Interpreting impulse response functions: Std dev or % ?
Moderators: EViews Gareth, EViews Moderator

 Posts: 9
 Joined: Fri Feb 12, 2016 3:33 am
Interpreting impulse response functions: Std dev or % ?
Hi to all,
I have a very basic question, but I am a bit stuck right now...
I am trying to estimate different VAR models to see the impact of an uncertainty shock on macro variables (industrial production, employment...) and FX (cf attached file).
My question is the following: how do I interpret the impulse response numbers (see for instance tri_graph or tri_coeff) in the file) ?
Are they directly interpretable in % terms ? Say, a 1 std variation in uncertainty has a 0.6% impact on FX (given that the variable is logdifferentiated) after 4 months?
It looks very small to me, so either my model has some issues, or I misinterpret the IRF.
Also, how do I know the SD used by Eviews for the shock ? In other words, say I want to calibrate the impact of an uncertainty shock wrt past events...say a variation of 200pts of my uncertainty index. If I want to know how many Eviews' SD it corresponds to, do I just compute the SD of my uncertainty index? Is it the same as the SD computed in the output file at the line "S.D dependent"?
Thanks for the help!
I have a very basic question, but I am a bit stuck right now...
I am trying to estimate different VAR models to see the impact of an uncertainty shock on macro variables (industrial production, employment...) and FX (cf attached file).
My question is the following: how do I interpret the impulse response numbers (see for instance tri_graph or tri_coeff) in the file) ?
Are they directly interpretable in % terms ? Say, a 1 std variation in uncertainty has a 0.6% impact on FX (given that the variable is logdifferentiated) after 4 months?
It looks very small to me, so either my model has some issues, or I misinterpret the IRF.
Also, how do I know the SD used by Eviews for the shock ? In other words, say I want to calibrate the impact of an uncertainty shock wrt past events...say a variation of 200pts of my uncertainty index. If I want to know how many Eviews' SD it corresponds to, do I just compute the SD of my uncertainty index? Is it the same as the SD computed in the output file at the line "S.D dependent"?
Thanks for the help!
 Attachments

 var brexit.wf1
 (528.07 KiB) Downloaded 1046 times
Re: Interpreting impulse response functions: Std dev or % ?
You can use the sirf addin. It will help you convert SD into %.

 Posts: 9
 Joined: Fri Feb 12, 2016 3:33 am
Re: Interpreting impulse response functions: Std dev or % ?
dakila wrote:You can use the sirf addin. It will help you convert SD into %.
thanks for your answer. Where can I find it though? I search on th eadd in page and did not see it.
thanks!
Re: Interpreting impulse response functions: Std dev or % ?
Sorry. You need to wait. I hope Eviews will post it soon on the web.

 Posts: 9
 Joined: Fri Feb 12, 2016 3:33 am
Re: Interpreting impulse response functions: Std dev or % ?
Ok, thanks.
Meanwhile, any other answers?
Meanwhile, any other answers?
Re: Interpreting impulse response functions: Std dev or % ?
I am sure Dakila's addin will be a lifesaver as usual. The following should clarify few points for you in the meantime:
Yes, they are. Your interpretation is correct and the magnitude of impact is quite plausible.
The very first number in the column of impulse variable (i.e. 37.389) is the SD used by EViews.
No. It comes from the Cholesky decomposition of the covariance matrix of the residuals of your VAR model.
IRF coefficients are linear. So you can scale the responses as you like. One standard deviation is 37 in your case and you want to compute your own impact shock of 200. Then you would need to multiply all the responses to that shock by 200/37.
lillumultipass wrote:Are they directly interpretable in % terms ? Say, a 1 std variation in uncertainty has a 0.6% impact on FX (given that the variable is logdifferentiated) after 4 months?
Yes, they are. Your interpretation is correct and the magnitude of impact is quite plausible.
lillumultipass wrote:Also, how do I know the SD used by Eviews for the shock ?
The very first number in the column of impulse variable (i.e. 37.389) is the SD used by EViews.
lillumultipass wrote:Is it the same as the SD computed in the output file at the line "S.D dependent"?
No. It comes from the Cholesky decomposition of the covariance matrix of the residuals of your VAR model.
lillumultipass wrote:say I want to calibrate the impact of an uncertainty shock wrt past events...say a variation of 200pts of my uncertainty index
IRF coefficients are linear. So you can scale the responses as you like. One standard deviation is 37 in your case and you want to compute your own impact shock of 200. Then you would need to multiply all the responses to that shock by 200/37.

