which coefficients for a gam(mgcv) model equation?

[Darren Norris]

Dear users,
I am trying to show the equation (including coefficients from the model estimates) for a gam model but do not understand how to.
Slide 7 from one of the authors presentations (gam-theory.pdf  URL:  http://people.bath.ac.uk/sw283/mgcv/) shows a general equation
log{E(yi )} = α+ ßxi + f (zi ) .
What I would like to do is put my model coefficients and present the equation used. I am an ecologist not a statistician - and have no access to statistical advice.
How would I use values from the model below to complete the equation:  log{E(yi )} = α+ ßxi + f (zi )?
Many thanks,
Darren

## from the help
library(mgcv)
set.seed(0) ## simulate some data...
dat <- gamSim(1,n=400,dist="normal",scale=2)
b<-gam(y~s(x0)+s(x1)+s(x2)+s(x3),data=dat)
summary(b)

 I have tried searching help forum with "gam equation" I was not able to find an answer from the 18 threads returned.
Also tried looking at "?predict.gam" and simply "predict.gam" however my understanding of R is not sufficient to understand the details to be able to answer my question.
I checked Simon Woods excellent presentations (http://www.maths.bath.ac.uk/~sw283/) however if the answer is there I am too stupid to see it. In truth the content assumes a level of basic understanding that I don't have. Unfortunately I do not have access to the book ( Wood S.N. (2006) Generalized Additive Models: An Introduction with R. Chapman and Hall/CRC Press.) or finances necessary to get it.
Any help would be much appreciated,

R version 2.10.1 (2009-12-14)
i386-pc-mingw32

locale:
[1] LC_COLLATE=English_United Kingdom.1252  LC_CTYPE=English_United Kingdom.1252  
[3] LC_MONETARY=English_United Kingdom.1252 LC_NUMERIC=C                          
[5] LC_TIME=English_United Kingdom.1252  

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base    

other attached packages:
[1] mgcv_1.6-1

loaded via a namespace (and not attached):
[1] grid_2.10.1        lattice_0.17-26    Matrix_0.999375-33 nlme_3.1-96      

[Gavin Simpson]

On Thu, 2010-03-04 at 14:49 -0800, doon75 wrote:
> Dear users,
> I am trying to show the equation (including coefficients from the model
> estimates) for a gam model but do not understand how to.
> Slide 7 from one of the authors presentations (gam-theory.pdf  URL:
> http://people.bath.ac.uk/sw283/mgcv/) shows a general equation
> log{E(yi )} = α+ ßxi + f (zi ) .

That would be the model for the Gamma GAM.

> What I would like to do is put my model coefficients and present the
> equation used. I am an ecologist not a statistician - and have no access to
> statistical advice.
> How would I use values from the model below to complete the equation:
> log{E(yi )} = α+ ßxi + f (zi )?
> Many thanks,
> Darren
>
> ## from the help
> library(mgcv)
> set.seed(0) ## simulate some data...
> dat <- gamSim(1,n=400,dist="normal",scale=2)
> b<-gam(y~s(x0)+s(x1)+s(x2)+s(x3),data=dat)
> summary(b)

coef(b)

will give you the coefficients for the smooth terms + the intercept for
that model. ?gamObject describes thus a little.

The coefficients for the smooth/spline terms here are of length 9 each,
and each set of 9 coefficients pertains to an f(), so the model in the R
code you gave would be something like

E(y_i) ~ alpha + f_1(x0_i) + f_2(x1_i) + f_3(x2_i) +f_4(x3_i)

or

E(y_i) = alpha + f_1(x0_i) + f_2(x1_i) + f_3(x2_i) +f_4(x3_i) + e, where
e ~ N(0, sigma)

The first number returned by coef(b) is the intercept term, the alpha
above. The next 9 numbers are the coefficients for the spline (f_1)
fitted to the variable x0, and so on. These coefficients relate to the
set of basis functions used to represent the spline for each covariate
(x0, x1 etc).

