I was wondering if there is a list of steps I can take to screen my variables for assumption violations prior to using them in my GWR. I have watched this video tutorial :

https://www.youtube.com/watch?v=plfCMZhROeQ and read this post:

Regression Model and Spatial Autocorrelation

However, I am still a little bit confused. I understand from logic and background information the types of explanatory variables I expect to explain my dependent variable (these account for about 4-7 from the >20 exp. variables I have. What I don't understand is how to convince myself I should not use certain ones.

The video tutorial suggesting using Scatterplots and OLS (ex VIF value, non-stationarity etc) prior to using GWR. It was my understanding from readings on this forum that I should not use OLS to screen variables for GWR. Thus, I am a little confused.

  • I am still looking for help with respect to using OLS assumption violation criteria for GWR. Any help would be appreciated. Commented Oct 1, 2014 at 15:43

1 Answer 1


The best reading (in my opinion) that lays out all you need to know about GWR is now available at no charge on JSTOR. It is in nice simple language a non-statistician such as myself can understand.

Sections 2, 2.1, and 2.4 address your least squares question.

My answer would be yes on the least squares. It is a common approach for selection of h (kernel bandwidth) when you cannot subjectively justify it.

"Another important issue in GWR is the choice of h-sometimes referred to as the kernel bandwidth. As stated earlier, this can greatly affect the properties of the 3-matrix. Again following the advice of Silverman (1986) when considering kernel density estimates, there are occasions when a subjective choice lends itself well to the problem in hand. If we have strong theoretically based prior beliefs about the value of h in a given situation, then it is reasonable to make use of them. However, there are many situations in which no such theoretical understandings exist, and in these cases some form of automatic data-led choice of h may be more appropriate. One method suggested here is that of least squares cross-validation. A common calibration technique is that of least squares." (Brunsdon et al. 1998).

Geographically Weighted Regression-Modelling Spatial Non-Stationarity Chris Brunsdon, Stewart Fotheringham and Martin Charlton. Journal of the Royal Statistical Society. Series D (The Statistician), Vol. 47, No. 3 (1998), pp. 431-443

  • Thank you. I guess a more direct way to phrase my question is this: do the violations my variables exhibit under OLS conditions apply to GWR? Commented Sep 30, 2014 at 20:57
  • 1
    See chapter one of this achsani.blog.mb.ipb.ac.id/files/2011/08/… Commented Sep 30, 2014 at 21:01
  • 1
    This answer, which is about selecting kernel bandwidths h, does not seem to address the question, which is about selecting explanatory variables.
    – whuber
    Commented Sep 30, 2014 at 23:12
  • @whuber Can you merge my 2 questions somehow? I would still love an answer wrt OLS and GWR variable assumptions. Commented Oct 1, 2014 at 2:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.