[STATSGRASS] gstat: problem with prediction value at point location

[Roger Bivand]

On Tue, 23 Jan 2007, [UTF-8] Jarosław Jasiewicz wrote:

> Roger Bivand napisał(a):
> > On Tue, 23 Jan 2007, Edzer J. Pebesma wrote:
> >
> >   
> >> My guess is that the technique you're referring to (and used by 
> >> Geostatistical Analyst, afaik) is [kriging in] cross 
> >> validation---observations are left out to obtain a kriging estimate for 
> >> which it makes sense to compare it with the observation. Definitely 
> >> distinct from kriging, which does not leave observations out.
> >>     
> >
> > When run head to head, gstat CV LOO (leave one out) on trend surfaces and 
> > GA give the same results, so, yes, GA seems by default to do LOO CV.

Leave-one-out cross validation is when you make a prediction for data 
point j, using the data for point i=1,...,n; i!=j (leave out j). That 
means that you can validly compare the observed value at j with its 
prediction made using all the rest of the data (and the same variogram 
model for all cases). Then you do RMSE on the LOO CV residuals. See 
?gstat.cv in the gstat package.

Roger

> >
> > Roger
> >
> >   
> Sorry I didn't uderstand that...
> Jarek
> 

-- 
Roger Bivand
Economic Geography Section, Department of Economics, Norwegian School of
Economics and Business Administration, Helleveien 30, N-5045 Bergen,
Norway. voice: +47 55 95 93 55; fax +47 55 95 95 43
e-mail: Roger.Bivand at nhh.no