[R] ANOVA between linear models.

[Rune Haubo]

Hi Richard

You are trying to compare two models, that are not nested. This means
that all usual asymptotics of the test statistics break down, hence
the (second) test you are attempting is not meaningful. Usually one
decides on the form of the response on other grounds such as residual
analysis or based on a box-cox analysis. You could take a look at
boxcox() in MASS (book as well as package).

That said it *is* possible to compare completely different models
using the likelihood or AIC in case of difference in number of
parameters although tests are seldom (if ever) available. In this case
you need to be sure to include the additive constants of the log
likelihood - see ?extractAIC.

In case the log-transform makes sense for your problem, you might want
to contemplate a gamma-GLM with a log link.

Hope this helps

/Rune

2008/5/15 Richard Martin <radiosity at gmail.com>:
> Hi All,
>
> I'm accustomed to performing an ANOVA to aid in choosing between
> linear models (for example y~x or y~x+x^2), however with different
> models I can't seem to do it. I'm trying to fit an exponential model
> of the form ye^(bt).
>
> Below is a code snippet that highlights what I'm trying to do
>
> s = rnorm(100, sd = 1, mean=10)
> s = s + seq(0.1,10,0.1)
> x = 1:length(s)
> model.lin = lm(s ~ x)
> model.poly = lm(s ~x  + I(x^2))
> model.exp = lm(log(s) ~ x)
>
> anova(model.lin, model.poly)
> #gives the correct outcome
>
> anova(model.lin, model.exp)
> #doesn't.
>
> This fails because of the transformation of the response variable. Can
> anyone give any advice as to how I should proceed - is there a
> different test to use in this instance, or some way of reversing the
> transform after the model has been fitted?
>
> Any help greatly appreciated!!
>
> Richard
>
> --
> Contendere, Explorare, Invenire, et non Cedere
>
> ______________________________________________
> R-help at r-project.org 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.
>