Depending on why you want to write out the equation, you may not need
the detail provide at the level of the coefficients. Are the equations
above sufficient for your needs? I'm an ecologist too and rarely do I
need to think about the models I am working with at the level of the
coefficients for the basis functions, but your needs may be different.
If so, perhaps you could provide more details on why you want the
equation for the model?

HTH

G

>
>  I have tried searching help forum with "gam equation" I was not able to
> find an answer from the 18 threads returned.
> Also tried looking at "?predict.gam" and simply "predict.gam" however my
> understanding of R is not sufficient to understand the details to be able to
> answer my question.
> I checked Simon Woods excellent presentations
> (http://www.maths.bath.ac.uk/~sw283/) however if the answer is there I am
> too stupid to see it. In truth the content assumes a level of basic
> understanding that I don't have. Unfortunately I do not have access to the
> book ( Wood S.N. (2006) Generalized Additive Models: An Introduction with R.
> Chapman and Hall/CRC Press.) or finances necessary to get it.
> Any help would be much appreciated,
>
> R version 2.10.1 (2009-12-14)
> i386-pc-mingw32
>
> locale:
> [1] LC_COLLATE=English_United Kingdom.1252  LC_CTYPE=English_United
> Kingdom.1252  
> [3] LC_MONETARY=English_United Kingdom.1252 LC_NUMERIC=C                          
> [5] LC_TIME=English_United Kingdom.1252  
>
> attached base packages:
> [1] stats     graphics  grDevices utils     datasets  methods   base    
>
> other attached packages:
> [1] mgcv_1.6-1
>
> loaded via a namespace (and not attached):
> [1] grid_2.10.1        lattice_0.17-26    Matrix_0.999375-33 nlme_3.1-96      

--
%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%
 Dr. Gavin Simpson             [t] [phone]
 ECRC, UCL Geography,          [f] [phone]
 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
%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%

______________________________________________
[hidden email] 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.

[Darren Norris]

I have spent a few days trying to figure this from the reply out but am still stuck!
I need the equation to reply to a request from a referee that was to: "show the specific estimating equation associated with the fitted line".
the model I am running is (I hope the data frame is not necessary as I think I am just not getting some basic concept, but it can be provided off list):
gam1<-gam(LR~s(Property_lg),data=property) ## use default family gaussian(link = "identity")

coef(gam1) returns:
     (Intercept) s(Property_lg).1 s(Property_lg).2 s(Property_lg).3 s(Property_lg).4 s(Property_lg).5
      44.1777350       -9.4673457       -1.5743877        0.5658906        2.2219434        0.4118942
s(Property_lg).6 s(Property_lg).7 s(Property_lg).8 s(Property_lg).9
       2.4477335       -0.6590291       14.6142365        3.4184510

so is the "estimating equation":
E(y_i) ~ 44.1777 + f_1(-9.467) + f_2(-1.574) + f_3(0.565) +f_4(2.221) + f_5(0.411) + f_6(2.447) + f_7(-0.659) +f_8(14.614)+f_9(3.418)

From the “predict” function I know the fitted value at x = 0 (intercept) is -25.5256255 and at x = 1 is -1.3417508. How do I calculate these values from the “estimating equation” above?
For x=1, I am doing the calculation below which is obviously incorrect, but how do I calculate the predicted values by hand, I must be missing something incredibly obvious?
 44.1777 + (1*-9.467) + (1*-1.574) + (1*0.565) + (1*2.221) + (1*0.411) + (1*2.447) + (1*-1*0.659) + (1*4.614)+ (1*3.418)

Many thanks for any further guidance,
Darren


<quote author="Gavin Simpson">
coef(b)

will give you the coefficients for the smooth terms + the intercept for
that model. ?gamObject describes thus a little.

