In this case the models being compared are really identical, and the
P-value is meaningless numerical noise.
If your main focus is hypothesis testing and you really need near exact
p-values, then do this sort of testing using unpenalized models. i.e.
don't have mgcv::gam estimate the EDF of the
Looks like a bug in mgcv::summary.gam when the model is strictly
parametric... I'll take a look and fix it. thanks, Simon
In my original message I mentioned a gam fit that turns out to be a
linear fit. By curiosity I analysed it with a linear predictor only
with mgcv package, and then as a
I think you probably should state more clearly the goal of your
analysis. In such situation, estimation and hypothesis testing are
quite different. The procedure that gives you the `best' estimate is
not necessarily the `best' for testing linearity of components. If your
goal is
Dear Denis,
Take a closer look at the anova table: The models provide identical fits to
the data. The differences in degrees of freedom and deviance between the two
models are essentially zero, 5.5554e-10 and 2.353e-11 respectively.
I hope this helps,
John
John
Fair enough, Andy. I thought I was getting both predictive ability
and confirmation that the phenomenon I was studying was not linear. I
have two projects, in one prediction is the goal and I don't really
need to test linearity. In the second I needed to confirm a cycle was
taking place
Hi John,
Le 05-10-05 à 09:45, John Fox a écrit :
Dear Denis,
Take a closer look at the anova table: The models provide identical
fits to
the data. The differences in degrees of freedom and deviance
between the two
models are essentially zero, 5.5554e-10 and 2.353e-11 respectively.
I
PROTECTED]
[mailto:[EMAIL PROTECTED] On Behalf Of Denis Chabot
Sent: Wednesday, October 05, 2005 9:04 AM
To: John Fox
Cc: R list
Subject: Re: [R] testing non-linear component in mgcv:gam
Hi John,
Le 05-10-05 à 09:45, John Fox a écrit :
Dear Denis,
Take a closer look at the anova
Thank you everyone for your help, but my introduction to GAM is
turning my brain to mush. I thought the one part of the output I
understood the best was r-sq (adj), but now even this is becoming foggy.
In my original message I mentioned a gam fit that turns out to be a
linear fit. By
PROTECTED]
Sent: Wednesday, October 05, 2005 3:33 PM
To: John Fox
Cc: R list
Subject: Re: [R] testing non-linear component in mgcv:gam
Thank you everyone for your help, but my introduction to GAM
is turning my brain to mush. I thought the one part of the
output I understood the best was r-sq