- I'm interested in doing a regression with spatially correlated errors. There

are several regressor variables. Usually I would use SAS PROC MIXED to

estimated all the parameters at once using maximum likelihood, but this dataset

has about 4000 observations. I'd be happy if I could just do estimated

generalized least squares given a known (assumed) covariance matrix, but I have

a strong feeling any program is going to balk at inverting such a large

covariance matrix. What do you all do in a situation like this?

Thanks,

Marcia

--

-----------------------------------------------------------------

Marcia Gumpertz phone: (919)515-1923

Box 8203 fax: (919)515-1169

Statistics Department e-mail: gumpertz@...

NCSU

Raleigh, NC 27695-8203

-----------------------------------------------------------------

--

*To post a message to the list, send it to ai-geostats@....

*As a general service to list users, please remember to post a summary

of any useful responses to your questions.

*To unsubscribe, send email to majordomo@... with no subject and

"unsubscribe ai-geostats" in the message body.

DO NOT SEND Subscribe/Unsubscribe requests to the list! - I am sorry that I cannot be helpful, but

I am interested in maximum likelihood regression analysis

by SAS applied to spatial dataset.

Could you elaborate it farther, I mean, are there

any geostatistical module for SAS?

"Marcia Gumpertz" <gumpertz@...> sama said:>I'm interested in doing a regression with spatially correlated errors. There

Yoshiro Nagao ( Y-Nagao )

>are several regressor variables. Usually I would use SAS PROC MIXED to

>estimated all the parameters at once using maximum likelihood, but this dataset

>has about 4000 observations. I'd be happy if I could just do estimated

>generalized least squares given a known (assumed) covariance matrix, but I have

>a strong feeling any program is going to balk at inverting such a large

>covariance matrix. What do you all do in a situation like this?

International Centre for Medical Research

Kobe University School of Medicine

--

*To post a message to the list, send it to ai-geostats@....

*As a general service to list users, please remember to post a summary

of any useful responses to your questions.

*To unsubscribe, send email to majordomo@... with no subject and

"unsubscribe ai-geostats" in the message body.

DO NOT SEND Subscribe/Unsubscribe requests to the list! - With regard to large datasets, I cannot comment on how to estimate spatial

autoregressions in SAS. I have estimated a lattice model in S-PLUS (SAR)

using 3,107 observations. This took around 20 minutes on a 200 Mhz Pentium

Pro.

I have published some papers about how to estimate large lattice models - I

have dealt with as many as 200K+ observations using some of these

techniques. I currently have two software packages I have written which I

am distributing for free (with correspondingly little emphasis on GUI).

I have a Fortran 90 based package with PC executable code. This relies on

nearest neighbors for specifying the spatial relations. One can choose the

number of neighbors and the weight given to them. To conduct maximum

likelihood you need the log of the determinant of the variance-covariance

matrix or its inverse. This package uses an approximation (which yields

confidence intervals) for the log-determinant. We have handled matrices as

large as 1M by 1M using this technique (see Barry and Pace, Linear Algebra

and its Applications, forthcoming).

Given the log of the determinant, we estimate via ML the model:

Y=X*B1+S*X*B2+alpha*S*Y+e, where S is the n by n spatial weight matrix (we

take the log-determinant of (I-alpha*S)). This corresponds to having

separately spatially lagged independent and dependent variables. It takes

under 10 seconds on a Pentium 233 MMX to find the neighbors, estimate the

log-determinant, and compute the maximum likelihood estimates for the 3,107

observation dataset. It yields profile likelihoods for the overall model

and many submodels. There are separate profile likelihoods for the lower

and upper bounds of the log-determinant and hence the model accounts for

this source of uncertainty. Hence, one can conduct likelihood ratio tests

easily.

Compressed this package with an example and some documentation takes under

1MB compressed and hence I can email it to whoever wishes to use it.

I have a second more comprehensive (at least for lattice models) package

written in Matlab. This does SAR, CAR, the model above, has Delaunay and

nearest neighbor weight matrices, and simulation routines. This also has

several example datasets and so forth. This package takes around 30MB and

cannot be so easily sent. If anyone truly wants it right away, I can burn a

CDROM and send it. However, I am planning on duplicating this on CDROM and

sending out some copies at the end-of-the month for those without a

pressing need.

Anyone who is interested in receiving these please send a message to

kpace@... and put "spatial package" as the first two words

of the subject. I will send you either or both of these toolboxes.

Kelley Pace

Real Estate Research Institute

E.J. Ourso College of Business

Louisiana State University

Baton Rouge, LA 70803

email: kpace@...

also

email kelleypace@...

phone: 504-388-6257

--

*To post a message to the list, send it to ai-geostats@....

*As a general service to list users, please remember to post a summary

of any useful responses to your questions.

*To unsubscribe, send email to majordomo@... with no subject and

"unsubscribe ai-geostats" in the message body.

DO NOT SEND Subscribe/Unsubscribe requests to the list!