The RATS Software Forum

The default is the 16-84% band. See the discussion about this in the User's Guide. You can check the options for @MCPROCESSIRF at
http://www.estima.com/forum/viewtopic.p ... 96---those are basically the same on @MCGRAPHIRF.

Is it standard in the literature to use 16-84 % band noticed some of the paper that replicated with RATS here use the 16-84 % band .. I just realized this today, when I modify 90 or 95% I don't get the nice results as before...The error bands are large and most of them are insignificant .. this is really disappointing.. any suggestion?

You should read the Sims and Zha paper cited in the manual---that's the source of the choice for the 16-84%. That gives you a better idea of what the typical type of response is, since it covers over 2/3 of the replications. However, if quite a few of the responses are insignificant when you do 95% bands then, quite a few of the responses are insignificant at the 5% level. Changing the band width doesn't change that.

How to estimate the variance decomposition for the model 2 ? keep in mind I'm using "MONTESVAR.RPF " I just couldn't figure out how to estimate that with over identified model. Thanks!

researcher84 wrote:How to estimate the variance decomposition for the model 2 ? keep in mind I'm using "MONTESVAR.RPF " I just couldn't figure out how to estimate that with over identified model. Thanks!

First off, I'm not a fan of doing error bands for variance decompositions---the results are often misleading and sometimes are out-and-out-strange. (Yes, referees sometimes ask for it. Yes, referees sometimes don't understand certain issues.)

That's a complicated calculation when you use importance sampling. If you switch it over to M-H, you can use the already existing @MCFEVDTABLE procedure. The Bayesian Econometrics course covers the various sampling techniques for SVAR's in detail.

Hey Tom,
I hope things are going well. I don't want any error bands for the variance decomposition, what I need is to estimate the variance decomposition for the over identified model I have. I know how to do it for Cholesky decomposition but not when I have some over identifying restrictions. Is there a code or a procedure for that? Thanks alot.


btw... what does H-M stand for ?

If you want point estimates, just do an ERRORS instruction with a FACTOR option equal to the one returned by CVMODEL. The "MONTE" is MONTESVAR.RPF is for the Monte Carlo integration which is special for SVAR's.

M-H is Metropolis-Hastings.

If I want to use two different error bands in the same impulse response like what many papers show, say I have the same model, same response functions but I need the same graph (impulse response) to show two different confidence intervals in same graph ... is that doable using RATS ? Thanks in advance...

Those graphs tend to look rather busy, but you can use @MCPROCESSIRF to get whatever collection of quantiles you want, then create a GRAPH instruction with the appearance that you want.

hi dear tom:

i have some question about montesvar.1) do we can use montesvar for 10 variables? 2)what is the limitation of using of monesvar ?
i used montesvar for 10 variables but give me this error Effective sample size NA and
## SR10. Missing Values And/Or SMPL Options Leave No Usable Data Points The Error Occurred At Location 2657, Line 180 of MCGRAPHIRF.please guid me how fix this errors. thanks so much.

shimarats wrote:hi dear tom:

i have some question about montesvar.1) do we can use montesvar for 10 variables? 2)what is the limitation of using of monesvar ?
i used montesvar for 10 variables but give me this error Effective sample size NA and
## SR10. Missing Values And/Or SMPL Options Leave No Usable Data Points The Error Occurred At Location 2657, Line 180 of MCGRAPHIRF.please guid me how fix this errors. thanks so much.

There's no hard limit. It's just that with any type of Metropolis draw procedure, as the size of a block (number of parameters) gets larger, it gets increasingly hard to get a process for the proposals which works well.

hi dear tom

i follow montesvar code and to get impulse responses .but there is a problem that is the graphs are not inside interval confidence bounds.what is the problem here.how i can put this graphs inside these interval confidence bounds.thanks so much for help me.i attached this graphs that you to see those.

how i can to solve this problem of graphs and how i can to get good graph according to the economic theory.please reply my questions thanks so much.

i run montesvar .when i run matrix of variables then give me Covariance Model-Marginal Posterior - Estimation by BFGS
NO CONVERGENCE IN 100 ITERATIONS. can you tell me what is the convergence?the run model has convergence or no convergence?please reply my questions thanks. i attach this question to show you.

If there's no convergence in 100 iterations, give it more. It seems like it's relatively near convergence.

I don't know how many draws you were doing, but an effective sample size of 8.78 doesn't sound very good.

Just offhand, have you tried doing a standard VAR (not structural) with your data? If the basic lag structure of the VAR isn't sound (due to data problems), it doesn't matter what you do with the structural model.