Output of VAR model with Generalized Impulse Response

[bok1234]

Dear Mr. Doan,

I have several questions about VAR model with Generalized Impulse Response.
I attached 5 files as below.

KEXQ&5Countries_3.xls : data file
IMPULSES.RPF_6 by 6 IRS graphs.txt : code file (generating 6 by 6 IRS graphs)
IMPULSES.RPF_6 by 6 IRS excel file.txt : code file (generating 6 by 6 IRS excel file)
6 by 6 IRS.PNG : 6 by 6 IRS graphs
IRS & C.I_log diff_lag 1.xlsx : the first row of 6 by 6 IRS graphs (my interest)

For reference, My codes are for the IRS of ‘Export Volume (KEXQ)’ to Shocks of 5 countries (EEE, JJJ, UUU, CCC, AAA)’s real GDPs. I used log-difference varibles, disregarding log variables.

Question 1.
I want to know the confidence interval (95% or 90%) of this model. And let me know that how I can change the option of C.I. (e.g. from 90% to 95%)

Question 2.
If you compare the 5 panels from (1,2) to (1,6) of the first row of ‘6 by 6 IRS.PNG’ with 5 graphs of ‘IRS & C.I_log diff_lag 1.xlsx’, then you can find that each graph is different. But they should be same. Of course, Although some of them such as ‘IRS to CCC’ in excel file is very similar to (1,5) panel of png file, they are not same. Most of all, every time 0 line in ‘extracted’ sheet of ‘IRS & C.I_log diff_lag 1.xlsx’ has zero value but 5 panels from (1,2) to (1,6) of the first row of ‘6 by 6 IRS.PNG’ do not have zero value.
I don’t think that this is just x-axis problem (whether lag begins from time 0 or time 1). I want to know why these discords happen.

Question 3.
the 5 panels from (1,2) to (1,6) of the first row of ‘6 by 6 IRS.PNG’ mean that IRS of ‘Export Volume (KEXQ)’ to EACH ONE STANDARD DEVIATION SHOCK of 5 countries (EEE, JJJ, UUU, CCC, AAA)’ real GDPs.
I want to know that the ‘% RESPONSE’ of KEXQ to ‘% CHANGE SHOCK’ of each country’s GDP – for example, if CCC’s GDP change 1%, then how does export volume change in % unit? To estimate this, should I need each 1 s.d. value of each country? For example, ( M : value of IRS = 1% : X ) => X = value of IRS / M
where M = 1 s.d. value of country CCC’ real GDP / total value of country CCC’ real GDP. Unfortunately, I am not sure about this.
I guess that you already have much smarter solution for this.

[TomDoan]

Dear Mr. Doan,

I have several questions about VAR model with Generalized Impulse Response.
I attached 5 files as below.

KEXQ&5Countries_3.xls : data file
IMPULSES.RPF_6 by 6 IRS graphs.txt : code file (generating 6 by 6 IRS graphs)
IMPULSES.RPF_6 by 6 IRS excel file.txt : code file (generating 6 by 6 IRS excel file)
6 by 6 IRS.PNG : 6 by 6 IRS graphs
IRS & C.I_log diff_lag 1.xlsx : the first row of 6 by 6 IRS graphs (my interest)

For reference, My codes are for the IRS of ‘Export Volume (KEXQ)’ to Shocks of 5 countries (EEE, JJJ, UUU, CCC, AAA)’s real GDPs. I used log-difference varibles, disregarding log variables.

Question 1.
I want to know the confidence interval (95% or 90%) of this model. And let me know that how I can change the option of C.I. (e.g. from 90% to 95%)

@MONTEVAR actually does a 68% (robust analog of 1 s.d.) confidence interval. Use the combination of @MCVARDODRAWS and @MCGRAPHIRF to create different confidence intervals. See, for instance, the MONTEVECM.RPF example.

