Data for GMM/SMM

Estimation SMM/GMM/IRF-Matching
1

Dear all, I hope you are well.
A question regarding the data file for GMM/SMM estimation.
What’s the recommendation for the data to be in the data file for each procedure? Meaning, should I use the raw data in (real or nominal), in differences, log differences, hp filtered, etc?
I checked the example files but I didn’t notice any recommendation for this estimation procedure.
Thanks in advance
FgB

2

That very much depends on the moments you want to match. Often, it will be log differences. HP-filtered moments are not yet supported.

3

Thanks, Johannes. Just to confirm I am doing the things right. When using the data in log differences, should I use the following in the .mod file?
, logdata

4

No, you would need that if you would use the loglinear-option as well.

5

Great, thanks, Johannes.

6

Dear, Johannes.
Following the example found in the wiki of the Ispra workshop, “method of moments examples”. I see that the variables are set in log, but I am not sure how the data file is set, whether it is the log difference or not, log(Y) - log(Y(-1)).
I am asking this because in some of your examples for estimation you set the differences explicitly in the vars section and then you match the moments.
link
Can I match the log difference based-moments with GMM by setting the variables in log only, or I need an explicitly variable that set the log difference in the mod file?

7

The general principle is comparing moments of the same objects in the model and the data. You would not compare growth rates in the data to a logged level in the model.

8

Thanks. Just to be sure about it and speaking just on the example given in the workshop wiki page.
Hence, is the data in RBC_Data_2 of the example is the log deviation of the data with respect to its log mean so that the mod file matches the moments with the log of the variables?

9

This is tricky. The example file uses simulated data from the model. Here, you can easily generate log deviations from the steady state. That is a problem in the data where you don’t know the trend and always have to think hard about matching model and data.

10

Thanks, Johannes.