turn on suggestions

Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.

Showing results for

Find a Community

- Home
- /
- Analytics
- /
- Stat Procs
- /
- spatial regression with large data sets?

Topic Options

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

02-17-2010 10:44 PM

Hi,

I've tried various methods to model spatially autocorrelated residuals from a logistic regression using lattice (grid cell) data. My data set is too large (>50,000) for GLIMMIX to handle. My machine has 4 gb ram, but I assume this is not enough as I believe a NxN matrix must be populated. Is there a technique that accounts for spatial correlation that can handle larger data sets? I see that HPMIXED has sparse matrix technology but it does not seem to allow for residual correlation as specified by a spherical model, for instance. I've heard that generalized estimating equations (GEE) may work, but I do not see a way to do this in SAS aside from specifying many clusters. The problem is that any one observation could be in multiple 'clusters' in a sense because of spatial distances. The correlation in residuals from a basic logistic regression ends at 700 m distance, so a zero could be in the correlation matrix at this distance, hence my thinking of using sparse matrices if possible. Any advice is very appreciated.

Seth

I've tried various methods to model spatially autocorrelated residuals from a logistic regression using lattice (grid cell) data. My data set is too large (>50,000) for GLIMMIX to handle. My machine has 4 gb ram, but I assume this is not enough as I believe a NxN matrix must be populated. Is there a technique that accounts for spatial correlation that can handle larger data sets? I see that HPMIXED has sparse matrix technology but it does not seem to allow for residual correlation as specified by a spherical model, for instance. I've heard that generalized estimating equations (GEE) may work, but I do not see a way to do this in SAS aside from specifying many clusters. The problem is that any one observation could be in multiple 'clusters' in a sense because of spatial distances. The correlation in residuals from a basic logistic regression ends at 700 m distance, so a zero could be in the correlation matrix at this distance, hence my thinking of using sparse matrices if possible. Any advice is very appreciated.

Seth