Wednesday, August 11, 2010

Model selection paper now online

The paper I've been working on with Stephanie Seifert on model selection in Maxent is now available as a preprint from Ecological Applications. The paper demonstrates that AIC and BIC can be useful in setting regularization, and also acts as the first demonstration of the utility of ENMTools' data simulation functions.

10 comments:

  1. 一個人的價值,應該看他貢獻了什麼,而不是他取得了什麼............................................................

    ReplyDelete
  2. Cool paper. Are the appendices available yet?

    ReplyDelete
  3. One more question. Would it be ok to infer from this study that AIC could be used to test sets of variables to model ecological niches, or should you do variable selection always by changing the regularization value?

    ReplyDelete
  4. You know, we didn't really mess around with variable selection very much (in terms of actually eliminating variables from model construction), but we did find that models of the appropriate level of complexity did the best job of determining the relative contribution of different variables. Given that, I don't see any reason why using AIC/BIC to select variables wouldn't be a good way to do it, and I have done so myself in a few recent studies. However, we don't have a direct test of whether or not AIC/BIC do a good job of comparing ENMs of approximately the same complexity that are based on different variables. I would hazard a guess that information criteria work fairly well for this, based on their performance in other modeling approaches and the fact that what little information we have indicates that they work well for ENMs, but once again that is just an intuition that has yet to be tested.

    That's the state of an awful lot of parts of the model construction process, however; there are a hundred little places where we all have rules of thumb that we use based on their intuitive appeal, but we have relatively few tests of the effects that these procedures have on modeling outcomes.

    ReplyDelete
  5. Dear Dan,
    Thanks for taking the time to answer my question. This is a very useful approach...I read the paper this afternoon and I'm already testing your method on my own data.
    Jorge, E&E Stony Brook University

    ReplyDelete
  6. thank you for sharing us such a nice and informative discussion. please will you provide me brief on this
    arc fashions

    ReplyDelete