[R-sig-eco] Clustering AIC derived variable importance scores?

[Chris McOwen]

Dear List,

I am unsure if this is the right forum to direct my question and if it is not i apologise. 

My query is more of the theoretical / "can it be done" type than practical.


I have modelled time series of fish catch in relation to a suite of environmental variables (SST, ChlA etc) using a dynamic factor analysis. 

The output from this is usual model output including a AIC value for each model. I then calculated AIC difference and chucked away candidate models with a AIC difference score of < 2 i.e. as per Burnham and Anderson.

I then calculated the relative variable importance from the Akaike weights.

I did this for 52 distinct sites.

So I have for site 1

Chla = 1 - implying very important
SST = 0.0 - - implying not very important
Depth = 0.3

site 2

Chla = 0.4
SST = 0.2
Depth = 0.3

Etc etc

Now my question - In my eyes it is reasonable to cluster these scores using vegan for example in order to understand which sites share similarities in relation to what is important in explain variance in fish catch. I have done it, it works, and the results are nice and make sense but is there something i am missing - i have not been able to find other studies using this approach?

Thanks

Chris


Chris Mcowen
Postdoctoral Scientist, Nippon Foundation Nereus Senior Fellow
Marine Assessment and Decision Support Programme

UNEP World Conservation Monitoring Centre
219 Huntingdon Road
Cambridge CB3 0DL
United Kingdom
Switchboard: +44 (0)1223 277 314
Fax: +44 (0)1223 277 136

www.unep-wcmc.org