The coefficients for the smooth/spline terms here are of length 9 each,
and each set of 9 coefficients pertains to an f(), so the model in the R
code you gave would be something like

E(y_i) ~ alpha + f_1(x0_i) + f_2(x1_i) + f_3(x2_i) +f_4(x3_i)

or

E(y_i) = alpha + f_1(x0_i) + f_2(x1_i) + f_3(x2_i) +f_4(x3_i) + e, where
e ~ N(0, sigma)


If so, perhaps you could provide more details on why you want the
equation for the model?

[Simon Wood-4]

Darren,

Sorry for the slow reply. This is probably much to late to be of use to you
now. I think that the referee is being unreasonable here. There are many
perfectly respectable ways of estimating GAMs for which no explicit
expression for the estimated smooth terms is available (See Hastie and
Tibshirani's GAM book). mgcv actually  does use smoothers that can be
represented explicitly, but it is completely un-illuminating to do so.

I've attached a paper describing exactly how the smooths you have used are
represented. You could provide the referee with expression (7), and the
description of how \delta_i and \alpha_i are obtained from the coefficients
actually estimated, but it's not going to help the paper's readers much.

best,
Simon

On Thursday 11 March 2010 16:22, Darren Norris wrote:

> I have spent a few days trying to figure this from the reply out but am
> still stuck!
> I need the equation to reply to a request from a referee that was to: "show
> the specific estimating equation associated with the fitted line".
> the model I am running is (I hope the data frame is not necessary as I
> think I am just not getting some basic concept, but it can be provided off
> list): gam1<-gam(LR~s(Property_lg),data=property) ## use default family
> gaussian(link = "identity")
>
> coef(gam1) returns:
>      (Intercept) s(Property_lg).1 s(Property_lg).2 s(Property_lg).3
> s(Property_lg).4 s(Property_lg).5
>       44.1777350       -9.4673457       -1.5743877        0.5658906
> 2.2219434        0.4118942
> s(Property_lg).6 s(Property_lg).7 s(Property_lg).8 s(Property_lg).9
>        2.4477335       -0.6590291       14.6142365        3.4184510
>
> so is the "estimating equation":
> E(y_i) ~ 44.1777 + f_1(-9.467) + f_2(-1.574) + f_3(0.565) +f_4(2.221) +
> f_5(0.411) + f_6(2.447) + f_7(-0.659) +f_8(14.614)+f_9(3.418)
>
> >From the “predict” function I know the fitted value at x = 0 (intercept)
> > is
>
> -25.5256255 and at x = 1 is -1.3417508. How do I calculate these values
> from the “estimating equation” above?
> For x=1, I am doing the calculation below which is obviously incorrect, but
> how do I calculate the predicted values by hand, I must be missing
> something incredibly obvious?
>  44.1777 + (1*-9.467) + (1*-1.574) + (1*0.565) + (1*2.221) + (1*0.411) +
> (1*2.447) + (1*-1*0.659) + (1*4.614)+ (1*3.418)
>
> Many thanks for any further guidance,
> Darren
>
>
>
> coef(b)
>
> will give you the coefficients for the smooth terms + the intercept for
> that model. ?gamObject describes thus a little.
>
> The coefficients for the smooth/spline terms here are of length 9 each,
> and each set of 9 coefficients pertains to an f(), so the model in the R
> code you gave would be something like
>
> E(y_i) ~ alpha + f_1(x0_i) + f_2(x1_i) + f_3(x2_i) +f_4(x3_i)
>
> or
>
> E(y_i) = alpha + f_1(x0_i) + f_2(x1_i) + f_3(x2_i) +f_4(x3_i) + e, where
> e ~ N(0, sigma)
>
>
> If so, perhaps you could provide more details on why you want the
> equation for the model?

--
> Simon Wood, Mathematical Sciences, University of Bath, Bath, BA2 7AY UK
> [phone]  www.maths.bath.ac.uk/~sw283

______________________________________________
[hidden email] 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.