1

Most of the publications I have reviewed, conduct both OLS and GWR for comparison, and prior to analysis, ensure dependent and explanatory variables are normally distributed. Can someone explain why this is necessary for GWR? Are both OLS and GWR parametric? Is there a non-parametric version of GWR?

3

No, GWR is a frequentist, linear method but there is flexibility in how you define the distributional form of the model (ie., Poisson, Binomial, Gaussian). The GWR is a local regression that emphases 2nd order variation whereas OLS is a first order model. The general motivation in running both is to draw inference about first (global) and second (local) order process but, more directly GWR is specified to account for nonstationarity.

As far a normality, heteroscedasticity is impossible to evaluate in a local regression and as such is effectively ignored. However, this is important in the OLS.

Please note that GWR is a very suspect method and unless you have very strong nonstationarity in your data I would highly recommend abandoning its application. Besides, in the absence of nonstationarity, there is really no justification for the specification of a local regression. In the absence of second-order variation GWR will likely just overfit the data.

An OLS can often be robust to weak first-order spatial autocorrelation and should provide a reasonable model. If not your next step should not be GWR but a model, such as mixed effects models or spatial autoregressive model, where a spatial term can be explicitly defined.

  • I have data that clusters, now I want to try and explain some of the clustering using explanatory variables - hence GWR. I got stuck when reviewing publications, which comment on transforming their data to be normal before OLS and GWR application. Thus I wanted to know if a) I should bother with an OLS model, given that I have local clustering, and 2) if I need to normalize my variables for GWR. – I Heart Beats Jun 11 '15 at 15:19
  • It is unlikely that GWR will describe a clustering process. Perhaps you should explore point process models which directly incorporate the intensity (density) funciton of the data into the model. – Jeffrey Evans Jun 11 '15 at 16:16
  • The clusters are not based on density. Basically there is regional variation among E. coli rates in private wells in a given study area. I will see if variables such as well depth, casing type etc explain the contamination. Hence GWR for localized regression since I cannot use OLS because I know my dependent variable clusters. – I Heart Beats Jun 11 '15 at 17:34
  • And what I mean by "not based on density" is that, SatScan was used to obtain the clusters, which does not cluster based on density. – I Heart Beats Jun 11 '15 at 17:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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