ERRORS instruction in SVAR

[Joung]

Dear Tom,

As shown in the attachment, I used the ERRORS instruction at the end to compute the FEVD while analyzing the SVAR model. However, I noticed that the results differ each time I run the code. Is this inevitable due to the Monte Carlo simulations? The issue is that the differences between the result sets are too large, making it inappropriate to report just one set of values in the paper. I would appreciate any guidance on possible solutions.

[TomDoan]

First, that's not an average across draws---it's just the results from a single draw (the last one you did). You need to use @MCFEVDTABLE to get summaries from the whole set of simulations.

Second, your effective sample size is very small---when I ran it, I got

Effective sample size 8.10956

on 10000 draws.

While it may be possible to tune the sampler better to get a higher yield, the main problem is likely that your second and third equations differ only in normalization (which shouldn't matter) and the presence of A24 in the second one. However, A24 is coming in small and quite insignificant, so those two equations really can't easily be separated in the data.

[Joung]

Dear Tom

I understand that the @MCFEVDTABLE has certain limitations for the reasons you mentioned. However, now that the simulation has been conducted, could you kindly advise if there is a method to compute the average and one standard deviation of the simulation results, such that the contributions of each shock sum up to 100? For example, is the method used on pages 82–83 of the attached paper a different approach? If so, is it possible to implement it in RATs?

-------------------------------------------------------------------------------
Re: Variance decomposition with BVAR model
Post by TomDoan » Tue Aug 20, 2013 7:18 am

If A, B and C are random variables and A+B+C=100, then their means must add up to 100. However, their medians (which is what @MCFEVDTABLE computes) don't have to. @MCFEVDTABLE doesn't compute means and standard errors because those are misleading due to the highly asymmetrical nature of the components; in particular, the mean can be outside the 5-95% band of values. The fact that you don't want to compute the error bands doesn't change the fact that the mean is highly misleading.
-------------------------------------------------------------------------------

[TomDoan]

@MCFEVDTABLE has a STDDEV option. If you use that, it will do the moment-based statistics (means, standard deviations). Just note that even a 1 std deviation band will often go outside [0,1].