I just wanted to ask you two general questions on the shock decomposition conducted after the estimation:
Is it possible to scale a shock and therefore lower its contribution? Or would this be bad practice?
Does it make sense to perform a shock decomposition only on one variable and ignoring the rest of the dataset? For example I estimate a model on data which includes GDP, investment, consumption, etc. However, after the estimation I only perform a shock decomposition on the GDP data and not taking into account the rest of the dataset, i.e. I restrict my information set? This would be equivalent of performing a shock decomposition on a variable using a calibrated model. Does this make sense, or would this be also considered bad practice?
Many thanks for your help!
I would not try your first suggestion… It would be more consistent to revise your priors, if you are not happy with the results, rather than change the values of the parameters after the estimation.
I have no problem with your second experience. It just means that you do not identify the parameters and the shocks with the same information set. That said, and I do not know your model, the problem you will probably have is that it is not possible to estimate uniquely all the shocks with only one observed variable (you don’t impose enough constraints on the endogenous variables).
What exactly do you want to do in 2.? It sounds as if you want to run
shock_decomposition on one series for a given parameter set that you estimated. That does not amount to restricting the information set. The data was used to estimate the parameters. For a given parameter set, you can decompose any variable you like while still having the full data as your information set.
@jpfeifer I am not sure to understand your point… If you compute the smoothed variables with different set of observed variables, for given values of the parameters, you will obtain different estimates of the latent variables. The information set used for the estimation of the parameters is of second order in this regard (except that different information sets will eventually change the parameter values, which in turn will affect the identification of the latent variables, so second order is most likely a too strong statement).
@stepan-a The question is which of the two cases are we talking about
y only, although there are more observed variables that have been used for estimation
varobs c y g;
- Run estimation on one set of observables, then change
varobs and run
shock_decomposition on the smaller data
varobs c y g;
I was talking about 1, you were talking about 2.
Okay then. I probably misread the initial post…
Thanks Stéphane and Johannes for your helpful comments! Sorry for not being clearer, as Johannes said I was talking about:
varobs c y g;
Would you agree with Stéphane’s point Johannes?
Yes, I fully agree with him for that case.