[Openspace] reduction of OLS model in Geoda
yud at mail.montclair.edu
Sun May 4 11:14:23 CDT 2008
Basically, you want to correctly specifiy your model before running
regression. There are many sources leading to misspecification. However,
as for answering your question, to reduce the unnecessary independent
variables, you need to check first the possible multicollinearity (which
is reported in GeoDa by the Multicollinearity Conditional Number. If
it's too big (that is, 20 seems to be the cutting number), you might
want to examine the relationships among your "independent" variables by
pair-ploting them, and remove one of the ones that are highly correlated
with each other (they contain repetitive information, anyhow). Then with
the reduced set of independents, you can run an OLS, and you might want
to remove the ones that are not statistically significant (via t-test).
However, sometimes it is quite possible that one independent variable
that is not significant in OLS running might turn out to be significant
(via z-score) in spatial autoregressive models. Hence your domain
knowledge shall dominate the process of removing or not any variables.
Hope this helps.
Ola Claësson wrote:
> Hi everyone!
> On what basis should one reduce a OLS regression model in Geoda. I am studying infant and child mortality in sweden during the eighteenth and nineteenth century, using data on parish level. As independent variables I have, population denisty, distance to highway etc. I start out with running an ordinary OLS with all possible variable. However, when reducing such model before running a spatial or an error model, what premises shall I base the reduction on? akaike? Schwarz? R2? significance of the independent variables?
> Openspace mailing list
> Openspace at sal.uiuc.edu
Danlin Yu, Ph.D.
Department of Earth & Environmental Studies
Montclair State University
Montclair, NJ, 07043
email: yud at mail.montclair.edu
More information about the Openspace