R help - Mar 2003 - locfit troubles

Dear R experts,

We've been playing with the locfit package and are experiencing a
problem I am hoping for some help on.

We have a dataset of 13k points, one dimension of which is "judgment",
either 0 or 1, and the other "score", an arbitrary scalar (in this
case
it's between .65 and .85). We use

 
locfit(judgement~score,kern="rect",deg=0,family="normal",alpha=c(alpha0,alpha1))

which is the simplest scenario we can think of (though not the one we
would like to use ultimately). The problem: under very small alpha1s
(e.g. 0.001), we see the warning "newsplit: out of vertex space" and
locfit produces a non-sensical output.

Does anyone know the source of this warning? More importantly, are we
mistaken in our usage of such a small alpha1? (The scores are very
tightly packed; there are no windows of size .001 without at least
a few datapoints).

Thank you for your help.

-- 
William T Morgan <wmorgan at mitre.org>
The Mitre Corporation

You should look at

Loader, C. (1999) Local Regression and Likelihood, Springer.

on which the package is based.  In particular, check out page 218.

locfit() evaluates the regression surface over a grid that's chosen
adaptively.  In small bandwidth situations, it can run out of memory for
storing the grid.  Try increasing the `maxk' argument.

-roger
_______________________________
UCLA Department of Statistics
rpeng at stat.ucla.edu
http://www.stat.ucla.edu/~rpeng

On 25 Mar 2003, William T Morgan wrote:
> Dear R experts,
> 
> We've been playing with the locfit package and are experiencing a
> problem I am hoping for some help on.
> 
> We have a dataset of 13k points, one dimension of which is
"judgment",
> either 0 or 1, and the other "score", an arbitrary scalar (in
this case
> it's between .65 and .85). We use
> 
>  
locfit(judgement~score,kern="rect",deg=0,family="normal",alpha=c(alpha0,alpha1))
> 
> which is the simplest scenario we can think of (though not the one we
> would like to use ultimately). The problem: under very small alpha1s
> (e.g. 0.001), we see the warning "newsplit: out of vertex space"
and
> locfit produces a non-sensical output.
> 
> Does anyone know the source of this warning? More importantly, are we
> mistaken in our usage of such a small alpha1? (The scores are very
> tightly packed; there are no windows of size .001 without at least
> a few datapoints).
> 
> Thank you for your help.
> 
> -- 
> William T Morgan <wmorgan at mitre.org>
> The Mitre Corporation
> 
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://www.stat.math.ethz.ch/mailman/listinfo/r-help
>

On Tue, 2003-03-25 at 17:15, Roger Peng wrote: 
> locfit() evaluates the regression surface over a grid that''s
chosen
> adaptively.  In small bandwidth situations, it can run out of memory for
> storing the grid.  Try increasing the `maxk'' argument.
That was it! We had gone over the book but never quite made it to page
218. Thank you for a timely and accurate reply.

-- 
William T Morgan <wmorgan at mitre.org>
The Mitre Corporation