Identifying VARs with sign restrictions

[smallick]

Dear RATS user,
I like your code which allows one to calculate cumulative responses and the multiplier as in Mountford and Uhlig (2009). When I try to integrate your additional bits into the replication code for MU (made available by Estima under resources), I cannot get these extra codes to run without error.
I would be grateful for your time if you can send me the steps for how to adapt the bits on cumulative responses and the multiplier to the newest version of the replication code that is currently available under resources.
Many thanks
Sushanta
http://webspace.qmul.ac.uk/skmallick/

Dear Tom,

Thank you so much for correcting my codes!....I am now trying to adjust the code to calculate the impulse responses of the model subject to a restricted 1% shock to revenue. I then move on to calculate the cumulative response and the multiplier....the shape of the responses seem very similar to those in Mountford and Uhlig (2008). but somehow the mean response of multiplier is a bit weird....Is this normal? (for your information, I follow closely the GUASS codes of Mountford and Uhlig (2008), available at the website of journal of applied econometrics)

Again, many thanks!

MC

[MC128]

Hi,

Here is my adaptation to the latest version of Mountford and Uhlig code. You need to change some of the input, like the the source programme for the VAR, the variables restricted etc......the numbers in the calculation of multipliers are the sample ratio of fiscal variables and GDP.

Hope it help.

MC

[TL]

Dear Tom,

I would like to add dummy variables for the period of the Asian crisis, from Oct 97 to Sep 98.

With the series ‘DUMMY’ that is 0 during such period,

Code: Select all

OPEN DATA Zero_Crisis_Dummy_All_4_South_Korea.xls
    CALENDAR 1989 2 12
    compute missc=1.0e+32
    data(format=xlS,org=columns) 1989:2 2008:11 CPI Y INT SS FF PP DUMMY

    set CPI = (CPI*100.0)*DUMMY
    set Y   = (Y*100.0)*DUMMY
    set SS  = (SS*100.0)*DUMMY
    set FF  = (FF*100.0)*DUMMY
    set PP  = (PP*100.0)*DUMMY
    set INT = (INT)*DUMMY
    *
    system(model=varmodel)
    variables Y CPI INT PP FF SS
    lags 1 to 6
    deterministic constant DUMMY
    end(system)
    estimate(noprint,resid=resids)

I am not sure if that code is correct or not. Could you please give me suggestions?

Thank you very much for your help and kindness.

Sincerely,

Tim

[TomDoan]

I'm not sure what you're trying to do with the dummies. You've zeroed out your data during that period, but those will still be, for instance, lags for the post dummy data, which will really distort your results. If you want to skip that entirely, use a SMPL option.

[smallick]

Hello! thank you very much for your help.
I have one quick question. Could you pl let me know why you have put (2-1) in the revenue multiplier and (1-1) in the expenditure multiplier in the steps below:

compute mult1(draw)(m) = %%Impplc(draw)(1+(1-1)*nvar,m)/(%%Impplc(draw)(2+(1-1)*nvar,m)*0.142)
compute mult2(draw)(m) = -%%Impplc(draw)(1+(2-1)*nvar,m)/(%%Impplc(draw)(3+(2-1)*nvar,m)*0.138)

Is this impact multiplier or present value multiplier? If it is impact multiplier, I am just wondering why the numbers (0.142 and 0.138) are used in the above calculation. If they are discount factors, could you please give me some hint how you calculated it.
I will appreciate your reply very much.
Regards
Sushanta
s.k.mallick@qmul.ac.uk

Hi,

Here is my adaptation to the latest version of Mountford and Uhlig code. You need to change some of the input, like the the source programme for the VAR, the variables restricted etc......the numbers in the calculation of multipliers are the sample ratio of fiscal variables and GDP.

Hope it help.

MC

[MC128]

Hi,

The formula corresponds to the present value multiplier. The (2-1) indicate that it is the second policy experiment, which is the deficit financed tax cut....
The (1-1) indicate that is the first policy experiment, which is the deficit financed expenditure increase.

Please let me know if you find some mistake in the calculation.

MC

[TL]

Dear Tom,

Thank you very much for your reply.

I would like to include dummy variable (exogenous control) for the period of the Asian crisis in the VAR model using sign restrictions.

The dummy variable takes a value of 1 for that period of Asian financial crisis (Oct 97 – Sep 98) and 0 otherwise.

In order to perform this, is it correct to just add the dummy variable at the DETERMINISTIC instruction?

Code: Select all

    system(model=varmodel)
    variables Y CPI INT PP FF SS
    lags 1 to 6
    deterministic constant DUMMY
    end(system)
    estimate(noprint,resid=resids)

Thank you very much for your help and kindness.

