Without seeing the data and results it's hard to say. mgcv::predict.gam is
already `safe' so that's not the issue. It's also pretty heavily
tested, so a
problem with that function wouldn't be the first place I'd look. How
`large
positive' are the predictions relative to the observed response? How well do
the x1,x2 cover the unit square? Smoothers often do not extrapolate well even
over quite modest distances...
Slight negative predictions are not really surprising, given that the model
you have fitted allows -ve fitted values. You could fix this by using a log
link (with gaussian or Gamma family).
If you think the results are not right, I can take a look at what's
happening
if you send me the data, R and mgcv version numbers and exact commands
generating the problem (`off line'). I would not use the data for any other
purpose of course. It may take a little while to get to however, as I've one
or two local difficulties here at the moment.
best,
Simon
On Wednesday 08 August 2007 16:58, Johnson, Elizabeth
wrote:> I am fitting a two dimensional smoother in gam, say junk >
gam(y~s(x1,x2)), to a response variable y that is always positive and
> pretty well behaved, both x1 and x2 are contained within [0,1].
>
>
>
> I then create a new dataset for prediction with values of (x1,x2) within
> the range of the original data.
>
>
>
> predict(junk,newdata,type="response")
>
>
>
> My predicted values are a bit strange (some negative and some large
> positive values).
>
>
>
> When I plot the predicted surface, it looks well behaved with no strange
> dips/etc.
>
>
>
> Could it be a problem with the predict command?
>
> Is there a "safe" version of predict for higher-dimensional
smoothers in
> gam?
>
>
>
> Elizabeth Johnson
>
> Research Associate
>
> Johns Hopkins Unviersity
>
> Department of Biostatistics
>
>
>
>
>
>
> [[alternative HTML version deleted]]
>
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
> http://www.R-project.org/posting-guide.html and provide commented, minimal,
> self-contained, reproducible code.
-- > Simon Wood, Mathematical Sciences, University of Bath, Bath, BA2 7AY UK
> +44 1225 386603 www.maths.bath.ac.uk/~sw283