AICc for fitspecaccum object

[NivDeMalach]

Hi,
I try to apply model selection approach for different models using the 'fitspecaccum' function.
When I apply AIC it works. but when I try AICc I get the following error message

Error in UseMethod("logLik") :
no applicable method for 'logLik' applied to an object of class "c('fitspecaccum', 'specaccum')"

Thanks,
Niv

[jarioksa]

When I try AICc, I get this message:

> AICc(mods)
Error: could not find function "AICc"

So it looks like R does not have function AICc in base or stats. I guess AICc is in some add-on package, and nobody has written a method to access fitspecaccum results for that package. However, judging from your error message it could be sufficient to write a logLik method for fitspecaccum. It seems to me that writing that logLik method is very simple, but I am not sure if this is sufficient: for that I would need to know in which package you have the AICc function.

The logLik looks this simple:

`logLik.fitspecaccum` <-
    function(object, ...)
{
    sapply(object$models, logLik, ...)
}

I don't know if this sufficient, and I don't know what would you do with this.

[jarioksa]

Please note that fitspecaccum models are not nested which means that you cannot use AIC or AICc to select between them. People do so, but it is not correct.

[NivDeMalach]

Thanks a lot for your help,

Indeed AICc comes from the "MuMIn' package...
After adding this logLik function I get a new error message 'Error in nobs.default(object) : no 'nobs' method is available'. Do you have another suggestions?

P.S. The assertion that AICc cannot be used for non-nested models is very controversial( but common among R users since it was proposed by Ripley). I believe that most of the experts in the field would disagree (see example below). Are you aware of any other method for comparing nonlinear models that can be used for 'fitspecaccum' (BIC also doesn't work...)

https://sites.warnercnr.colostate.edu/anderson/wp-content/uploads/sites/26/2016/11/AIC-Myths-and-Misunderstandings.pdf

[jarioksa]

I assumed that mere logLik.fitspecaccum would not suffice. Obviously you also need a nobs method:

nobs.fitspecaccum <-
    function(object, ...)
{
    sapply(object$models, nobs, ...)
}

Once we get a list of all needed method functions, I'll add them to vegan.

[NivDeMalach]

Thanks again!
Now, I don't get an error message anymore but something else is wrong I get the value
"numeric(0)"

[jarioksa]

It seems that we need the following functions:

`logLik.fitspecaccum` <-
    function(object, ...)
{
    out <- sapply(object$models, logLik, ...)
    ## sapply() strips attributes: get'em back
    attr(out, "df") <- 1L + length(coef(object$models[[1]]))
    attr(out, "nobs") <- nobs(object$models[[1]])
    class(out) <- "logLik"
    out
}

`nobs.fitspecaccum` <-
    function(object, ...)
{
    sapply(object$models, nobs, ...)
}

[NivDeMalach]

It works!

Thanks a lot

[jarioksa]

Fine! Now I can give an easier way of doing the same:

## mod is a fitspecaccum result object
sapply(mod$models, AICc)

This does not need any new vegan functions. I didn't want to give that at once because I wanted someone to test that the new functions also work and are sufficient...

[jarioksa]

Closed with commit 4759e1f

[gavinsimpson]

@jarioksa

Please note that fitspecaccum models are not nested which means that you cannot use AIC or AICc to select between them. People do so, but it is not correct.

I hear both sides of this.

Reading Simon Wood's new 2nd Edition of Generalized Additive Models: an Introduction with R, he claims that comparing models via AIC doesn't require the models to be nested, but if they are some terms in the approximation of the expected K-L divergence cancel.

[jarioksa]

It is just the same for me how you do if you are aware of what you do.