Question 2.
If you compare the 5 panels from (1,2) to (1,6) of the first row of ‘6 by 6 IRS.PNG’ with 5 graphs of ‘IRS & C.I_log diff_lag 1.xlsx’, then you can find that each graph is different. But they should be same. Of course, Although some of them such as ‘IRS to CCC’ in excel file is very similar to (1,5) panel of png file, they are not same. Most of all, every time 0 line in ‘extracted’ sheet of ‘IRS & C.I_log diff_lag 1.xlsx’ has zero value but 5 panels from (1,2) to (1,6) of the first row of ‘6 by 6 IRS.PNG’ do not have zero value.
I don’t think that this is just x-axis problem (whether lag begins from time 0 or time 1). I want to know why these discords happen.

This is from your "xlsx" targeting program. You're not including the FFUNCTION option on the @MCVARDODRAWS, so it's doing Cholesky factors.

@mcvardodraws(model=canmodel,draws=4000)
@mcprocessirf(model=canmodel,lower=lower,upper=upper,center=median,irf=irfs)

Question 3.
the 5 panels from (1,2) to (1,6) of the first row of ‘6 by 6 IRS.PNG’ mean that IRS of ‘Export Volume (KEXQ)’ to EACH ONE STANDARD DEVIATION SHOCK of 5 countries (EEE, JJJ, UUU, CCC, AAA)’ real GDPs.
I want to know that the ‘% RESPONSE’ of KEXQ to ‘% CHANGE SHOCK’ of each country’s GDP – for example, if CCC’s GDP change 1%, then how does export volume change in % unit? To estimate this, should I need each 1 s.d. value of each country? For example, ( M : value of IRS = 1% : X ) => X = value of IRS / M
where M = 1 s.d. value of country CCC’ real GDP / total value of country CCC’ real GDP. Unfortunately, I am not sure about this.
I guess that you already have much smarter solution for this.

In IRF terminology, one s.d. is one s.d. of the VAR residuals, not of the raw data.

[bok1234]

Dear Mr. Doan,


About my Question 2, should I replace ‘model=canmodel’ in my code with ‘model=varmodel’ to use FFunction as your example below?

@mcvardodraws(model=varmodel,draws=mcdraws,steps=nsteps)
@mcgraphirf(model=varmodel,$
center=input,impulses=baseirfs,percent=||.025,.16,.84,.975||,$
footer="Figure 2. Pointwise 68 and 95% Posterior Bands, Y-M Model")


About my Question 3, sorry for a wrong (silly) question and thank you for your correction. I changed the estimate option from noprint to print, and then got result as below. Although that I have heard that VAR model’s coefficients are not very important, the result shows that 1% increase of EEE influences –0.72% of KEXQ after 1 quarter in this case (if ignoring the significant level of coefficients). Can I think like this and is this all about elasticity between KEXQ and RGDP_EEE in this model?


Dependent Variable LD_KEXQ
Mean of Dependent Variable 0.0209222772
Std Error of Dependent Variable 0.0400270404
Standard Error of Estimate 0.0343209515
Sum of Squared Residuals 0.1130810601
Durbin-Watson Statistic 2.0893
Variable Coeff Std Error T-Stat Signif
************************************************************************************
1. LD_KEXQ{1} -0.151045482 0.102265521 -1.47699 0.14295049
2. LD_RGDP_EEE{1} -0.724522795 0.495159432 -1.46321 0.14667612
3. LD_RGDP_JJJ{1} -0.701607765 0.416440248 -1.68477 0.09528060
4. LD_RGDP_UUU{1} 1.584921695 0.660112495 2.40099 0.01827758
5. LD_RGDP_CCC{1} 1.225006673 0.210740011 5.81288 0.00000008
6. LD_RGDP_AAA{1} 0.574795553 0.306355023 1.87624 0.06366022
7. Constant -0.012774011 0.007195136 -1.77537 0.07900630
F-Tests, Dependent Variable LD_KEXQ
Variable F-Statistic Signif
*******************************************************
LD_KEXQ 2.1815 0.1429505
LD_RGDP_EEE 2.1410 0.1466761
LD_RGDP_JJJ 2.8385 0.0952806
LD_RGDP_UUU 5.7647 0.0182776
LD_RGDP_CCC 33.7896 0.0000001
LD_RGDP_AAA 3.5203 0.0636602

[TomDoan]

Dear Mr. Doan,


About my Question 2, should I replace ‘model=canmodel’ in my code with ‘model=varmodel’ to use FFunction as your example below?

