st: Thread-Index: AczQiJRFzC33P8gmT9aFpgNN1jWtLw==

[Clyde B Schechter <clyde.schechter@einstein.yu.edu>]

Thanks to Joerg Luedicke and Austin Nichols for responding to my question about a GLM with spatially correlated errors.
 
I don't think a simple crossed-effects model, as suggested by Joerg Luedicke, will work here.  First, the "neighborhood" effect is completely confounded with the individual effect, as there will be few subjects living in the same home (we will probably select one at random from any such clusters), so while individual is nested in school, it is not nested in neighborhood--it defines neighborhood.  More important, even if this were otherwise a straightforward crossed effects model, there remains the problem that the covariance structure of the "neighborhood" effects is not captured by any of the garden-variety structures (exchangeable, autoregressive, etc.), and most certainly not by an independence structure.
 
I think Austin Nichols' approach will work for us.  I became fixated on ways to model the spatial correlation.  But the spatial correlation of effects is merely a nuisance factor for our purposes--it is not something we are specifically trying to study or quantify.   We just need to overcome the effect of non-independent observations on standard errors.  So using the robust standard error clustered on some other geographical unit will probably get us what we need.  Thanks!
 
Thanks, too, for the caution about family effects.  We're going to look into that further and I'm sure it will strengthen our study design.  I have heard of the Moving to Opportunity experiment, but have never actually read any of its studies.  Looks like that goes to the top of my to-do list.
 
Clyde Schechter
Dept. of Family & Social Medicine
Albert Einstein College of Medicine
Bronx, NY, USA
 

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/