R help - Jun 2010 - package mgcv inconsistency in help files? cyclic P-spline "cs" not cyclic?

Dear all,

I'm a bit stunned by the behaviour of a gam model using cyclic
P-spline smoothers.  I cannot provide the data, as I have about 61.000
observations from a time series.

I use the following model :
testgam <-
gam(NO~s(x)+s(y,bs="cs")+s(DD,bs="cs")+s(TT),data=Final)

The problem lies with the cyclic smoother I use for seasonal trends.
The variable Final$y is a numerical variable, going from 1 to 366,
representing the day of the year. I have hourly data from 2003 until
2009, so each day is represented 168 times in the dataset (apart from
366, that one only 48). DD is the wind direction, going from 1 to
3600, and is also modeled with the same cyclic smoother. Yet, if I
check the predictions, the smoother for y is far from cyclic.

I checked the help files ?smooth.terms, and found about 10 lines apart :

bs="cs" specifies a shrinkage version of "cr".

bs="cs" gives a cyclic version of a P-spline.

When I use the (bs="cc") option, I get the results as I want them, so
I keep with the cyclic cubic splines for now. Yet, I find the
behaviour of bs="cs" puzzling, and I'm wondering whether I missed
something, or if this really is an inconsistency in the package.

I currently run mgcv 1.6-1 on R 2.10.1

A small example showing what I experience. Mind you that here x is in
fact NOT cyclic, whereas in my data I'm sure it has to be :
y <- rep(1:20,200)
x <- 1:4000
DD <- sample(1:360,4000,replace=T)
TT <- sample(-10:10,4000,replace=T)
NO <- TT^2 + (10-y+2)^2 + 10*sin(DD*2*pi/360) - 0.002*sqrt(x)
+rnorm(4000,0,100)

model <- gam(NO~s(x)+s(y,bs="cs")+s(DD,bs="cs")+s(TT))
plot(model)

model <- gam(NO~s(x)+s(y,bs="cc")+s(DD,bs="cc")+s(TT))
plot(model)


Cheers
Joris

-- 
Ghent University
Faculty of Bioscience Engineering
Department of Applied mathematics, biometrics and process control

tel : +32 9 264 59 87
Joris.Meys at Ugent.be
-------------------------------
Disclaimer : http://helpdesk.ugent.be/e-maildisclaimer.php

On Fri, 2010-06-04 at 15:56 +0200, Joris Meys wrote:
> Dear all,
> 
> I'm a bit stunned by the behaviour of a gam model using cyclic
> P-spline smoothers.  I cannot provide the data, as I have about 61.000
> observations from a time series.
<snip />
> 
> I checked the help files ?smooth.terms, and found about 10 lines apart :
> 
> bs="cs" specifies a shrinkage version of "cr".
> 
> bs="cs" gives a cyclic version of a P-spline.
This is a typo in the help file ?smooth.terms. ?p.spline indicates that
bs = "cp" is what is needed for a cyclic P spline.

In ?smooth.terms

?bs="cs"? gives a cyclic version of a P-spline.

should read

?bs="cp"? gives a cyclic version of a P-spline.

in in the P Spline section.

By specifying bs = "cs", you weren't getting a cyclic spline at
all, it
is a cubic regression spline with shrinkage (so a smooth can potentially
be penalized out of the model entirely during the fitting).

If cc'd the author & maintainer of mgcv (Simon Wood) so that this
doesn't get overlooked amongst the other R-Help traffic.

HTH

G
> 
> When I use the (bs="cc") option, I get the results as I want
them, so
> I keep with the cyclic cubic splines for now. Yet, I find the
> behaviour of bs="cs" puzzling, and I'm wondering whether I
missed
> something, or if this really is an inconsistency in the package.
> 
> I currently run mgcv 1.6-1 on R 2.10.1
> 
> A small example showing what I experience. Mind you that here x is in
> fact NOT cyclic, whereas in my data I'm sure it has to be :
> y <- rep(1:20,200)
> x <- 1:4000
> DD <- sample(1:360,4000,replace=T)
> TT <- sample(-10:10,4000,replace=T)
> NO <- TT^2 + (10-y+2)^2 + 10*sin(DD*2*pi/360) - 0.002*sqrt(x)
+rnorm(4000,0,100)
> 
> model <-
gam(NO~s(x)+s(y,bs="cs")+s(DD,bs="cs")+s(TT))
> plot(model)
> 
> model <-
gam(NO~s(x)+s(y,bs="cc")+s(DD,bs="cc")+s(TT))
> plot(model)
> 
> 
> Cheers
> Joris
> 
-- 
%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%
 Dr. Gavin Simpson             [t] +44 (0)20 7679 0522
 ECRC, UCL Geography,          [f] +44 (0)20 7679 0565
 Pearson Building,             [e] gavin.simpsonATNOSPAMucl.ac.uk
 Gower Street, London          [w] http://www.ucl.ac.uk/~ucfagls/
 UK. WC1E 6BT.                 [w] http://www.freshwaters.org.uk
%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%