[R] {Spam?} RE: {Spam?} RE: {Spam?} Re: mgcv, testing gamm vs lme, which degrees of freedom?
Joris Meys
jorismeys at gmail.com
Tue Jun 22 15:47:07 CEST 2010
Hi Carlos,
there is no possible way you can compare both models using a classical
statistical framework, be it ML, REML or otherwise. The assumptions
are violated. Regarding the df, see my previous mail.
In your case, I'd resort to the AIC/BIC criteria, and if prediction is
the main focus, compare the predictive power of both models using a
crossvalidation approach. Wood suggests in his book also a MCMC
approach for more difficult comparisons.
Cheers
Joris
On Tue, Jun 22, 2010 at 1:31 AM, Carlo Fezzi <c.fezzi at uea.ac.uk> wrote:
> Hi Christos,
>
> thanks for your kind reply, I agree entirely with your interpreation.
>
> In the first model comparison, however, "anova" does seem to work
> according to our interpretation, i.e. the "df" are equal in the two model.
> My intuition is that the "anova" command does a fixed-effect test rather
> than a random effect one. This is the results I get:
>
> anova(f1$lme,f2$lme)
>
> Model df AIC BIC logLik
> f1$lme 1 5 466.6232 479.6491 -228.3116
> f2$lme 2 5 347.6293 360.6551 -168.8146
>
> Hence I was not sure our interpretation was correct.
>
> On your second regarding mode point I totally agree about the appealing of
> GAMs... howver, I am working in a specific application where the quadratic
> function is the established benchmark and I think that testing against it
> will show even more strongly the appeal of a gamm approach. Any idea of
> which bases could work?
>
> Finally thansk for the tip regarding gamm4, unfortunately I need to fit a
> bivariate smooth so I cannot use it.
>
> Best wishes,
>
> Carlo
>
>
>
>
--
Joris Meys
Statistical consultant
Ghent University
Faculty of Bioscience Engineering
Department of Applied mathematics, biometrics and process control
tel : +32 9 264 59 87
Joris.Meys at Ugent.be
-------------------------------
Disclaimer : http://helpdesk.ugent.be/e-maildisclaimer.php
More information about the R-help
mailing list