Diagnostic Testing of SVAR using Monte Carlo Integration

[sanjeev]

Respected Sir


I am using Structural VAR model with Monte Carlo Integration to derive my results. My question is: whether I need to do "diagnostic testings" in this case.

if yes, then which tests need to be applied?.

Regards

[TomDoan]

Diagnostic tests of what? If you're doing some form of Gibbs sampling, you would want to do standard diagnostics on the chain to see if has a reasonable degree of "forgetfulness". If you're talking about diagnostic tests for the model itself, that's not what MC integration is designed to do.

[sanjeev]

Yes I am talking of about diagnostic tests for the model itself.

I explained to my supervisor that in case of MC integration, diagnostic tests for the model is not required. She asked me to support my argument with some reference . Sir,

Can you please suggest me some paper or book so that I can support the point that "in case of MC integration, diagnostic tests for the model is not required".

Thanks and Regards

[TomDoan]

Yes I am talking of about diagnostic tests for the model itself.

I explained to my supervisor that in case of MC integration, diagnostic tests for the model is not required. She asked me to support my argument with some reference . Sir,

Can you please suggest me some paper or book so that I can support the point that "in case of MC integration, diagnostic tests for the model is not required".

Thanks and Regards

I can't because that's a complete misinterpretation of what I said. If the model is wrong, MC integration isn't some magic bullet that fixes it---you're just doing simulation analysis around a flawed model. Formal hypothesis testing using MC Bayesian analysis is extremely difficult and requires that you are able to do MC analysis on a range of alternative models. You should do the testing based upon ML estimation and LR tests. If the LR test on CVMODEL strongly rejects your model, fix the model.

[sanjeev]

Yes Sir, I have already applied LR test based on ML estimation, that did not reject my model.


I just want to know whether I now need to do diagnostic testing i.e. test for autocorrelation and test for hetroscendasticity ???

Thanks and Regards

[TomDoan]

If you used a reasonable procedure for picking lag length, there probably is no reason to test for residual autocorrelation, though it certainly doesn't hurt. Generally, you don't test a VAR for heteroscedasticity simply because there isn't much that you can do about it if you find it.