probably a problem with the modified harmonic mean estimator.
Log data density is Inf.
and I don’t know if this is a problem when refereeing my project, as for example in econometrics that the Rsquared is important or the levels of significance, etc.
I am attaching the corresponding files to see if I can get some feedback, I would greatly appreciate it
Database.xlsx (66.9 KB) Final.mod (21.7 KB)
There are a couple of issues.

identification
tells you
[varpi,nu] are PAIRWISE collinear!
 Your prior is suboptimal:
Prior distribution for parameter aleph has unbounded density!
Prior distribution for parameter aleph has two modes!

mode_compute=5
returns
Final value of minus the log posterior (or likelihood):603.434118
where many parameters are at the bounds. Such corner solutions create problems.
The advisable thing for the first case would be to fix a parameter and estimate another I guess.
The second case I noticed that I wanted to give variability to the distribution, but I think I should limit the variance.
The last problem I suppose it will be a better choice of the means of the distributions, or perhaps correcting the first cases does not yield corner solutions.
Thanks for your answer, I will try those cases to see what happens.
I corrected the problem with the collinearity of the parameters by setting “varpi” to 0.80 as in Bouakez & Rebei (2006) and estimating “nu”. Then, I reduced the variance of “aleph” to 0.25 and ran the mean and variance values of the AR (1) parameters to 0.80 and 0.10, respectively. I ran the estimation with mode_compute = 6 and I keep getting log data density inf, and most of all I would like to know how did to get the message of the identification problem previously mentioned?
I attach the updated files in case you like to see them
Database.xlsx (66.9 KB) Final.mod (21.6 KB)
Please provide the _mode.mat
file.
I was able to correct the problem by setting some parameters that are frequent in the literature (such as EIS). I got the following results, but I see that my theoretical moments are not as close as they were when I was getting the “log data density inf” error. And my question is, is he still a good estimate and can I use the model to do simulations? or what criteria are necessary? Thank you in advance. I attach the .mod file with the data and the estimation log, as well as the mode file.
database.mat (4.3 KB) Final.mod (21.6 KB) Final.log (34.3 KB) Final_mode.mat (9.2 KB)