Imposing Restrictions on SVAR

For technical questions regarding estimation of single equations, systems, VARs, Factor analysis and State Space Models in EViews. General econometric questions and advice should go in the Econometric Discussions forum.

Moderators: EViews Gareth, EViews Moderator

EViews Matt
EViews Developer
Posts: 158
Joined: Thu Apr 25, 2013 7:48 pm

Re: Imposing Restrictions on SVAR

Postby EViews Matt » Fri Jul 21, 2017 9:14 am

Choosing starting values is primarily a matter of experimentation to determine what works. Sensitivity to starting values may just be an unfortunate property of your data and model, but it could also indicate that the model is not well specified or that some of the implicit assumptions are false. Some of the estimated coefficients, particularly C(2) and C(9), are noticeably non-zero but have much larger standard errors (hence low z values), which makes me suspect a model/data issue is present.

nasa
Posts: 39
Joined: Tue Aug 02, 2016 12:41 am

Re: Imposing Restrictions on SVAR

Postby nasa » Sat Jul 22, 2017 3:15 am

Ok, thanks

Is that a problem when the confidence bands(the red line) very close to the upper line in the black box as it can be seen from the attachment? What it indicates us that?

Best,
Attachments
Impulse response.docx
(19.2 KiB) Downloaded 25 times

EViews Matt
EViews Developer
Posts: 158
Joined: Thu Apr 25, 2013 7:48 pm

Re: Imposing Restrictions on SVAR

Postby EViews Matt » Mon Jul 24, 2017 9:29 am

No, though if you look at the Monte Carlo SEs at a longer period you can see that many error bands grow ever wider, which is different from the analytic SEs. Looking at the VAR estimates, I notice that five of the six equations have very high R^2 values and there's one big residual in the equation for interest_rate that dominates the residual covariance matrix. That may be the cause of the SVAR model's numerical sensitivity.

nasa
Posts: 39
Joined: Tue Aug 02, 2016 12:41 am

Re: Imposing Restrictions on SVAR

Postby nasa » Tue Jul 25, 2017 6:40 am

Thanks for your excellent explanation.

I noticed also similar problem that the interest rate equation is causing the problem.
So, how does that problem emerge and what possible solutions can you suggest for this,please?

Thanks a lot.

nasa
Posts: 39
Joined: Tue Aug 02, 2016 12:41 am

Re: Imposing Restrictions on SVAR

Postby nasa » Tue Jul 25, 2017 7:02 am

and plus i think such problem of numerical sensitivity is not appearing in the recursive short run impulse responses but in the A and B model.

nasa
Posts: 39
Joined: Tue Aug 02, 2016 12:41 am

Re: Imposing Restrictions on SVAR

Postby nasa » Tue Jul 25, 2017 7:59 am

One last thing,when I change the level specification of the VAR in to a difference specification,I will get a relatively low R2 as in the second equation for the rest five equations and when I use log specification for the interest_rate variable(even though it is in a percent form at level),I will get low residual for it like in the other cases.So,considering all these cases,would you give me some remedial for this sensitivity problem of the model,please?

Best

EViews Matt
EViews Developer
Posts: 158
Joined: Thu Apr 25, 2013 7:48 pm

Re: Imposing Restrictions on SVAR

Postby EViews Matt » Tue Jul 25, 2017 4:30 pm

Even under the recursive short-run SVAR model, drawing from the standard uniform normal fails to converge about 25% of the time. While using the log of interest_rate in place of interest_rate numerically helps the residual covariance matrix, and thus the SVAR model, I think there are still issues with the VAR model. There's a residual spike in interest_rate (or log(interest_rate)) in 1998Q2, plus there seems to be some autocorrelation among the residuals, most notable the residuals of dcpi and interest_rate(-1). These properties suggest to me that the VAR model needs to be redesigned a bit.

nasa
Posts: 39
Joined: Tue Aug 02, 2016 12:41 am

Re: Imposing Restrictions on SVAR

Postby nasa » Wed Jul 26, 2017 12:46 am

Ok excellent observations

But the thing is including spike or shift dummy variable for the interest rate(1998Q2) does not solve the residual autocorrelation problem plus increasing the number of lags to 2 may solve the autocorrelation problem(neglecting dummyb and only incorporating c,@trend and dummya for the exogenous case) but we may loose degrees of freedom as we are using small number of observations plus the graph result of the impulse response functions are not looking good when we use 2 lags. So, I would appreciate if you suggest me a specific solution in redesigning the VAR model.

Thanks a lot.

EViews Matt
EViews Developer
Posts: 158
Joined: Thu Apr 25, 2013 7:48 pm

Re: Imposing Restrictions on SVAR

Postby EViews Matt » Wed Jul 26, 2017 10:42 am

It sounds like you've experimented with two good approaches for addressing the issues in the VAR. A few comments...

