backward stepwise selection based on Wald-tests in the package unmarked

[istir...@gmail.com]

Hi

I am reading a paper that used the package unmarked.

They state: "We performed backward stepwise model selection starting with the full model including all abundance and detection covariates, and dropped insignificant terms until all remaining terms were significant on the 0.1 alpha level.
Then we applied backward stepwise model selection based on Wald-tests to remove non-significant terms from the model."

I would like to try doing something similar. However, drop1 and step1 do not seem to work?

Does anybody know how they would have done this?


Thanks

Ingrid

[Jeffrey Royle]

hi Ingrid,

 Who wrote the paper? Maybe they are on the unmarked list here.

--
You received this message because you are subscribed to the Google Groups "unmarked" group.
To unsubscribe from this group and stop receiving emails from it, send an email to unmarked+u...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Kery Marc]

Hi Ingrid,

 

I have not written that paper (I believe) but I think that with a not too great number of covariates, you can easily do something like this „by hand“.

 

In the literature, step-wise model selection is often advised against, however, many fancier approaches are not so readily implemented with more complex hierarchical models such as those in unmarked. So I feel that it may be acceptable to do it (and I am doing it myself).

 

Usually, if you do a model selection exercise in a hierarchcial model, you start with one part of the model (e.g., detection), while keeping abundance constant (or whatever your state), and then you select the abundance part of the model in a 2^nd step, while keeping the detection part in the structure identified as „best“ in step 1.

 

Kind regards  --- Marc 

[Richard Schuster]

Hi Ingrid,

We have done something along those lines Marc is suggesting and created an R function for this. Its (in my opinion) a bit more appropriate than standard stepwise model selection as it combines stepwise model selection with multi-model inference. Have a look at the attached paper and R code. If you think that is something you might be interested in let me know and I can help you get the function to work in case its not straight forward to use. I should also mention that we have created an updated version of the function as well where instead of only retaining one 'best' model per iteration one can specify an AIC cutoff and all models that are within that cutoff will be the 'base' models for the next iteration (f.AIC_cut.occu.sig.R).

If you have a small set of covariates and would be interested to use a multi-model inference approach package MuMIn allows you to create the full list of models with all possible covariate combinations. The main reason why I created the attached function was that I had a bigger set of covariates which made use of MuMIn not practical as well as some of my models were not stable so I  wanted more control over the model selection.

Best regards,
Richard