Sincerely,

Tim

[TomDoan]

Dear Tom,

Thank you very much for your reply.

I would like to include dummy variable (exogenous control) for the period of the Asian crisis in the VAR model using sign restrictions.

The dummy variable takes a value of 1 for that period of Asian financial crisis (Oct 97 – Sep 98) and 0 otherwise.

In order to perform this, is it correct to just add the dummy variable at the DETERMINISTIC instruction?

Code: Select all

    system(model=varmodel)
    variables Y CPI INT PP FF SS
    lags 1 to 6
    deterministic constant DUMMY
    end(system)
    estimate(noprint,resid=resids)

Thank you very much for your help and kindness.

Sincerely,

Tim

You're not really answering the question about what you want the dummy to do. If you add the dummy as indicated, you're saying that you want a shift to the intercept (only) during that period. That's a single shift for the entire period and the lag coefficients stay the same. Is that what you want?

[TL]

Dear Tom,

Thank you very much for your reply.

I understand that that is what literature does (to include a discrete intercept dummy variable). However, would it be possible to also have a change in the lag coefficients?

I am sorry for not answering the question about what I want the dummy to do. Essentially, I am following literature, capturing an effect of the crisis through an introduction of the crisis dummy.

Thank you very much for your help and kindness.

Sincerely,

Tim

[TomDoan]

Dear Tom,

Thank you very much for your reply.

I understand that that is what literature does (to include a discrete intercept dummy variable). However, would it be possible to also have a change in the lag coefficients?

I am sorry for not answering the question about what I want the dummy to do. Essentially, I am following literature, capturing an effect of the crisis through an introduction of the crisis dummy.

Thank you very much for your help and kindness.

Sincerely,

Tim

I'd be hard-pressed to imagine how a constant dummy shift in a VAR could somehow be the correct way to handle that. But if that's standard practice, then you have that coded correctly.

Dummying out the lag coefficients as well would be equivalent to using a SMPL which excludes that period from the sample. Note that that does not exclude the entries in the crisis period from being in the lags for entries immediately afterwards.

[TL]

Dear Tom,

I have used the following codes with the data set (Jan 1995 – Nov 2008). Impulse responses acquired are unusual. Please see pdf attached.

When I use longer data set (Feb 1989 – Nov 2008), results are fine.

I am wondering whether this is because the data set is too short. Could you please give me suggestions on this?

Thank you very much for your help and kindness.

Sincerely,

Tim

[TomDoan]

Dear Tom,

I have used the following codes with the following data set (Jan 1995 – Nov 2008). Impulse responses acquired are unusual. Please see pdf attached.

When I use longer data set (Feb 1989 – Nov 2008), results are fine.

I am wondering whether this is because the data set is too short. Could you please give me suggestions on this?

Thank you very much for your help and kindness.

Sincerely,

Tim

Offhand, the only thing that might be described as "unusual" is the rapid widening towards the end of the IRF. That's not, in fact, that unusual; if you take almost any model, even one estimated with quite a bit of data, and run the IRF out too far, you'll eventually see that. This is due to the fact that the impulse responses will eventually be dominated by the largest root, which will typically be just a bit larger than 1. With a fairly small data set, the largest root is less well-determined than with a larger one, hence you will see the exploding confidence bands at shorter horizons. If you cut the IRF off at 36 rather than 60, the graphs won't look as odd.

[MC128]

Dear Tom,

Do you know how to modify your replication codes for the paper Mountford and Uhlig (2009) such that the size of the basic government spending and revenue shocks are normalised to 1% in the initial period?

Thank you so much!

MC

[luching]

I am also trying to identify multiple shocks using sign restrictions. I think I understand the codes above. But to begin with, I was wondering why it may be useful/necessary to identify multiple shocks SIMULTANEOUSLY. As in, what value additions do we get by identifying the shocks simultaneously instead of identifying each shocks SEQUENTIALLY using the original (one shocks) Uhlig codes? That is, run the Uhlig codes with different sign restrictions for each shock each time.

Any thoughts?

[TomDoan]

The two are quite different. If you run the procedure twice with two different sign restrictions, you isolate a pair of shocks which are very likely to be correlated with each other---there's nothing preventing them from being correlated, and, in fact, there may be nothing preventing them from being the same (it depends upon what restrictions you use). On the other hand, if you use this procedure with multiple sign restrictions, the shocks are uncorrelated by construction, which is what you would like if you're trying to identify separate fundamental shocks.