<BR><BR>----- Original Message -----<BR>From: Roger Bivand <Roger.Bivand@nhh.no><BR>Date: Saturday, June 20, 2009 12:32 am<BR>Subject: Re: [R-sig-Geo] Spatial Regression<BR>To: youngbin <youngbin@ua.t.u-tokyo.ac.jp><BR>Cc: r-sig-geo@stat.math.ethz.ch<BR><BR>> On Fri, 19 Jun 2009, youngbin wrote:<BR>> <BR>> > Hi,<BR>> ><BR>> > 1. While conducting spatial regression models, R does not <BR>> directly provide<BR>> > the Rsquared values. Does anybody have an idea how to get the Pseudo<BR>> > Rsquared values in spatial regression models?<BR>> <BR>> The models are fitted with maximum likelihood, so R squared is <BR>> not a very <BR>> suitable measure, although I'm sure you can find various ways of <BR>> computing <BR>> them. On the other hand, you can also get the AIC and log-<BR>> likelihood for <BR>> OLS and some other models, and they also provide a way of <BR>> comparing <BR>> models.<BR>> <BR>As Roger pointed out, R-squared isn't a very sutable measure in spatial regression. I actually alway suggest against calculating R-squared in spatial regressions. R-squared is calculated based on the ratio between explained and unexplained (residual) variation, which requires the residual to be independent to each other. While the reason of using spatial regression is that we found there is spatial autocorrelation in the residual. R-squared in such scenario is hence really of not much use (how can the explained and unexplained variations be separated in such scenario?)<BR> <BR>Hope this helps.<BR> <BR><BR>> > 2. Regarding spatial regression models, how to conduct the <BR>> general spatial<BR>> > model which both the lag and error are included?<BR>> ><BR>> <BR>> This is not provided, and is not even well understood in spatial <BR>> statistics (there are very complicated interactions between the <BR>> lag and <BR>> error components). Spatial Durbin models do provide a general <BR>> structure <BR>> within which both lag and error models nest. If both spatial <BR>> coefficients <BR>> are significant in a general model, you know with little chance <BR>> of mistake <BR>> that your model is badly misspecified, I'm afraid. The only <BR>> possible <BR>> alternative is that you have well-motivated behavioural models <BR>> for both <BR>> processes and their interactions.<BR>> <BR>> Hope this helps,<BR>> <BR>> Roger<BR>> <BR>> > Thanks<BR>> ><BR>> > youngbin<BR>> ><BR>> > _______________________________________________<BR>> > R-sig-Geo mailing list<BR>> > R-sig-Geo@stat.math.ethz.ch<BR>> > https://stat.ethz.ch/mailman/listinfo/r-sig-geo<BR>> ><BR>> <BR>> -- <BR>> Roger Bivand<BR>> Economic Geography Section, Department of Economics, Norwegian <BR>> School of<BR>> Economics and Business Administration, Helleveien 30, N-5045 Bergen,<BR>> Norway. voice: +47 55 95 93 55; fax +47 55 95 95 43<BR>> e-mail: Roger.Bivand@nhh.no<BR>> <BR>> _______________________________________________<BR>> R-sig-Geo mailing list<BR>> R-sig-Geo@stat.math.ethz.ch<BR>> https://stat.ethz.ch/mailman/listinfo/r-sig-geo<BR><BR>___________________________________________ <BR>Danlin Yu, Ph.D. <BR>Assistant Professor <BR>Department of Earth & Environmental Studies <BR>Montclair State University <BR>Montclair, NJ, 07043 <BR>Tel: 973-655-4313 <BR>Fax: 973-655-4072 <BR>email: yud@mail.montclair.edu<BR><BR>