accumulate option in mcgraphirf

[indrani_5]

Can anyone provide an example of how to give accumulate option in mcgraphirf.

Thanks

[TomDoan]

See Ceccheti and Rich for example. (It does ACCUM on @MCPROCESSIRF, but it's the same idea). There are other examples of the use of ACCUM, which you can find using Help-Find in Files looking for ACCUM=

[indrani_5]

Impulse responses are quite weird (unusual) after I use the accumulate option in mcgraphirf. What is the right way to do it? I tried even mcprocessirf, it produces same results. files are attached.

[TomDoan]

What's your effective sample size? I'm going to guess that it's really small (like maybe under 10?)

[indrani_5]

data is from 1985:1 to 2016:4. Why would effective sample size shrink to below 10? And how do I fix it?

[TomDoan]

The effective sample size is a measure on the behavior of the importance sampler. See the discussion starting at the bottom of https://estima.com/docs/RATS%209%20User ... f#page=544. A very low effective sample size means that the importance function is seriously missing the actual shape of the likelihood---in particular, that some regions which have a fairly high likelihood aren't being sampled often. That's a fairly big parameter set---if the parameters aren't very sharply estimated, it could be difficult to match the shape of the likelihood based upon the usual quadratic approximation. I would be particularly concerned about the top left corner of your model, which may be coming close to failing to identify the first three shocks.

[indrani_5]

I changed the specification at the top left hand corner subject to of course RWZ rule. But there is no change in impulses and effective sample size. If that is the case, then only recursive ordering would be the best that satisfies all rules.

Also how can i change the number of iterations. Because what i see is that there is no convergence in 100 iterations.

[TomDoan]

With the ITERATIONS option (same for all non-linear estimation instructions).

The RWZ counting rules apply to just-identified models. However, there's no reason to avoid over-identified models when you have an A or B model. It's models with short-and-long-run restrictions that need to be just-identified. Don't throw in unneeded parameters to bring it up to just-identified---the fewer parameters you try to estimate, the easier it is likely to be.

[indrani_5]

The top left hand corner of the matrix looks ok now. But effective sample size is still only 0.24%. How do I resolve this?

[TomDoan]

First, you have a large VAR relative to the amount of data. (You're using about 1/3 of the degrees of freedom). You also have quite a few coefficients in the SVAR (most of them in fact) that aren't significant and some are pretty close to being negligible, both economically and statistically. (This isn't really a surprise because the residuals aren't very strongly correlated). You're too concerned with the MC results before you get your model straight. However, it will probably help to lower the degrees of freedom on the t (and maybe scale up the multiplier) to get fatter tails.

[TomDoan]

There is also very little theory on how inference works on "B" form models. However, it will work better to use a delta value of -1.5. (This pushes the variances down somewhat rather than up).