🚀 go-pugleaf

Thanks to Shinichi  Sakata (and other  references),  I now know  that  SAS has
added  an "ACOV" option to  proc   reg after about  version 6.06.  Hallelujah!
Furthermore, in the test statement of Proc Reg, the option SPEC allows  one to
use Heteroskedasticity consistent estimates.

/ivo


In article <1992Dec28.222415.19246@mic.ucla.edu> iwelch@agsm.ucla.edu (Ivo Welch) writes:
>
>Is there  a  routine  that  produces  White   Heteroskedasticity
>consistent standard errors for regressions in SAS?
>
>/ivo welch
>
>PS: Actually, this  is a pretty basic  correction, used  in many
>papers   in  my area.   Without  something equivalent (e.g.    a
>jack-knife), SAS regression results are almost unpublishable.

Well, let me correct my last statement(s).

	Yes, SAS can compute the White (1980) heteroskedasticity-adjusted
	covariance matrix. However, it can

	neither	[a] output this matrix to a data set for further processing
	nor	[b] use it to compute White se's or T-stats

Essentially,  this procedure is  infeasible    for anything but   small   data
problems.  If you  have 2,000  coefficients, be  prepared to  wade through   4
million numbers, and compute the square root of 2,000 of them by hand.

(Of course, I could  start learning IML programming,  and  compute  everything
myself.  However, if I  have to  program, my  preference  is to do so in other
languages [e.g. C, Fortran, Gauss,...] instead.)

/ivo welch

In article <1993Jan4.112016.12341@mic.ucla.edu> iwelch@agsm.ucla.edu (Ivo Welch) writes:
>
>Well, let me correct my last statement(s).
>
>	Yes, SAS can compute the White (1980) heteroskedasticity-adjusted
>	covariance matrix. However, it can
>
>	neither	[a] output this matrix to a data set for further processing
>	nor	[b] use it to compute White se's or T-stats
>
>Essentially,  this procedure is  infeasible    for anything but   small   data
>problems.  If you  have 2,000  coefficients, be  prepared to  wade through   4
>million numbers, and compute the square root of 2,000 of them by hand.
>
>(Of course, I could  start learning IML programming,  and  compute  everything
>myself.  However, if I  have to  program, my  preference  is to do so in other
>languages [e.g. C, Fortran, Gauss,...] instead.)
>
>/ivo welch

My order of preference would be Gauss, C++, Fortran.  And you can order
Gauss with LINREG which computes Het.-Cons. standard errors for single
and multiple equation regression, and MAXLIK which will do the same for
user provided likelihood functions.


-- 
Ronald Schoenberg                    fax: 206-727-6521
University of Washington           email: rons@u.washington.edu