Re: [Gretl-devel] [Gretl-users] multivariate lrvar?
Am 16.04.2018 um 17:35 schrieb Riccardo (Jack) Lucchetti:
On Mon, 16 Apr 2018, Sven Schreiber wrote:
> Hmm, but look at section 19.3 of the gretl guide. I take it that
we
> refer to X'\Omega X = Sigma as the XOX matrix (leaving hats aside).
> This _is_ divided by T because all the Gammas that it is made of have
> the 1/T factor. And consequently Sigma is explicitly referred to as
> the long-run covariance (of X'u), irrespective of whether it's
> sandwiched or not.
> In contrast, in HAC_XOX I only see running sums, not the 1/T factor
> that makes those things an average. (But I'm not good at reading the
> gretl C code, so I may have missed something.)
The difference comes from the fact that you have to scale by T if what
you want is a consistent estimator of the long-run covariance matrix of
a bunch of observable variables.
Conversely, when you use the "sandwich" estimator for the covariance
matrix a vector of estimated parameters, the scaling is unnecessary.
Yes, I see what Allin and you mean, but could you please look at the
section 19.3 ("Time series data and HAC covariance matrices") and verify
that the first five formulas/equations are correct? Because it seems to
me there's a contradiction between what you are saying and what is shown
there.
I mean, either the factor is right or it's wrong, it cannot be
"unnecessary" unless the omission is balanced by something else.
thanks,
sven