|
Activity Number:
|
292
|
|
Type:
|
Invited
|
|
Date/Time:
|
Tuesday, August 4, 2009 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Statistics and the Environment
|
| Abstract - #302807 |
|
Title:
|
Adding Spatially Correlated Errors Can Mess Up the Fixed Effect You Love
|
|
Author(s):
|
James Hodges*+ and Brian J. Reich
|
|
Companies:
|
The University of Minnesota and North Carolina State University
|
|
Address:
|
2221 University Ave SE, Minneapolis, MN, 55455,
|
|
Keywords:
|
spatial analysis ; fixed effects ; collinearity ; conditionally autoregressive model
|
|
Abstract:
|
In analyzing spatially-referenced data, when the primary interest is the outcome's association with a fixed effect, spatial correlation is included to get appropriate standard errors or posterior standard deviations. Our collaborator Vesna Zadnik did such an analysis with data describing stomach-cancer incidence and socioeconomic status (SES) in Slovenia's 194 municipalities. When she did, the negative association between these two measures---clear in maps and highly significant in a non-spatial analysis---disappeared. This talk explains why, giving two equivalent explanations: implicit collinearity; and implicitly inflated error variance of the data contrasts containing most of the relevant information. The latter implies this can happen with any spatial correlation structure, and we give examples. We also present a remedy that is quite general and thus perhaps dangerous.
|
