Jan,
What mgcv version are you using, please? (Older versions have a poor
p-value approximation when select=TRUE, but of course it's possible that
you've managed to break the newer approximation as well)
The 'select=TRUE' option adds a penalty to each smooth, to allow it to
be penalized out of the model altogether via optimization of the
smoothing parameter selection criterion. Usually it is better to use
REML for smoothing parameter selection in this case using
'method="REML"' as an option to gam. This is because REML is less prone
to undersmoothing than GCV. So 'select=TRUE' is not selecting on the
basis of the p-values, themselves, but obviously this sort of
discrepancy should not be happening.
best,
Simon
On 17/04/13 15:50, Jan Holstein wrote:
I have 11 possible predictor variables and use them to model quite a few
target variables.
In search for a consistent manner and possibly non-manual manner to identify
the significant predictor vars out of the eleven I thought the option
"select=T" might do.
Example: (here only 4 pedictors)
first is vanilla with "select=F"
fit1<-gam(target~s(mgs)+s(gsd)+s(mud)+s(ssCmax),family=quasi(link=log),data=wspe1,select=F)
summary(fit1)
Family: quasi
Link function: log
Formula:
target ~ s(mgs) + s(gsd) + s(mud) + s(ssCmax)
Parametric coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -34.57 20.47 -1.689 0.0913 .
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Approximate significance of smooth terms:
edf Ref.df F p-value
s(mgs) 2.335 2.623 0.260 0.829
s(gsd) 6.868 7.506 13.955 < 2e-16 ***
s(mud) 8.990 9.000 11.727 < 2e-16 ***
s(ssCmax) 6.770 6.978 6.664 7.68e-08 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
R-sq.(adj) = 0.402 Deviance explained = 40.4%
GCV score = 8.8563e+05 Scale est. = 8.8053e+05 n = 4511
then turn select=TRUE
fit2<-gam(target~s(mgs)+s(gsd)+s(mud)+s(ssCmax),family=quasi(link=log),data=wspe1,select=TRUE)
summary(fit2)
Family: quasi
Link function: log
Formula:
target ~ s(mgs) + s(gsd) + s(mud) + s(ssCmax)
Parametric coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.1585 1.7439 0.091 0.928
Approximate significance of smooth terms:
edf Ref.df F p-value
s(mgs) 2.456 8 24.50 <2e-16 ***
s(gsd) 7.272 9 14.33 <2e-16 ***
s(mud) 7.678 9 20.38 <2e-16 ***
s(ssCmax) 6.556 9 14.36 <2e-16 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
R-sq.(adj) = 0.397 Deviance explained = 40%
GCV score = 8.9209e+05 Scale est. = 8.8715e+05 n = 4511
I seem to not fully understand how to work with "select".
The predictor "mgs" is obviously not significant, as seen from "fit"
(above), yet here it appears as significant. Why was it not dropped? How are
not-significant predictors are identified?
--
View this message in context:
http://r.789695.n4.nabble.com/mgcv-how-select-significant-predictor-vars-when-using-gam-select-TRUE-using-automatic-optimization-tp4664510.html
Sent from the R help mailing list archive at Nabble.com.
______________________________________________
R-help@r-project.org 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 Science, University of Bath BA2 7AY UK
+44 (0)1225 386603 http://people.bath.ac.uk/sw283
______________________________________________
R-help@r-project.org 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.