@mcvardodraws(model=varmodel,draws=mcdraws,steps=nsteps)
@mcgraphirf(model=varmodel,$
center=input,impulses=baseirfs,percent=||.025,.16,.84,.975||,$
footer="Figure 2. Pointwise 68 and 95% Posterior Bands, Y-M Model")

I have no idea. I copied those lines straight out of your "excel..." program, which uses CANMODEL for the model name. The point is that you're comparing a GIRF with a Cholesky.

About my Question 3, sorry for a wrong (silly) question and thank you for your correction. I changed the estimate option from noprint to print, and then got result as below. Although that I have heard that VAR model’s coefficients are not very important, the result shows that 1% increase of EEE influences –0.72% of KEXQ after 1 quarter in this case (if ignoring the significant level of coefficients). Can I think like this and is this all about elasticity between KEXQ and RGDP_EEE in this model?


Dependent Variable LD_KEXQ
Mean of Dependent Variable 0.0209222772
Std Error of Dependent Variable 0.0400270404
Standard Error of Estimate 0.0343209515
Sum of Squared Residuals 0.1130810601
Durbin-Watson Statistic 2.0893
Variable Coeff Std Error T-Stat Signif
************************************************************************************
1. LD_KEXQ{1} -0.151045482 0.102265521 -1.47699 0.14295049
2. LD_RGDP_EEE{1} -0.724522795 0.495159432 -1.46321 0.14667612
3. LD_RGDP_JJJ{1} -0.701607765 0.416440248 -1.68477 0.09528060
4. LD_RGDP_UUU{1} 1.584921695 0.660112495 2.40099 0.01827758
5. LD_RGDP_CCC{1} 1.225006673 0.210740011 5.81288 0.00000008
6. LD_RGDP_AAA{1} 0.574795553 0.306355023 1.87624 0.06366022
7. Constant -0.012774011 0.007195136 -1.77537 0.07900630
F-Tests, Dependent Variable LD_KEXQ
Variable F-Statistic Signif
*******************************************************
LD_KEXQ 2.1815 0.1429505
LD_RGDP_EEE 2.1410 0.1466761
LD_RGDP_JJJ 2.8385 0.0952806
LD_RGDP_UUU 5.7647 0.0182776
LD_RGDP_CCC 33.7896 0.0000001
LD_RGDP_AAA 3.5203 0.0636602

That would be the case if you were doing unit shocks. If you're doing any shocks which allow for correlations, no.

[bok1234]

Dear Mr.Doan,

About my 1st question, you replied that @MONTEVAR actually does a 68% (robust analog of 1 s.d.) confidence interval.
Some persons told me that because 68% C.I. is too low, so this result cannot be acceptable.
And they say that 90% or 95% C.I. are common criteria.
I think that they could be right.
But you might have some reason to contrive your GIR model based on 68% C.I. as a default.
Should 68% C.I. be criticized or ignored?
Although this could be a little subjective judgement problem.
I want to know your opinion on '68% C.I.' - actually I prefer your expertise to my colleagues' knowledge.

[TomDoan]

The use of the 68% was recommended in Sims, C.A. and T. Zha (1999), “Error Bands for Impulse Responses”, Econometrica, vol. 67, pp. 1113-1156. The point is that it shows a more representative collection of responses.

[bok1234]

Thank you Mr.Doan. Your opinion and the paper which you recommended helped me a lot.