SVAR estimation

[donihue]

Hello,

I am puzzled by an apparent difference in the estimation methodology between EV6 and EV7, as it concerns SVARs.

Thus, in the attached workfile, you will find two small VARs.
"Test1" was estimated with EV6 on 12/09/09. This worked fine, and will produce IRFs based on the structural decomposition.
"Test2" is identical (created by Object/Copy Object) but I am unable to obtain estimates for it under EV7.1, Mar 15 2011 build.
All estimation attempts result in the error message "Hessian matrix is near-singular at final iteration parameter values".

This may just represent some peculiar quirk in my system [I have found, for example, that when I have two instances of EV7 open, in one I can
estimate IRFs based on the structural decomposition, but in the other instance an otherwise identical SVAR will not produce such IRFs (the
"radio button" for IRFs based on the structural decomposition is "greyed out" ....)] but perhaps it is something more?

Regards
Donihue

[EViews Gareth]

Test2 is not a structural VAR, it is a standard VAR. You will have to estimate it via Proc->Structural Factorisation before you can perform the structural impulse responses.

Note that Test1 has starting values set to draw from a Standard Normal. You will have to set that on Test2 as well for it to estimate.

[donihue]

Thanks for your quick reply, Gareth.

Actually, if you look under Proc/Est Str Fact, you will see that I have indeed included the required restrictions for both Test1 and Test2; it is just that when one opens them after the structural estimation, the view always returns to that of the standard VAR.

But you have given me the answer - it is the "starting values" which make the difference. When I reset "Test2" to starting values from Standard Normal, the estimation goes through.

May I pursue on another SVAR-related issue: it appears that neither the "system" nor the "model" created from Proc/Make System or Proc/Make Model after the structural factorisation actually takes that factorisation into account: what is produced appears to be just the standard VAR. Is this correct? If so, would it be possible to change the procedure to obtain the factorised model?

Regards
Donihue

[EViews Gareth]

Yes, that is correct, it is not easy for us to change it.

[donihue]

Unfortunate!

In smallish systems, one can edit the equations by hand and then re-estimate the system by FIML, generate a model from that and then solve to get IRFs in a very round-about way.

I have done that for the small system I posted and although the actual FIML estimation ("sys_svar_fiml_test") does not work (since "WARNING: Singular covariance - coefficients are not unique"), the new coefficients appear in the model made from the re-estimated system ("mod_svar_fiml_test") - albeit with values entirely different from those estimated via the structural factorisation ("tab_svar_1"). OLS of course works ("sys_svar_ols_test"), but again with different coefficients, presumably since it does not take into account the cross-equation linkages.

Regards
Donihue

[alex_hk90]

Unfortunate!

In smallish systems, one can edit the equations by hand and then re-estimate the system by FIML, generate a model from that and then solve to get IRFs in a very round-about way.

I have done that for the small system I posted and although the actual FIML estimation ("sys_svar_fiml_test") does not work (since "WARNING: Singular covariance - coefficients are not unique"), the new coefficients appear in the model made from the re-estimated system ("mod_svar_fiml_test") - albeit with values entirely different from those estimated via the structural factorisation ("tab_svar_1"). OLS of course works ("sys_svar_ols_test"), but again with different coefficients, presumably since it does not take into account the cross-equation linkages.

Regards
Donihue

I've been trying to do similar things with EViews (manually estimate SVARs) and I think the reason why you get different coefficients to the 'Structural Factorisation' when using FIML is that (as far as I can tell) you probably need to restrict the residual variance-covariance matrix (E[uu')), to either the identity matrix or at least a diagonal matrix (these seem to be the only ones I have come across in the SVAR literature).

[donihue]

Thanks for your comment.

OLS of course gives a diagonal covariance matrix (look under View==>Residuals==>Covariance Matrix) yet as indicated the coefficients are somewhat different from those of the SVAR, so that does not seem to be the source of the problem.

Have you managed to get FIML to work?

Regards
Donihue

[alex_hk90]

Thanks for your comment.

OLS of course gives a diagonal covariance matrix (look under View==>Residuals==>Covariance Matrix) yet as indicated the coefficients are somewhat different from those of the SVAR, so that does not seem to be the source of the problem.

"View - Residuals - Covariance Matrix" for the (reduced form) VAR, or for the system? Unless I'm missing something, neither is a diagonal matrix in the file you posted (lsz_test3).

As a clarifcation, in that file (lsz_test3), what is the relationship between svar_1 and sys_svar_ols_test? In other words, how have you generated the latter from the former? It looks like you've applied the identifying restrictions in some way to get the SVAR as a system.

Have you managed to get FIML to work?

Unfortunately not, and I think it's a similar issue to the simple example I've posted here: http://forums.eviews.com/viewtopic.php?f=4&t=4019.

[alrai17]

As a relatively new user of EViews, I'm attempting to estimate IRFs from a SVAR estimated using the 'Model' object in EViews 6. In particular, I am trying to replicate Stock and Watson (2001) -- "Vector Autoregressions", Journal of Economic Perspectives -- who estimated a 3-variable SVAR, consisting of unemployment, inflation, and the Fed Funds rate. The Taylor rule (augmented with AR terms) is the Fed Funds equation in the SVAR. I estimated the SVAR and am now trying to construct the IRFs.

Do I need to first recover the VARMA(inf) representation to get the IRFs (and if so how do I do this)? I checked for a 'Impulse' tab ala the 'VAR' object, but couldn't find one.

Any advice/assistance you could offer me on this would be greatly appreciated.

Regards
Alan

[arthur]

Hello,

gentleman, please help me with a matter! I've constructed a 7 variable SVAR and created 7x7 contemporaneous relationship matrix. when I estimate it, the program sets all the values to 0.15, which is the starting value in optimization control tab. I cannot undrstand what what wrong it is with my model.

Thank You in advance

[startz]

Maybe because you only have 25 observations?

[arthur]

hi all

gentleman, an important question!!!!
If I estimate a SVAR model and then in impulse responce analysis I choose "user specified" innovations, can I consider the results still being structural?
And the second question- how can I give my estimated SVAR model 1% shocks, instead of one std?

[frank]

Hi all,

I want to know how to put 'sign restrictions' on my 'patc' long run restrictions matrix.

So instead of just having zeroes in the patc matrix I want to have '<0' etc. Eviews doesn't want to take the sign restrictions in that form and I was wondering if anyone knows how to do it?

[EViews Gareth]

I don't believe you can.

[arthur]

Hi,

gentlemen, I have estimated SVAR, and I want to make my model in excel. So should I multiply my coefficients matrix from VAR with A matrix or inverse A matrix?

Thanks and regards