Panel VAR

[luching]

Hi Tom: I am trying to estimate a panel VAR using RATS standard "montevar" routine. I created a panel dataset as follows: I first loaded the data as individual series, then transformed into a panel using PFORM. Then, I used "montevar". But it seems the estimation does not take into account the panel structure of the data. The results seem to be very similar to a pooled version. Also, I did a quick panel VAR estimation using Eviews and the results there seem quite different from the RATS version. I suspect it's because RATS is not factoring in the panel structure. Any hint will be of great help. Thank you.

[TomDoan]

What do you mean by "panel VAR"? That's not a description of any specific technique, but one of possibly many depending upon what you assume to be homogeneous and what you assume to be heterogeneous. See

https://estima.com/forum/viewtopic.php? ... 803&p=3740

and the remainder of that thread.

Panel VAR's have a separate chapter in the Panel/Group Data course. This is the introduction to that chapter:

A Vector Autoregression is a close relative of the dynamic panel model examined
in Chapter 7. And, as with those models, issues arise when a VAR is
applied to panel data with a time dimension substantially smaller than would
typically be used in standard time series analysis.

Before we talk about that, we need to first determine if a standard VAR is the
appropriate choice for work with panel data. After all, a VAR isn’t a structural
model—it’s a reduced form. As such, what do we think will be homogeneous
and what will be heterogeneous? The dynamics of a VAR system are complicated
functions of the full set of parameters. If the lag parameters are homogeneous,
the responses to shocks will be identical for each cross section. Is that
too strong an assumption? Perhaps they can be assumed to be similar, but not
identical. That would require a more complicated estimation process, such as
those in Chapter 12. We’ll look at those in Sections 13.2 and 13.3. In a large
N-small T data set, there is one other possibility, which is lag systems which
are homogeneous across individuals, but differ across time periods.

One thing to note is that many panel VAR techniques ignore the bias problem
that was the emphasis in Chapter 7 (ed: on dynamic panels). Any method designed to apply to data sets
which have a reasonably large T dimension will have a negligible bias problem.