  • I wouldn't expect an additional dummy for the interest rate spike in 1998Q2 to solve the autocorrelation issue, but it should resolve the spike in the interest rate residuals. I'd guess that a dummy with 1's for either one, two, or three quarters, starting at 1998Q2, would be effective, but you may need to experiment.
  • Regarding the autocorrelation issue, increasing the VAR(1) to a VAR(2) may indeed be problematic with your small data set, as you've pointed out. However, in EViews 10 you can place restrictions on the VAR coefficients, which means you can include only those terms of the VAR equations that you want. For example, since the residuals of dcpi and interest_rate(-1) are correlated it would be nice to add an interest_rate(-2) term to the equation for dcpi. You much switch to a VAR(2) to achieve this, but since you don't want all the other lag 2 terms you can restrict them to be zero. Basically, you can take a VAR(1) and added one VAR(2) term to it instead of all 6 * 6 = 36 VAR(2) terms. Adding only one additional coefficient to the model should be fine with your limited data.

    Since restricted VARs are a new feature, let me walk you through the GUI steps for this example. First, go to your VAR specification and change it to a VAR(2), i.e., lags "1 2". Next, go to the "VAR Restrictions" tab. This tab is very similar to the SVAR restrictions tab, but with pattern matrices for the VAR's endogenous lag terms (matrices L1 and L2) and pattern vectors for the exogenous terms (C, @TREND, etc.). We want to restrict the L2 matrix to be all zeros except for the element that corresponds to interest_rate(-2) in the dcpi equation. That is element <2,3> (hover the cursor over that element to double check). Leave that element as "NA" (unrestricted) and change all the other elements to zero. Alternatively, you could create a matrix object with this pattern and specify it by name in the dialog. You're now ready to estimate the VAR.
I hope experimenting with the above improves your results.

nasa
Posts: 39
Joined: Tue Aug 02, 2016 12:41 am

Re: Imposing Restrictions on SVAR

Postby nasa » Thu Jul 27, 2017 12:45 am

Thank you very much.

I did the VAR restriction,however,sensitivity of the results is still there.Do you have any more solution for this VAR model,please?

Best,

nasa
Posts: 39
Joined: Tue Aug 02, 2016 12:41 am

Re: Imposing Restrictions on SVAR

Postby nasa » Thu Jul 27, 2017 7:46 am

One last question,

Is it a problem when confidence bands obtained from monte carlo simulations are getting wider? What does it signifies?

Thanks a lot.

EViews Matt
EViews Developer
Posts: 158
Joined: Thu Apr 25, 2013 7:48 pm

Re: Imposing Restrictions on SVAR

Postby EViews Matt » Thu Jul 27, 2017 11:27 am

Widening Monte Carlo confidence bands indicates that the impulse response is sensitive to either changes in the estimated VAR and/or SVAR coefficients. In other words, a relatively small change in a coefficient may cause a large change in the impulse response. This gives you some idea of how the uncertainly in the estimates affects the uncertainty in the impulse responses.

Regarding your VAR, I believe I'm out of advice. You may just be using an SVAR model that is numerically challenging. If the estimated coefficients are the same in the cases when it does converge, that's a good sign. If the coefficients change significantly then there could be a problem with identification in the SVAR model.

nasa
Posts: 39
Joined: Tue Aug 02, 2016 12:41 am

Re: Imposing Restrictions on SVAR

Postby nasa » Fri Jul 28, 2017 6:24 am

Ok I appreciate all your support

But I want to send these impulse responses to the referee(as he ordered me to use monte carlo rather than analytic),hence what is your opinion on this matter? I mean what the referee can say regarding the widening monte carlo confidence bands? Do you have any solution for this?

Best,

nasa
Posts: 39
Joined: Tue Aug 02, 2016 12:41 am

Re: Imposing Restrictions on SVAR

Postby nasa » Fri Jul 28, 2017 6:27 am

and one point my SVAR estimated coefficients are the same in all cases when they do converge,so that is a good sign as you said.

EViews Matt
EViews Developer
Posts: 158
Joined: Thu Apr 25, 2013 7:48 pm

Re: Imposing Restrictions on SVAR

Postby EViews Matt » Fri Jul 28, 2017 9:45 am

I just graphed the raw data and it looks like you have a mix of I(0) and I(1) processes (a quick ADF test appears to confirm this). As an alternative to what we've done previously, I quickly tried respecifying the VAR to use d(gdppc), dcpi, log(interest_rate), d(ner), d(spendpc), and taxpc1 with no additional dummy term for the spike in interest_rate (but still c, @trend, dummya, and dummyb) and no VAR(2) terms. This seems to result in a more numerically stable SVAR with well-behaved impulse responses. There may still be some benefit in including the additional dummy and the interest_rate(-2) term, but I'll leave that experimentation to you.


Return to “Estimation”

Who is online

Users browsing this forum: No registered users and 5 guests