I would like to request your help with the model I´m trying to estimate. My main concern is about the persistence of the government spending shock. As far as I know, whenever the estimation yields a shock with near unit root features, it could be due to model misspecification. For the government spending shock, I got a value of 0.993 for the posterior mean (I didn´t use government spending data in the estimation). Even though the estimation results in apparently reasonable IRF´s, the variance decomposition shows that the government spending shock is responsible for a great share of the variability of the forecast error of consumption and output, which seems strange for me. I you could just take a look. Thanks in advance. norges2.xls (23.5 KB) norges_data2.m (2.22 KB) model (2).mod (10.2 KB)

All parameters are identified in the model (rank of H).

WARNING !!!
The rank of J (moments) is deficient!

[xi_rf,xi_b] are PAIRWISE collinear (with tol = 1.e-10) !
[rho_rf,rho_b] are PAIRWISE collinear (with tol = 1.e-10) !

[/quote]

You definitely need to fix this.

Also, your data does not seem to be mean 0, but your model variables are. Finally, do not use the HP filter for estimation. See [Small Question about Bayesian Est)

Thank you. The data doesn´t have mean 0 because originally I was trying to estimate for a larger sample (in which the average is zero), but when I tried to estimate for a smaller sample I did´t notice that in this case the mean wouldn´t be zero anymore.
I have a question about you last sentence, about the HP filter. Many times I got confused. You say I shouldn´t use the standard HP filter, instead of the one-sided HP filter (remark 12 of your guide)? But generally people speak and use the standard HP filter, no ?
Or are you referring to the prefilter=1 option in the estimation command (which only demeans the data, if I understood correctly) or about the treatment of hours? Another question: if the data already have zero mean, does this prefilter=1 command make any difference ?

I inferred from this that you stationarized your data using a two-sided HP filter. You shouldn’t do that. You can use the two-sided HP filter for model evaluation (do the moments of the simulated model fit the data), but you shouldn’t use it for data treatment prior to estimation.
If data and model are both mean 0, prefiltering does not do anything. Just drop it (there are some bugs related to prefiltering in 4.4.3)

Thank you a lot, very helpful. I have an additional question: how problematic is to get very persistent shocks. Because now Dynare indicate that all the parameters pass the identification exam, but I´m still getting persistent shocks (though slightly below 0.99). Even hypothetically assuming that I get shocks with persistence larger than 0.99, passing the identification command, would that be a source of concern? Thank you in advance.

That depends. Sometimes close to unit roots are not a problem, sometimes they result in hugely implausibe IRFs. It all depends on whether the results look sensible.