[R-sig-eco] Removing the effect from a qualitative variable into a GAM

[Nicholas Lewin-Koh]

Hi Clement,
I am not sure exactly what you are proposing. Is there any data
or is this all simulated? My first question is why is years ordinal?
It would seem that richness would vary smoothly from year to year.
The second question is why is spatial location lumped in with gradient?
wouldn't a "better" model be something like

rich.fit <- gam(Richness~s(Lat,Long) + s(Gradient) + s(Time),
family=Gamma())

Using the gamma, or Poisson if you want to stay discrete, is probably
closer to the observed distribution of species richness, which tends to
be skewed.

Now what do you mean by indirect effects from GCM? These could manifest
as a lagged
time effect, or second order spatial effects or both or they will
manifest in the errors.
Basically you wantto predict 100 years from a model, I would probably
model more of the
effects explicitly, so the assumptions are clearer. If you do have data
for a few years
I would look at partially specified models (Simon Wood) where some of
your effects are from smooths
and the rest are modeled as pde's.

Hope this helps.

Nicholas

 ----------------------------------------------------------------------
> 
> Message: 1
> Date: Mon, 28 Sep 2009 12:21:17 +0200
> From: Cl?ment Tisseuil <tisseuil at cict.fr>
> Subject: [R-sig-eco] Removing the effect from a qualitative variable
> 	into a	GAM
> To: r-sig-ecology at r-project.org
> Message-ID:
> 	<8f656ccd0909280321i5f55526ai76f1eed1161d6d86 at mail.gmail.com>
> Content-Type: text/plain
> 
> Dear all,
> 
> I'm investigating the spatial and temporal variability for some
> biological
> features of stream fish communities (e.g. richness, diversity), modelled
> under future climate change according different 5 climate model outputs
> (GCM) and 3 emissions scenarios (SRES)...
> 
> One of my goal is to highlight graphically the spatial (characterized by
> "Gradient" as a quantitative predictor) and temporal (characterized by
> "Years" from 2005 to 2100 as an ordinal predictor) differences in species
> richness (Richness) according to the 3 future scenarios. One idea was to
> fit, individually for each scenario, a generalized additive model, namely
> Rich.fit=gam(Richness ~ s(gradient) + s(year)), so that the Richness
> could
> then be interpolated on a regular xy grid o highlight spatio-temporal
> patterns of variability.
> 
> The 'problem' is that I expect some indirect effects in the modelled
> patterns of richness related to GCM. In this context, I would be
> interested
> to remove the effect from these GCM in my models to get a more consistent
> ideas about the 'real' patterns of differences between scenarios... Do
> you
> have any idea for doing that?
> 
> Best regards,
> 
> Clement
> 
> -- 
> Clement Tisseuil - PhD student
> 
> Laboratoire  "Evolution et Diversité Biologique" (EDB)
> UMR 5174 - Université Paul Sabatier CNRS
> 118 route de Narbonne, Bât 4R3, Porte 112
> 31062 Toulouse Cedex 9 - France
> Phone : +33 5 61 55 67 35
> Fax: +33 5 61 55 67 28
> webpage: http://www.clement-tisseuil.eu
> 
> 	[[alternative HTML version deleted]]
> 
> 
>