Dear S-plussers
I have fitted a non-linear regression in the form of
a logistic growth function to experimental data. I found significant effects of
my treatment (a gradient in temperature increase at soil surface) on the f1 parameter
describing the asymptote (Asym.). This fixed-effect was best modelled as a
quadratic term within each of the three species subjected to experimentation:
Fixed effects: list(Asym ~
Species:fati^2, xmid ~ Species:fati + micro, scal ~ 1)
Value Std.Error DF t-value p-value
Asym.(Intercept)
8.60263485 0.5743440284 567 14.97819151 <.0001
Asym.SpeciesCeraI(fati^2)
-0.02873649 0.0218307696 567 -1.31632989 0.1886
Asym.SpeciesPolyI(fati^2)
-0.07381521 0.0218344296 567 -3.38067979 0.0008
Asym.SpeciesSaxiI(fati^2)
-0.04759965 0.0218275779 567 -2.18071160 0.0296
The reported standard errors are for the coefficients
of the model, but my follow-up research question regards what temperature increase results in a significant
response of the species, and for this I reckon that I need the pointwise
confidence intervals. I have not been able to find coverage of the issue of pointwise
error estimation in the lme/nlme documentation and also not in the Pinheiro
& Bates book.
I am very little of a statistical expert - I can manage
to calculate the pointwise CI if I assume that I can treat the fixed-effects
model as a simple linear regression and forget that it is embedded in a
non-linear model. But my confidence in doing so is not sufficient for a
scientific publication! I also guess that it is less simple – else, it
would have been natural to implement the se.fit=T option in the predict.nlme-function
just like predict.lm. (?)
I would appreciate your advice on this.
Regards
Rasmus
Rasmus Ejrnæs
Associate Professor
Arctic Station
3953 Qeqertarsuaq
www.nat.ku.dk/as