Estimating shock standard error

Dynare help (legacy posts)
1

Hello,

I am trying to apply LOWW (2005) with the help of Dynare. In particular, I am doing Bayes MCMC estimation.

I have encountered a problem of estimating the standard error of shocks. Their estimated values depend hugely on the choice of prior distribution. For instance, I get the following result from ML before Bayes MCMC:

      prior mean   mode    s.d.        t-stat    prior   pstdev

theta 0.7 0.1399 0.0514 2.72 beta 0.1500
…
sigma_a 1.0 0.4607 0.1881 2.4495 invg NaN
sigma_i 1.0 0.4607 0.1881 2.4495 invg NaN
sigma_u 0.5 0.2303 0.0940 2.4495 invg NaN

Structured parameters seem to be estimated well. But the three standard deviation of shocks (sigma’s) are clearly strange. This happens irrespective the form of prior distribution (e.g. inverted gamma and uniform distribution). The following Bayes MCMC does not function well, though it is not so serious as the above.

I am not an expert of econometrics.
I would be grateful if you could tell me the reason.

Best wishes,
Kozo

2

This is indeed strange. Could you post or send me the *.mod file and the data?

Kind regards

Michel

3

Thanks for the files. Now I understand.

What you see is just the prior mode, because the data have no influence on the standard error parameters.

Your error is the following. For example,

  1. you declare nu_g as an exogenous variable
  2. you declare sigma_g as a parameter
  3. in shocks, you declare
    var nu_g; stderr sigma_g
  4. in estimated_params
    sigma_g, inv_gamma2_pdf,0.4,inf

This doesn’t work because the ‘shocks’ block is evaluated only once and isn’t updated thru the estimation process. The correct syntax is much simpler

  1. declare nu_g as an exogenous variable
  2. in estimated_params:
    stderr nu_g,inv_gamma2_pdf,0.4,inf

By the way, because you estimate a standard error it make more sense to use an inv_gamma than an inv_gamma2 that is usually associated with a variance

Kind regards

Michel

4

Dear Michel,

Thank you very much indeed.
Now my problem has been solved.

Best,
Kozo