 Posts: 9
 Joined: Fri Feb 12, 2016 3:33 am
Re: Interpreting impulse response functions: Std dev or % ?
thanks Trubador, this is much clearer!
However, I do have a few additional questions:
1. ok for that, but what about the same variable in logs (not log differenced) ? In levels?
2. I am asking this because looking at the literature on these issues (i.e., impact of uncertainty on macro variables), I see widely different ways of modeling, i.e., people either logdifference the variables, or detrend them using an HP filter, or use variables in levels (given that most of them, such as industrial production, are not stationary). I know that it is not always necessary to have stationary variables in a VAR if you are mostly interested in the IRF...but I don't know how to discriminate between the different "methods". And then, I am not sure how to interpret the IRF when I am in levels terms and in log (levels).
However, I do have a few additional questions:
1. ok for that, but what about the same variable in logs (not log differenced) ? In levels?
2. I am asking this because looking at the literature on these issues (i.e., impact of uncertainty on macro variables), I see widely different ways of modeling, i.e., people either logdifference the variables, or detrend them using an HP filter, or use variables in levels (given that most of them, such as industrial production, are not stationary). I know that it is not always necessary to have stationary variables in a VAR if you are mostly interested in the IRF...but I don't know how to discriminate between the different "methods". And then, I am not sure how to interpret the IRF when I am in levels terms and in log (levels).
Re: Interpreting impulse response functions: Std dev or % ?
Be careful that responses are always in the original variables and they are in whatever are the units of the variable. You cannot interpret the responses as percentage changes unless the data are put into the model in logs. In your case, response variables are log differenced, but you are interested in accumulated response which yields a similar interpretation to log level.
Transforming the variables before putting them into a VAR may have a number of reasons, which may differ among the research questions studied or hypotheses tested in those papers you have seen. And there is nothing wrong to try all these approaches and see for yourself how (and why) the results differ with respect to one another.
Transforming the variables before putting them into a VAR may have a number of reasons, which may differ among the research questions studied or hypotheses tested in those papers you have seen. And there is nothing wrong to try all these approaches and see for yourself how (and why) the results differ with respect to one another.

 Posts: 9
 Joined: Fri Feb 12, 2016 3:33 am
Re: Interpreting impulse response functions: Std dev or % ?
Thanks again Trubador !
However, I may be a bit dense today, but am not sure I follow you when you say "response variables are log differenced, but you are interested in accumulated response which yields a similar interpretation to log level."
So I took a screen capture of the 3 versions of my VAR: basically it relates an uncertainty index, industrial production, inflation and interest rates.
I look at the IRF from a shock on UNCER:
In uncertainty 1, all variables are in log (except the interest rate, which always appear as a difference in levels) ; in uncertaintybis, all variables are logdifferenced and in the last one, in level terms.
In uncertainty 1, it says that the impact of a 1 SD rise in uncer on LIP more or less peaks at 6 months (0,4), whereas the impact on Logip is much more shortlived (2 months and only 0,1%) and finally the impact on IP in levels is more or less the same as on LIP but in units of industrial production, which are hard to interpret.
So basically, if I understand correctly, in the 2nd case, I can talk about % but, in terms of interpretation, do I have to say that IP will fall by at most 0,1% after 2 months and then the effect will disappear, or, do I have to say that the growth rate of IP will be 0,1% lower after 2 months ? I think the latter right?
I see the interpretation for the 3rd case but I am unsure about the 1st one. I guess in that case I can say that IP will fall by 0,1% to 0.4% each month over the 1st 810 months, and, if I accumulate responses, by about 9% after 24 months.
Is that right?
I promise, I will stop asking dumb questions after that
Thanks immensely for your help!
However, I may be a bit dense today, but am not sure I follow you when you say "response variables are log differenced, but you are interested in accumulated response which yields a similar interpretation to log level."
So I took a screen capture of the 3 versions of my VAR: basically it relates an uncertainty index, industrial production, inflation and interest rates.
I look at the IRF from a shock on UNCER:
In uncertainty 1, all variables are in log (except the interest rate, which always appear as a difference in levels) ; in uncertaintybis, all variables are logdifferenced and in the last one, in level terms.
In uncertainty 1, it says that the impact of a 1 SD rise in uncer on LIP more or less peaks at 6 months (0,4), whereas the impact on Logip is much more shortlived (2 months and only 0,1%) and finally the impact on IP in levels is more or less the same as on LIP but in units of industrial production, which are hard to interpret.
So basically, if I understand correctly, in the 2nd case, I can talk about % but, in terms of interpretation, do I have to say that IP will fall by at most 0,1% after 2 months and then the effect will disappear, or, do I have to say that the growth rate of IP will be 0,1% lower after 2 months ? I think the latter right?
I see the interpretation for the 3rd case but I am unsure about the 1st one. I guess in that case I can say that IP will fall by 0,1% to 0.4% each month over the 1st 810 months, and, if I accumulate responses, by about 9% after 24 months.
Is that right?
I promise, I will stop asking dumb questions after that
Thanks immensely for your help!
Re: Interpreting impulse response functions: Std dev or % ?
In the first case, the interpretation of the impact should be in percentages since the response variables are all in logarithms. One SD shock to UNCER leads to a 0.004 units decrease in the logarithm of LIP after 6 months, which corresponds to 0.4% drop in the LIP when translated into original level.
In the second case, however, one SD shock to UNCER leads to a 0.001 units decrease in the change in the logarithm of LIP after 4 months, which corresponds to 0.1 percentage points drop in the growth rate of LIP (e.g. from 2% to 1.9%). The only way to translate the impact into original level is to accumulate the responses, which can then be interpreted similar to the first case.
Third case is more straightforward. One SD shock to UNCER leads to 0.5 units decrease in the original level of LIP after 6 months. And that’s it.
In the second case, however, one SD shock to UNCER leads to a 0.001 units decrease in the change in the logarithm of LIP after 4 months, which corresponds to 0.1 percentage points drop in the growth rate of LIP (e.g. from 2% to 1.9%). The only way to translate the impact into original level is to accumulate the responses, which can then be interpreted similar to the first case.
Third case is more straightforward. One SD shock to UNCER leads to 0.5 units decrease in the original level of LIP after 6 months. And that’s it.

