 |
 |
Dear Sir/Madam, Pursuant to my query on advantage of KL vs. LRT, GOF KS, ...., I would like to share the responses that I have obtained from ANZStat and AllStat. Acknowledging all members replied to me, but my question is still unanswered! Below you can find all the replies I received so far. Regards, Ali ================1====================== I think from the principal of KL distance, we get AIC and Likelihood ratio test statistic. So it is not justified to compare KL with AIC or LRT. However, BIC is totally different than these KL/AIC/LRT. I personally prefer BIC, because if there are more than good model then BIC selects the model with less parameters. However, AIC and other method does not certify this. However AIC is efficient than BIC. But concept of efficiency is totally driven; so may be fraudulent idea. =============2====================== we found it useful for a dynamic model selection issue in "Semi-parametric dynamic time series modelling with applications to detecting neural dynamics" Fabio Rigat and Jim Q. Smith. it's here: http://www.imstat.org/aoas/next_issue.html the operational advantage of KL in our context is that it is an integrated ratio which can be computed or approximated easily. we use it as a test statistic and the slight complication is in the approximation of the null region, but it can be done by MC or MCMC. hope this helps ========== 3============================ One of the biggest advantages I see in KL-based approaches is the ability to consider and compare non-nested models. =============4============ Thats a good question Ali. I think there are tests and then there are selection criteria fot selcting between ,odels. I think KL is like AIC and BIC in this respect. ==================5============ .....I suspect that you will not get a reply as you have posted what seems like a "homework" type question. The ANZStat (or ALLStat) mailing lists are not provided for such questions. If this is not a homework question, perhaps re-word you questions and make it into a discussion providing your thoughts on the matter. ===========6============= I did not reply to your query as I felt you were asking the wrong question. You needed to give the background to your query. There may be times when your question is relevant, but first and foremost you must state clearly what you are trying to model and why. In other words, what do you plan to use your model for? If it is for prediction then you need to be very confident you are embodying all that might be known physically about the process you are modelling. The degree to which you have a good fit will help to subjectively justify any limited extrapolation, but there is no "one size fits all" answer to which fit is the best fit. The real answer is "it depends" and you need to explore what it depends on. One of the most important areas you need to explore is the risk associated with making an incorrect prediction. If for example getting it wrong means a plane falls from the sky or a patient dies on the operating table, then your prediction needs to be pretty good before you can "use" it. If on the other hand the purpose of the model is to try to understand the process better, then my approach would be to look at all fits and see how they look! After all, in this circumstance, what it looks like is probably going to give you the best clue as to what might be going on. Hope this is some help. ===================7================= It is on the list, but sorry I can not help on this one. ================8============== I just wanted to confirm that I have received your query however unfortunately I am unable to help with you. ==================9=================== Interesting question to which I don't know the answer either. Would you mind forwarding me the responses when you get them?? =============10====================== I've been wondering the answer to this questions as well, for a while now. If you receive a response, may you kindly forward it on, please. _________________________________________________________________ Hotmail: Free, trusted and rich email service. http://clk.atdmt.com/GBL/go/171222984/direct/01/ You may leave the list at any time by sending the command SIGNOFF allstat to [log in to unmask], leaving the subject line blank.