R help - Jun 2007 - interpretation of F-statistics in GAMs

dear listers,
I use gam (from mgcv) for evaluation of shape and strength of relationships
between a response variable and several predictors.
How can I interpret the 'F' values viven in the GAM summary? Is it
appropriate to treat them in a similar manner as the T-statistics in a
linear model, i.e. larger values mean that this variable has a stronger
impact than a variable with smaller F?
When I run my analysis for two different response varables (but identical
predictors), is there a way to compare the F values among tests (like to
standardize them by teh sum of F within each test?) I append two summaries
below.
Thanks in advance,
Robert


### example 1 ###

Family: gaussian
Link function: identity

Formula:
dep[sel, i] ~ s(date, k = 3) + s(depth, k = kn) + s(temp, k = kn) +
    s(light, k = kn) + s(PO4, k = kn) + s(DIN, k = kn) + s(prop.agpla,
    k = kn)

Parametric coefficients:
            Estimate Std. Error t value Pr(>|t|)
(Intercept)   5.1048     0.0384   132.9   <2e-16 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05
'.' 0.1 ' ' 1

Approximate significance of smooth terms:
                edf Est.rank      F  p-value
s(date)       1.669        2 12.161 1.07e-05 ***
s(depth)      1.671        2 36.125 4.85e-14 ***
s(temp)       1.927        2  6.686  0.00156 **
s(light)      1.886        2 12.604 7.20e-06 ***
s(PO4)        1.676        2  3.237  0.04143 *
s(DIN)        1.000        1 38.428 3.41e-09 ***
s(prop.agpla) 1.405        2 15.987 3.79e-07 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05
'.' 0.1 ' ' 1

R-sq.(adj) =  0.687   Deviance explained = 70.5%
GCV score = 0.31995   Scale est. = 0.30076   n = 204

### example 2 ###
Family: gaussian
Link function: identity

Formula:
dep[sel, i] ~ s(date, k = 3) + s(depth, k = kn) + s(temp, k = kn) +
    s(light, k = kn) + s(PO4, k = kn) + s(DIN, k = kn) + s(prop.agpla,
    k = kn)

Parametric coefficients:
            Estimate Std. Error t value Pr(>|t|)
(Intercept)  7.13588    0.05549   128.6   <2e-16 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05
'.' 0.1 ' ' 1

Approximate significance of smooth terms:
                edf Est.rank      F  p-value
s(date)       1.944        2 15.997 3.67e-07 ***
s(depth)      1.876        2 25.427 1.52e-10 ***
s(temp)       1.000        1  2.866   0.0921 .
s(light)      1.751        2  4.212   0.0162 *
s(PO4)        1.950        2 10.632 4.14e-05 ***
s(DIN)        1.805        2 10.745 3.73e-05 ***
s(prop.agpla) 1.715        2  2.674   0.0715 .
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05
'.' 0.1 ' ' 1

 R-sq.(adj) =  0.479   Deviance explained = 50.9%
GCV score = 0.6863   Scale est. = 0.64348   n = 209

On Friday 15 June 2007 08:06, robert.ptacnik at niva.no
wrote:
> dear listers,
> I use gam (from mgcv) for evaluation of shape and strength of relationships
> between a response variable and several predictors.
> How can I interpret the 'F' values viven in the GAM summary? Is it
> appropriate to treat them in a similar manner as the T-statistics in a
> linear model, i.e. larger values mean that this variable has a stronger
> impact than a variable with smaller F?
- I'd be a bit cautious about this (even for T-statistics and linear models 
it's not quite clear to me what `impact' means if judged this way).
These gam
F statistics are only meant to provide a rough and ready means of judging 
approximate significance of terms, and I'm unsure about interpreting a  
comparison of such F ratios: for example the F statistics can be based on 
differerent numbers of degrees of freedom, depending on the term concerned...
> When I run my analysis for two different response varables (but identical
> predictors), is there a way to compare the F values among tests (like to
> standardize them by teh sum of F within each test?) I append two summaries
> below.
- Again, I don't really known how this would work. I'd be more inclined
to
compare the plotted terms and associated CIs (and maybe the p-values), 
especially if you are using GAMs in a quite exploratory way (e.g. if the 
assumption of an additive structure is really a convenience, rather than 
being something that is suggested by the underlying science). 

best,
Simon
>
>
> ### example 1 ###
>
> Family: gaussian
> Link function: identity
>
> Formula:
> dep[sel, i] ~ s(date, k = 3) + s(depth, k = kn) + s(temp, k = kn) +
>     s(light, k = kn) + s(PO4, k = kn) + s(DIN, k = kn) + s(prop.agpla,
>     k = kn)
>
> Parametric coefficients:
>             Estimate Std. Error t value Pr(>|t|)
> (Intercept)   5.1048     0.0384   132.9   <2e-16 ***
> ---
> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05
'.' 0.1 ' ' 1
>
> Approximate significance of smooth terms:
>                 edf Est.rank      F  p-value
> s(date)       1.669        2 12.161 1.07e-05 ***
> s(depth)      1.671        2 36.125 4.85e-14 ***
> s(temp)       1.927        2  6.686  0.00156 **
> s(light)      1.886        2 12.604 7.20e-06 ***
> s(PO4)        1.676        2  3.237  0.04143 *
> s(DIN)        1.000        1 38.428 3.41e-09 ***
> s(prop.agpla) 1.405        2 15.987 3.79e-07 ***
> ---
> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05
'.' 0.1 ' ' 1
>
> R-sq.(adj) =  0.687   Deviance explained = 70.5%
> GCV score = 0.31995   Scale est. = 0.30076   n = 204
>
> ### example 2 ###
> Family: gaussian
> Link function: identity
>
> Formula:
> dep[sel, i] ~ s(date, k = 3) + s(depth, k = kn) + s(temp, k = kn) +
>     s(light, k = kn) + s(PO4, k = kn) + s(DIN, k = kn) + s(prop.agpla,
>     k = kn)
>
> Parametric coefficients:
>             Estimate Std. Error t value Pr(>|t|)
> (Intercept)  7.13588    0.05549   128.6   <2e-16 ***
> ---
> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05
'.' 0.1 ' ' 1
>
> Approximate significance of smooth terms:
>                 edf Est.rank      F  p-value
> s(date)       1.944        2 15.997 3.67e-07 ***
> s(depth)      1.876        2 25.427 1.52e-10 ***
> s(temp)       1.000        1  2.866   0.0921 .
> s(light)      1.751        2  4.212   0.0162 *
> s(PO4)        1.950        2 10.632 4.14e-05 ***
> s(DIN)        1.805        2 10.745 3.73e-05 ***
> s(prop.agpla) 1.715        2  2.674   0.0715 .
> ---
> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05
'.' 0.1 ' ' 1
>
>  R-sq.(adj) =  0.479   Deviance explained = 50.9%
> GCV score = 0.6863   Scale est. = 0.64348   n = 209
>
> ______________________________________________
> 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