[R-sig-Geo] Geographically weighted regression

[Roger Bivand]

On Sun, 24 Feb 2019, f-c-b using web.de wrote:

> Hello Roger,
>  
> thank you for your answer.
>  
> I tried:
> > gwrGp <- gwr(formula=Ziel~ Var1 + Var3 + Var4, data = Daten90,
> + bandwidth = bwG, gweight = gwr.Gauss,hatmatrix = TRUE,
> + fit.points=Daten10, predictions=TRUE, fittedGWRobject=gwrG)
>  
> Now the prediction works, but when I give out the result, I get this error:
> 
> Error in sqrt(x$results$sigma2.b) :
>   non-numeric argument to mathematical function
> How can I solve this problem?

I am not looking over your shoulder. Always provide the code verbatim, if 
I guess which mess you are in, I may get the wrong mess.

Generally, I advise against GWR in all settings, it should only ever be 
used for exploring data for non-stationarity.

Roger

>  
> Thank you,
> Christoph
> 
> 
> > Thank you for your answer.
> >  I tried to do the transformations before the calculation, but the
> > Problem is still the same. I think the problem lies in the factor. But
> > in my model I need the factor, because there is no linear influence of
> > the year.
> >  
> > Is there any solution for this Problem?
> >  https://www.dropbox.com/s/fbhwsy3sd333ung/Example.zip?dl=0 The zip-file
> > in the link contains a picture of str(Daten90) and str(Daten10) and also
> > my data and R-skript. Var4 is log(Var2).
> >  
> 
> I see:
> 
> > gwrGp <- gwr(formula=Ziel~ Var1 + Var3 + Var4, data = Daten90,
> + bandwidth = bwG, gweight = gwr.Gauss,hatmatrix = TRUE,
> + fit.points=Daten10, predictions=TRUE)
> Warning message:
> In gwr(formula = Ziel ~ Var1 + Var3 + Var4, data = Daten90, bandwidth =
> bwG, :
> standard errors set to NA, normalised RSS not available
> 
> but using your gwrG object and:
> 
> > gwrGp <- gwr(formula=Ziel~ Var1 + Var3 + Var4, data = Daten90,
> + bandwidth = bwG, gweight = gwr.Gauss,hatmatrix = TRUE,
> + fit.points=Daten10, predictions=TRUE, fittedGWRobject=gwrG)
> 
> I think the model matrix handling of the factor is OK. The problem in the
> GWmodel::gwr.predict() approach is that the full model matrix approach is
> used on the formula and data arguments, but not on predictdata, which is
> coerced to data frame but never regularised by going through model.matrix,
> hence the error message:
> 
> Browse[2]>
> debug: if (any((inde_vars %in% names(predictdata)) == F)) stop("All the
> independent variables should be included in the predictdata")
> Browse[2]> inde_vars
> [1] "Var12006" "Var12007" "Var12008" "Var12009" "Var3" "Var4"
> Browse[2]> names(predictdata)
> [1] "OBJECTID" "Rechtswert" "Hochwert" "Var1" "Var2"
> [6] "Var3" "Ziel" "Var4" "coords.x1" "coords.x2"
> 
> Had the proper approach been used, the names would have been the same.
> 
> The relevant part of spgwr::gwr() in R/gwr.R is:
> 
> if (predictions) {
> t1 <- try(slot(fit.points, "data"), silent=TRUE)
> if (class(t1) == "try-error")
> stop("No data slot in fit.points")
> predx <- try(model.matrix(delete.response(mt), fit.points))
> if (class(predx) == "try-error")
> stop("missing RHS variable in fit.points")
> if (ncol(predx) != ncol(x))
> stop("new data matrix columns mismatch")
> }
> 
> (lines 71-80)
> 
> which uses model.matrix() and uses try() to catch mis-matches.
> 
> Hope this helps,
> 
> Roger
> 
> > Christoph
> >  
> >  
> >  
> >  
> > The first step should be to look at
> >
> > str(Daten90)
> > str(Daten10)
> >
> > and if that doesn't solve the problem, then consider a reproducible
> > example, or at the very least posting the results of the above to this
> > list.
> >
> > Sarah
> >
> > On Fri, Feb 22, 2019 at 7:38 AM <f-c-b using web.de> wrote:
> > >
> > > Dear all,
> > >
> > > I am currently working out a geographically weighted regression, in
> which
> > 90% of the data set the model should be calculated and for 10% of the
> values
> > to be predicted. For the prediction I use the function gwr.predict from
> the
> > package GWModel:
> > >
> > > Erg<-gwr.predict(formula=Ziel~ as.factor(Var1) + log(Var2, base =
> exp(1))
> > + Var3, data = Daten90,predictdata = Daten10,bw = bwG, kernel =
> > "gaussian",adaptive = FALSE, p = 2, theta = 0, longlat = FALSE)
> > >
> > > I always get this error, although Daten10 and Daten90 have the same
> > structure:
> > > Error in gwr.predict(formula = Ziel~ as.factor(Var1) + log(Var2, base =
> > exp(1)) + Var3, :
> > > All the independent variables should be included in the predictdata.
> > >
> > > Can you tell me what the problem with this code is?
> > > Or is there any other way for a GWR and the prediction?
> > >
> > > Thank you,
> > > Christoph
> > > _______________________________________________
> > > R-sig-Geo mailing list
> > > R-sig-Geo using r-project.org
> > > https://stat.ethz.ch/mailman/listinfo/r-sig-geo
> >
> >
> >
> > --
> > Sarah Goslee (she/her)
> > http://www.numberwright.com
> >
> >
> 
> --
> Roger Bivand
> Department of Economics, Norwegian School of Economics,
> Helleveien 30, N-5045 Bergen, Norway.
> voice: +47 55 95 93 55; e-mail: Roger.Bivand using nhh.no
> https://orcid.org/0000-0003-2392-6140
> https://scholar.google.no/citations?user=AWeghB0AAAAJ&hl=en
> 
>

-- 
Roger Bivand
Department of Economics, Norwegian School of Economics,
Helleveien 30, N-5045 Bergen, Norway.
voice: +47 55 95 93 55; e-mail: Roger.Bivand using nhh.no
https://orcid.org/0000-0003-2392-6140
https://scholar.google.no/citations?user=AWeghB0AAAAJ&hl=en