Kim and Roubini (2000)

[researcher84]

Dear Tom and RATS visitors,

I just started looking at different codes in RATS and really new with the software. I am trying to use both Kim and Roubini (2000) along with Cushman and Zha (1997). Is it possible to estimate this model using RATS ? where should I start (any codes for the paper) ? is it a problem that we have within the identification restriction both interest rate and money supply along with interest rate and exchange rate response simultaneously to each others ? is it possible to obtain the variance decomposition in this case?? I am really in need for help to get started with my research. Thanks to all of you and especially to Tom for all the good work he puts in this site.

[TomDoan]

Sure. That's a CVMODEL type A similar to the second model in Example 7.4 in the RATS v8 User's Guide.

Code: Select all

nonlin b21 b31 b41 b43 b53 b54 b56 b61 b65 b67 b71 b72 b73 b74 b75 b76
dec frml[rect] afrml
frml afrml = || 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0|$
               -b21, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0|$                  
               -b31, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0|$
               -b41, 0.0,-b43, 1.0, 0.0, 0.0, 0.0|$
                0.0, 0.0,-b53,-b54, 1.0,-b56, 0.0|$
               -b61, 0.0, 0.0, 0.0,-b65, 1.0,-b67|$
               -b71,-b72,-b73,-b74,-b75,-b76, 1.0||

Estimate the standard VAR, then use a CVMODEL with A=AFRML as is done in the second case in that example. Whether the all-zero guess values for the off-diagonals will work isn't clear with a model with the "simultaneity" shown in this one with the B56-B65 and B67-B76 pairs, so you might have to do a bit of experimenting. That's why Example 7.4 uses PMETHOD=GENETIC, which does a fairly broad search.

[terrya]

My experience with the KR model on a data set different from KR's reinforces your comments. I updated my data set from 2010(Q3) to 2013(Q3) and ran the model again. The old data set resulted in plausible solutions. Every estimation with the new data set produced different outcomes from the others, including failures to converge.

I've also used EViews 8 to attempt to estimate the model. Every estimation failed.

[TomDoan]

You can try with the DMATRIX=IDENTITY option (which requires you to replace the 1's on the diagonal with free parameters). The problem with those bottom three equations is that the normalization to 1's on the diagonal aren't necessarily going to hold when the model is estimated. If one of those normalizations is wrong (that is, it'son a coefficient which actually is truly near zero), then you run into convergence problems since the normalization forces the other non-zero coefficients in that row to be huge.

[terrya]

Thanks. I'll try this when I get back to my computer (I'm using an iPad now). In the estimations I've done, some of parameters were very large.

[terrya]

Thanks for you suggestion.

I used it and got convergence and the same estimates for the bs each time I estimated the model. The IRFs appeared implausible. I think that probably the model is inappropriate for the data (small country) I'm using.

[researcher84]

Thank you to both of you, Tom and Terrya .. As I mentioned I am really new to this program and just ordered a copy ... I will try what you suggested Tom but Terrya do you mind sending me your codes? I want to see how to set up the model and most likely I will modify the original model to include more variables..Thanks in advance and hope you help me with that Terrya. Thank you all.

[terrya]

Hi

All I did was follow the instructions in the user/reference guides and various example programmes and posts on this forum (including Tom's in this thread, which is identical with what I did initially before I applied his suggestion concerning the dmatrix). That'll be far more useful to you for learning than me providing you with exactly the same realisation using the afrml.

Also, look at the example programme for the cvmodel (cvmodel.rpf) that's in the guide and supplied with RATS.

[researcher84]

Hi there,

I am following Kim and Robin (2000) and I have few questions that I would appreciate any help or feedback on them:

1) are there any codes available for that paper now?

2) As you might know, KR uses the log of levels of the variables in their model and so I follow that. The impulse response I get do not go to zero as shown in the attached images. I looked also in KR results and also some of their impulse response do not go to zero. Any feedback on how to justify the results would be greatly appreciated. I am only interested in the qualitative analysis ( increase or decrease). One might say this is because of unit root, KR do not test for unit root or mentioned that. Thanks

Here is an example of the results from KR:

http://postimg.org/image/e8yql9o3r/

Here is results from my paper:

http://postimg.org/image/q4ld53ozz/

http://postimg.org/image/7oc3yuzvf/

http://postimg.org/image/beqlsfvqt/

[TomDoan]

Apparently, the authors couldn't find the original data. The program was done with RATS version 3---I assume with the old @BERNANKE procedure since CVMODEL didn't exist at the time. It's a straightforward structural VAR, though the presence of the "simultaneity" in variables six and seven in the formulation in https://estima.com/forum/viewtopic.php?p=8338#p8338.

Regarding stationarity, see footnote 10 in the paper.

An observation on non-stationarity and co-integration. Our statistical inference is not affected
by presence of non-stationarity since we follow a Bayesian inference (see Sims, 1988; Sims and Uhlig,
1991). Some theoretical models predict that some variables may be cointegrated or integrated. If we
strongly believe that such a relation exists, then we have to impose such a restriction; it is however,
not clear whether the variables in the system are indeed co-integrated or integrated. Moreover, if we
impose false restrictions, our inference would be incorrect.

[researcher84]

Hi Tom,

Does the impulse responses I attached look fine to you?

[TomDoan]

There's nothing obviously wrong with them. I can't say anything beyond that---it's your job to figure out if they make sense theoretically.

[researcher84]

There's nothing obviously wrong with them. I can't say anything beyond that---it's your job to figure out if they make sense theoretically.

Dear Tome,

Thanks for your reply. I do have the theoretical justification of the response and it makes very sense. What worries me is that the response does not go to zero. If you look at the response to shock 3 in the following graph, the effect of the shock does not go to zero. Any take on that?

Thanks a lot for your help!
http://postimg.org/image/beqlsfvqt/

[TomDoan]

First of all, 19 steps doesn't really determine the limit. But did you read the quote? They didn't test for unit roots or cointegration because they rely on the results that those are unnecessary, but these are (all) series which are generally conceded to have unit roots.