Mountford & Uhlig JAE 2009 replication files

[TomDoan]

Use @MCPROCESSIRF (suggested two posts earlier in this thread). That does the calculation of the bands from the raw simulations (as the LOWER and UPPER options) so you can print or export them.

[pascal]

Dear Tom,
Thank you for your reply.
Do you mean, in the case of mu2009x2,

Code: Select all

@MCGraphIRF(model=varmodel,varlabels=vl,shocklabels=sl,$
  page=byshock,center=median,columns=2,include=||1,2,10,4,6,8,3,9,5,7||)

replace to:

Code: Select all

@MCPROCESSIRF(model=varmodel,varlabels=vl,shocklabels=sl,$
  page=byshock,center=median,columns=2,include=||1,2,10,4,6,8,3,9,5,7||)

?
But it shows: ## OP3. This Instruction Does Not Have An Option VAR
>>>>varmodel,varlabels=<<<<
I am a new user of RATS programme. Would you please instruct me how to solve this problem.
I would be grateful if you would help me.
Thank you very much
Pascal

[TomDoan]

You're copying a bunch of options off the @MCGRAPHIRF which are for doing the graphs (labels, page option, etc.). @MCPROCESSIRF doesn't do the graphs---it does the summary calculations and saves them so you can do whatever you want with them. See the example in the description of the procedure.

[arneskjaeveland]

Hi Tom.

I am currently trying to estimate the fiscal effects on output in Norway by using Sign restrictions. I have run the 2009B1 code and it works fine, but I have a couple of clarification questions.

First, the model being used in the setup file is as far as I can see specified as a standard OLS var model, but in the original paper they specify that they use a VAR with a bayesian approach (BVAR). Does the additional Monte Carlo integration + draws from the posterior make it bayesian? If so, is it unecessary to do "classical model" tests on the VAR model (Augmented Dfuller test, lagrange multiplier test, jbarca test etc.) to clearify if the model contains a unit root. I have read that bayesian models are less restricted to the issues with Unit roots, and degrees of freedom.

Second, to the ever lasting question on rescaling IRFs. I am guessing that the 100*log(variable) is to scale standard deviations in the model to percentages? Meaning that the IRFs can be understood originally as a 1 std shock to the variable (e.g log(spending)) in question, yields a 0.002 std response in e.g. log(GDP). Is this correct? If so, is anything done in the model to keep this linear relationship when using sign restrictions? (is the IRF outputs comparable to a traditional SVAR model with cholesky decomposition)

Lastly, I have noticed that different papers display their calculated fiscal multipliers in different ways to obtain dollar/dollar changes:

E.g. blanchard & Perrotti/Caldara & Kamps -->"the original impulse responses are first divided by the standard deviation of the fiscal shock in order to have shocks of size 1 per cent. These impulse responses are then divided by the sample mean of the ratio of the macroeconomic variable of interest and the shocked fiscal variable"

Mountford & Uhlig --> see attachment

Do you know if there is "one" correct way to calculate the impact multipliers for sign restricted models such as the Mu example?


Best regards,
Arne Skjæveland

[TomDoan]

Hi Tom.

I am currently trying to estimate the fiscal effects on output in Norway by using Sign restrictions. I have run the 2009B1 code and it works fine, but I have a couple of clarification questions.

First, the model being used in the setup file is as far as I can see specified as a standard OLS var model, but in the original paper they specify that they use a VAR with a bayesian approach (BVAR). Does the additional Monte Carlo integration + draws from the posterior make it bayesian? If so, is it unecessary to do "classical model" tests on the VAR model (Augmented Dfuller test, lagrange multiplier test, jbarca test etc.) to clearify if the model contains a unit root. I have read that bayesian models are less restricted to the issues with Unit roots, and degrees of freedom.

I believe they may have used a weak Minnesota prior for the lag coefficients, which would slightly modify the sampling procedure (see the GIBBSVAR.RPF program) but would likely not materially affect the results. Unit root behavior of the data will have no effect on the short-term responses that are being analyzed. (Note that they don't do unit root tests despite most of the series being strongly trending).

Second, to the ever lasting question on rescaling IRFs. I am guessing that the 100*log(variable) is to scale standard deviations in the model to percentages? Meaning that the IRFs can be understood originally as a 1 std shock to the variable (e.g log(spending)) in question, yields a 0.002 std response in e.g. log(GDP). Is this correct? If so, is anything done in the model to keep this linear relationship when using sign restrictions? (is the IRF outputs comparable to a traditional SVAR model with cholesky decomposition)

Yes. There is no change is the linearity of scaling (and any other linear transformations). The shocks in the MU analysis are in natural scale (based upon the values of the covariance matrix) just like the one standard deviation Cholesky shocks.

Lastly, I have noticed that different papers display their calculated fiscal multipliers in different ways to obtain dollar/dollar changes:

E.g. blanchard & Perrotti/Caldara & Kamps -->"the original impulse responses are first divided by the standard deviation of the fiscal shock in order to have shocks of size 1 per cent. These impulse responses are then divided by the sample mean of the ratio of the macroeconomic variable of interest and the shocked fiscal variable"

Mountford & Uhlig --> see attachment

Do you know if there is "one" correct way to calculate the impact multipliers for sign restricted models such as the Mu example?

I'm not sure that those are really all that different. The "average fiscal variable share" in MU and sample mean of the ratio... in BP sound like the same thing. BP are doing a structural VAR, so there is just a single calculation, while MU are doing a median value of the same thing.