Hello, a had another thought about which model to select when reading Bob O'Hara's and Michelle Scardi's messages which echos Wirt's ideas with which I fully agree. I think considering whether you have a process pased model which actually tries to get to the fundamental cause and effect relationships in the sense of fundamental relationships, or something like a neural network, which is sort of a better classificator, is important. In the first case, I would argue, you can achieve some understanding and may dare to venture with your modelling a little bit beyond the range of the data you have used to calibrate the model. If you are reasonable confident that your model captures the essential fundamental relationships. I would not recommend that with a neural network, which is in principle more limited to the range of data used for calibration Clearly, these are somewhat broad brush generalisations, but if I had two models, one trying to capture the essence of the system under consideration and another black be/neural network, I would accept some penalty in the fit of the process based model and opt for that model for the understanding and the predictions it allows. Clearly this is even more subjective than any criterion, but I believe this is something we should not forget. Btw, I would even be a little more critical than Bob, who used the word "explanation" in the statistical sense perfectly, but I would argue, that even though we say "factor X explains"... statistical models actually explain very little at best, in the broader sense of the term. I do not mean to imply that Bob wanted to say otherwise - I an sure he didn't - just a thought about the usage of "explain" Cheers, Joerg
________________________________ From: Ecological Society of America: grants, jobs, news on behalf of Michele Scardi Sent: Mon 3/6/2006 10:38 To: [email protected] Subject: Re: AIC Monday, March 6, 2006, 1:04:33 AM, Wirt Atmar wrote: WA> ...Gareth's comments do allow me however an opportunity to expand WA> a little bit on my previous posting. I personally hold David WA> Anderson and Ken Burnham in very high regard, but I worry that the WA> AIC is being oversold to the ecological community -- for two WA> different reasons. ... I really enjoyed reading Wirt Atmar's insightful opinion about the role AIC is supposed to play in (ecological) modeling, and I entirely agree with him about the arbitrariness of its formulation. Of course, this doesn't mean that AIC is useless. In fact, it's actually as good as Akaike's personal opinion and therefore it can be used as one of the (many) possible criteria for selecting models. However, Wirt Atmar's post leads to a more general consideration about indices and the way they are used. For instance, let's think about biotic indices and the practical consequences of their uncritical application in terms of environmental management. As ecologists, we should be used to deal with complexity, but many of us just can't refrain from turning that complexity into a single value, especially if a pre-compiled scale is available for interpreting that value as excellent, good, average, etc. Like in the AIC case, of course, some biotic indices are probably very smart, but they are still inherently subjective. So, are we doing good science when we base our conclusions on them? I don't think so, but I'm afraid that some ecologists don't even ask themselves this question. WA> ... Nevertheless, let me also say at this point that this WA> scattershot method has also received a measure of high acceptance WA> in the scientific community of late. The most exquisite example of WA> the simultaneous engineering utility and scientific meaningless of WA> the procedure exists in the training of neural networks. ... As for the "simultaneous engineering utility and scientific meaningless" Wirt Atmar mentioned, I agree with him: neural networks can be regarded as a dumb (although practically useful) tool. However, if properly trained, they're able to capture relevant relationships in very complex, non-linear systems. Of course, this is possible because some "knowledge" gets implicitly embedded into a neural network during its training. So, our problem is to extract (i.e. to understand) at least some of that knowledge. Basically, a properly trained neural network can be regarded as a simplified (but still very complex) model of a real system. However, we can "play" with it more easily than with the real thing. For instance we can do sensitivity analyses and try to figure out which stimuli (i.e. independent variables, using a regression-based analogy) are relevant with respect to each response (e.g. dependent variables). In other words, we can do experiments with the neural network model and make inferences about the properties of the real system, and then plan further research on the real system on the basis of those inferences. And this can be definitely meaningful from a scientific point of view. Cheers, Michele -------------------------------- Michele Scardi Associate Professor of Ecology Department of Biology University of Rome "Tor Vergata" Via della Ricerca Scientifica 00133 Roma Italy http://www.mare-net.com/mscardi --------------------------------
