Prev: Gibbs sampling
Next: FTP Binary File and Verify
From: David L Cassell on 27 Oct 2006 18:44
mkekrou(a)YAHOO.CO.UK wrote:
>
>Hi,
>
>I use the spec option to test for heteroscedasticity and I get a warning
>message on the log that the average covariance matrix for the SPEC test
>has been deemed singular which violates an assumption of the test and that
>the results should be interpreted with caution.
>
>Is there any other option to test for heteroscedasticity in this case?
>Instead of using the acov option straight away, I wanted to check if there
>is heteroscedasticity in the first place.
>
>Ta,
>
>Marina

The SPEC option and the ACOV option both depend on your data
meeting all the assumptions of White' model. And that's a lot more
than you have for Ordinary Least Squares Regression. If you check
White's 1980 paper, you'll find 7 different technical assumptions
required to make this work properly. What you hit - the average
covariance matrix being singular - is a violation of the requirements
for the test. And for White's correction using the ACOV option.

So you should NOT use the SPEC option to test for heteroskedasticity,
and you should not use the ACOV option to correct for it.

Instead, you need to investigate why your errors are not i.i.d
(Independent and Identically Distributed). Are your errors really
all normal, all with mean 0 (no outliers), and independent? If not,
then heteroskedasticity is not the only problem you need to address.

Perhaps, if you write back to SAS-L and explain more about your data
and your data sources, then someone can provide more assistance.

HTH,
David
--
David L. Cassell
mathematical statistician
Design Pathways
3115 NW Norwood Pl.
Corvallis OR 97330

_________________________________________________________________
Add a Yahoo! contact to Windows Live Messenger for a chance to win a free
trip!
http://www.imagine-windowslive.com/minisites/yahoo/default.aspx?locale=en-us&hmtagline
From: Marina Kekrou on 29 Oct 2006 10:17
Hi,

Thanks for your reply.

I include year and industry dummies in my model. When I drop the dummies,
there is no problem at all when using the spec test. However, I want to
include year and industries effects. So I checked my data by year and
industry and dropped years/industries with relatively few observations to
see if this is causing the problem but the average covariane matrix is
still singular.

Any ideas?

Thanks.
From: Marina Kekrou on 29 Oct 2006 12:45
further to my previous message...

as I understand sas doesn't take into account that the square of a dummy
variable is the variable itself hence, the singularity problem when I use
the spec test. Taking into account that the results (parameter estimates
and standard errors) will change if I don't include the year and industry
dummies, do you have any suggestions of how to adapt the sas code in order
to drop the redundant cross products from the regression? (i.e the ones'
resulting from dummy variables)?

Thanks
From: David L Cassell on 6 Nov 2006 00:49
mkekrou(a)YAHOO.CO.UK wrote:
>
>Hi,
>
>Thanks for your reply.
>
>I include year and industry dummies in my model. When I drop the dummies,
>there is no problem at all when using the spec test. However, I want to
>include year and industries effects. So I checked my data by year and
>industry and dropped years/industries with relatively few observations to
>see if this is causing the problem but the average covariane matrix is
>still singular.
>
>Any ideas?
>
>Thanks.

[1] I think you are making a tactical error by including dummy
variables for year and industry, instead of analyzing your data as
a time series with panel data. This is suited to PROC TSCSREG or
PROC PANEL, but *NOT* to PROC REG. You cannot adjust for
the cross-correlations between industries in a given year, and
you cannot adjust for the time series nature of the year-to-year data
by an adjustment for heteroskedasticity.

[2] Your large number of dummy variables is probably what is causing
the singularity of the averaged covariance matrix. But it would not
matter.

[3] You are violating *other* of the fundamental assumptions
underlying White's methodology, in addition to this one. So you should
*NOT* be using SPEC as a test - there is no guarantee it will return
valid results. And you should *NOT* use ACOV, whether SPEC
indicates heteroskedasticity or not.

Sorry,
David
--
David L. Cassell
mathematical statistician
Design Pathways
3115 NW Norwood Pl.
Corvallis OR 97330

_________________________________________________________________
Try the next generation of search with Windows Live Search today!
http://imagine-windowslive.com/minisites/searchlaunch/?locale=en-us&source=hmtagline
 | 
Pages: 1
Prev: Gibbs sampling
Next: FTP Binary File and Verify