 Posts: 9
 Joined: Fri Feb 12, 2016 3:33 am
Re: Interpreting impulse response functions: Std dev or % ?
Ok, great, that's what I thought.
The 1st case is the most straightforward to interpret I would say, and I guess the reason why most people use logs in the first place.
Thanks again for taking the time to answer and for your patience
The 1st case is the most straightforward to interpret I would say, and I guess the reason why most people use logs in the first place.
Thanks again for taking the time to answer and for your patience

 Posts: 8
 Joined: Wed Aug 17, 2016 3:29 pm
Re: Interpreting impulse response functions: Std dev or % ?
Hello,
I want to have scaled IRFs as well, for my VAR.
I have installed the stir addin. So, I select the VAR and click on the "Scaled IRF" in the addin menu.
Now, I have a message box that asks me 3 things:
 Number of scale factor
 Impulse variable
 Number of perdod
Except the last one, I don't know what to put inside these boxes. So I try the following:
 Number of scale factor > it's 1 by default, so I leave it as 1
 Impulse variable > I put the name of one of the variables in the VAR
 And I choose 10 for the number of periods.
When I click "OK",I get the following message:
"0 is not a valid index for vectorseriescoefficient FAC_DIAG01"
What should I do? Could you please help me?
Thank you,
I want to have scaled IRFs as well, for my VAR.
I have installed the stir addin. So, I select the VAR and click on the "Scaled IRF" in the addin menu.
Now, I have a message box that asks me 3 things:
 Number of scale factor
 Impulse variable
 Number of perdod
Except the last one, I don't know what to put inside these boxes. So I try the following:
 Number of scale factor > it's 1 by default, so I leave it as 1
 Impulse variable > I put the name of one of the variables in the VAR
 And I choose 10 for the number of periods.
When I click "OK",I get the following message:
"0 is not a valid index for vectorseriescoefficient FAC_DIAG01"
What should I do? Could you please help me?
Thank you,
Re: Interpreting impulse response functions: Std dev or % ?
First, you have to identify the structural shock (use Eviews impulse function).
After that use the sirf addin.
After that use the sirf addin.

 Posts: 8
 Joined: Wed Aug 17, 2016 3:29 pm
Re: Interpreting impulse response functions: Std dev or % ?
dakila wrote:First, you have to identify the structural shock (use Eviews impulse function).
After that use the sirf addin.
Yes, I did that. But with the sirf addin, I get the following message, when I want to have the scaled IRFs:
"0 is not a valid index for vectorseriescoefficient FAC_DIAG01"
Re: Interpreting impulse response functions: Std dev or % ?
I think you did not insert the impulse variable properly. For example you put off_sa instead of d(off_sa)
Return to “Econometric Discussions”
Who is online
Users browsing this forum: No registered users and 3 guests